Journal of Political Economy | Macro Paper Warehouse

Debiasing and T-Tests for Synthetic Control Inference on Average Causal Effects

Mon, 01 Jan 0001 00:00:00 +0000

Chernozhukov, Wüthrich, and Zhu propose a debiased synthetic control (SC) estimator and an accompanying self-normalized t-test for making inferences on the average treatment effect on the treated (ATT) in aggregate panel data settings with one treated unit. The inferential target is the time-averaged treatment effect τ = (1/T1) Σ_{t=T0+1}^{T} (Y0t(1) − Y0t(0)), a one-number summary of the overall causal impact that admits standard-form confidence intervals, in contrast to per-period effects (which cannot be consistently estimated with one treated unit) and sharp null hypotheses (which do not inform effect magnitude).

The method addresses two structural challenges in SC inference. First, the canonical SC estimator τ_SC is biased because the weights are estimated from high-dimensional pre-treatment data, and the bias can be substantial under misspecification. Second, even if true weights were known, constructing standard errors requires estimating the long-run variance (LRV), for which classical estimators such as Newey-West are unreliable in the small samples typical of SC applications.

The debiasing procedure is a K-fold cross-fitting scheme applied to the pre-treatment period. The pre-treatment sample is split into K consecutive blocks. For each fold k, SC weights w_(k) are estimated on the leave-one-block-out pre-treatment data H_{(-k)}, and a component estimator τ_k is formed as the difference between the post-treatment SC residual (using w_(k)) and the in-block pre-treatment SC residual. The latter serves as an estimator of the bias, which under the model assumptions is stable across the pre- and post-treatment periods. The final estimator τ_hat is the average of τ_k across folds. A self-normalized t-statistic T_K = sqrt(K)(τ_hat − τ)/σ_τ is constructed using the cross-fold variance; its asymptotic distribution is t_{K-1}, so no LRV estimation is required and (1−α) confidence intervals take the textbook form τ_hat ± t_{K-1}(1−α/2) × σ_τ/sqrt(K).

The t-test is proven valid with both stationary and non-stationary data. With stationary data (Theorem 2), it is valid under arbitrary misspecification. With non-stationary data, validity holds either when all units share a common nonstationarity (Theorem 3, also misspecification-robust) or when units deviate from a common nonstationarity under restrictions on the magnitude and heterogeneity of deviations but SC is correctly specified (Theorem 4). The latter covers heterogeneous deterministic time trends and certain cointegration structures. Researchers therefore need not pre-test for unit roots and select inference procedures accordingly.

A formal efficiency result (Section 3.3) shows that the asymptotic variance of the debiased SC estimator is no larger than that of difference-in-differences (DID), because SC minimizes prediction error and w* dominates the equal-weight DID vector. The relative asymptotic efficiency (RAE) of the t-test versus DID rises with K: K=3 yields RAE of 63.56%; K=5 yields 82.08%; K=10 yields 92.25%.

Simulations calibrated to Andersson’s (2019) Swedish carbon tax application — T0=30, T1=16, N=14, Gaussian AR(1) errors — show that the t-test at K=3 achieves coverage close to the nominal 90% level across correct-specification and misspecification DGPs, while Newey-West standard errors produce substantial undercoverage (coverage = 0.72–0.84) at moderate to high AR(1) coefficients. The method performs comparably to or better than subsampling (Li, 2020) and synthetic DID (Arkhangelsky et al., 2021), and avoids bandwidth selection.

In the empirical application, the debiased SC t-test (K=3) applied to annual CO2 emissions from transport across Sweden (treated, 1990) and 14 OECD control countries over 1960–2005 yields a negative and statistically significant ATT, with a 90% confidence interval lying entirely below zero, implying approximately an 11% average reduction in per capita CO2 emissions from transport attributable to the Swedish carbon tax over 1990–2005. The pre-treatment AR(1) coefficient of SC residuals is approximately 0.31, supporting K=3 as appropriate. These findings corroborate and extend Andersson’s (2019) permutation-based results by providing a confidence interval for the magnitude of the average effect. The method is implemented in the R package scinference.

Q: What is the primary inferential target and why is it preferred over per-period effects or sharp nulls? A: The target is the ATT τ = (1/T1) Σ_{t=T0+1}^{T} (Y0t(1)−Y0t(0)), the time-averaged treatment effect on the treated unit over the post-treatment period. Per-period effects cannot be consistently estimated when there is only one treated unit, yielding wide and uninformative confidence intervals. Sharp nulls (e.g., of no effect whatsoever) are useful starting points but do not inform policy decisions about effect magnitude. The ATT provides an interpretable one-number summary and admits standard-form confidence intervals.

Q: What are the two main inferential challenges that the paper addresses? A: First, the canonical SC estimator τ_SC is biased due to estimation error in the high-dimensional weights, even under correct specification, and the bias can be substantial under misspecification. Second, even with known true weights, standard error estimation requires the long-run variance (LRV), for which classical estimators such as Newey-West (1987) and Andrews (1991) are not sufficiently accurate in the small samples typical of SC applications.

Q: How does the K-fold cross-fitting procedure debias the SC estimator? A: The pre-treatment period is divided into K consecutive blocks H1,…,HK. For each fold k, SC weights w_(k) are estimated using leave-one-block-out pre-treatment data H_{(-k)}. The component estimator τ_k subtracts the in-block pre-treatment SC residual (an estimator of the bias in period Hk) from the post-treatment SC residual (using w_(k)). Because the bias is assumed stable across pre- and post-treatment periods, this subtraction removes it. The final estimator τ_hat averages τ_k across k=1,…,K.

Q: How does the self-normalized t-statistic avoid LRV estimation? A: The statistic T_K = sqrt(K)(τ_hat − τ)/σ_τ uses σ_τ = sqrt(1 + Kr/T1) × sqrt[(1/(K−1)) Σ_k (τ_k − τ_hat)^2], which is the cross-fold standard deviation of the component estimators scaled by a factor reflecting the ratio of pre- to post-treatment block lengths. Under the asymptotic theory, T_K converges to a t_{K-1} distribution, which is pivotal and requires no bandwidth or kernel choice. The cross-fold structure acts as a self-normalizer analogous to the fixed-b approach in the LRV literature.

Q: What does the paper prove about validity with non-stationary data? A: Theorem 3 establishes that when all units share a common nonstationarity (Assumption 4: Yt(0) = Vt(0)+θt and Xt = Zt+1_N·θt where {Vt(0),Zt} is stationary and θt is unrestricted), T_K → t_{K-1} under arbitrary misspecification. Theorem 4 establishes validity when units deviate from common nonstationarity (Assumption 5) under restrictions on the magnitude and heterogeneity of deviations, but requires SC to be correctly specified. These results jointly imply that researchers need not pre-test for unit roots before applying the t-test.

Q: How does the paper formally show that debiased SC is more efficient than DID? A: The pseudo-true SC weights w* minimize mean squared prediction error over W_SC, so the residual variance σ^2_* = E(Yt(0)−Xt’w*)^2 ≤ E(Yt(0)−Xt’w_DID)^2 = σ^2_DID, where w_DID = (1/N,…,1/N)’ is the equal-weight DID vector. This inequality holds regardless of whether SC is correctly specified or not, so the efficiency gain over DID is unconditional. The t-test is also valid when the parallel trends assumption underlying DID is violated, making it more robust.

Q: What is the trade-off in choosing K, and what does the paper recommend? A: A larger K produces shorter confidence intervals (higher RAE: 63.56% at K=3 versus 92.25% at K=10) but may reduce coverage accuracy in finite samples because the t_{K-1} approximation improves with K while each block becomes smaller. The paper recommends K=3 as a starting point for typical SC applications where T0 is small, based on simulation evidence showing excellent 90% coverage at K=3. When T0 is moderate or large, K can be increased without loss of coverage accuracy.

Q: What do the simulations show about the performance of Newey-West standard errors versus the t-test? A: In simulations calibrated to the Swedish carbon tax application (T0=30, T1=16, N=14, AR(1) errors), the t-test at K=3 achieves coverage close to the nominal 90% level across both correct-specification and misspecification DGPs. Newey-West standard errors produce coverage of only 0.72–0.84 when the AR(1) coefficient of the error process is moderate to high. DID achieves nominal coverage when parallel trends hold but is biased and has poor coverage under violations of parallel trends.

Q: How does the method compare with Li (2020) subsampling and synthetic DID (Arkhangelsky et al., 2021)? A: Compared with Li (2020), the t-test allows N to grow with (T0,T1) rather than treating N as fixed, directly corrects for SC estimation bias via cross-fitting, avoids the need to pre-process data for stationarity, and does not require a subsampling bandwidth choice. Compared with SDID (Arkhangelsky et al., 2021), the t-test is simpler, does not require homoskedasticity across units as SDID’s placebo variance estimator does, and is developed under a linear prediction model rather than a factor model. Simulations show the t-test performs comparably to or better than both alternatives in the application-calibrated DGP.

Q: What are the empirical findings for the Swedish carbon tax application? A: Using annual CO2 emissions from transport for Sweden and 14 OECD control countries over 1960–2005, with T0=30 (1960–1989) and T1=16 (1990–2005), the debiased SC t-test at K=3 yields a negative and statistically significant ATT. The 90% confidence interval lies entirely below zero. The estimated average effect is approximately an 11% reduction in per capita CO2 emissions from transport attributable to the carbon tax over 1990–2005. The pre-treatment SC residuals show an estimated AR(1) coefficient of approximately 0.31, confirming moderate persistence and supporting the use of K=3.

Q: When does the paper recommend against using the t-test? A: The paper advises against the t-test when T1 is very small (T1 < 8–10), as asymptotic approximations may be inaccurate; when there are structural breaks shortly after T0 (making the ATT ill-defined); and when SC fit is poor because the treated unit is very different from controls. The method requires T0, T1, N → ∞ for asymptotic validity, and T1 ≥ 10–15 is suggested for reliable finite-sample performance.

Q: How does the paper cover higher-order improvements in finite samples? A: Appendix D formally establishes that the coverage error of the confidence interval I_K(1−α) is O(1/T) rather than O(1/sqrt(T)), analogous to the fixed-b approach in the LRV literature. This provides a formal justification for the excellent finite-sample coverage observed in the simulations and distinguishes the t-test from Gaussian approximations whose coverage error is of larger order.

K-fold cross-fitting debiasing: A procedure that splits the pre-treatment period into K consecutive blocks, estimates SC weights on the leave-one-block-out pre-treatment data for each fold, and subtracts the in-block pre-treatment prediction error as an estimator of the bias. Under the model, the bias is assumed stable across pre- and post-treatment periods, so this subtraction removes it from the final estimator.

Self-normalized t-statistic: A scale-free test statistic T_K = sqrt(K)(τ_hat − τ)/σ_τ whose denominator is the cross-fold standard deviation of the K component estimators, scaled to account for the ratio of pre-treatment block length to post-treatment period length. The statistic converges to a t_{K-1} distribution without requiring any LRV estimation.

Average treatment effect on the treated (ATT): The target parameter τ = (1/T1) Σ_{t=T0+1}^{T} (Y0t(1)−Y0t(0)), representing the time-averaged causal effect of the treatment on the treated unit over the post-treatment period. It provides an interpretable one-number summary that admits standard-form confidence intervals, in contrast to per-period effects (not consistently estimable with one unit) and sharp null hypotheses (informative about presence but not magnitude of effect).

Common nonstationarity: The condition (Assumption 4) that all units share the same nonstationary component θt — formally, Yt(0) = Vt(0)+θt and Xt = Zt+1_N·θt with {Vt(0),Zt} stationary and θt unrestricted. Under this condition, the t-test is valid under arbitrary misspecification of SC weights, without requiring the researcher to specify or pre-test the type of nonstationarity.

Relative asymptotic efficiency (RAE): The ratio of the asymptotic expected confidence interval length of the debiased SC t-test to a benchmark (taken as K→∞), quantifying the cost in interval length from using a finite K. At K=3, RAE = 63.56%; at K=5, RAE = 82.08%; at K=10, RAE = 92.25%.

Long-run variance (LRV): The quantity that governs the asymptotic variance of time-averaged quantities in settings with serially correlated data. The paper argues that classical LRV estimators (Newey-West, Andrews) are insufficiently accurate in the small samples typical of SC applications, motivating the self-normalization approach that avoids LRV estimation entirely.

Pseudo-true SC weights: The population minimizer w* = argmin_{w ∈ W_SC} E(Yt(0)−Xt’w)^2, defined as the best linear predictor of the treated unit’s counterfactual outcome within the SC simplex constraint. These weights exist and satisfy the efficiency bound even under model misspecification, providing the foundation for the efficiency comparison with DID.

Environmental Consequences of Hydrocarbon Infrastructure Policy

Mon, 01 Jan 0001 00:00:00 +0000

Covert and Kellogg study policies that aim to “keep carbon in the ground” by blocking fossil fuel infrastructure investment, with the Dakota Access Pipeline (DAPL) as their empirical application. DAPL moves more than 500,000 barrels per day of oil from the Bakken Shale of North Dakota to the U.S. Gulf Coast and was completed in June 2017 amid substantial opposition. The central research question is whether blocking pipeline construction actually keeps oil in the ground or merely shifts transport to alternative modes — specifically crude-by-rail — and what the net environmental and economic consequences are.

The paper develops a two-period model of crude oil production and transportation mode choice. In the model, oil shippers decide in period 1 whether to commit to pipeline capacity under ship-or-pay contracts, then in period 2 allocate flows between the committed pipeline and the more flexible but costlier railroad alternative. Pipeline construction is an irreversible sunk cost with zero ongoing marginal cost; rail involves no sunk cost but substantial ongoing marginal costs including quadratic adjustment costs that capture capital investment in rail cars and loading/unloading facilities. Equilibrium pipeline capacity is determined by a shippers’ indifference condition: expected per-barrel returns from pipeline access equal the FERC-regulated tariff.

The empirical model is estimated using monthly Bakken oil production and transportation data, price differentials across three coastal destinations (Gulf, East, West), and drilling productivity data. Crude-by-rail marginal costs are estimated via 2SLS, yielding static marginal cost intercepts of $9.49/bbl to the East Coast, $12.64/bbl to the Gulf Coast, and $8.69/bbl to the West Coast, plus a dynamic adjustment cost of $1.28/bbl per mbbl/d of flow change. The upstream supply model follows Anderson, Kellogg, and Salant (2018), with old-well production following exponential decline (estimated decay parameter β = 0.955) and new-well drilling responding to current and lagged prices with a total long-run elasticity of 1.32. Shippers’ beliefs about future oil prices are calibrated to an AR(1) process fit to historical price volatility (persistence φ₁ = 0.9925, volatility σ_G = 0.098). Model validation confirms a predicted expected return to pipeline commitment of $6.17/bbl against DAPL’s actual tariff of $5.50–$6.25/bbl.

The main counterfactual asks what would have happened had DAPL’s construction been enjoined. In expectation, blocking DAPL reduces pipeline flows by 306 mbbl/d. Expected crude-by-rail flows increase by 248 mbbl/d, offsetting 81% of the pipeline reduction. Bakken oil production falls by only 58 mbbl/d, a 4% reduction. The modal shift from pipeline to rail worsens local environmental outcomes: per-barrel local pollution damages from rail transport substantially exceed those from pipelines, dominated by locomotive NOx emissions in populated areas. Foreclosing DAPL increases net local pollution damages by $444,000 per day (the decrease in pipeline-related harm of $144,000/day is more than offset by the increase from rail of $588,000/day). The total cost of blocking DAPL is $45/tonne of CO2 abated — $28/tonne from lost producer surplus and $17/tonne from increased local pollution damages — a figure comparable to the contemporaneous U.S. government social cost of carbon estimate of $42/tonne.

An upstream production tax achieving the same CO2 reduction costs only $1.01–$2.68/tonne CO2 abated, an order of magnitude less, because it does not induce the distortionary modal shift to rail. Two caveats apply: if 57% of Bakken production reductions leak to other basins, the cost of blocking DAPL rises from $45/tonne to $104/tonne; and if reductions represent production delays rather than permanent reductions, effective abatement is further diminished. The analysis is scoped to Bakken crude oil and land transportation alternatives. The finding that blocking infrastructure increases local pollution is atypical of CO2 abatement policies, which usually generate local pollution co-benefits.

Q: What is the core economic mechanism by which blocking a pipeline can keep oil in the ground? A: When a pipeline is foreclosed, crude oil can still move by railroad, but rail transport involves substantial ongoing marginal costs. These costs create a wedge between upstream (Bakken) and downstream (Gulf Coast) prices that depresses upstream supply. Only when downstream prices are high enough to cover both rail marginal cost and this wedge will rail fully substitute for the pipeline; at lower prices, some production is uneconomical and stays in the ground. In the model, this price-depressing wedge is the mechanism that reduces production — but it operates only partially, since rail can substitute for much of the pipeline’s flow.

Q: How much of the blocked pipeline flow substitutes to rail versus stays in the ground? A: In expectation, blocking DAPL reduces pipeline flows by 306 mbbl/d. Expected crude-by-rail flows increase by 248 mbbl/d, offsetting 81% of the pipeline reduction. Bakken oil production falls by only 58 mbbl/d, or approximately 4%. In a specific simulated month (December 2019), 348 mbbl/d (67%) of the 520 mbbl/d of foregone pipeline flows would still move by rail.

Q: How are crude-by-rail costs estimated, and what is the role of adjustment costs? A: The authors estimate a 2SLS model of rail flows on price differentials, allowing for quadratic adjustment costs to capture investments and disinvestments in rail cars and loading facilities. Static marginal costs are $9.49/bbl (East Coast), $12.64/bbl (Gulf Coast), and $8.69/bbl (West Coast). The adjustment cost parameter γ is estimated at $1.28/bbl per mbbl/d, meaning a 10 mbbl/d monthly increase in rail flows raises marginal shipping cost by $12.76/bbl — a substantial share of total rail costs. Adjustment costs are necessary to reconcile the model with the sluggish observed response of rail flows to price differentials.

Q: What is the structure of the upstream oil supply model and what are its key parameter estimates? A: The model distinguishes “old” production from pre-existing wells, which follows exponential decline with estimated decay parameter β = 0.955, and “new” production from newly drilled wells, which is price-responsive with a total long-run elasticity of 1.32 — comparable to the 1.1–1.2 estimated by Newell and Prest (2019) across major U.S. shale plays. This structure implies that total production is highly inelastic in the short run (dominated by old wells) but responds to persistent price shocks over the long run through changes in drilling rates.

Q: How do the local pollution damages of rail compare to those of pipeline transport? A: At a social cost of carbon of $100/tonne, local air pollution damages from rail transport to the Gulf Coast are $1.66/bbl (plus $0.73/bbl in spill/accident costs), versus only $0.35/bbl local pollution (plus $0.11/bbl spills) for pipelines. Locomotive NOx emissions are the dominant factor, both because locomotives have high NOx emission factors and because these emissions often occur in densely populated areas. CO2 damages at $100/tonne SCC are roughly similar across modes ($0.79–0.83/bbl), so local pollution is the key differentiator.

Q: What is the net welfare impact of foreclosing DAPL, and how is it decomposed? A: Foreclosing DAPL reduces producer surplus by $716,000/day, increases net local pollution damages by $444,000/day (the $588,000/day increase from rail more than offsets the $144,000/day decrease from pipeline), and reduces CO2 emissions by 25.2 mtonnes/day from the 58 mbbl/d production reduction. The cost per tonne of CO2 abated is $28/tonne from lost producer surplus and $17/tonne from increased local pollution damages, totaling $45/tonne — broadly comparable to the U.S. government’s contemporaneous SCC estimate of $42/tonne. This means the policy’s abatement cost is approximately equal to the social value of each tonne abated, leaving little or no net social gain even before accounting for leakage.

Q: How does the model validate against observed data and institutional parameters? A: The model predicts an expected return to committed DAPL pipeline shipment of $6.17/bbl, which closely matches the actual DAPL tariff for committed shippers of $5.50–$6.25/bbl. The authors also validate simulated crude-by-rail flows against actual flows across destinations. The close match on the tariff is particularly meaningful because it tests the model’s equilibrium condition for pipeline capacity investment rather than a within-sample fit.

Q: How does an upstream production tax compare to blocking DAPL as a policy instrument? A: A production tax normalized to achieve the same CO2 reduction requires only $3.68/bbl if imposed after shippers have committed to DAPL (holding capacity fixed), or $3.24/bbl if announced before commitments are made (reducing pipeline capacity to 443 mbbl/d). The production tax reduces combined producer surplus and government revenue by only $96,000–$109,000/day versus $716,000/day under the DAPL ban, and reduces local pollution damages by $82,000/day rather than increasing them. The resulting cost per tonne CO2 abated is $1.01–$2.68 — an order of magnitude smaller than the $44.63/tonne for blocking DAPL.

Q: What is the production leakage caveat and how large is its effect? A: If blocking DAPL causes Bakken production to fall, production from other U.S. or global oil basins may increase, partially or fully offsetting the CO2 reduction. Following Prest (2022) and Prest et al. (2023), the authors note that if 57% of the Bakken production reduction leaks to other basins, the cost of blocking DAPL rises from $45/tonne to $104/tonne. Leakage would increase the cost per tonne for the upstream tax as well, but the relative advantage of the tax over the pipeline ban is unaffected by this caveat.

Q: What is the production delay caveat? A: Even absent leakage, the paper cautions that production reductions from either policy may represent production delays rather than permanent reductions — oil not extracted today may be extracted later as prices rise or technology improves. To the extent that reductions are temporary, the effective carbon abatement is smaller than the authors compute, and the cost per tonne of CO2 abated is correspondingly higher. The paper does not quantify this effect but flags it as a material caveat.

Q: What institutional features drive pipeline capacity investment and risk allocation? A: Pipelines are irreversible investments subject to ex-post holdup, so construction financing requires firm ship-or-pay commitments from shippers before construction and before future prices are known, meaning oil price risk is borne primarily by shippers rather than the pipeline owner. Pipeline tariffs are regulated by FERC on a cost-of-service basis. In the DAPL case, shippers executed binding ten-year ship-or-pay contracts in June 2014, and shippers’ beliefs about future oil prices at that date — calibrated to historical price volatility using an AR(1) process with estimated persistence φ₁ = 0.9925 and volatility σ_G = 0.098 — determine equilibrium capacity investment.

Q: How does the paper’s finding relate to the typical co-benefit structure of climate policies? A: Most CO2 abatement policies generate local pollution co-benefits (reduced NOx, SOx, particulates), so the abatement cost is partially offset by local pollution gains. Blocking DAPL reverses this: the pipeline-to-rail modal shift increases local pollution damages, making local pollution a cost rather than a co-benefit of the policy. The authors note this is atypical but not unprecedented — urban densification and post-combustion emissions controls in fossil fuel boilers also present CO2–local pollution trade-offs.

Infrastructure foreclosure policy: A “keep it in the ground” strategy that blocks construction of specialized fossil fuel transportation infrastructure (pipelines) with the aim of inhibiting production of the fuels that would have been transported, without requiring direct acquisition or buyout of mineral rights.
Ship-or-pay agreement: A firm, up-front capacity commitment in which a pipeline shipper agrees to pay for reserved pipeline capacity whether or not they ultimately use it, made before construction and before future prices are realized; the institutional mechanism by which oil price risk is transferred from pipeline owners to shippers.
Crude-by-rail adjustment costs: Quadratic costs modeled as linear in the period-to-period change in rail volumes to a given destination, capturing capital investments and disinvestments in rail cars, loading facilities, and unloading terminals needed to expand or contract crude-by-rail capacity; estimated at $1.28/bbl per mbbl/d of monthly flow change.
Production leakage: The partial or full offset of production reductions in one oil basin (Bakken) by production increases in other U.S. or global basins in response to the same price signals; at 57% leakage, the cost of blocking DAPL rises from $45/tonne to $104/tonne of CO2 abated.
Old-well vs. new-well production dynamics: The distinction between production from pre-existing wells (which follows an exponential decline path insensitive to current prices, β = 0.955) and production from newly drilled wells (which responds to current and lagged upstream prices with long-run elasticity 1.32); this structure makes total short-run supply highly inelastic while allowing substantial long-run price responsiveness through drilling adjustments.
Local pollution damages from NOx: The dominant component of environmental harm from crude-by-rail transport, arising from locomotive NOx emissions that are both large in magnitude and concentrated in densely populated areas along rail corridors; at $100/tonne SCC, monetized local pollution damages from rail exceed CO2 damages for all three coastal destinations, whereas for pipelines CO2 damages exceed local pollution costs.
Cost per tonne of CO2 abated: The authors’ metric for comparing infrastructure foreclosure to alternative policies; computed as the sum of lost producer surplus and net change in local pollution damages divided by the quantity of CO2 emissions avoided from reduced oil production and consumption; equals $45/tonne for blocking DAPL versus $1.01–$2.68/tonne for an equivalent upstream production tax.

Exchange Rates and Asset Prices in a Global Demand System

Mon, 01 Jan 0001 00:00:00 +0000

The paper develops an asset demand system to analyze, jointly and across all countries, how international portfolio holdings and flows, exchange rates, short-term rates, long-term yields, and equity prices are determined in equilibrium. The authors specify a nested logit model of asset demand (substitution across countries within an asset class, and across asset classes) and introduce a new instrumental-variables identification strategy based on the size distribution of countries and bilateral distances; estimating on portfolio-holdings data for 37 countries and three asset classes from 2003 to 2020, they find demand is relatively inelastic, with mean demand elasticities of 27.9 (s.e. 1.9) for short-term debt, 3.2 (0.4) for long-term debt, and 1.2 (1.1) for equity. A variance decomposition attributes 82% of exchange-rate variation, 86% of short-term-rate variation, and 60% of log market-to-book equity variation to ’latent demand’ (the residual demand shifter), while portfolio flows (54%) and macro variables (43%) dominate long-term yields. Applying the framework to the European sovereign debt crisis, latent demand explains essentially all of the Italian long-term-yield variation and 74% of the Portuguese, whereas macro fundamentals are relatively more important for Greece (46% vs. 32% for latent demand), which the authors read as consistent with Greece being insolvent while Italy and Portugal were solvent but perceived as vulnerable. Estimating the convenience yield on US assets, they find, in units of expected annual returns, 1.41% on the US dollar, 2.71% on US long-term debt, and 0.50% on US equity. All estimates are specific to their sample, model, and identification assumptions.

Summary of a forthcoming paper, AI-assisted and human-reviewed. See the linked original for the authoritative claims and full conditions.

In depth

Q1. What is a ‘global demand system’ and what does it explain?

The authors represent the equilibrium of an international macro model as an asset demand system and replace traditional optimal portfolios with estimated asset demand functions that match observed international portfolio holdings, so that portfolio flows and shifts in asset demand explain all movements in exchange rates and asset prices. This lets them reinterpret the exchange rate disconnect (Meese and Rogoff 1983) as the finding that shifts in asset demand through macro variables explain much less variation than portfolio flows and latent demand, and to identify which countries’ latent demand matters for exchange rates and asset prices.

Q2. What is the nested logit model of asset demand?

Asset demand follows a nested logit model with substitution across countries in the inner nest and across asset classes in the outer nest, where demand depends on expected returns (asset prices or yields and real exchange rates), macro variables (GDP, GDP per capita, inflation, equity volatility, sovereign rating), bilateral distance (the gravity effect), a domestic-ownership indicator (home bias), and latent demand. The nested structure gives more flexible substitution than the logit model of Koijen and Yogo (2019), while latent demand captures heterogeneous beliefs about risk exposure across investors and assets.

Q3. How are the demand elasticities identified?

The authors develop an instrumental-variables strategy in which an exogenous component of one investor group’s demand shifters generates variation in residual supply that identifies another group’s demand elasticity, isolating cross-sectional variation in residual supply from the size distribution of countries and the bilateral distances between them. Intuitively, smaller issuer countries in close proximity to larger investor countries have lower residual supply and thus higher asset prices and/or real exchange rates (the example contrasts Dutch with Australian long-term debt).

Q4. What are the estimated demand elasticities, and why do they matter?

Averaged across years and issuer countries, the mean demand elasticities are 27.9 (s.e. 1.9) for short-term debt, 3.2 (0.4) for long-term debt, and 1.2 (1.1) for equity — so, e.g., a country’s aggregate equity demand falls about 1.2% per 1% rise in its price. The authors present these as empirical targets for international macro models that rely on inelastic demand and demand shocks unrelated to fundamentals to resolve long-standing puzzles, and they note the estimates are broadly consistent with prior, more granular estimates for narrower sets of countries and asset classes once differences in aggregation and identification are accounted for.

Q5. What does the variance decomposition reveal?

Latent demand is relatively more important for exchange rates, short-term rates, and equity prices — explaining 82% of exchange-rate variation (of which foreign-exchange reserves explain 10%), 86% of short-term-rate variation, and 60% of log market-to-book equity variation — whereas portfolio flows (54%) and macro variables (43%) are relatively more important for long-term yields (latent demand explains only about 3%). For equity, North American investors explain 13% and European investors 26% of the log market-to-book variation.

Q6. How does the framework interpret the European sovereign debt crisis?

Applied to extreme long-term-yield movements in Greece, Italy, and Portugal, the decomposition shows macro variables are relatively more important for Greece (46% vs. 32% for latent demand), while latent demand explains all of the Italian and 74% of the Portuguese yield variation, with European investors alone explaining 98% of the Italian and 65% of the Portuguese movements. The authors read this as consistent with the narrative that Greece was insolvent while Italy and Portugal were solvent but perceived as vulnerable.

Q7. What are the estimated convenience yields on US assets?

Computing counterfactual prices that remove the special demand for US assets, the authors estimate convenience yields, in units of expected annual returns, of 1.41% on the US dollar, 2.71% on US long-term debt, and 0.50% on US equity. In the absence of special status, a value-weighted US-dollar exchange rate would be 5.23% higher, the US long-term yield 0.73% higher, and US market-to-book equity 3.35% lower, consistent with the view that the dollar is the global reserve currency and US Treasury debt the global safe asset.

Q8. How does the framework connect to monetary policy?

The authors note in their conclusion that, because unconventional monetary policy fundamentally concerns changes in the supply of long-term debt and its impact on exchange rates and asset prices through substitution effects, the demand-system approach is suited to study the simultaneous and cumulative impact of conventional and unconventional monetary policy across many countries — and they flag this as a direction for future research rather than a result of the current paper. This scope condition matters: the present paper estimates the demand system and its decompositions, not the effects of monetary policy itself.

Key concepts

asset demand system / demand system asset pricing : an approach (introduced in Koijen and Yogo 2019 and here extended to international finance) that estimates asset demand functions on portfolio holdings data and analyzes the equilibrium relation between holdings/flows and prices, in place of traditional optimal portfolios.

nested logit asset demand : the specific functional form for demand, with substitution across countries in the inner nest and across asset classes in the outer nest, allowing flexible substitution patterns.

latent demand : the residual component of demand shifters — capturing heterogeneous beliefs about risk exposure — that, together with portfolio flows and macro variables, accounts for movements in exchange rates and asset prices; it is the dominant driver of exchange rates and short-term rates in the decomposition.

demand elasticity (inelastic markets) : the percentage change in a country’s aggregate asset demand per 1% change in its price; the paper’s low estimates (especially 1.2 for equity) are offered as empirical targets for ‘inelastic markets’ macro-finance models.

convenience yield : the extra demand for (and hence lower expected return on) US assets owing to their special status as global reserve currency and safe asset; measured here as 1.41% (USD), 2.71% (US long-term debt), and 0.50% (US equity) in expected-annual-return units.

gravity effect and home bias : the empirical regularities that portfolio holdings decline with bilateral distance (gravity) and are tilted toward domestic assets (home bias), which the demand system captures via distance and a domestic-ownership indicator.

Exorbitant Privilege Gained and Lost: Fiscal Implications

Mon, 01 Jan 0001 00:00:00 +0000

Layer 1: Overview

This paper studies three centuries of U.K. fiscal history to understand the fiscal implications of safe asset supplier status — what the authors call “exorbitant privilege” — and how it can be gained and lost. Using the discounted cash flow approach to fiscal capacity developed in Jiang, Lustig, Van Nieuwerburgh, and Xiaolan (2019), the paper measures the present discounted value of expected future primary surpluses (inclusive of convenience yield seigniorage) and compares it to the observed market value of outstanding government debt. The central finding is a sharp historical discontinuity: before World War I, when the U.K. was the world’s dominant safe asset supplier and its gilts served as the global reserve asset, roughly only three-quarters of U.K. debt was backed by future surpluses even after accounting for convenience yields earned from global safe asset demand. After World War II, when the U.K. lost its safe asset supplier status to the U.S., the U.K.’s debt became fully backed by surpluses and fiscal capacity became closely tied to its own macro fundamentals. By contrast, the U.S. after World War II shows a pattern similar to the pre-war U.K. but more extreme: less than one-third of outstanding U.S. Treasury debt is backed by future surpluses according to the paper’s estimates, with the gap between debt and estimated fiscal capacity growing sharply over recent decades.

In depth

Q1. How does the paper measure fiscal capacity?

The paper follows the Jiang-Lustig-Van Nieuwerburgh-Xiaolan (2019) methodology, expressing the market value of outstanding government debt as the present risk-adjusted discounted value of expected future primary surpluses under the government’s intertemporal budget constraint — the no-arbitrage condition that rules out rational debt bubbles. The market value of the government debt portfolio equals the present value of tax revenues minus the present value of government spending. A Vector AutoRegression (VAR) imposing cointegration of GDP with tax revenues and government spending is used to forecast the joint dynamics of the surplus. The paper uses the market or output risk premium as the discount rate, imputing the risk properties of GDP to spending and tax revenue claims.

A key methodological challenge is handling structural breaks: before World War I, U.K. fiscal policy was pre-Keynesian — acyclical spending and taxes (except during wars) — so spending and tax revenue as shares of output inherit the risk properties of output, and the market risk premium is the appropriate discount rate. After World War II, spending becomes counter-cyclical and taxes pro-cyclical in the Keynesian framework; the paper argues that applying the market risk premium in this regime produces an upper bound on the PDV of surpluses. For the U.K., this methodology is validated: the correlation of fiscal capacity with the debt/output ratio is 0.90 in the pre-WW-I sample and remains high after WW-II, despite the fiscal regime change.

Q2. What are the quantitative findings for the U.K.?

The paper finds that before World War I, U.K. fiscal capacity fell systematically short of the observed market value of U.K. debt: the average debt/GDP ratio was 87.06% while the estimated fiscal capacity was only 69.32%, with the ratio of fiscal capacity to debt averaging 74.32% — implying roughly 26% of U.K. debt was not backed by future surpluses even after including convenience yield seigniorage. The U.K. earned average long-term convenience yields of approximately 100 basis points per annum from 1873 to 1931 (translating to approximately 0.47% of GDP in annual seigniorage), reflecting its dominant position as the world’s safe asset supplier and the quasi-monopoly position of gilts in global securities markets (U.K. national debt accounted for more than half of the world’s traded securities around 1815). Despite these convenience yields, the gap between fiscal capacity and debt persisted throughout the 19th and early 20th century.

After World War II, the picture reverses: the U.K.’s average post-war fiscal capacity of 82.03% of GDP exceeds its average debt/GDP ratio of 53.42%, leaving more than 50% of fiscal capacity unborrowed. The correlation with debt dynamics persists but the sign changes — U.K. borrowing is now constrained by own macro fundamentals rather than extended by global coordination.

Q3. What do the authors find for the United States?

The U.S. experience mirrors the pre-war U.K. after World War II but with much larger magnitudes: the paper’s estimates indicate that less than one-third (32.20%) of post-war U.S. Treasury debt is backed by future surpluses, with the gap between fiscal capacity and debt growing sharply toward the end of the sample to exceed U.S. GDP. Before World War I, the U.S. did not earn convenience yields — it was forced to borrow at higher rates than the U.K. despite having lower debt-to-output ratios — and its fiscal capacity exceeded its debt, with the ratio of capacity to debt averaging 169.36%. After World War II, when the U.S. became the global safe asset supplier under the Bretton-Woods architecture, the relationship inverted: average U.S. fiscal capacity of 13.20% of GDP represents only 32.20% of outstanding debt. The gap is increasingly large in recent decades as U.S. debt has grown while surplus projections have not expanded commensurately.

Q4. Why does safe asset supplier status allow a country to borrow beyond its fiscal capacity?

The paper argues that global investors coordinate on a single safe asset issuer based on relative macro fundamentals; this coordination is self-reinforcing because each additional investor holding the asset reduces rollover risk and renders the debt safer for all others, creating strategic complementarities that concentrate global fiscal capacity in one country beyond what its own surpluses would warrant. Unlike domestic convenience yields (arising from household demand for safe assets to insure idiosyncratic risks), the global safe asset effect creates a form of extra-fiscal capacity that depends on investors’ common belief about which country is the hegemon. The measured seigniorage from convenience yields — about 0.47% of U.K. GDP before WW-I and 0.36% per year for the U.S. post-war — does not fully capture this coordination benefit; the remaining gap between fiscal capacity and debt reflects the additional “license to borrow” that comes with global hegemon status.

The transition from U.K. to U.S. hegemony illustrates the mechanism: as U.K. macro fundamentals deteriorated relative to U.S. fundamentals after the world wars, investors shifted the concentration of fiscal capacity toward the U.S. The U.K. lost its license to borrow beyond its fundamentals; the U.S. gained it. The paper notes that the U.K. debt/output ratio exceeded 200% after WW-II — a level associated with the loss of hegemony.

Q5. What historical data and institutional context does the analysis use?

The paper uses annual data for the U.K. from 1729 to 2020 (from the Bank of England’s Millennium of Macroeconomics dataset and the Ellison-Scott dataset on individual bond market values from 1694 onward) and for the U.S. from 1791 to 2020 (from Hall-Sargent and CRSP), constructing primary surpluses, tax revenues, spending, GDP, and convenience yields consistently over nearly three centuries. U.K. convenience yields before WW-I are measured as the interest rate differential between U.K. government securities and otherwise comparable bonds from countries on the gold standard; the sample average is approximately 147 basis points at the short end and 110 basis points at the long end, with the spread declining at longer maturities (the opposite of what default risk would predict), providing evidence that convenience yield rather than residual default risk drives the differential. U.S. post-war convenience yields are constructed from the spread between the 3-month Treasury yield and the 3-month CD rate (or bankers’ acceptance rate before 1964), averaging 36 basis points per year from 1947 to 2020.

Q6. What are the implications for models of fiscal capacity and debt sustainability?

The results favor models in which the safe asset supplier’s fiscal capacity is determined partly by relative macro fundamentals (which country the global financial system coordinates on) rather than solely by absolute fundamentals (its own surpluses), and challenge models that treat the transversality condition as a binding constraint at all times for all countries. The finding that a large fraction of U.S. Treasury debt is not backed by future surpluses — even when the market risk premium is used to discount — has implications for debt sustainability analyses: standard present-value-of-surpluses calculations will understate the true fiscal capacity of the safe asset supplier, while overstating it for others. The paper’s framework suggests this extra capacity depends on maintaining relative macro fundamentals and global investor coordination, and can be lost — as the U.K. experience demonstrates — when relative fundamentals deteriorate.

Key Concepts

fiscal capacity: the present risk-adjusted discounted value of a government’s expected future primary surpluses, computed from the government’s intertemporal budget constraint under no-arbitrage; in this paper, inclusive of seigniorage revenue from convenience yields earned on government debt.
exorbitant privilege: the ability of the safe asset supplier country to borrow at below-market interest rates due to global demand for its government debt as a safe asset; quantified in this paper as the gap between the market value of debt and estimated fiscal capacity from surpluses alone, which exceeds fiscal capacity for the pre-WW-I U.K. and post-WW-II U.S.
convenience yield: the yield reduction (below comparable risky borrowing rates) that the safe asset supplier earns from global safe asset demand; measured as approximately 100 bps long-term for the pre-WW-I U.K. and approximately 36 bps on average for the post-WW-II U.S., contributing 0.47% and 0.36% of GDP annually in seigniorage revenue respectively.
transversality condition (TVC): the condition ruling out rational government debt bubbles, requiring the expected discounted value of outstanding debt to approach zero at long horizons; the paper imposes the TVC and finds that for the pre-WW-I U.K. and post-WW-II U.S., the observed debt level exceeds fiscal capacity even under this constraint, interpreted as evidence of the extra borrowing license conferred by safe asset supplier status.

From Doubt to Devotion: Trials and Learning-Based Pricing

Mon, 01 Jan 0001 00:00:00 +0000

This paper studies a dynamic mechanism design problem in which an informed seller sells an experience good to a skeptical buyer who learns about the product through consumption. The central question is: how does a seller leverage proprietary data about product-buyer match quality together with the buyer’s ability to learn, and what are the welfare implications in equilibrium?

The model features a seller who privately observes a binary match quality (theta in {H, L}) between their service and the buyer. The buyer does not observe match quality and has an initially unknown private value v for the good, drawn from a Myerson-regular distribution F with support [v_low, v_high] and normalized mean E[v] = 1. If the match is high, the buyer receives instantaneous utility rewards according to a Poisson process with flow rate lambda*I, where I in [0,1] is the seller-controlled access level. Upon receiving the first reward, the buyer perfectly learns both match quality theta and their own value v. The seller commits to a dynamic mechanism over time horizon T = [0, T] specifying access and prices conditional on reported histories. Both parties are risk-neutral and there is no discounting in the baseline.

Two benchmark cases show the first-best is attainable absent both key features simultaneously. If trade is static (prices set only at time 0) or if the seller is uninformed about theta, the seller achieves first-best revenue of lambdamu_0T by selling the entire service upfront. Proposition 1 establishes both cases; this implies that consumer data on theta is not required for maximizing social welfare, and it is weakly dominant for a seller to never collect consumer data in static environments.

The central result is that the combination of dynamic pricing and seller private information breaks the first-best. A high-type seller can deviate by offering a “Myersonian free trial”: provide full access up to time tM (defined as argmax_t {(1 - exp(-lambdat))(T - t)}), then offer the remaining service at post-trial price lambdavM(T - tM), where vM is the Myerson monopoly price. The buyer accepts the trial regardless of beliefs (participation is weakly dominant) and purchases the post-trial service if and only if v >= vM. This deviation yields payoff pi_F = (1 - exp(-lambdatM))(1 - F(vM))lambdavM*(T - tM). Proposition 2 states that the first-best cannot be implemented in any equilibrium if and only if pi_F > lambdamu_0T. Corollary 1 shows this condition holds for sufficiently large T, since pi_F grows proportionally with T while the first-best also grows with T but the ratio converges to a constant less than 1 only for some parameter configurations and exceeds 1 for others.

Theorem 1 (the main mechanism design result) characterizes the boundary of the IC-IR feasible payoff set: any mechanism on this boundary is outcome-uniquely implemented by a trial mechanism, defined by a triple (v0, t0, p0) — a trial length, a post-trial value threshold, and a trial price. During [0, t0] uninformed buyers receive full access; after t0 only buyers who received a reward with v >= v0 continue at a premium. Trial length t0 is weakly increasing in the weight placed on the low-type seller and in the prior mu_0; post-trial threshold v0 is weakly decreasing in the same objects (Proposition 3).

Equilibrium payoffs (Proposition 5) are precisely the IC-IR feasible pairs satisfying pi_H >= pi_F, implemented by pooling trial mechanisms in which both seller types propose identical mechanisms and the buyer updates beliefs only through private consumption signals. Under the D1 refinement (Proposition 6), only mechanisms with trial length tM and post-trial threshold vM survive. These have the shortest trial and highest post-trial price of all equilibrium mechanisms, minimize social surplus, and may leave both seller types strictly worse off than in a world without private information — directly contrasting the static informed principal result of Koessler and Skreta (2016) where data always helps the seller.

When the seller can control service quality q in addition to access I (Section 6), the relevant equilibrium mechanisms become dynamic tiered pricing rather than binary trials: a low-quality, high-ad-load free tier provides learning opportunities while reducing information rents; convinced buyers upgrade to a premium ad-free tier. Counterintuitively, enriching the seller’s screening technology can reduce both revenue and social efficiency in equilibrium because additional instruments create additional signaling opportunities that distort outcomes further.

Q: What is the core tension that prevents the first-best from being an equilibrium?

A: When the seller is privately informed and pricing is dynamic, the high-type seller anticipates a greater likelihood of the buyer receiving a utility shock than the buyer’s own prior implies. This belief gap makes it profitable for the high-type seller to deviate from a proposed first-best mechanism by offering a free trial that “proves” high match quality and then extracting rent from convinced buyers. Because this deviation is profitable — yielding pi_F > lambdamu_0T under some parameters — the first-best pooling contract unravels. The interaction of both ingredients (dynamic pricing and informed seller) is necessary: either ingredient alone is insufficient to break the first-best (Proposition 1).

Q: What exactly is the Myersonian free trial and why does the buyer always accept it?

A: The Myersonian free trial provides full service access up to time tM = argmax_t {(1 - exp(-lambdat))(T - t)} at (approximately) zero price, then offers the remaining service at price lambdavM(T - tM) where vM is the Myerson monopoly price. The buyer accepts the trial regardless of their prior belief about match quality because the trial itself is free and provides non-negative payoff. After the trial, the buyer purchases the post-trial service if and only if they received a reward with v >= vM; otherwise they exit. The deviation payoff is pi_F = (1 - exp(-lambdatM))(1 - F(vM))lambdavM*(T - tM).

Q: Under what parametric conditions can the first-best not be supported in equilibrium?

A: By Proposition 2, the first-best cannot be implemented if and only if pi_F > lambdamu_0T. Corollary 1 states that for sufficiently large T this always fails, since as T grows, pi_F grows proportionally (the post-trial term (T - tM) dominates) while tM converges to a finite value. More precisely, for large T, pi_F / (lambdamu_0T) converges to (1 - exp(-lambda*tM)) * (1 - F(vM)) * vM / mu_0, which exceeds 1 under appropriate parameter configurations. Conversely, when mu_0 is high or the service horizon is short, the first-best may remain implementable.

Q: What is a trial mechanism and how does Theorem 1 characterize it?

A: A trial mechanism is defined by a triple (v0, t0, p0): uninformed buyers receive full access on [0, t0] and no access thereafter; a buyer who reports a reward of value v >= v0 at time t receives full service for the remainder [t, T] at a price increment of lambdav0(T - t0); the trial itself is priced at p0. Theorem 1 states that any payoff pair on the boundary of the IC-IR feasible set is outcome-uniquely attained by such a trial mechanism with appropriately determined (v0, t0, p0). The proof uses a relaxed problem retaining only two key constraint families: local incentive constraints on value reporting (IC-V) and a global intertemporal constraint preventing buyers from hiding the arrival of rewards forever (IC-U).

Q: How does the trial length respond to changes in prior belief mu_0 and distributional spread?

A: Proposition 3 states that t0 is weakly increasing in mu_0: as market belief becomes more optimistic, both seller types extract higher revenue from the trial, so the mechanism designer extends the trial. Proposition 4 adds that for a uniform distribution on [1-delta, 1+delta], trial length t0 is weakly increasing in delta (greater spread). The post-trial threshold v0 is weakly decreasing in mu_0, meaning that a more optimistic prior leads to a less exclusive post-trial cutoff.

Q: What are the equilibrium payoffs and how does the high-type seller’s free-trial option constrain them?

A: Proposition 5 states that (pi_L, pi_H) is an equilibrium payoff if and only if it lies in the IC-IR feasible set and pi_H >= pi_F. The lower bound pi_H >= pi_F reflects the high-type seller’s outside option: they can always deviate to the Myersonian free trial. Corollary 4 then shows that all “reasonable” equilibrium payoffs (those with pi_H >= pi_L, surviving a mild off-path refinement) are implemented by trial mechanisms with complete pooling — both seller types propose the same mechanism and the buyer updates beliefs only through private consumption signals, not the mechanism’s structure.

Q: What does the D1 refinement select and why do it lead to worse outcomes?

A: Proposition 6 shows that the only equilibrium trial mechanisms surviving the D1 criterion have trial length tM and post-trial threshold vM — the Myersonian free trial parameters. These have the shortest trial and highest post-trial price among all equilibrium mechanisms, resulting in the minimum social surplus. The intuition is that the high-type seller signals credibly by proposing mechanisms that generate high revenue from post-trial price discrimination (which the low type cannot profit from), pushing toward maximum learning-based discrimination. All D1-surviving payoffs are Pareto dominated by the point H (the unconstrained IC-IR optimum) for any prior mu_0, and Pareto dominated by point B when mu_0 is small.

Q: Can having consumer preference data hurt the seller, and under what conditions?

A: Yes. The distortion from signaling incentives can be so large that both seller types earn strictly less in the D1-surviving equilibrium than they would if neither possessed private information (where the first-best is attained). This result holds when the condition of Proposition 2 is satisfied — i.e., when pi_F > lambdamu_0T. This contrasts sharply with the static result of Koessler and Skreta (2016), in which the ex-ante profit-maximizing mechanism is always supportable in equilibrium and data always (weakly) helps sellers.

Q: How do trial mechanisms differ from the prior literature on signaling through introductory prices?

A: The earlier literature (Milgrom and Roberts 1986; Bagwell 1987; Bagwell and Riordan 1991; Judd and Riordan 1994) uses two-period models with no seller commitment, so all pricing behavior is necessarily trial-like by model restriction. The present model instead allows the seller full flexibility to design any dynamic mechanism — including selling everything ex-ante, which would prevent buyers from gaining information rent. Trials emerge endogenously as the equilibrium outcome rather than being imposed by the model structure, and the paper provides new economic content on what determines trial length and price thresholds.

Q: What happens when the seller controls service quality in addition to access?

A: Section 6 extends the baseline by allowing the seller to choose (I, q) from a subset of [0,1]^2, where I governs the Poisson arrival rate and q scales the reward value (utility from a reward is v*q). Theorem 2 shows that the relevant equilibrium mechanisms now take the form of dynamic tiered pricing: a low-quality tier (interpreted as high ad load) provides learning opportunities while reducing information rents; once convinced, buyers upgrade to a premium high-quality tier. Enriching the screening technology in this way can reduce both revenue and social efficiency in equilibrium, because additional instruments create additional signaling opportunities that distort outcomes further from the revenue-maximizing benchmark.

Q: What are the two sources of welfare loss relative to the first-best in D1-surviving equilibria?

A: The welfare analysis in Appendix F identifies two sources. First, exclusion inefficiency: buyers with values v in [v_low, vM) who would generate positive surplus are excluded from post-trial service. Second, service truncation inefficiency: service access is cut off after trial length tM for buyers who were never convinced (theta = L type realizations and high-type buyers with v < vM), reducing total surplus below the first-best of mu_0 * lambda * T. Both losses are minimized (welfare is maximized) among trial mechanisms by longer trials and lower post-trial cutoffs, precisely the opposite of what D1 selects.

Q: Does the model extend to continuous seller types or multiple buyer types?

A: Appendix K outlines an extension to continuous seller types theta drawn from a distribution G on [theta_low, theta_high], where rewards arrive at rate lambdaItheta. The main economic forces persist: higher seller types anticipate faster buyer learning and have stronger incentives to offer trials. The main results generalize: equilibrium mechanisms are trial mechanisms, and under D1, pooling equilibria with maximum post-trial discrimination are selected. Appendix G similarly notes that the multiple-buyer-type extension preserves complete pooling and the D1 selection result.

Q: What is the role of the “global intertemporal constraint” (IC-U) in the proof of Theorem 1?

A: The canonical approach to dynamic mechanism design (Eso and Szentes 2007; Pavan, Segal, and Toikka 2014) relaxes the problem to only local incentive constraints on the initial report. This fails here because the informed seller causes buyer and seller to disagree on the evolution of buyer beliefs, making the timing of trade matter and requiring tracking of incentive constraints at every point in time. The paper identifies two key binding constraints in the relaxed problem: (IC-V) the buyer does not misreport their reward value, and (IC-U) the buyer does not remain silent about the arrival of a reward forever. Retaining only these two constraint families yields a tractable bang-bang solution for the optimal access policy, which is then verified to satisfy all original IC-IR constraints.

Q: What are the implications for platform design and data collection strategy?

A: The results imply that the value of consumer data depends critically on market dynamics. In static markets, collecting data about consumer match quality is weakly beneficial for sellers (Proposition 1, first point). In dynamic markets with buyer learning and sufficiently long service horizons, the same data can strictly reduce seller revenue by enabling a deviation that unravels first-best pricing. This suggests platforms in dynamic digital markets should weigh whether possessing and acting on proprietary match data improves or worsens their equilibrium position, and that regulatory attention to consumer data collection in dynamic markets may have welfare-ambiguous effects.

Trial mechanism: A dynamic mechanism parameterized by (v0, t0, p0) in which the seller provides full service access during [0, t0] for uninformed buyers, offers continued service after t0 only to buyers who received a reward with value v >= v0, and charges a post-trial price of p0 + lambdav0(T - t0) for those who qualify. In the paper’s usage, this is the unique outcome-implementing mechanism on the boundary of the IC-IR feasible payoff set.

Myersonian free trial: The limiting trial mechanism as the trial price epsilon approaches zero, with trial length tM = argmax_t {(1 - exp(-lambdat))(T - t)} and post-trial threshold vM equal to the Myerson monopoly price. It yields payoff pi_F = (1 - exp(-lambdatM))(1 - F(vM))lambdavM*(T - tM) to the high-type seller, and constitutes the binding outside option constraining equilibrium payoffs.

Belief gap: The divergence between the seller’s and buyer’s beliefs about the rate at which the buyer will receive Poisson rewards. Because the high-type seller knows theta = H, they anticipate a higher probability of reward arrival than the buyer’s prior implies. This gap makes the buyer’s belief process non-martingale from the seller’s perspective, breaking the standard dynamic mechanism design approach and creating profitable deviation incentives.

IC-IR feasible payoff set: The set of seller payoff pairs (pi_L, pi_H) achievable by mechanisms satisfying both incentive compatibility (for seller type reports and buyer learning reports) and individual rationality (non-negative ex-ante payoffs for all parties). Theorem 1 establishes that the boundary of this set is uniquely implemented by trial mechanisms.

Dynamic tiered pricing: The equilibrium mechanism form that emerges when the seller controls both access I and service quality q. It features a low-quality tier (high ad load) providing learning opportunities at reduced information rent, and a premium tier offering full quality to buyers convinced of high match quality. This generalizes trial mechanisms to settings with richer screening technology.

Global intertemporal constraint (IC-U): The constraint requiring that, upon receiving a Poisson reward, the buyer finds it suboptimal to remain silent about its arrival forever. Together with the local value-reporting incentive constraint (IC-V), these two constraints constitute the binding restrictions in the paper’s relaxed mechanism design problem, replacing the full continuum of incentive constraints that would otherwise be intractable.

D1 criterion: A standard equilibrium refinement from signaling games applied here to the space of mechanism proposals. Among all pooling equilibrium trial mechanisms, D1 selects only those with parameters (tM, vM) — the shortest trial length and highest post-trial threshold — because the high-type seller has a strictly larger set of buyer responses for which deviation to a high-discrimination mechanism is profitable. These surviving mechanisms Pareto dominate no other equilibrium mechanism and minimize social surplus.

Health Shocks, Health Insurance, Human Capital, and the Dynamics of Earnings and Health

Mon, 01 Jan 0001 00:00:00 +0000

Capatina and Keane build and calibrate a life-cycle model of labor supply and savings for U.S. men that incorporates health shocks, endogenous human capital accumulation via learning-by-doing, employer-sponsored health insurance (ESHI), means-tested social insurance, and endogenous medical treatment decisions. The model is calibrated to White males using the Medical Expenditure Panel Survey (MEPS) for 2000–2013, supplemented by CPS, HRS, and PSID data; separate calibrations are presented for Black and Hispanic men with high school or less education.

The paper’s central research question is how health shocks affect labor supply, earnings, and earnings inequality over the life cycle, and through which mechanisms. Four channels are identified and quantified: (1) the direct labor supply effect — sick days and reduced tastes for work caused by health shocks; (2) the human capital effect — reduced work experience from health-shock-induced employment exits, which deteriorates future job and wage offers in a snowball dynamic; (3) the health-productivity effect — reduced functional health directly lowering wage offers; and (4) the behavioral effect — anticipation of health risk induces low-skill workers lacking ESHI to curtail labor supply to maintain means-tested transfer eligibility.

The key quantitative findings from eliminating serious health shocks for working-age men (ages 25–64) are: the expected present value of lifetime earnings (PVE) for White men rises by 11% on average, and inequality in PVE falls by 12% (coefficient of variation). For White men with high school or less education the increase in PVE is 17.9%. For the typical White male the four channels contribute 5.7%, 2.7%, 1.4%, and 0.8% respectively. For low-skill White high school men the same channels contribute 10.7%, 14.8%, 1.3%, and 9.8% — with the human capital and behavioral effects dramatically larger for the low-skill group. For comparison, a severe health shock at age 40 reduces the present value of remaining lifetime earnings by 5.6% (approximately $53.9k) for a typical college man and by 11.5% (approximately $55.0k) for a typical high school man.

Human capital amplification operates through employment persistence: a major health shock causes full-time employment to drop by 12 percentage points one year after the shock for the average man, and by 20 percentage points for high school men, with recovery still incomplete eight years later (employment remains 7.8 pp and 10 pp below baseline, respectively). Holding human capital fixed as in the pre-shock baseline causes employment to recover quickly, confirming that persistent wage-offer deterioration is the mechanism.

On health insurance policy, the model evaluates providing public insurance to all workers lacking ESHI. This substantially increases medical utilization, improves health and life expectancy (survival to age 65 rises from 82% to 87% when health shocks are eliminated, as a related benchmark), reduces Medicaid and free-care costs, and raises labor supply among low-skill workers by weakening means-tested transfer incentives. The net program cost in a balanced budget simulation is modest, and all agent types are ex ante better off. By contrast, expanding Medicaid access creates perverse labor supply disincentives — workers reduce labor supply to maintain eligibility — does little to improve health, and makes almost all agents worse off in a balanced budget scenario.

Scope conditions: the primary calibration covers non-institutionalized civilian White males; results for Blacks and Hispanics are presented only for the high school or less education group due to small samples. The model period ends at 2013, before ACA implementation.

Q: What is the model’s overall estimate of how much health shocks reduce lifetime earnings for White men? A: Eliminating serious health shocks at working ages (25–64) would increase the expected present value of lifetime earnings (PVE) for the average White male by 11% and reduce inequality in PVE by 12% as measured by the coefficient of variation. For White men with high school or less education the PVE gain is larger at 17.9%.

Q: What are the four channels through which health shocks affect earnings, and how large is each for the average White male versus a low-skill high school male? A: The four channels are (1) direct labor supply via sick days and reduced tastes for work, (2) human capital deterioration from lost work experience worsening future job/wage offers, (3) reduced health productivity lowering wage offers, and (4) behavioral responses to health risk reducing labor supply to preserve transfer eligibility. For the average White male the contributions to PVE are 5.7%, 2.7%, 1.4%, and 0.8%, respectively. For low-skill White high school men the same channels contribute 10.7%, 14.8%, 1.3%, and 9.8% — the human capital and behavioral effects are roughly five to twelve times larger for the low-skill group.

Q: Why is the human capital effect so much larger for low-skill high school men than for college men? A: Low-skill high school men are much more likely to exit full-time employment following a major health shock and are slow to return. Lifetime work years decline by 1.89 for the typical high school man versus only 0.84 for the typical college man following a major shock at age 40. Because job offer probabilities depend on lagged employment, absence from the labor market creates a snowball effect that persistently depresses offer quality; human capital accounts for 42% of the earnings decline for high school men versus 34% for college men.

Q: How does the paper characterize the persistent employment effects of a major health shock? A: For the average man, full-time employment drops by 12 percentage points one year after a severe shock and remains 7.8 pp below baseline after eight years. For high school men the initial drop is 20 pp, still 10 pp below baseline after eight years; for college men the figures are 7 pp and 3 pp. When human capital is held fixed at the pre-shock baseline — so wage and job offers do not deteriorate due to lost experience — employment recovers quickly for workers of all skill levels, confirming the human capital mechanism drives the persistence.

Q: How does the behavioral effect operate for low-skill workers? A: Workers without ESHI who face health risk have an incentive to maintain sufficiently low income and assets to qualify for means-tested social insurance, which provides a consumption floor approximating Medicaid, Food Stamps, SSDI, and SSI. This perverse incentive leads low-skill workers to curtail labor supply preemptively. When health risk is eliminated, this incentive disappears and labor supply rises, generating the behavioral effect of 9.8% of PVE for low-skill high school men versus only 0.8% for the average White male.

Q: How does the paper correct for under-reporting of health shocks among the uninsured? A: The measurement model assumes health shocks are correctly measured for the treated, but uninsured workers who do not seek treatment only record a shock with a shock-specific probability less than one. A key identifying assumption is that, conditional on health status, risk factors, age, and education, the true frequency of health shocks does not differ by insurance status per se — ruling out ex ante moral hazard. The measurement model parameters are calibrated to match observed frequencies of health shocks and high risk in MEPS for the uninsured.

Q: What does the model estimate regarding the effect of a severe health shock on cumulative earnings relative to existing reduced-form evidence? A: The model predicts an average cumulative (non-discounted) earnings loss of $42.8k over ten years following a severe shock for men aged 50, compared with Smith’s (2004) estimate of $37k from the HRS. The paper argues Smith’s estimate identifies effects on workers who actually experience shocks, who are a selected sample with low baseline earnings (as untreated shocks are more likely to be severe, and non-treaters tend to have low earnings). The model’s “average effect” — comparing a world where everyone experiences the shock to one where no one does — yields a substantially higher loss of $59.8k.

Q: What are the key findings from the public insurance experiment (providing insurance to the uninsured)? A: Providing public insurance to all workers lacking ESHI substantially increases medical utilization among the previously uninsured, who are intrinsically less healthy. This improves health and life expectancy, raising Social Security costs. However, it also generates positive labor supply incentives for low-skill workers (reducing their reliance on means-tested transfers), substantially reduces Medicaid and free-care costs, and increases tax revenue. On balance, the net program cost in a balanced budget simulation is modest, and all types of workers are ex ante better off.

Q: Why does expanding Medicaid access produce perverse results in contrast to providing public insurance? A: Medicaid is means-tested, so expanded access requires workers to maintain sufficiently low income and assets to remain eligible. This creates disincentives to work and save — workers reduce labor supply to preserve eligibility. The result is reduced earnings, lower tax revenue, little improvement in health (as access to care depends on maintaining low income), and almost all agents being worse off in a balanced budget scenario.

Q: What role does insurance play beyond consumption smoothing in this model? A: Beyond lowering out-of-pocket (OOP) costs and smoothing consumption, insurance grants access to care: in the US system, proof of insurance is often required before treatment, so uninsured workers may not have the option to treat at all. The model captures three distinct option sets for the uninsured — all options available, treatment not available, or default not available — each motivated by different real-world contexts. Non-treatment worsens health transition probabilities, so the access-granting role of insurance independently affects health trajectories beyond its cost-reducing role.

Q: What explains the observed positive association between education, income, insurance, and health transitions in the data, and how does the model generate this without education entering the health production function directly? A: The association between education and health is largely driven by the positive correlation between education and latent health types; controlling for latent health type in a descriptive logit largely eliminates the education coefficient. The association between insurance and health transitions is driven by the fact that the insured are more likely to receive treatment; controlling for treatment and true shocks eliminates the insurance coefficient. Education affects health indirectly through its effects on treatment decisions — via wages, job offers with ESHI, and consumption capacity — without appearing as a direct argument in the health production function.

Q: How large are the effects of health shocks on key population health statistics according to the model? A: Eliminating serious health shocks at working ages would increase the fraction of working-age men in good health from 60% to 75% and raise the probability of survival to age 65 from 82% to 87%. Average annual sick days of 16.42 would be eliminated, implying a 6% increase in work days for employed workers and an employment rate increase from 88% to 91%. Average annual medical costs would fall from $4,618 to $1,132.

Q: How do the results for Black and Hispanic men compare to White men? A: The results are qualitatively similar, but the magnitudes for Black men are somewhat larger. Eliminating health shocks would raise PVE for Whites, Blacks, and Hispanics with high school or less education by 17.9%, 23.7%, and 17.7%, respectively. Separate access-to-care probabilities are calibrated for each group, reflecting racial disparities in access that explain part of the observed differences in health outcomes and treatment rates.

Q: What is the role of the consumption floor (means-tested social insurance) in shaping equilibrium outcomes for low-skill workers? A: The consumption floor guarantees a minimum household consumption level approximating Medicaid, Food Stamps, SSDI, and SSI. It shields low-skill workers from the full cost of health shocks, reducing both the consumption-smoothing value of ESHI and precautionary saving incentives. However, it also creates a powerful disincentive for low-skill workers without ESHI to work, as earning above the eligibility threshold would eliminate benefits. This mechanism amplifies earnings inequality by generating perverse labor supply behavior concentrated among low-skill, uninsured workers.

Functional Health (H): A discrete stock variable (Poor, Fair, or Good) measuring aspects of health that directly affect worker productivity and tastes for work; distinguished from asymptomatic health risk. Transitions depend on lagged health, latent health type, age, persistent health shocks, and whether shocks are treated.

Asymptomatic Health Risk (R): A binary state (low or high) capturing risk factors such as obesity, high cholesterol, and hypertension that increase the probability of future health shocks but do not affect current productivity.

Human Capital Effect: The channel by which health shocks reduce lifetime earnings not directly but indirectly — by causing employment exits that slow work experience accumulation, which in turn deteriorates future job offer probabilities and wage offers in a persistent, self-reinforcing (snowball) dynamic.

Behavioral Effect: The reduction in labor supply — and associated earnings loss — that occurs because workers facing health risk and lacking ESHI have an incentive to keep income and assets low enough to maintain eligibility for means-tested social insurance, even absent any contemporaneous health shock.

Tied Wage-Hours-Insurance Offer: The model’s labor market structure in which employment offers jointly specify a wage rate, hours (no offer, part-time, or full-time), and whether the offer includes ESHI; workers accept or reject the bundle rather than choosing hours and insurance independently.

Source Text Origin: The paper’s own term distinguishing how the full text of a paper was obtained (PDF, OA-HTML, or abstract-only); used in the summarization pipeline. [Note: this concept is from the summarization pipeline metadata, not from the paper itself — omitting.]

Treatment/Payment Options: The set of decisions available to a worker after a health shock occurs — whether to seek treatment and, if treated, whether to pay the out-of-pocket cost or default on bills. The available choice set differs by insurance status and context: the uninsured may face denial of access (option to treat unavailable) or required prepayment (default unavailable), or may have all options including free care.

Latent Health Type: An unobserved permanent individual characteristic capturing innate biological resilience and pre-age-25 health investments; determines baseline transition probabilities for functional health conditional on shocks. Positively correlated with latent skill type within education groups.

Illiquid Lemon Markets and the Macroeconomy

Mon, 01 Jan 0001 00:00:00 +0000

The paper develops a quantitative capital-accumulation model in which capital trades in illiquid markets with asymmetric information — sellers know the quality of their capital but buyers do not. It combines this model with microdata on nonresidential capital units listed for trade to measure the degree of information asymmetry and quantify its macroeconomic effects.

Model: The economy features heterogeneous capital units characterized by observed quality ω (e.g., size, location, age — observable to both buyers and sellers) and unobserved quality a (known only to the seller). Capital trades in directed-search markets: sellers post a price and a target submarket; buyers direct their search; a matching function determines trade probabilities. Buyers observe announced quality and have an inspection technology that reveals true quality with probability ψ (“lemon detection probability”); with probability 1−ψ a low-quality unit goes undetected. In equilibrium, sellers of high-quality capital signal their type by listing at higher prices and accepting lower trading probabilities (the Guerrieri-Shimer-Wright 2010 competitive search separating equilibrium, adapted to the capital accumulation setting). The key model prediction is that the residual price — the component of a listed price orthogonal to observed characteristics — is positively correlated with duration on the market, with the slope increasing as the degree of asymmetric information (1−ψ) rises.

Data: Idealista, Spain’s largest online real estate platform, provides monthly listings for all nonresidential structures (retail, office, and industrial space) listed for sale from 2005 to 2018 — approximately 8.9 million property-month observations from over 1.15 million distinct capital units. The average listed price per square foot is $162 (2017 dollars); the average duration on the market is 10.5 months; each listing receives on average 800 views, 45 clicks, and 3 emails per month from prospective buyers.

Empirical facts (Section 4): Two cross-sectional regularities confirm the model’s predictions:

Predicted price (from a hedonic regression on observable characteristics) is negatively correlated with duration — units with better observable characteristics sell faster, consistent with full-information competitive search (higher buyer valuation → higher matching rate)
Residual price (orthogonal to observables) is positively correlated with duration — estimated slope coefficient ŷq ≈ 0.148 — consistent with asymmetric-information signaling (high-quality capital sellers post high residual prices to separate from low-quality sellers, accepting lower trading probabilities)
The residual-price/duration slope exhibits strong countercyclical variation, roughly doubling during the Euro crisis (peak slope ≈ 0.38, compared to baseline ≈ 0.148), consistent with asymmetric information worsening during downturns

Calibration (monthly frequency, Table 4 fixed; Table 5 fitted):

Fixed parameters: β = 0.9966 (annual rate of time preference 4%), α = 0.35 (capital share), δ = 0.0074/month (8.5% annual nonresidential depreciation), γ = 1.004 (1.6% annual TFP growth), γn = 1.0027 (1% annual population growth), ϕ = 0.0027 (3.2% annual firm exit rate), η = 0.8 (matching curvature), φ = 0.5 (seller bargaining power)
Fitted to four data moments (slope ŷq, SD of predicted prices, SD of residual prices, mean duration): ψ = 0.9795 (probability a lemon goes unnoticed = 2% per inspection); σω = 0.72 (SD observed quality); σa = 0.58 (SD unobserved quality); m̄ = 0.267 (matching efficiency)
Model-simulated moments match targets essentially exactly (Table 5); untargeted relationship between duration and predicted prices is also well-matched (Table 6)

Steady-state output effects (Table 7, relative to full-information benchmark):

Total output: −1.22% in baseline (ψ = 0.9795)
Effective capital input: −2.55% (main driver of output loss)
Capital stock: −1.12% (32% of output effect — reduced returns to producing new capital)
Capital unemployment rate: +1.0 pp above full-information rate of 5% (25% contribution — high-quality capital remains listed longer)
Allocation channel: 16% contribution — information asymmetries disproportionately reduce trading of high-quality capital, lowering average quality of employed capital
Labor input: −0.5% (26% contribution — reduced capital input lowers labor demand)
Moving to full information (ψ → 1): output gain of +1.5% — modest at baseline, indicating the baseline economy is not far from full information
Moving to Euro-crisis level (ψ = 0.96): output decline of ~2% — large response because the economy’s output elasticity to ψ is high

Crisis experiment (Section 5.3): An unexpected 2 percentage-point decline in ψ (to 0.96, calibrated to match the observed increase in the residual-price/duration slope during the Euro crisis), lasting 3 years and reverting with persistence ρψ = 0.94:

Output contraction on impact: 2%
Time to recover half the output decline: more than 5 years (slow recovery driven by persistent capital underinvestment)
Primary mechanism: lower inspection accuracy → high-quality capital sellers reduce trading probability to signal quality → capital unemployment rate rises (especially for high-quality units) → expected return to producing new capital falls → investment contracts → capital input declines persistently
Secondary interaction: at higher steady-state asymmetric information (ψ = 0.96), other shocks (TFP, exit rate, discount factor) are amplified — e.g., the cumulative output response to an exit rate shock is 26% larger than in a full-information economy

Scope conditions: The model abstracts from aggregate uncertainty (the baseline is steady-state analysis), financial intermediaries, and endogenous information technology. The dataset covers Spain’s nonresidential real estate market 2005–2018; the measurement of ψ from listed prices and duration assumes that residual prices fully reflect unobserved capital quality (Proposition 5’s small-search-cost approximation). The quantitative results are robust to alternative bargaining protocols (TIOLI), higher firm exit rates, inelastic labor supply, and narrower observable-characteristic sets.

Summary of a forthcoming paper, AI-assisted and human-reviewed. See the linked original for the authoritative claims and full conditions.

In depth

Q1. Why does asymmetric information generate a positive correlation between residual prices and duration?

In the model’s separating equilibrium, sellers of high-quality capital choose prices and targeting strategies that prevent low-quality sellers from mimicking them; since low-quality sellers have a lower marginal cost of accepting lower trading probabilities (their capital is worth less to them in continued use), high-quality sellers can separate by listing at higher residual prices paired with lower market tightness and lower matching rates. The correlation between residual price and duration is therefore a direct measure of the degree of asymmetric information: the slope coefficient ŷq increases monotonically as ψ decreases (Proposition 5 and Figure 4), allowing the researcher to back out ψ from the micro data.

Q2. Why is the residual-price/duration slope countercyclical?

The data show that the slope roughly doubled during Spain’s 2008–2013 downturn and euro crisis, consistent with the model’s prediction that asymmetric information (1−ψ) worsens during economic contractions. The paper interprets this as evidence that buyers’ ability to evaluate capital quality deteriorates when economic uncertainty rises — for example, during crises it is harder to assess the profitability of retail or office space based on observable characteristics alone. This countercyclical pattern motivates the crisis experiment in Section 5.3, where a 2pp increase in 1−ψ (the degree of information asymmetry) replicates the observed slope dynamics.

Q3. Why is the 2% crisis output contraction slow to recover?

The sluggishness of recovery operates through the investment channel: when high-quality capital sellers reduce trading probabilities to signal their type, they slow the transfer of used capital from sellers (firms that exit) to buyers (firms that expand), reducing the effective capital input; this lower capital input reduces the expected marginal return to producing new capital, depressing investment; because capital accumulates gradually, the output recovery inherits the slow pace of investment recovery. The persistence parameter ρψ = 0.94 (monthly) adds further sluggishness from the slow normalization of the information environment itself.

Q4. Why are the steady-state output losses modest while the crisis response is large?

The economy features a moderate baseline degree of asymmetric information (ψ = 0.9795 — only 2% lemon-detection failure), so the steady-state distortion is small (−1.22% output relative to full information); however, the economy has a large elasticity of output to ψ, so even a small deterioration in information quality (2pp) generates large output effects (−2%). This high sensitivity arises because the effects of asymmetric information are highly nonlinear: at low levels of information frictions, small increases in the lemon probability generate proportionally large increases in the required signaling by high-quality sellers, sharply reducing their trading probabilities.

Q5. How does asymmetric information interact with other shocks?

At the baseline degree of asymmetric information (ψ = 0.9795), the aggregate responses to standard shocks (TFP, discount factor, exit rate) are similar to an economy with full information; however, at the Euro-crisis level (ψ = 0.96), the cumulative output response to an exit rate shock is 26% larger than under full information. The mechanism is that asymmetric information taxes the reallocation of capital: when more capital must be reallocated (due to higher firm exit), more of it passes through the illiquid, distorted lemon market, amplifying the output effect of the underlying shock.

Q6. What policies can reduce the distortions from asymmetric information?

The paper notes two broad policy directions: (1) policies that improve information transparency — making previously private capital characteristics public, e.g., mandatory disclosure or standardized quality certification — directly raise ψ and shift the economy toward full information, eliminating the signaling distortion; (2) policies that reduce the incentive for mimicking — for example, by allowing post-transaction renegotiation after quality is revealed (the TIOLI bargaining extension in Table 8) — have similar quantitative effects to the baseline. The paper leaves the welfare analysis of specific information-provision policies for future research.

Q7. What is the role of the data in identifying the model parameters?

The four targeted moments — slope of duration on residual prices, standard deviation of predicted prices, standard deviation of residual prices, and mean duration — jointly identify the four structural parameters {ψ, σω, σa, m̄} (Proposition 5); the key insight is that ψ and m̄ are separately identified because ŷq and mean duration respond differently to each: ψ and m̄ both affect ŷq positively, but m̄ reduces mean duration while ψ increases it, providing orthogonal variation. The calibration achieves an essentially exact match of the four targeted moments (Table 5) and also matches the untargeted negative slope between duration and predicted prices (Table 6), providing an overidentification check.

Key concepts

lemon market : a secondary market for heterogeneous assets in which sellers have private information about quality; following Akerlof (1970), lemons (low-quality assets) crowd out high-quality assets unless high-quality sellers can credibly signal their type; in the paper, signaling takes the form of higher listed prices paired with lower trading probabilities.

residual price : the component of a capital unit’s listed price orthogonal to its observable characteristics (the residual from a hedonic regression); the paper’s key empirical variable, theoretically shown to be positively correlated with unobserved capital quality and with duration under asymmetric information.

inspection technology : a buyer’s technology that reveals the true quality of a capital unit with probability ψ before (or after) purchase; the accuracy ψ governs the degree of asymmetric information in the economy — lower ψ implies worse information, requiring more costly signaling by high-quality sellers.

countercyclical asymmetric information : the empirical finding that the slope between residual prices and duration roughly doubles during the Euro crisis, interpreted as deterioration in buyers’ ability to evaluate capital quality during economic downturns; motivates the crisis experiment.

three channels of output loss : the three mechanisms through which asymmetric information reduces output: (i) lower capital stock (reduced investment incentives); (ii) higher capital unemployment rate (high-quality capital remains listed longer); (iii) adverse allocation effect (high-quality capital trades less frequently, lowering average quality of employed capital).

Linking Social and Personal Preferences: Theory and Experiment

Mon, 01 Jan 0001 00:00:00 +0000

This paper asks whether an individual’s attitude toward risk in the personal domain (choices affecting only oneself) can be linked to that same individual’s attitude toward risk in the social domain (choices affecting both oneself and others). The authors provide a theoretical answer in the form of necessary and sufficient conditions, and then test those conditions experimentally.

The formal model posits a decision maker (DM) with a preference relation over lotteries on a set of social states, where a distinguished subset of states are personal (consequences for the DM alone). The authors assume preferences satisfy Completeness, Transitivity, Continuity, and State Monotonicity — the last being equivalent to respect for First-Order Stochastic Dominance (FOSD), a condition weaker than the Expected Utility Independence Axiom and satisfied by virtually all extant decision theories including Weighted Expected Utility, Rank-Dependent Utility, and Prospect Theory. The key theoretical result (Theorem 1) establishes that the full preference relation over all social lotteries can be uniquely deduced from the partial observations of (i) riskless social choices and (ii) risky personal choices if and only if the DM finds every social state indifferent to some personal state. When this condition fails, there exist social lotteries whose ranking cannot be recovered from the partial data.

For two empirically relevant preference types, this condition generates directly testable predictions: for selfish subjects (who allocate nothing to others in deterministic social choices), risky personal preferences must coincide with risky social preferences; for impartial subjects (who treat self and other symmetrically in deterministic social choices), riskless social preferences must coincide with risky social preferences.

The experiment was conducted at the University of Bergen and NHH Norwegian School of Economics with 276 undergraduate subjects. Each subject faced 50 budget-line choice problems in each of three domains: Personal Risk (equiprobable binary lotteries over own payoffs only), Social Choice (deterministic splits between self and an anonymous other), and Social Risk (equiprobable binary lotteries over symmetric payout pairs for self and other). The graphical interface of Choi et al. (2007b) was used throughout. One randomly selected decision per domain was paid out; each token was worth 1.2 NOK (approximately 0.2 USD), with average earnings of approximately 270 NOK.

Within-domain consistency, measured by the Critical Cost Efficiency Index (CCEI), is high: mean CCEIs are 0.959, 0.952, and 0.902 in the Personal Risk, Social Choice, and Social Risk domains respectively. At the CCEI > 0.90 threshold, 89.9%, 85.9%, and 69.9% of subjects pass in the three domains. Using a 0.95 share-to-self threshold, 103 subjects (37.3%) are classified as selfish; using revealed-preference criteria at the 5% significance level, 33 subjects (12.0%) are classified as impartial.

Testing is done via an individual-level nonparametric permutation test that draws 10,000 random data sets per subject and compares simulated CCEI distributions to actual cross-domain CCEIs, with Bonferroni correction. At the 1% significance level, the null that Personal Risk and Social Risk preferences coincide is rejected for only 5.9%–9.3% of selfish subjects (varying by classification threshold), compared with 14.7%–16.3% rejection rates for non-selfish subjects. For impartial subjects at the 1% level, the null that Social Choice and Social Risk preferences coincide is rejected for 0.0%–11.1%, compared with 19.8%–26.8% for non-impartial subjects. The theory’s predictions are thus supported for a large majority of both selfish and impartial subjects.

A theoretical extension (Theorem 2) shows that if one additionally observes comparisons between social states and personal lotteries, unique deduction of the full preference relation requires that preferences in both personal and social domains satisfy Expected Utility (Independence Axiom) and that every social state is indifferent to some personal lottery — a strictly stronger set of conditions.

Q: What is the central theoretical question and why does it matter? A: The paper asks whether preferences over risky social choices (lotteries over outcomes for self and others) can be deduced from observing only riskless social choices and risky personal choices. This matters because people frequently observe or predict the risky social choices of leaders and representatives, but may have access only to those leaders’ personal risk-taking behavior and their expressed social preferences under certainty.

Q: What is the main theoretical result (Theorem 1)? A: Under Completeness, Transitivity, Continuity, and State Monotonicity, the unique extension of the partial preference relation (over social states and personal lotteries) to the full domain of social lotteries exists if and only if every social state is indifferent to some personal state. When this condition is not met, multiple distinct preference relations can extend the partial observations, making deduction impossible.

Q: What is State Monotonicity and how does it relate to standard axioms? A: State Monotonicity requires that if each social state in one lottery dominates the corresponding state in another lottery, then the first lottery is weakly preferred. The paper shows this is equivalent to respect for First-Order Stochastic Dominance (FOSD) given the other axioms, and is strictly weaker than the von Neumann–Morgenstern Independence Axiom. It is satisfied by Weighted Expected Utility, Rank-Dependent Utility, and Prospect Theory, making it a broadly applicable assumption.

Q: What are the testable predictions for selfish subjects? A: Proposition 2 establishes that if a subject’s Social Choice preferences are selfish — meaning any bundle (x, y) is indifferent to (0, y), so the subject is indifferent between keeping x for self and giving it to other — then preferences in the Personal Risk domain must coincide with preferences in the Social Risk domain. In the experiment, selfish subjects are those allocating more than 95% of tokens to themselves in the Social Choice domain (103 of 276 subjects, or 37.3%).

Q: What are the testable predictions for impartial subjects? A: Proposition 3 establishes that if a subject’s Social Choice preferences are symmetric — meaning (x, y) is indifferent to (y, x) for all pairs — then preferences in the Social Choice domain must coincide with preferences in the Social Risk domain, implying risk neutrality toward social lotteries. The intuition is that such a subject treats self and other identically, so risky splits are evaluated by expected value alone. In the experiment, 33 subjects (12.0%) are classified as impartial by the revealed-preference criterion at the 5% significance level.

Q: How does the experiment measure within-domain rationality? A: Choices within each domain are evaluated using the Critical Cost Efficiency Index (CCEI, following Afriat 1967), which measures how much a budget constraint must be relaxed to remove all GARP violations. Mean CCEIs are 0.959 (Personal Risk), 0.952 (Social Choice), and 0.902 (Social Risk). At the CCEI > 0.90 threshold, 248 subjects (89.9%), 237 (85.9%), and 193 (69.9%) pass in the three domains respectively, compared to a simulated mean CCEI of only 0.585 for subjects randomizing uniformly.

Q: How does the cross-domain test work and why is it nonparametric? A: The test uses individual-level permutation inference: under the null that preferences in domains I and J are identical, any 50-element subset drawn from the pooled 100 choices should satisfy GARP as well as the actual domain-specific choices. For each subject, 10,000 such random draws are generated, their CCEI scores are computed, and the distribution is compared to the actual cross-domain CCEI with Bonferroni correction. The test makes no functional form assumptions about utility and accommodates the observed within-domain errors without parametric error modeling.

Q: What are the rejection rates for the selfish-subject prediction? A: At the 1% significance level, the null that Personal Risk and Social Risk preferences coincide is rejected for only 5.9%–9.3% of selfish subjects (range across four classification thresholds from 0.99 to 0.90 share-to-self), compared to 14.7%–16.3% for non-selfish subjects. At the 5% level, rejection rates rise to 20.4%–25.6% for selfish and 22.4%–31.8% for non-selfish subjects.

Q: What are the rejection rates for the impartial-subject prediction? A: At the 1% significance level, the null that Social Choice and Social Risk preferences coincide is rejected for 0.0%–11.1% of impartial subjects (range depending on threshold and classification method), compared to 19.8%–26.8% for non-impartial subjects. At the 5% and 10% levels, rejection rates for impartial subjects range from 0.0% to 22.2%.

Q: Does the theory predict how risk aversion should map across domains for non-selfish, non-impartial subjects? A: The theory does not directly produce testable cross-domain predictions for subjects who are neither selfish nor impartial without additional parametric assumptions, because the specific personal-state equivalent of each social state depends on the form of preferences. The paper restricts its nonparametric tests to the two polar cases where the equivalence mapping is determinate from social choice behavior alone.

Q: What is the extended result (Theorem 2) and what stronger conditions does it require? A: When one additionally observes comparisons between social states and personal lotteries (not just within each domain separately), unique deduction of the full preference relation is possible if and only if preferences in both the personal and social domains are consistent with an Expected Utility representation and every social state is indifferent to some personal lottery. This requires the Independence Axiom — a strictly stronger condition than State Monotonicity — highlighting that the main Theorem 1 result exploits the weaker observational structure.

Q: What is the distribution of social preferences in the sample? A: Of 276 subjects, 103 (37.3%) are classified as selfish at the 0.95 share-to-self threshold. Only 6 subjects (2.2%) kept fewer than 0.45 of tokens on average, making purely altruistic subjects rare. In the Personal Risk domain, 41 subjects (14.9%) allocated more than 95% to the cheaper account (consistent with risk neutrality), while 9 (3.3%) allocated fewer than 55% (consistent with infinite risk aversion). In the Social Risk domain, 30 subjects (10.9%) are consistent with utilitarianism in money and 9 (3.3%) with Rawlsianism in money.

Q: How does the Social Risk domain compare to the Personal Risk and Social Choice domains in terms of rationality scores? A: The Social Risk domain shows lower consistency than the other two: mean CCEI is 0.902 versus 0.959 and 0.952, and only 69.9% of subjects exceed the 0.90 threshold versus 89.9% and 85.9%. The CCEI distribution is shifted left for Social Risk, suggesting the novel combined dimension of social and risky choice introduces more decision complexity or error.

Q: What is the relationship to the prior experimental literature on social and risk preferences? A: The Personal Risk domain replicates the symmetric risk experiment of Choi et al. (2007a), and the Social Choice domain replicates the linear two-person dictator experiment of Fisman et al. (2007). The Social Risk domain is new to this paper. The theoretical framework connects to Saito (2013) on social preferences under risk, and to the preference extension literature of Grant et al. (1992) and Nishimura et al. (2017).

State Monotonicity: The axiom requiring that if each social state in one lottery weakly dominates the corresponding social state in another lottery, the first lottery is weakly preferred. The paper proves this is equivalent to respect for First-Order Stochastic Dominance given Completeness, Transitivity, and Continuity, and distinguishes it from the stronger Independence Axiom by noting that Independence compares lotteries over lotteries while State Monotonicity only compares lotteries over states.

Selfish preferences (in the paper’s sense): Preferences in the Social Choice domain such that (x, y) is indifferent to (0, y) for all bundles — the subject is indifferent between receiving x themselves versus giving x to the other person. Operationally measured as allocating more than a threshold share (e.g., 95%) of tokens to self across Social Choice decisions.

Impartial preferences (in the paper’s sense): Preferences in the Social Choice domain such that (x, y) is indifferent to (y, x) for all bundles — the subject treats self and other symmetrically. Operationally identified by the revealed preference criterion that choices in the Social Choice domain satisfy GARP and are consistent with symmetric treatment.

Unique extension (deducibility): The property that there exists exactly one complete preference relation over all social lotteries that is consistent with the axioms and agrees with the observed partial relation over social states and personal lotteries. Theorem 1 identifies the necessary and sufficient condition for unique extension under State Monotonicity.

Personal state indifference condition: The condition that for every social state omega in Omega minus P, there exists some personal state in P to which the DM is indifferent. This is the necessary and sufficient condition in Theorem 1 for deducibility of the full preference relation. Interpreted as: for every proposed social allocation, there exists a “bribe” — a personal allocation with nothing for others — that the DM finds equally desirable.

Critical Cost Efficiency Index (CCEI): A measure of how much budget constraints must be scaled down to eliminate all GARP violations in a dataset of choices from budget lines (following Afriat 1967). A CCEI of 1 indicates perfect rationality; the paper uses 0.90 as a practical threshold. Mean values are 0.959, 0.952, and 0.902 in the Personal Risk, Social Choice, and Social Risk domains respectively.

Nonparametric permutation test: The individual-level test used to assess consistency across choice domains. Under the null that preferences are identical in domains I and J, any random 50-element draw from the pooled 100 choices should achieve CCEI scores no worse than the actual domain scores. The test draws 10,000 permuted datasets per subject and uses the Bonferroni correction for multiple comparisons, making no assumptions about the functional form of utility.

Markups: A Search-Theoretic Perspective

Mon, 01 Jan 0001 00:00:00 +0000

What this paper finds — and why it matters

Across macroeconomics, market power is almost always modelled with the Dixit–Stiglitz (1977) monopolistic-competition framework, in which a seller’s markup is pinned down by how substitutable buyers perceive its variety to be. This paper instead derives a closed-form formula for the equilibrium distribution of markups in the search-theoretic model of imperfect competition of Butters (1977), Varian (1980) and Burdett–Judd (1983), where a seller has market power not because its good lacks substitutes but because search and information frictions leave some buyers unable to reach the cheapest seller. In this model markups are strictly positive even though all sellers’ varieties are perfect substitutes, are dispersed even when all sellers operate the same technology, and — once sellers differ in marginal cost — can be increasing, decreasing, or constant in a seller’s size; yet the equilibrium is efficient. Menzio proves an “anything-goes” result: any twice-differentiable markup function can arise as an equilibrium for an appropriate choice of parameters, so a Dixit–Stiglitz model can always reproduce the search model’s markups — but only with reduced-form buyer preferences that depend on the search model’s deep parameters and are therefore unstable to policy changes (a Lucas-critique problem), and that would (incorrectly) read those markups as symptoms of inefficiency and a case for corrective subsidies. The paper’s central and deliberately modest claim is a cautionary one for macroeconomics: because two well-established models can both match observed markups yet imply opposite conclusions about welfare, optimal policy, and counterfactuals, markup data alone cannot identify the macroeconomic consequences of market power — one also needs evidence on the origin of that market power. The results are theoretical (unit demand, constant returns to scale, a Poisson contact process); the sharp comparative statics are derived for a log-uniform cost distribution, and the same logic extends to labor-market markdowns in the Burdett–Mortensen (1998) model.

Summary of a forthcoming paper, AI-assisted and human-reviewed. See the linked original for the authoritative claims and full conditions.

In depth

Q1. What two theories of market power does the paper compare, and how do they differ at root?

The paper contrasts the Dixit–Stiglitz (1977) monopolistic-competition framework, in which market power comes from product differentiation, with the search-theoretic framework of Butters (1977), Varian (1980) and Burdett–Judd (1983), in which market power comes from buyers’ limited choice sets. In Dixit–Stiglitz, “every seller is a monopolist of its own product variety,” and the size of markups “is determined by the substitutability of different varieties in the buyers’ utility function.” In the search-theoretic framework, by contrast, “a seller has market power not because it carries a good that has no perfect substitutes, but because (some) buyers do not have every seller in their choice set due to informational frictions … or physical frictions,” so markups are instead “determined by the distribution of the size of buyers’ choice sets.” Menzio motivates the second view with retail examples (e.g., the same bottle of Heinz ketchup sold at many stores at different markups), where it strains credulity that buyers see one store’s bottle as a poor substitute for the identical bottle elsewhere.

Q2. What is the equilibrium markup formula when all sellers are identical?

With homogeneous sellers, a seller at quantile x of the price distribution charges a gross markup μ(x) = 1 + (u/c − 1)·e^(−λ(1−x)), the product of a monopoly markup and a rank-dependent discount factor. Here u is the buyer’s valuation, c the common marginal cost, and λ the Poisson coefficient for the number of sellers a buyer contacts — “the average number of sellers with which a buyer is in contact, and, in this sense, … a measure of the extent of competition in the market.” The term u/c − 1 is “the net markup for a monopolist.” The discount factor e^(−λ(1−x)) “is equal to 1 for the seller at the top of the price distribution” (no discounting) and falls to its minimum e^(−λ) for the seller at the bottom; a higher λ makes markups decline more steeply down the price ranking. The equilibrium price distribution and its support are derived in closed form (F(p) and the lowest price p_ℓ = c + e^(−λ)(u − c)), and the equilibrium is shown to exist, be unique, and be efficient (Proposition 1).

Q3. Why are markups positive and dispersed even when goods are perfect substitutes and technology is identical?

Markups are positive because search frictions leave some buyers “captive” — in contact with only one seller — which forces equilibrium profits, and hence prices, strictly above marginal cost; markups are dispersed for the same reason there is price dispersion in these models — non-captive buyers prevent any mass point in the price distribution. As Menzio puts it, “sellers meet a positive measure of buyers that are captive, in the sense that these buyers cannot purchase from any other seller,” so “prices must be strictly above marginal cost”; simultaneously, the positive measure of non-captive buyers “implies that the price distribution cannot have any mass points above marginal cost.” The two facts together require sellers to post different prices and therefore charge different markups, despite identical goods and identical technology.

Q4. In the homogeneous-seller case, how do markups relate to a seller’s price and size?

With identical sellers, markups are increasing in a seller’s price and decreasing in a seller’s size. Because μ(x) and the posted price p(x) both rise with rank x while quantity sold q(x) = bλ·e(−λx) falls with x, “markups are increasing in the seller’s price” and “decreasing in the seller’s size.” Menzio notes this is the opposite of “Marshall’s second law of demand,” and that it implies larger sellers face a higher elasticity of demand. He stresses this counterfactual pattern (empirically, larger firms tend to charge higher markups) is exactly why the paper goes on to add cost heterogeneity.

Q5. What changes when sellers differ in marginal cost?

With heterogeneous marginal costs, the markup formula gains an extra term reflecting that higher-ranked (higher-cost) firms put less competitive pressure on a seller, and equilibrium markups need no longer be decreasing in size — they can be increasing, decreasing, or hump-shaped. A seller’s price is a strictly increasing function of its cost (Lemma 3), so its rank in the price distribution equals its rank in the cost distribution. The generalized markup (eq. 3.22) adds, to the monopoly-times-discount term, “the additional markup that the seller can charge because the firms ranked above it in the price distribution produce at higher marginal cost,” with the excess cost of nearer-ranked firms weighted more heavily. Using a phase-diagram (nullcline) analysis, Menzio shows the markup function μ(x) can be strictly increasing, strictly decreasing, or hump-shaped in rank depending on parameters. The heterogeneous-cost equilibrium is again shown to exist, be unique, and be efficient (Proposition 2).

Q6. What is the “anything-goes” theorem, and why does it matter?

Menzio proves (Theorem 3) that any twice-continuously-differentiable markup function μ(x) > 1 can be generated as an equilibrium of the search-theoretic model, given an appropriate contact intensity λ and cost distribution c(x).* Concretely, for any target markup schedule there is a λ and a quantile cost function c(x) (given in closed form) that deliver it as the equilibrium outcome. The consequence is sharp: “the search-theoretic model of market power can rationalize any pattern of markups observed in the data,” so “markup data cannot be used to reject the search-theoretic model.” Combined with the fact that the Dixit–Stiglitz model can reproduce the same markups, both theories are consistent with any markup evidence — which is the crux of the paper’s identification argument.

Q7. Can a Dixit–Stiglitz model reproduce these markups, and at what cost?

Yes — a Dixit–Stiglitz model can always reproduce the search model’s markups, but only with reduced-form buyer preferences that depend on the search model’s deep parameters (λ, u, c, b) and are therefore unstable. Menzio constructs the buyer utility function v(q) (its marginal utility solves a differential equation, eq. 2.24) that makes a Dixit–Stiglitz seller choose the same markups and quantities as in the search model. That reduced-form utility has v’(q) decreasing (so varieties look like imperfect substitutes, rationalizing positive markups) and an elasticity of demand that rises with q (rationalizing markups that fall with size). Critically, “the reduced-form utility function depends on the parameters of the search-theoretic model” and so “is unstable, in the sense that changes in the environment and counterfactual experiments lead to changes in the reduced-form utility function” — meaning any policy or counterfactual exercise that holds these preferences fixed “would not produce valid predictions,” i.e., is subject to the Lucas critique.

Q8. Why would reading these markups through the Dixit–Stiglitz lens give the wrong welfare and policy conclusions?

Because in Dixit–Stiglitz positive and heterogeneous markups signal inefficiency and call for subsidies, whereas the search-theoretic equilibrium that generated those very markups is efficient. Through the Dixit–Stiglitz lens, positive net markups imply “sellers produce an inefficiently small quantity,” and heterogeneous markups imply misallocation across sellers, leading an analyst to “recommend the introduction of consumption subsidies” and “finely-tuned production subsidies that reallocate inputs and consumption from low to high-markup sellers.” “None of these welfare and policy implications are, however, correct, since the equilibrium of the search-theoretic model … is efficient.” The root of the error is the demand curve’s interpretation: the quantity q(p) − q(c) a seller does not sell is, in Dixit–Stiglitz, lost gains from trade (an inefficiency), but in the search model it is “equally valuable trades that the buyers make with other sellers,” and so is not an inefficiency.

Q9. What determines the level and shape of the markup distribution?

For a log-uniform cost distribution (Theorem 4), markups decrease with the extent of competition λ, increase with the buyers’ valuation u, decrease with the highest marginal cost c_h, and increase with the rate κ at which marginal costs decline across sellers; the sign of the markup–size relationship flips at parameter thresholds. Specifically, the markup function is strictly decreasing in rank x (markups rising with size) when competition is weak (λ below a cutoff λ*), constant when λ = λ*, and strictly increasing in x (markups falling with size) when λ > λ*; analogous thresholds u* and κ* govern the slope’s sign as u and κ vary. The intuition: when λ is low, sellers rarely compete for the same buyers and low-cost sellers face little pressure, so markups are high and higher for low-cost (large) sellers; when λ is high, low-cost sellers are pushed toward marginal-cost pricing while high-cost sellers — facing no pressure from above — retain markups near u/c_h. Menzio notes the monotone-level results (markups decreasing in λ and c_h, increasing in u and in κ(x) = c’(x)/c(x)) generalize beyond the log-uniform family to arbitrary cost distributions, while the slope-sign results are stated for the log-uniform case.

Q10. What is the bottom-line claim for macroeconomics?

Markup data alone are insufficient to draw conclusions about the welfare, policy, and counterfactual consequences of market power; identifying those consequences requires evidence on the source of market power — product differentiation versus search/information frictions. The paper frames this as “a cautionary note to the macroeconomic literature that uses the Dixit–Stiglitz framework to model market power and markups” — a literature spanning monetary policy (e.g., Blanchard–Kiyotaki 1985; Christiano, Eichenbaum and Evans 2005; Golosov and Lucas 2007), misallocation and aggregate TFP (Hsieh and Klenow 2009), and the gains from trade (Krugman; Melitz 2003). In Dixit–Stiglitz estimations, markup heterogeneity is “quantitatively important” for the welfare cost of inflation in sticky-price models (Galí 1995), the gains from trade (Dhingra and Morrow 2019), and the cost of market power (Boar and Midrigan 2024); Menzio’s point is that “neither the level nor the dispersion of markups observed in the data are necessarily symptomatic of any inefficiency.”

Q11. Does the paper claim the search-theoretic model is the correct one?

No — the paper explicitly does not argue that the search-theoretic model is closer to the truth than monopolistic competition; it makes the “more modest, but not unimportant” claim that two sensible, well-established models fit the same markup data yet imply very different welfare, policy, and counterfactual conclusions. Menzio notes both theories “are likely to be overly simplified descriptions of the world,” and that the existence of still other models generating the same markups “only strengthens” the point. The constructive takeaway he poses is an empirical identification question: “How much of the downward sloping demand curve facing a seller is due to the heterogeneity in buyer’s outside options and how much is it due to preferences?”

Q12. Does the argument extend beyond product markets?

Yes — the same logic applies to the labor market: in the Burdett–Mortensen (1998) search model one can derive a closed-form formula for equilibrium markdowns that are positive even when employers are perfect substitutes to workers, are heterogeneous even with identical technology, and may be increasing, decreasing, or constant in firm size, with the equilibrium again efficient. Menzio concludes that “the same caution that I recommend using when interpreting markups should be applied to the interpretation of markdown data.”

Q13. What are the scope conditions, and what does the paper not do?

The results are theoretical, derived under unit buyer demand, constant returns to scale, and a Poisson process for the number of sellers each buyer contacts; the closed-form comparative statics of Theorem 4 assume a log-uniform marginal-cost distribution; and the paper offers no empirical calibration or estimation. Menzio notes the efficiency result depends on the model’s assumptions — relaxing unit demand or adding externalities could make the equilibrium inefficient — but argues this does not weaken the core identification point. A companion paper (Menzio 2024b, NBER WP 33253) shows the efficiency of the search-theoretic equilibrium extends to a general-equilibrium setting with endogenous firm entry. The paper’s contribution is an analytical characterization and a cautionary/identification argument, not a quantitative welfare estimate.

Key concepts

Search-theoretic model of imperfect competition : The Butters (1977)/Varian (1980)/Burdett–Judd (1983) framework in which sellers carry identical (perfectly substitutable) goods, and market power arises because buyers contact only a random subset of sellers — so some buyers are “captive” to a single seller. Markups are determined by the distribution of buyers’ choice-set sizes, not by preferences over differentiated varieties.

Dixit–Stiglitz monopolistic competition : Any model in which each seller is the sole producer (monopolist) of its own variety, sets its price, and is too small to affect the aggregate; the size of markups is governed by the substitutability of varieties in buyers’ utility (CES, VES, translog, or Kimball preferences all qualify in the paper’s usage).

Gross / net markup : The gross markup μ is the ratio of a seller’s posted price to its marginal cost (p/c); the net markup is μ − 1.

Captive vs. non-captive buyers : A captive buyer is in contact with only one seller and so cannot shop around (the source of strictly positive markups); a non-captive buyer is in contact with several sellers and buys from the cheapest (the source of price dispersion and the absence of mass points in the price distribution).

λ (extent of competition) : The coefficient of the Poisson distribution governing how many sellers a buyer contacts — equivalently the average number of contacts per buyer; higher λ means more competition and lower markups.

Reduced-form preferences / Lucas critique : The buyer utility function a Dixit–Stiglitz modeller would infer to rationalize the search model’s markups; because it depends on the search model’s deep parameters (λ, u, c, b), it shifts whenever the environment or policy changes, so counterfactuals computed holding it fixed are invalid.

Efficiency (of the equilibrium) : The search-theoretic equilibrium maximizes the sum of buyer and seller payoffs — every contacted buyer buys (since valuation u exceeds cost c) and, with heterogeneous costs, buys from the lowest-cost contacted seller — so the positive, dispersed markups are not symptoms of any inefficiency.

Markdown : The labor-market analogue of a markup — the gap between a worker’s marginal product and the wage — which in the Burdett–Mortensen (1998) search model has the same qualitative properties (positive, heterogeneous, size-dependent, efficient) as product-market markups here.

Mis(sed) Diagnosis: Physician Decision Making and ADHD

Mon, 01 Jan 0001 00:00:00 +0000

This paper develops and estimates a structural model of ADHD diagnosis to decompose the mechanisms driving the observed 2.3:1 male-to-female diagnostic difference in the United States. The research question is: to what extent does the large gender gap in ADHD diagnosis reflect true differences in symptom prevalence, versus patient-side utilization costs, versus physician decision-making under uncertainty? The setting is particularly well-suited to this question because DSM-V diagnostic guidelines for ADHD are explicitly gender-neutral, making any gender difference in physician thresholds a detectable deviation from uniform clinical rules.

The data come from de-identified electronic health records from a large Arizona healthcare system covering January 2014 through September 2017. The sample encompasses 36,193 unique encounters for approximately 11,070 pediatric patients. The raw male-to-female diagnostic ratio in the data is 2.32:1 (7.2% of males vs. 3.1% of females receive a clinical ADHD diagnosis). This gap persists after controlling for demographics, general healthcare utilization, and mental health utilization in reduced-form regressions, motivating the structural approach.

Because two key variables — whether a patient received a behavioral assessment (Qi) and the ADHD match signal observed by the physician (xi) — are not directly recorded in the EHR, the author constructs them from clinical doctor note text. A random forest machine learning classifier trained on labeled appointments predicts behavioral assessment take-up for unlabeled encounters; approximately 20.8% of children are predicted to have received a behavioral assessment (23.2% of males vs. 18.3% of females). The ADHD match signal is constructed via an adjusted Bag-of-Words cosine similarity measure comparing each patient’s aggregated note text to the DSM-V symptom list, rescaled to [0,1]. The average signal is 0.319 overall, with males averaging 0.326 and females 0.311.

The structural model has three stages. First, patients/caregivers decide whether to schedule a behavioral assessment, a function of underlying latent ADHD risk (vi) and mental healthcare utilization costs (ci). Second, conditional on assessment, the physician receives a noisy signal of vi and updates beliefs via Bayesian learning; signal quality ρ governs diagnostic uncertainty. Third, the physician diagnoses ADHD if posterior risk exceeds a gender-specific diagnostic threshold τ. Population mean ADHD risk (μ) is identified using regression-adjusted initial primary care provider referral rates as a quasi-exogenous cost-shifter — patients of high-referral-rate providers select into assessment less selectively, so their observed signals approach population mean risk. This extrapolation approach follows Arnold et al. (2022).

The structural parameter estimates reveal that male and female children have similar but slightly different mean ADHD risk (μm = 0.290 vs. μf = 0.262) and similar mean utilization costs (cm = 0.116 vs. cf = 0.109). The most striking differences are in physician parameters: signal quality is lower for male patients (ρm = 0.479 vs. ρf = 0.552), indicating higher diagnostic uncertainty for boys; and diagnostic thresholds are substantially lower for male patients (τm = 0.257 vs. τf = 0.312), meaning physicians are willing to diagnose ADHD in boys with lower posterior risk.

Counterfactual decomposition simulations attribute approximately 20–25% of the 2.32:1 diagnostic gap to underlying differences in ADHD risk, approximately 20% to differences in selection into behavioral assessments, and the remaining majority — approximately 55–60% — to physician decision-making. Within physician decision-making, differences in diagnostic thresholds alone account for roughly two-thirds of the overall diagnostic gap.

The paper offers economic rationales for why gender-specific thresholds may be consistent with physician rationality despite uniform guidelines: higher diagnostic uncertainty for boys justifies lower thresholds under Bayesian updating; hyperactive/impulsive symptoms predominant in boys impose larger classroom externalities (Aizer, 2008); and female patients show higher rates of internalizing co-morbidities (anxiety, depression) that may reduce the marginal benefit of an additional ADHD diagnosis. A type-specific threshold extension finds that for male patients the threshold for hyperactive/impulsive symptoms is significantly lower than for inattentive symptoms, consistent with salience of externally disruptive behaviors. These rationalizations do not vindicate the gap as fully guideline-consistent, but suggest physicians may be responding to real heterogeneity in external costs and co-morbidity patterns.

Q: What is the main research question and why is ADHD a useful setting? A: The paper asks what mechanisms produce the 2.3:1 male-to-female ADHD diagnostic difference: true symptom prevalence, patient utilization costs, or physician decision-making. ADHD is well-suited because (1) clinical guidelines (DSM-V) are explicitly gender-neutral and require the same symptom count threshold regardless of sex; (2) diagnosis is based on subjective behavioral assessment rather than objective testing, creating substantial physician discretion; and (3) both missed and excess diagnosis carry meaningful costs — missed diagnosis limits educational accommodations; excess diagnosis exposes children to Schedule II controlled substances.

Q: What data does the paper use and what are the key descriptive facts? A: The data are de-identified electronic health records from a large Arizona healthcare system, 2014–2017, covering 36,193 encounters for 11,070 pediatric patients aged 5 and above. Overall ADHD diagnosis rate is 5.2%, with males at 7.2% and females at 3.1%, a 2.32:1 ratio that matches national levels. Approximately 49.5% of the sample is Hispanic, which the author notes contributes to a below-national-average overall diagnosis rate. The gender diagnostic gap persists even after controlling for demographics, general healthcare utilization, and mental health utilization in reduced-form regressions.

Q: How does the paper construct the behavioral assessment indicator (Qi) and the ADHD match signal (xi)? A: Qi is constructed using a random forest classifier trained on doctor notes from appointments where assessment status is known with near-certainty (ADHD diagnosis or DSM-V comorbid diagnosis = positive; non-mental-health diagnosis code for patients with no mental health history = negative). The classifier uses 41 features including note length and top-20 word frequencies for each label class. xi is constructed via an adjusted Bag-of-Words cosine similarity between each patient’s combined behavioral assessment notes and the DSM-V symptom list, separately for inattentive and hyperactive/impulsive sub-types, taking xi = max{xi1, xi2}. The average xi is 0.319 (males 0.326, females 0.311) in the behavioral assessment subsample.

Q: What is the identification strategy for recovering population mean ADHD risk (μ)? A: Because xi is observed only for endogenously selected patients, the observed sample mean overestimates population mean risk. The author uses regression-adjusted referral rates of each patient’s initial primary care provider (IPCP) as a quasi-exogenous cost-shifter satisfying (a) relevance — IPCP referral intensity lowers patient scheduling costs — and (b) independence from patient ADHD risk vi, since IPCPs are typically chosen before behavioral symptoms develop and only 28% of IPCPs in the sample ever diagnose ADHD themselves. Population mean risk is then recovered by extrapolating the relationship between IPCP referral propensity and average observed xi to propensity = 1, following Arnold et al. (2022). The maximum observed IPCP referral propensity is only about 0.75, so the estimate requires extrapolation beyond the observed support.

Q: What are the estimated structural parameters and what do they imply? A: Mean ADHD risk is μm = 0.290 vs. μf = 0.262 — males have modestly higher underlying risk. Mean utilization costs are cm = 0.116 vs. cf = 0.109 — nearly identical across genders. Signal quality (diagnostic certainty) is lower for males: ρm = 0.479 vs. ρf = 0.552, indicating physicians face more diagnostic uncertainty when assessing boys. Most importantly, diagnostic thresholds are lower for males: τm = 0.257 vs. τf = 0.312, meaning physicians diagnose ADHD in boys at a lower required posterior risk level, consistent with viewing missed diagnosis as relatively more costly for male patients.

Q: How much of the 2.32:1 diagnostic gap can be attributed to each mechanism? A: Counterfactual simulations decompose the gap as follows: differences in underlying ADHD risk distribution account for approximately 20–25% of the diagnostic difference; differences in selection into behavioral assessments (utilization costs operating through assessment rates) account for approximately 20%; and physician decision-making differences account for the remaining majority, approximately 55–60%. Within physician factors, differences in diagnostic thresholds (τm < τf) are the single largest contributor, explaining roughly two-thirds of the overall male/female diagnostic gap.

Q: What do the type-specific threshold estimates reveal? A: When the baseline model is extended to allow separate diagnostic thresholds for inattentive vs. hyperactive/impulsive symptom sub-types, male patients show significantly lower thresholds for hyperactive/impulsive symptoms relative to inattentive symptoms (τ^HI_m < τ^Inatt_m). This is consistent with the hypothesis that more externally salient and disruptive symptoms carry larger classroom externalities, which physicians may implicitly factor into diagnosis decisions (following Aizer, 2008). For female patients, the threshold differences across symptom types are smaller and less statistically significant.

Q: What economic rationales does the paper offer for gender-specific diagnostic thresholds despite uniform guidelines? A: Three mechanisms are identified. First, higher diagnostic uncertainty for males (lower ρm) implies that under symmetric costs, Bayesian-rational physicians should set lower thresholds when the signal is noisier — this alone partially rationalizes the threshold gap. Second, hyperactive/impulsive symptoms predominant in boys impose greater externalities on classroom peers (Aizer, 2008), increasing the social benefit of diagnosis for boys on the margin. Third, females show substantially higher rates of co-morbid internalizing conditions (anxiety, depression) whose treatment may mitigate ADHD-related behaviors or whose interaction with stimulant medication makes the marginal ADHD diagnosis less beneficial for girls (Currie et al., 2014). These factors together suggest physicians may be responding to genuine heterogeneity in net diagnosis benefits, even if their behavior deviates from gender-neutral clinical guidelines.

Q: What share of the 2.3:1 national diagnostic gap is consistent with genuine symptom prevalence differences? A: Simulations indicate that only about 20–25% of the 2.32:1 male/female diagnostic difference can be explained by the underlying difference in ADHD risk distributions. The majority — roughly 75–80% — reflects factors beyond true prevalence: selection into care and, most substantially, physician decision-making differences including both signal quality and diagnostic thresholds.

Q: What are the policy implications? A: The findings suggest that targeted interventions in physician awareness and clinical training are likely more effective than generic awareness campaigns, since the dominant driver of the diagnostic gap is physician threshold-setting rather than symptom prevalence. Structured decision support tools or updated training that make physicians aware of gender-specific diagnostic patterns could reduce medically unwarranted diagnostic differences. Policies targeting patient-side access barriers (the ~20% explained by selection) remain relevant but secondary. The roughly 20–25% of the gap attributable to genuine symptom prevalence differences is, by construction, guideline-consistent and should not be targeted for elimination.

Q: What are the methodological contributions? A: The paper makes three methodological contributions. First, it develops a structural model of mental health diagnosis that explicitly incorporates endogenous patient selection — a feature absent from standard physician decision-making models — which is shown empirically important. Second, it applies machine learning and NLP to clinical doctor note text to construct key unobserved clinical variables (behavioral assessment indicator and ADHD match signal) that are unavailable as structured data in EHRs. Third, the identification of population mean health risk uses a quasi-exogenous variation approach (IPCP referral rates) analogous to Arnold et al. (2022)’s method for measuring racial discrimination in bail decisions, adapted here to a continuous health risk setting with endogenous selection.

Diagnostic threshold (τ_θ): The gender-specific posterior ADHD risk level above which a physician chooses to diagnose ADHD. Set ex-ante, it reflects the physician’s perceived tradeoff between the costs of over-diagnosis (misdiagnosis) and under-diagnosis (missed diagnosis). A lower threshold implies the physician views missed diagnosis as relatively more costly for that patient group. By construction, uniform clinical guidelines imply a single threshold independent of patient gender.

ADHD match signal (x_i): A physician-observed, noisy signal of a patient’s true latent ADHD risk (v_i), observed only conditional on the patient receiving a behavioral assessment. In estimation, it is proxied via a cosine similarity measure between the patient’s aggregated clinical doctor note text and the DSM-V symptom list, constructed separately for inattentive and hyperactive/impulsive sub-types.

Signal quality / diagnostic uncertainty (ρ_θ): The correlation between the physician’s observed ADHD match signal and the patient’s true ADHD risk. Higher ρ means the physician’s signal is more informative and diagnostic uncertainty is lower. In the Bayesian updating framework, higher ρ implies the physician places more weight on the observed signal relative to the prior.

Mental healthcare utilization cost (c_i): The composite of all patient/caregiver factors that affect the decision to schedule a behavioral assessment net of child symptom level. Includes non-monetary barriers such as time constraints, distance, stigma, and information from primary care providers during wellness visits; does not include monetary out-of-pocket costs since insurance typically covers behavioral assessments.

Initial Primary Care Provider (IPCP) referral rate: The regression-adjusted share of a given PCP’s patients who ultimately receive a behavioral assessment at some point in the sample. Used as a quasi-exogenous cost-shifter that influences patient scheduling costs without being correlated with patient ADHD risk, enabling identification of population mean ADHD risk via extrapolation.

Latent ADHD risk (v_i): An unobserved continuous measure of a child’s underlying ADHD-related behavioral symptoms, drawn from a gender-specific normal distribution N(μ_θ, σ²_θ). A child’s true ADHD status is Si = 1(v_i > v̄), where v̄ is the DSM-V minimum symptom threshold, defined identically for boys and girls.

Adjusted Bag-of-Words (BOW) cosine similarity: The NLP method used to construct the ADHD match signal proxy. Patient notes are tokenized into uni-grams and bi-grams after preprocessing (spell check, abbreviation replacement, part-of-speech tagging, synonym replacement), and tf-idf weighted. The cosine similarity between the resulting document vector and the DSM-V symptom text vector is computed separately for each ADHD sub-type and rescaled to [0,1].

On the Nature of Entrepreneurship

Mon, 01 Jan 0001 00:00:00 +0000

This paper uses a novel longitudinal administrative dataset drawn from U.S. Internal Revenue Service (IRS) and Social Security Administration (SSA) records to characterize income dynamics and the determinants of entrepreneurial entry for pass-through business owners — sole proprietors, partners, and S corporation owners — who collectively account for over 50 percent of all U.S. business net income. The sample covers 2000–2015 and includes up to 1.3 billion person-year observations for individuals aged 25–65. The authors construct balanced panels using birth cohorts 1950–1975, impute education (college attainment) and skill (cognitive, interpersonal, manual) via machine-learning classifiers trained on CPS and O*NET data, and estimate life-cycle income profiles using a three-component model that separates individual fixed effects, group-specific time effects, and group-cohort-specific age effects.

The paper’s central departure from prior work is coverage of the full income distribution, including the high-earning right tail that household surveys such as the CPS misrepresent due to top-coding and small samples. When the IRS and CPS samples are compared on a consistent classification basis, median self-employment income is lower in the IRS data at all ages, consistent with the survey literature’s emphasis on the “typical” self-employed individual. However, mean incomes diverge sharply: the IRS shows mean self-employment income rising from $23 thousand at age 25 to $93 thousand at age 55, whereas the CPS (with incorporated owners reclassified) shows a rise from only $41 thousand to $73 thousand. Roughly 80 percent of self-employment income in the IRS data accrues to individuals above the $100 thousand threshold, compared to 42–53 percent in the CPS. The IRS-CPS gap is dominated by the right tail and concentrated in professional services and health care. For paid-employed individuals, the IRS and CPS medians and means are close at all ages, confirming the discrepancy is specific to self-employment.

The life-cycle estimation finds that individuals who have “tried self-employment” — a group earning virtually all self-employment income — start at similar average incomes to primarily paid-employed peers at age 25 but reach $134 thousand by age 55, compared with $79 thousand for paid-employed peers with the same observable characteristics. Age effects for the self-employed are 63 percent higher than for the paid-employed at age 26 and remain elevated until age 55. Time effects show dramatically greater cyclical volatility for the self-employed: income growth declined by $9,655 (2008) and $8,785 (2009) for the self-employed versus $373 and $1,583 for paid-employed in the same years, concentrated in real estate and construction.

On the determinants of entry, the paper finds: (i) no evidence that house-price appreciation raises entry rates, contra collateral-constraint hypotheses; (ii) most entrants have lower asset incomes than future entrants with the same characteristics, arguing against a liquid-wealth precondition; (iii) most entrants have higher prior labor income than future entrants, consistent with entry being driven by on-the-job experience rather than fallback from low-paid work; (iv) almost all founders report positive individual tax income in their first year of operation despite negative business net income and no external debt financing. Self-employed income growth exhibits greater dispersion — a 10th-to-90th percentile range roughly 2.5 times wider than for the paid-employed — and a Kelly skewness about 0.1 higher. A standard consumption-risk model calibrated with household-finance estimates of risk aversion rationalizes the patterns if individuals are insured against the most adverse downside shocks. Entry and exit rates are stable across the sample period, including the Great Recession, and the entrepreneurship share does not decline.

The subgroup congruent with non-pecuniary motivation — primarily self-employed individuals earning less than paid-employed peers with matching characteristics — comprises roughly 57 percent of primarily self-employed by count but earns only 16 percent of total self-employment income.

In depth

Q1. Why do IRS and CPS data give such different pictures of self-employment income?

The CPS suffers from top-coding of high incomes and small samples that underrepresent high earners in key industries. The IRS-CPS mean income gap for the self-employed is dominated by the right tail: in the main IRS sample, individuals above the $100 thousand threshold earn roughly 80 percent of all self-employment income, versus 42 percent in the comparable CPS sample. The average income of top earners above $100 thousand is $355 thousand in the IRS versus $218 thousand in the CPS. The gap is concentrated in professional services and health care and persists across all income thresholds and sample definitions tested. No analogous discrepancy exists for paid-employed individuals, where IRS and CPS medians and means are close at all ages.

Q2. What does the comparison look like at the median versus the mean?

At the median, IRS self-employment income is lower than both CPS samples at all ages, with the gap largest for younger owners and those with incorporated businesses — a pattern consistent with the survey-based “self-employment discount” narrative. At the mean, the IRS shows much higher income at older ages: by age 55, IRS mean self-employment income is $93 thousand versus $73 thousand in the CPS sample that includes reclassified incorporated-owner wages. The divergence arises because the mean is sensitive to the right tail, which the CPS systematically underrepresents.

Q3. How does the paper estimate life-cycle income profiles while separating age, time, and cohort effects?

Individual income is decomposed into an individual fixed effect (permanent latent ability and preferences), a group-specific time effect (business-cycle fluctuations common to a group), and a group-cohort-specific age effect (life-cycle income growth). Identification exploits the overlapping cohort structure of the 16-year panel: age effects are assumed equal across cohort bins of size at least two, allowing time and age effects to be separately identified. The model is estimated in levels rather than logs to accommodate business losses. Groups are defined as a Cartesian product of 32,256 subgroups based on education, three skill dimensions, industry (21 two-digit NAICS codes), demographics (gender, cohort, marital status, children), and employment-status history.

Q4. What are the headline life-cycle income profile findings for self- versus paid-employed?

Among the “primarily employed” group, those who have tried self-employment and those who are primarily paid-employed have similar average incomes at age 25. By age 55 the self-employed reach an estimated $134 thousand (2012 dollars) versus $79 thousand for paid-employed peers with identical observable characteristics. The estimated age effect for the self-employed is 63 percent higher than for the paid-employed at age 26 and remains higher through age 55. These gaps would widen further if incomes were adjusted upward for the BEA-estimated net misreporting rates of 46 percent for unincorporated owners and 14 percent for S corporation owners.

Q5. How large is the group consistent with non-pecuniary motivation, and how much income does it earn?

The non-pecuniary subgroup — primarily self-employed individuals (at least 12 years in self-employment) who earn less on average than primarily paid-employed peers matched on gender, education, skills, and other characteristics — is numerically larger, comprising approximately 57 percent of primarily self-employed by count. However, this group earns only 16 percent of total self-employment income. Adjusting for paid-employed fringe benefits and self-employed income misreporting can change the group’s size but does not alter the finding that it accounts for a small income share. The paper concludes that non-pecuniary motives may guide occupational choice for many individuals but are not the driver of the typical dollar earned in self-employment.

Q6. How does idiosyncratic income risk compare between self- and paid-employed?

Self-employed income changes are substantially more dispersed: the 10th-to-90th percentile range of income growth is roughly 2.5 times wider for the self-employed than for the paid-employed. Income changes for the self-employed are also more right-skewed, with a Kelly skewness difference of approximately 0.1. When a standard consumption-risk model — augmented with a lower bound on consumption growth to allow for external insurance — is parameterized with risk-aversion estimates from the household finance literature, the observed patterns are rationalized if individuals are insured against the most adverse downside shocks, i.e., the attractive aspect of self-employment is large potential upside with insured downside.

Q7. What happened to self-employed income and exit rates during the Great Recession?

Time effects show steep income growth declines for the self-employed of -$9,655 in 2008 and -$8,785 in 2009, compared with much more modest declines of -$373 and -$1,583 for paid-employed peers. The aggregate income declines are concentrated in cyclically sensitive self-employed subgroups in real estate and construction, with their paid-employed counterparts experiencing only modest declines. Despite these large income shocks, exit rates from self-employment showed little change during the Great Recession, either in aggregate or in the cyclically sensitive sectors. Entry rates were likewise stable, and the share of entrepreneurs in the population did not decline over the full sample period.

Q8. Does the evidence support collateral constraints as a binding barrier to entrepreneurial entry?

No. The paper tests the hypothesis, standard in the liquidity-constraints literature, that entry rates should be higher for homeowners experiencing house-price appreciation (which raises collateral value). The IRS data do not support this prediction. Separately, comparing asset incomes (interest, dividends, capital gains) of current entrants and future entrants with the same characteristics, the paper finds that most current entrants have lower asset incomes and less liquid wealth than those who switch later, which also argues against a liquid-wealth precondition for entry.

Q9. What does prior labor income reveal about why people enter self-employment?

Current entrants have higher prior labor income than matched future entrants with the same characteristics, indicating they enter with accumulated on-the-job experience rather than being pushed into self-employment as a fallback after failure in paid work. This is consistent with self-employment being a deliberate, experience-driven career transition for most entrants rather than a last resort for low earners. The paper interprets this as positive evidence for the role of experience-based human capital in driving entrepreneurial choice.

Q10. How do founders finance startup costs if most have negative business net income in early years?

Almost all founders in the sample report positive income on their personal (individual) tax form in the first year of operation, even though most report negative business net income and carry no external debt financing. This pattern suggests founders rely on personal income sources — prior savings, part-time paid employment, or spousal income — to cover startup costs rather than external debt, implying that formal credit-market financing constraints are not the primary barrier to entry for most entrants in the sample.

Q11. What are the scope conditions and key limitations?

The sample covers pass-through owners (sole proprietors, partners, S corporation owners) and excludes C corporation shareholders, whose entrepreneurial income does not flow to individual returns until distributed. Income measures exclude most employer fringe benefits; capital gains are excluded from self-employment income, and the authors note their inclusion would strengthen the main findings. The analysis covers 2000–2015 for cohorts born 1950–1975, and income is reported before taxes and transfers. Baseline estimates are not adjusted for misreporting, though BEA-implied adjustments of 46 percent for unincorporated owners and 14 percent for S corporation owners would widen the income gaps further.

Pass-through business owner: An individual who owns a sole proprietorship, partnership, or S corporation, such that business net income flows directly onto the owner’s personal tax return; excludes C corporation shareholders whose income appears only upon dividend or capital-gains distributions.

Tried self-employment: The paper’s primary self-employed comparison group within the “primarily employed” category — individuals with any years in self-employment (including frequent switchers and those with most years in self-employment) — who collectively earn virtually all self-employment income.

Group-specific age effect: The paper’s estimate of how individual income changes with age within a defined subgroup (determined by education, skill, industry, demographics, and employment history), identified by exploiting overlapping birth cohorts in the 16-year panel and separated from individual fixed effects and business-cycle time effects.

Primarily employed: Individuals with at least 12 of 16 sample years in either self- or paid-employment, with at most one intermediate year of non-employment; the paper’s main analytical focus for life-cycle income comparisons.

SOI Databank: The Statistics of Income Databank, a de-identified balanced panel combining SSA demographic records with IRS tax filing data for all living U.S. individuals with a Social Security number over 1996–2015; the paper’s primary data source providing Schedule C, K-1, W-2, and related filing information.

Kelly skewness: A robust measure of distributional asymmetry used by the paper to characterize income growth; the paper reports that Kelly skewness of self-employed income changes exceeds that of paid-employed by approximately 0.1, indicating greater right-skewness in self-employment income dynamics.

Non-pecuniary motivation subgroup: Primarily self-employed individuals who earn less on average than primarily paid-employed peers matched on observable characteristics, taken by the paper as consistent with non-wage job amenities (autonomy, flexibility) driving occupational choice; found to be 57 percent of primarily self-employed by count but earning only 16 percent of total self-employment income.

Online Business Models, Digital Ads, and User Welfare

Mon, 01 Jan 0001 00:00:00 +0000

Acemoglu, Huttenlocher, Ozdaglar, and Siderius develop a two-sided platform model to study the welfare consequences of digital advertising as an online business model. The platform intermediates between a firm selling a horizontally differentiated product and a continuum of users who derive utility from both entertaining content and informative signals about product quality embedded in ads. Users have a two-dimensional type: a sophistication dimension (sophisticated with probability lambda, naïve with probability 1-lambda) and a product-quality dimension (high quality with prior probability q). The central departure from the standard informational-advertising literature is that sophisticated users hold the correct model of the ad signal process, while naïve users underestimate the false-positive rate — the probability that a low-quality product generates a positive ad signal (phi_0). Naïve users perceive this false-positive rate to be phi_{0,N} = omega_N * omega_P * phi_0, where omega_N <= 1 captures inherent naïveté and omega_P <= 1 captures failure to understand personalized targeting, so phi_{0,N} < phi_0. The equilibrium concept is Berk-Nash equilibrium (Esponda and Pouzo 2016), meaning all agents are Bayesian given their subjective model.

The platform chooses ad load alpha (Poisson rate of ad displays), subscription fees, and the monetary transfer from the firm; the firm sets product price p after observing the platform’s contract. The central finding (Proposition 2) is that when the objective false-positive rate phi_0 exceeds a threshold phi-hat_0(lambda, phi_1, phi_{0,N}) — which is increasing in lambda and phi_{0,N} and decreasing in the true-positive rate phi_1 — the unique equilibrium is an advertising-based plan that fully segments the market: naïve users receive an ad load that extracts all their surplus, while sophisticated users are excluded entirely. In this regime the firm charges a strictly higher price p-hat* > p-bar*, where p-bar* = (beta*q + c)/2 is the monopoly price without advertising. The ad-based equilibrium emerges precisely when ads are more misleading (larger gap between phi_0 and phi_{0,N}), not when they are more informative — a comparative static the authors describe as paradoxical.

Welfare consequences (Proposition 4) are unambiguous in the advertising regime: both naïve and sophisticated users are strictly worse off than the baseline without any platform. Naïve users over-purchase due to inflated posteriors from misread signals; sophisticated users are harmed through the price channel — the firm’s higher profit-maximizing price p-hat* applies to all buyers. In the fully rational benchmark (phi_{0,N} = phi_0), the unique equilibrium is subscription-based and user welfare equals the no-platform baseline (Proposition 3).

These results extend to richer menus (Proposition 5), mixed subscription-plus-advertising plans (Proposition 7), and to multi-firm and multi-platform competition (Propositions 9-12). Digital ads soften Bertrand competition by generating endogenous horizontal differentiation among otherwise identical firms, so equilibrium prices can exceed marginal cost even with two competing firms. Platform competition similarly fails to restore welfare: platforms compete away subscription fees but both adopt ad-based plans targeting naïfs when phi_1 exceeds a threshold, maintaining the welfare loss.

On policy, the first best (planner observes types) cannot be decentralized because naïve users prefer more ads than is socially optimal, inverting the usual self-selection constraint. The second best (planner subject to incentive-compatibility constraints) is a single pooling plan with an intermediate ad load alpha^{SB} in [alpha^{FB}_N, alpha^{FB}_S] and yields average welfare above the no-platform baseline, though below first best (Proposition 13). This second best can be decentralized with a nonlinear digital ad tax, a per-unit product subsidy, and a platform subscription subsidy (Proposition 14). A simpler flat tax on digital ad revenues — above a threshold gamma-bar < 1 — also improves welfare relative to the ad-based equilibrium, though it does not restore the second best (Proposition 15).

Four robustness extensions are developed: endogenous manipulation (platform always chooses the most manipulative environment, lowest phi_{0,N}); naïve learning dynamics (learning raises the sophisticate share in steady state, making ad-based models less profitable but not overturning the main results); imperfect price discrimination by the firm (naïfs are unambiguously worse off, threshold for advertising equilibrium shifts down); and an added price-sensitivity dimension (the platform runs a 2x2 menu separating by both sophistication and price sensitivity, preserving the result that naïve users tolerate and receive more ads than sophisticates in every stratum).

Q: What is the key asymmetry between naïve and sophisticated users that drives the main results? A: Sophisticated users hold the correct Bayesian model of the ad signal process and thus correctly account for the false-positive rate phi_0 when updating beliefs from positive ad signals. Naïve users perceive the false-positive rate as phi_{0,N} = omega_N * omega_P * phi_0 < phi_0, so they treat positive signals as stronger evidence of high product quality than they actually are. Because naïve users overestimate the informativeness of ads, their (interim) subjective valuation of an ad-based plan is higher, making them more tolerant of ad loads and more willing to join platforms with heavy advertising. This asymmetry is what makes it profitable to target naïfs with high ad loads while excluding or charging subscription fees to sophisticates.

Q: Why does advertising to sophisticated users generate no additional firm profit, while advertising to naïve users does? A: Lemma 1 establishes that with linear-quadratic utility the firm extracts no surplus from advertising to sophisticates: because sophisticated agents are fully Bayesian, their expected posterior equals the prior (E_S[pi_i] = q), so expected demand after advertising is identical to demand before advertising. By contrast, Lemma 2 shows that the firm’s profit from naïve agents is positive and strictly increasing in ad load alpha, because naïve users’ average demand curve drifts upward as alpha rises — their inflated perceived informativeness of ads causes them to over-update on positive signals, systematically raising their willingness to pay. The platform captures this surplus from the firm via the advertising transfer m*.

Q: What is the threshold condition determining whether the equilibrium is subscription-based or advertising-based? A: Proposition 2 identifies a threshold phi-hat_0(lambda, phi_1, phi_{0,N}) that is increasing in the sophisticate share lambda and in the naïve false-positive perception phi_{0,N}, and decreasing in the true-positive rate phi_1. When the objective false-positive rate phi_0 is below this threshold, the profit-maximizing business model is subscription-based with price P* = T - v and product price p* = p-bar* = (betaq + c)/2. When phi_0 exceeds the threshold, the advertising model dominates: the platform sets a high ad load alpha-hat that makes naïve users exactly indifferent between participating and their outside option v, excludes sophisticates, and the firm charges p-hat* > p-bar*. The threshold falls with phi_1, meaning more informative ads expand the range of phi_0 over which the advertising equilibrium obtains.

Q: How does allowing the platform to offer menus change the results relative to the baseline two-plan case? A: Proposition 5 shows that with menus the platform can simultaneously serve both user types: sophisticates receive a subscription plan at P* = T - v and naïve users receive an ad-based plan with the same high load alpha-hat* as in the baseline. The threshold for the advertising equilibrium shifts down to phi*0(lambda, phi_1, phi{0,N}) < phi-hat_0, so advertising business models arise for a strictly larger set of parameters. Welfare consequences are unchanged (Corollary 1): when phi_0 > phi*_0, both types have welfare strictly below the no-platform baseline. Proposition 6 further shows consumer welfare is monotonically decreasing in both phi_0 and phi_1: higher phi_1 (more informative true-positive signals) also reduces welfare because any surplus from greater informativeness is fully captured by the platform.

Q: What is the welfare ranking across the three regimes: no platform, advertising equilibrium, and subscription equilibrium? A: In the subscription equilibrium (regime (a) of Proposition 2 or 4), user welfare for both types equals the no-platform base case W_base(tau) — the platform captures all surplus it creates and users are no better or worse off. In the advertising equilibrium (regime (b)), both naïve and sophisticated users are strictly worse off than with no platform: W-hat*(tau) < W_base(tau) for both tau in {S, N}. The first-best, where a planner controls ad loads separately by type, yields W^{FB}(tau) > W_base(tau) for both types because informative ads can genuinely improve sophisticated users’ decisions and a constrained amount improves naïve users’ decisions too.

Q: How does firm-level competition interact with digital advertising to affect prices and welfare? A: Without advertising, two ex ante identical firms compete à la Bertrand and price at marginal cost (p*_1 = p*_2 = c). Proposition 9 establishes that when phi_1 > phi^F_1 and phi_0 >= phi^F_0(phi_1), the platform offers an ad-based plan and equilibrium prices p-hat*_1 and p-hat*_2 are both strictly above p-bar* — the monopoly price without advertising. The mechanism is endogenous horizontal differentiation: users who see positive ad signals for one firm’s product form higher valuations for that product, so the two products become differentiated in the eyes of consumers even though they are ex ante identical, breaking Bertrand logic. Example 1 further illustrates that advertising can be more prevalent with competition than without: a second firm’s entry can push the equilibrium from no-advertising to separating.

Q: Does platform competition protect users from the welfare losses associated with digital advertising? A: Not fully. Proposition 11 shows that with two competing platforms (M=2, N=1) and no advertising, platforms compete away both subscription fees and ad loads, and welfare reaches the fully rational benchmark. However, when phi_1 exceeds threshold phi^P_1, both platforms adopt ad-based plans targeting naïve users, charge no subscription fees, and the product price rises to p-hat*_P > p-bar* (Proposition 12). Competition reduces subscription fees to zero but does not eliminate the incentive to target naïfs with heavy ads, because naïve users’ over-valuation of ads means they remain willing to join ad-heavy plans. The fundamental inefficiency from naïve users’ misspecified model persists under platform competition.

Q: Why is the first-best allocation not implementable as a decentralized equilibrium? A: Proposition 13 explains the obstacle: the social planner would ideally offer naïve users fewer ads (alpha^{FB}_N) than sophisticated users (alpha^{FB}_S), with alpha^{FB}_N <= alpha^{FB}_S. However, naïve users have a higher subjective valuation for ads than sophisticates because they believe ads are more informative. If offered a menu with both options, naïve users would self-select into the plan with the higher ad load alpha^{FB}_S — the exact opposite of what the planner wants. The incentive-compatibility constraints therefore force the planner toward a single pooling plan with an intermediate ad load alpha^{SB} in [alpha^{FB}_N, alpha^{FB}_S]. Average welfare under the second best exceeds the no-platform baseline, confirming that some advertising is socially valuable, but falls short of the first best whenever alpha^{FB}_N > 0.

Q: How does a flat digital ad tax improve welfare, and what are its limitations? A: Proposition 15 establishes that whenever the equilibrium features an ad-based plan, a flat tax on digital ad revenues at rate gamma > gamma-bar < 1 improves welfare by discouraging advertising-based business models and inducing the platform to shift toward subscription-based plans. The mechanism is that taxing ad revenue reduces the platform’s marginal gain from increasing ad load, making the subscription plan relatively more profitable. However, the flat tax does not achieve the second best because it operates linearly rather than targeting the nonlinear distortion: the optimal nonlinear tax-subsidy scheme (Proposition 14) requires a threshold-style ad tax at rate mu > mu-bar combined with a per-unit product subsidy delta* and a platform subscription subsidy eta > eta-bar.

Q: What happens when the platform can endogenously choose how manipulative its ads are? A: Proposition 16 shows that a profit-maximizing platform always chooses the lowest feasible phi_{0,N} = phi-bar — the most manipulative environment. Two reinforcing channels drive this: the pricing channel (lower phi_{0,N} amplifies naïve demand shifts per positive signal, so the downstream firm raises price and sales, increasing ad revenues extracted by the platform) and the participation channel (lower phi_{0,N} raises naïve users’ perceived informational value of ads, relaxing their participation constraint and permitting a higher ad load alpha). Platform competition constrains the equilibrium ad load through tighter participation constraints but does not alter the choice of phi_{0,N} = phi-bar, so competition limits ad quantity but not ad manipulativeness.

Q: How do naïve learning dynamics affect the main results? A: Proposition 17 introduces a birth-death environment where exposure to disconfirming evidence gradually converts naïve agents to sophisticates. A unique steady-state sophisticate share lambda*(alpha_N, phi_0) exists; both higher ad load alpha_N and higher phi_0 accelerate the conversion of naïfs, raising future sophisticate share and reducing future ad revenues. This creates a new intertemporal trade-off that constrains the platform’s choice of ad loads relative to the static case. The key result (part ii) is that the main characterization of Proposition 7 carries through under a modified cutoff phi-tilde^{dynamic}0 >= phi-tilde_0(lambda-tilde, phi_1, phi{0,N}), so learning dynamics make the ad-based business model less likely but do not overturn the fundamental welfare results.

Q: How does imperfect price discrimination by the firm affect naïve users? A: Proposition 18 considers a firm that observes a user’s sophistication type with probability kappa in [0,1]. With price discrimination, the firm sets type-specific prices satisfying p*_N >= p* >= p*_S, moving toward the type-specific monopoly levels. Naïfs are unambiguously worse off: when identified (with probability kappa), they face the higher price p*_N and a higher equilibrium ad load. The threshold for the advertising equilibrium also shifts down relative to the baseline, meaning advertising business models emerge for a larger parameter range when price discrimination is possible.

Q: How does the paper define and measure user welfare, and why is ex post rather than interim welfare the relevant concept? A: User welfare W(tau_i) is defined as ex post utility, which depends on the actual product quality theta_i realized after consumption, not on interim beliefs formed after viewing ads. Naïve users’ interim assessment inflates expected product quality, but their ex post utility depends on whether the product is genuinely high quality for them (theta_i = 1 with probability q, theta_i = 0 with probability 1-q). Because naïve users over-purchase due to misread signals — consuming more than optimal when theta_i = 0 — their ex post utility is strictly lower than their interim expected utility, and strictly lower than the no-platform baseline in the advertising equilibrium. The ex post welfare concept is the relevant one precisely because it captures the actual material consequences of manipulation, not the subjectively perceived gains from ads.

Naïve vs. Sophisticated Users: The paper’s primary user heterogeneity dimension. Sophisticated users hold the correct model of the ad signal process, setting phi_{0,S} = phi_0 (the true false-positive rate). Naïve users hold a misspecified model with phi_{0,N} = omega_N * omega_P * phi_0 < phi_0, underestimating the probability that a low-quality product generates a positive ad signal, due to inherent naïveté (omega_N) and failure to understand personalized targeting (omega_P).

Ad Load (alpha): The Poisson rate at which ads are displayed to a user per unit time. Total ad displays follow a Poisson(alpha*T) distribution. Higher ad load means less time on entertaining content — expected entertainment time is (1-alpha)T — and a higher probability (1 - exp(-alphaT)) that the user sees the ad at least once. The platform chooses alpha as its primary instrument for extracting surplus from naïve users.

False-Positive Rate (phi_0): The objective probability that a low-quality product (theta_i = 0) generates a positive (“good”) ad signal. The gap between phi_0 (objective) and phi_{0,N} (naïve users’ perceived rate) is the key parameter driving all welfare results: a larger gap implies greater de facto manipulation and a stronger incentive for the platform to adopt an advertising-based model.

Berk-Nash Equilibrium: The solution concept from Esponda and Pouzo (2016), used to model agents with misspecified subjective models. All agents are Bayesian conditional on their own subjective model. Sophisticates’ subjective model equals the objective model (standard Bayesian), while naïfs update using the misspecified phi_{0,N}. Perfection requires sequential rationality at each information set given beliefs.

De Facto Manipulation: The paper’s term for a situation in which the platform and firm exploit naïve users’ misspecified model to boost demand and extract surplus, without requiring any outright deception in the formal sense. It arises because naïve users voluntarily choose high-ad-load plans (believing ads to be highly informative) and voluntarily over-purchase (having updated on what they mistakenly think are strong positive signals). The manipulation is “de facto” because it operates through the users’ own rational (but misspecified) decision-making.

Separating Equilibrium: An equilibrium in which naïve and sophisticated users self-select into distinct platform plans. In the advertising equilibrium, naïve users join an ad-heavy plan (extracting all their surplus via inflated willingness to pay for ads) while sophisticated users are either excluded or placed on a subscription plan. This separation is the vehicle through which the platform maximizes revenue from naïf manipulation while limiting the disciplining force of sophisticates.

Second-Best Allocation: The welfare-maximizing allocation subject to the incentive-compatibility constraints that users self-select into plans. Because naïve users prefer more ads than sophisticated users (the inverse of what the planner desires), the second best is a single pooling plan with an intermediate ad load alpha^{SB} in [alpha^{FB}_N, alpha^{FB}_S]. This is strictly worse than the first best but achieves average welfare above the no-platform baseline, and can be decentralized with a nonlinear ad tax, product subsidy, and platform subscription subsidy.

Optimal Fiscal Policy with Heterogeneous Agents and Capital: Overturning Chamley-Judd

Mon, 01 Jan 0001 00:00:00 +0000

The Chamley-Judd result (1986) states that the optimal long-run capital income tax rate is zero in representative-agent models. This paper shows that introducing heterogeneous agents — specifically, agents with uninsurable idiosyncratic income risk who use precautionary saving — overturns this result. When agents differ in their wealth and income realizations, a capital income tax serves as a form of insurance that representative-agent models cannot provide. The paper derives a tractable analytical characterization of the optimal capital tax in an Aiyagari-type heterogeneous-agent model and finds that the optimal rate lies in the range of 10–30 percent at the steady state — strictly positive, in direct contradiction to Chamley-Judd. The magnitude of the optimal tax depends on the degree of idiosyncratic risk and the availability of alternative redistribution instruments: when other redistributive tools are limited, the optimal capital tax is higher.

Summary of a forthcoming paper, AI-assisted and human-reviewed. See the linked original for the authoritative claims and full conditions.

In depth

Q1. Why does heterogeneity overturn Chamley-Judd?

In representative-agent models, all agents hold the same capital stock, so a capital tax distorts intertemporal decisions identically and the Ramsey planner finds it optimal to zero out the distortion in the long run. With heterogeneous agents and uninsurable risk, the capital tax has an additional insurance role: taxing capital income and redistributing it reduces consumption variance across agents, generating welfare gains that outweigh the intertemporal distortion costs. The insurance benefit makes the optimal tax positive at the steady state because the tax-and-redistribute mechanism provides risk-sharing that incomplete markets cannot.

Q2. How tractable is the analytical result?

The paper derives closed-form expressions for the optimal tax rate as a function of the degree of idiosyncratic risk, the wealth distribution’s spread, and the available redistribution instruments, enabling comparative statics that go beyond what purely computational approaches provide. This tractability distinguishes the result from earlier numerical work that demonstrated positive optimal capital taxes without clear analytical structure.

Q3. What is the quantitative magnitude?

The optimal steady-state capital income tax is in the range of 10–30 percent, substantially above zero but well below confiscatory rates, in the paper’s benchmark calibration matched to U.S. income and wealth inequality. The range reflects the sensitivity to available redistribution instruments: the lower bound applies when the government has a rich set of redistribution tools, the upper bound when capital taxation is the only available instrument.

Key concepts

Chamley-Judd result : the proposition that the optimal long-run capital income tax is zero in representative-agent Ramsey taxation models; overturned in this paper once heterogeneous agents with uninsurable risk are introduced.

insurance role of capital taxation : the mechanism by which a capital income tax reduces consumption inequality in a heterogeneous-agent economy, generating welfare gains that outweigh the intertemporal distortion costs and making the optimal capital tax positive.

Optimal Tests Following Sequential Experiments

Mon, 01 Jan 0001 00:00:00 +0000

This paper addresses a practical gap in the inference literature for sequential and adaptive experiments: while the design of such experiments has been studied extensively, there is little theory characterizing which tests are optimal once the experiment concludes. Adusumilli asks what the best hypothesis test looks like after a sequential experiment — a costly sampling design, a group sequential trial, or a bandit experiment — and whether the complexity of the adaptive protocol can be reduced to a manageable set of sufficient statistics for inference purposes.

The methodological core is the derivation of two Asymptotic Representation Theorems (ARTs). The first ART applies to stopping-time experiments, where the sampling rule is fixed in advance but the stopping time is fully adaptive (updated after every observation). The second ART allows the sampling rule itself to be adaptive, but requires that both the sampling and stopping decisions are updated only a finite number of times after observing batches of data. Both ARTs establish that the asymptotic power function of any test in the original sequential experiment can be matched by a test in a limit experiment in which a Gaussian process is observed for each treatment and inference is made on the drifts of those processes.

The key sufficiency result is a dimension reduction: regardless of the number of batches or the complexity of the adaptive protocol, any candidate test’s asymptotic power can be reproduced by a test that depends only on a fixed, finite set of statistics. For stopping-time experiments, the sufficient statistics are the stopped value of the score process (parametric) or the efficient influence function process (non-parametric), together with the stopping time. For batched experiments with adaptive sampling, the sufficient statistics are the final allocation proportions for each treatment (q_1, q_0) and the final values of the influence function processes (x_1, x_0) — a fixed dimension of 2d+2 regardless of the number of batches. This stands in contrast to the earlier ART of Hirano and Porter (2023), whose state variables grow linearly with the number of batches.

The paper then characterizes optimal tests within the limit experiment under several criteria. Under no restriction, the Neyman-Pearson lemma yields the uniformly most powerful (UMP) test for a point alternative. For testing linear combinations of the parameter vector, a further dimension reduction applies and a UMP test exists in the limit experiment, depending only on a scalar projection of the sufficient statistic. Under unbiasedness, any valid test must satisfy an orthogonality condition on the stopped process. Under an alpha-spending constraint — where the overall size alpha is pre-allocated across stages — optimal stage-specific thresholds are derived. Under a weighted average power criterion, the optimal test takes the form of a likelihood ratio statistic integrated against the weight function.

Three application classes are treated with explicit optimal procedures. For horizontal boundary designs (stopping when a test statistic crosses a fixed threshold, including the SPRT and the Neyman-allocation design from Adusumilli 2022), the most powerful asymptotically unbiased test rejects when the stopping time falls below a specific quantile of its null distribution. Monte Carlo simulations show the test achieves nominal 5% size even for small n, while the standard two-sample test has actual size near 9% in the same setting. For group sequential trials (including O’Brien-Fleming designs with T=2 stages), the paper derives stage-specific critical values satisfying the alpha-spending constraint, with numerical simulations confirming the asymptotic approximation is close to nominal for small n, though accuracy degrades for larger values of the null mean. For bandit experiments run with a batched Thompson-sampling algorithm (K=2 treatments, J=10 batches), the paper constructs the power envelope and shows it is asymmetric: distinguishing (a, 0) from (0, 0) is easier than distinguishing (-a, 0) from (0, 0) for a > 0, because Thompson sampling directs more observations to the arm with higher estimated mean, reducing informativeness from the other arm. Simulations confirm the asymptotic approximation is accurate for as few as n=20 observations per batch (200 total).

The framework covers both parametric and non-parametric models. The non-parametric setting replaces the score process with the efficient influence function process, and the asymptotic power bound translates directly. Results also apply to conditional power given the stopping time.

Q: What is the core methodological contribution of the paper? A: The paper derives two Asymptotic Representation Theorems (ARTs) showing that the asymptotic power function of any test following a sequential experiment can be matched by a test in a Gaussian-diffusion limit experiment. The first ART covers stopping-time experiments with fully adaptive stopping rules; the second covers batched experiments with adaptive sampling rules. These ARTs reduce the infinite-dimensional adaptive experiment to a tractable limit object.

Q: What are the sufficient statistics for inference, and why does this matter? A: For stopping-time experiments, the sufficient statistics are the stopped value of the score (parametric) or efficient influence function (non-parametric) process, together with the stopping time. For batched experiments with adaptive sampling over K treatments, the sufficient statistics are the final allocation fractions (q_1, q_0) and the final influence function process values (x_1, x_0), a fixed dimension of 2d+2. This matters because it establishes that all the adaptive complexity of the protocol can be discarded: a test that uses only these statistics is asymptotically as powerful as any test that uses the full sample path.

Q: How does this paper extend or differ from Hirano and Porter (2023)? A: Hirano and Porter (2023) derive an ART for batched sequential experiments whose state variables grow linearly with the number of batches, making the limit experiment increasingly complex. Adusumilli shows that only a fixed number of sufficient statistics (2d+2) are needed to match unconditional asymptotic power, irrespective of the number of batches. The paper also extends to non-parametric models, derives optimal conditional tests given stopping times, and covers fully adaptive stopping-time experiments via a different route (Le Cam 1979) that does not require the batching restriction.

Q: What is the result for testing linear combinations of the parameter? A: When the null hypothesis is H0: a^T h = 0 in the limit experiment, a further dimension reduction applies: the UMP test depends only on a scalar projection x-tilde(tau) = sigma^{-1} a^T I^{-1/2} x(tau) and the stopping time tau. Because under the null this projection is a standard Brownian motion evaluated at the stopping time, the test is pivotal and uniformly most powerful for the composite hypothesis, regardless of the nuisance components of h.

Q: What is the unbiasedness condition in the limit experiment? A: A test phi is unbiased if its power exceeds its size under all alternatives. In the Gaussian limit experiment, Proposition 2 shows that any unbiased test of H0: h=0 vs H1: h≠0 must satisfy the moment condition E_0[x(tau) phi(tau, x(tau))] = 0, which is obtained by differentiating the power function at h=0 and applying the unbiasedness constraint. This condition restricts which tests can be considered, and the optimal unbiased test is characterized within this class.

Q: What is the alpha-spending criterion and what does the paper show about it? A: Alpha-spending (introduced by Gordon Lan and DeMets, 1983) pre-allocates the total size alpha across T stages via a spending vector (alpha_1, …, alpha_T) with sum equal to alpha, and requires that the conditional rejection probability at stage t not exceed alpha_t. Theorem 2 shows that for discrete stopping times, the asymptotic conditional power beta_n(h|t) converges to beta(h|t) in the limit experiment on subsequences, enabling the derivation of optimal stage-specific thresholds satisfying the spending constraint.

Q: What is the key finding for horizontal boundary designs with a fixed sampling rule? A: For experiments that stop when the influence function process first crosses a fixed threshold gamma — including the SPRT and the Neyman-allocation costly-sampling design of Adusumilli (2022) — Lemma 1 establishes that the UMP asymptotically unbiased test of H0: mu_1 = mu_0 is the test that rejects when the stopping time tau-hat falls below the alpha-quantile of its null distribution. Monte Carlo evidence shows this test achieves nominal 5% size even for small n, while a naive two-sample test ignoring the adaptive stopping rule has actual size near 9%.

Q: What does the power envelope look like for Thompson-sampling bandit experiments, and why is it asymmetric? A: For Thompson-sampling bandit experiments with K=2 arms and J=10 batches, the power envelope for testing H0: (mu_1, mu_0) = (0, 0) is asymmetric: it is easier to distinguish the alternative (a, 0) from the null than to distinguish (-a, 0) for the same a > 0. The mechanism is that Thompson sampling allocates more observations to the arm with the higher estimated mean, so a positive treatment effect leads to more data for treatment arm 1 and less for arm 0, making the joint test more informative in one direction than the other.

Q: How accurate are the asymptotic approximations in finite samples? A: For horizontal boundary designs, Monte Carlo simulations show size is close to nominal 5% even for small n. For group sequential trials with an O’Brien-Fleming design (T=2 stages), the approximation is close to nominal for small n but degrades for larger values of the null mean mu-bar. For Thompson-sampling bandit experiments with K=2 arms and J=10 batches, the approximation is accurate for as few as n=20 observations per batch (200 total observations).

Q: How does the paper handle non-parametric models? A: In non-parametric settings, the sufficient statistic is the efficient influence function process x_n(t) = (sigma^{-1}/sqrt(n)) sum_{i=1}^{floor(nt)} psi(Y_i), where psi is the efficient influence function for the functional of interest and sigma^2 = E[psi^2]. Proposition 3 establishes that the asymptotic power of any test is bounded above by the power envelope in the Gaussian limit experiment indexed by this process. The non-parametric and linear-combination parametric cases share the same limit structure.

Q: What are the open questions identified by the author? A: Two main limitations are noted. First, the ART for adaptive sampling rules is established only for batched experiments; whether it extends to fully adaptive (non-batched) sampling rules without loss of power is conjectured but not formally verified. Second, for fully adaptive experiments, the alpha-spending characterization is not yet available, and the author suggests exploring invariance restrictions or conditional inference as alternative optimality criteria.

Asymptotic Representation Theorem (ART): A result showing that the asymptotic power function of any test in the original sequential experiment can be matched by that of a test in a Gaussian-diffusion limit experiment; used to transfer optimality results from the limit to the original problem.

Limit experiment (Gaussian diffusion): The limiting statistical model in which one observes a Gaussian process x(t) = I^{1/2} h t + W(t) for each treatment, with unknown drift vector h; inference on h in this experiment characterizes optimal tests in the original sequential experiment.

Sufficient statistics (for sequential inference): The finite set of statistics that, in the limit experiment, capture all power-relevant information from the adaptive experiment: for stopping-time experiments, the stopped score/influence function process value and the stopping time; for batched adaptive experiments, the final allocation fractions (q_a) and final influence function values (x_a) for each treatment arm.

Alpha-spending constraint: A strengthened size requirement in group sequential trials that pre-allocates the total Type I error alpha across stages via a spending vector (alpha_1, …, alpha_T); requires that conditional rejection probability at each stage t not exceed alpha_t, and sum alpha_t = alpha.

Efficient influence function process: In a non-parametric model, the partial-sum process x_n(t) = (sigma^{-1}/sqrt(n)) sum_{i=1}^{floor(nt)} psi(Y_i), where psi is the efficient influence function for the target functional; this process is the non-parametric analogue of the score process and serves as the sufficient statistic for non-parametric sequential inference.

Stopping-time experiment: A sequential experiment in which the sampling rule (how to allocate observations across treatments) is fixed before the experiment begins but the stopping rule (when to terminate) is fully adaptive and updated after every observation.

Power envelope: The supremum of the asymptotic power function over all tests of a given size; computed in the limit experiment via the Neyman-Pearson lemma and the Girsanov theorem, and serves as an upper bound on the power of any feasible test in the original sequential experiment.

Peer Effects and Rank Concerns in the Classroom

Mon, 01 Jan 0001 00:00:00 +0000

This paper investigates the mechanisms behind peer effects in the classroom using exogenous variation in study disruptions generated by the 2010 Maule mega-earthquake in Chile (magnitude 8.8, the seventh-largest ever instrumentally recorded). The central research question is why classroom peers can shape academic achievement — specifically, whether beyond production complementarities and a desire to conform, a desire to compete for classroom rank can drive peer influence on learning.

The author constructs a novel dataset linking administrative and survey data from Chile’s Ministry of Education (SIMCE test scores, GPA, curriculum coverage, and school expenditure records) for two cohorts of roughly 150,000 eighth-grade students — one measured in 2009 before the earthquake, one measured in 2011 roughly 20–22 months after — to newly constructed measures of housing damage. Damage to each student’s home is built in three steps: (1) ground-shaking intensity using an established attenuation formula for the 2010 earthquake; (2) seismic vulnerability of each student’s home inferred from a latent-class-analysis model trained on census data linking housing construction materials to vulnerability classes; and (3) a combined expected “damage ratio” (fraction of home that needs to be rebuilt). Identification uses a difference-in-differences strategy that exploits the differential correlation between pre-existing seismic vulnerability and outcomes across the pre- and post-earthquake cohorts, controlling for socioeconomic composition.

The main findings, holding fixed a student’s own earthquake exposure, are as follows. (1) Own home damage reduced test scores by 0.03 standard deviations (SD) per SD increase in damages (a 4.4 percentage-point increase in collapsed home fraction, approximately USD 3,600) and raised self-reported cost of study effort. GPA effects (–0.02 SD) are statistically insignificant. (2) A 1 SD increase in the mean damage among classroom peers raised test scores by 0.05 SD and GPA by 0.04 SD. School expenditure data (available for the 42% of schools in the preferential subsidy program) show schools responded by reallocating funds away from administrative activities toward educational and psychological support, accounting for this positive effect. (3) A 1 SD increase in the within-classroom standard deviation of peer damages lowered test scores and GPA by approximately 0.085 SD on average, but with sharply heterogeneous effects across the prior-achievement distribution: it lowered test scores and GPA of high-prior-achievement students by 0.08–0.11 SD and raised achievement of low-prior-achievement students, without corresponding changes in those students’ GPA rank. Neither curriculum-coverage data nor school spending data show significant responses to damage dispersion, pointing to peer-to-peer interactions rather than school mediation.

The null effect on GPA rank despite heterogeneous GPA effects is the pivotal empirical finding motivating the paper’s theory. The author argues that high-achieving students reduced effort in response to a less threatening competitive environment while maintaining their classroom standing — consistent with rank concerns driving effort decisions. Direct survey evidence shows a majority of students agreed they like to do better than classmates.

Motivated by this evidence, the paper introduces a game-of-status model where each student chooses effort to maximize a utility function combining academic achievement and classroom GPA rank, with rank weighted by a preference parameter lambda > 0. The model admits a unique symmetric Bayesian Nash equilibrium. The model rationalizes all four main empirical patterns: positive mean-damage effects (school compensation); heterogeneous dispersion effects (rank competition changes the density of nearby competitors); null dispersion effects on GPA rank (simultaneous equilibrium adjustment preserves rank ordering); and the survey evidence on competitive preferences.

The study is confined to Chilean public and subsidized private schools in earthquake-affected, non-coastal regions, with outcomes measured at the 8th grade. The pre/post cohort design removes schools that closed or received earthquake evacuees. Findings apply to a context where classroom rank is observable to peers (GPA) and where competitive preferences are prevalent among students.

Q: What is the core identification strategy and why does it avoid the usual confounds in peer-effects research? A: The paper uses a difference-in-differences estimator that exploits the differential relationship between pre-existing seismic vulnerability and outcomes across a pre-earthquake cohort (outcomes measured in 2009) and a post-earthquake cohort (outcomes measured in 2011). Because identification relies on variation in peer disruptions rather than in peer characteristics — and because students did not reallocate across classrooms or schools in response to the earthquake in the estimation sample — the strategy avoids the reflection problem and selection confounds that typically plague peer-effects identification. The identifying assumption is that the relationship between seismic vulnerability and outcomes would have been the same across cohorts absent the earthquake.

Q: What evidence supports the identifying assumption? A: The paper provides three pieces of supporting evidence. First, the fraction of students switching schools or classrooms between grades 7 and 8 is identical across the pre- and post-earthquake cohorts in the estimation sample, indicating no earthquake-induced reallocation. Second, pre-trend tests show precise zero effects of own damage, mean peer damage, and SD of peer damage on lagged (4th-grade) test scores and GPA. Third, placebo tests using students in regions unaffected by the earthquake show no significant differential relationships between seismic vulnerability measures and outcomes across cohorts.

Q: How was housing damage measured, and why does this matter for identification? A: Damage is estimated in three steps: ground-shaking intensity at the student’s town is calculated from a validated attenuation formula; seismic vulnerability of the home is predicted using a latent-class-analysis model trained on pre-earthquake census housing data and then applied to student records; and the two are combined into a damage ratio (fraction of home to be rebuilt) using structural engineering damage-grade distributions. This constructed measure is not self-reported and is determined by physical and housing-quality factors largely predetermined before the earthquake, which supports exogeneity. Coastal towns are excluded because the accompanying tsunami caused damages not captured by the damage-ratio formula, and results are robust to different definitions of coastal proximity.

Q: What were the effects of damage to a student’s own home on achievement? A: A 1 SD increase in own home damages (corresponding to a 4.4 percentage-point increase in the collapsed fraction of the home, or roughly USD 3,600) reduced test scores by 0.03 SD. GPA fell by 0.02 SD but this was not statistically significant. Survey data show that own-home damages raised students’ self-reported cost of study effort, suggesting this effort channel may mediate the achievement effects. These negative effects did not vary significantly across the baseline achievement distribution.

Q: What were the effects of mean peer damage on own achievement, and what mechanism explains them? A: A 1 SD increase in mean peer home damage raised own test scores by 0.05 SD and GPA by 0.04 SD. School spending data from SEP-program schools (42% of the sample) show that schools responded to higher average student damage by reallocating expenditures away from administrative activities (recruitment of non-teaching staff, equipment purchases) toward educational support and psychological support activities. This reallocation more than offset potential negative peer-environment effects, generating positive net achievement effects that were approximately uniform across the prior-achievement distribution.

Q: What were the effects of within-classroom damage dispersion on achievement, and how do they vary across students? A: A 1 SD increase in the within-classroom standard deviation of peer damages lowered average test scores and GPA by approximately 0.085 SD. These average effects mask sharp heterogeneity: high-prior-achievement students experienced losses of 0.08–0.11 SD in test scores and GPA, while low-prior-achievement students saw gains. For some students the dispersion effect was comparable to or larger than the effect of damage to their own home.

Q: Why is the null effect of damage dispersion on GPA rank theoretically important? A: Students with high prior achievement experienced drops in GPA in classrooms with more dispersed damages, but without an accompanying drop in their GPA rank. The paper argues this is inconsistent with students passively absorbing a changed study environment: instead, students appear to have adjusted effort precisely enough to maintain their classroom standing. This equilibrium pattern — GPA changes that leave rank ordering intact — is the paper’s key empirical signature of rank-motivated competition as a mechanism for peer influence.

Q: What direct survey evidence is presented on rank concerns? A: Survey data from the post-earthquake cohort show that a majority of students agreed with the statement “I like to do better than my classmates in school,” providing direct evidence that students value classroom rank. Additionally, students with higher initial achievement reported reductions in self-reported ability to engage with course content in classrooms with more dispersed damages, consistent with these students reducing effort when the competitive environment became less threatening to their rank.

Q: Do schools mediate the damage-dispersion spillovers? A: The available data on curriculum coverage and school spending do not show statistically significant responses to within-classroom damage dispersion (as distinct from mean damage). Emergency reconstruction funds were also allocated by schools based on overall damage severity, not its within-classroom dispersion. This absence of a detectable school-mediation channel for dispersion effects strengthens the interpretation that the heterogeneous achievement effects of dispersion reflect peer-to-peer interactions rather than differential school responses.

Q: How does the game-of-status model rationalize the empirical findings? A: In the model, each student maximizes a utility function over academic achievement and GPA rank, with rank weighted by lambda > 0. Students choose effort simultaneously, and their cost-of-effort type is shaped by prior test scores, socioeconomic characteristics, and earthquake damage. The model admits a unique symmetric Bayesian Nash equilibrium. In this equilibrium: schools’ compensating inputs in response to mean damage raise achievement uniformly (rationalizing positive mean-damage effects); changes in damage dispersion alter the density of nearby types differently for high- and low-cost-effort students, changing the marginal benefit of exerting effort to overtake competitors (rationalizing heterogeneous GPA effects); and because all students adjust effort simultaneously, the rank ordering is approximately preserved (rationalizing null rank effects).

Q: What is the mechanism by which damage dispersion produces heterogeneous effort incentives? A: The key mechanism is that when students derive utility from rank, the marginal benefit of a unit of additional effort depends on how many competitors are “nearby” in the effort-cost distribution. When dispersion increases, the density of types just below a high-achiever (low-cost-effort student) decreases, reducing the gain from exerting more effort to maintain rank over nearby rivals; high-achievers therefore reduce effort and GPA falls. Conversely, when dispersion increases, low-achievers face a distribution where they can more effectively compete for higher ranks, raising their effort incentive and GPA.

Q: How does this paper’s theory differ from prior theories of peer influence? A: Prior theories have emphasized two mechanisms: production complementarities (peer ability directly improves own learning) and a desire to conform (students prefer to match their peers’ effort or achievement). Both rationalize a linear-in-means model that captures only mean peer characteristics. This paper’s theory is the first in the peer-effects literature to rationalize why higher-order moments of the peer distribution (specifically dispersion) affect learning, through a competitive rank-concern mechanism that is parsimonious and does not require extensions to production technology or preferences beyond adding rank to the utility function.

Q: What are the policy implications of the competitive-motive theory? A: The theory implies that classroom composition policies affecting the dispersion of student ability — such as ability tracking, gifted programs, or reshuffling policies — can have heterogeneous and potentially perverse effects: policies that reduce ability dispersion may concentrate competitive incentives in ways that harm some students while benefiting others. Standard linear-in-means models of peer effects, which capture only mean peer characteristics, would not predict these distributional consequences. The author argues this means the competitive mechanism has been largely unexplored despite its intuitive appeal, and calls for structural estimation and policy analysis in future work.

Q: What is the scope of the empirical findings? A: The findings apply to 8th-grade students in Chilean public and private subsidized schools located in earthquake-affected, non-coastal regions, with outcomes observed approximately 20–22 months post-earthquake. The sample excludes schools that closed due to the earthquake and schools that received evacuees. The paper notes that while the theory is formulated around an earthquake shock, the competitive-motive mechanism applies whenever the dispersion of students’ cost-of-effort types changes — including through classroom assignment policies or other shocks — and is not specific to the natural-disaster context.

Damage ratio: The fraction of a student’s home that needs to be rebuilt, constructed by combining geocoded ground-shaking intensity (via the Astroza et al. attenuation formula for the 2010 Chilean earthquake) with the predicted seismic vulnerability class of the home (derived from a latent-class-analysis model trained on census housing data). Used as the paper’s measure of disruption to each student’s environment.

Exogenous peer effect (in the sense of Manski 1993): The reduced-form impact on a student’s outcome of a change in the distribution of an exogenous characteristic — here, earthquake damage — among classroom peers, holding fixed the student’s own characteristics. Distinguished in the paper from endogenous peer effects (best-response functions).

Rank concern: Students’ utility derived from their position (rank) in the classroom GPA distribution, irrespective of whether that rank is formally rewarded. The paper treats rank concern as a preference parameter (lambda > 0 in the utility function) and identifies it as a mechanism for peer influence.

Game-of-status model: The paper’s theoretical framework, in which students simultaneously choose study effort to maximize utility over own academic achievement and GPA rank. The model admits a unique symmetric Bayesian Nash equilibrium. The central insight is that the density of nearby competitors in the effort-cost distribution determines the marginal benefit of effort, generating heterogeneous incentives when peer cost-of-effort types become more dispersed.

Effort-cost type: Each student’s marginal cost of exerting study effort, shaped by prior test scores, socioeconomic characteristics, and earthquake damages to the student’s own home. The key primitive of the model that links individual disruptions to equilibrium effort choices.

SEP (Subvencion Escolar Preferencial): Chile’s preferential school subsidy program for disadvantaged students, which requires participating schools (42% of the sample) to submit detailed annual spending reports to the Ministry of Education. The paper uses these reports to identify school spending responses to mean and dispersed peer damages.

Seismic vulnerability class: A classification of a home’s resistance to earthquake damage based on its construction materials (exterior walls, roof, floor), assigned using a logistic latent-class-analysis model estimated on census data. Found to align strongly with household socioeconomic status, enabling prediction of housing vulnerability from administrative student records.

Quota Mechanisms: Finite-Sample Optimality and Robustness

Mon, 01 Jan 0001 00:00:00 +0000

Ball and Kattwinkel study quota mechanisms — linking mechanisms that impose aggregate constraints on agents’ reports across multiple decision problems — and provide the first theoretical analysis under realistic finite-sample conditions with uncertainty about the type distribution. The canonical examples are mandatory grading curves, prescription drug monitoring programs, storable votes procedures, and lifetime assistance caps (TANF). Prior literature (Jackson and Sonnenschein 2007; Matsushima et al. 2010) established only asymptotic results under the assumption that the designer knows the exact population distribution, leaving the practical rationale for quotas incomplete.

The paper works in the Jackson–Sonnenschein (2007) decision framework: a principal and n agents face K independent copies of a primitive collective decision problem with independent private values and additively separable utilities. A quota mechanism requires each agent’s K reported type distributions to average to a fixed quota; in each problem copy the social choice function is applied to independently sampled types from the submitted distributions. The key methodological innovation is a reformulation of each agent’s best-response as an optimal transport problem, enabling tight bounds.

The central result (Theorem 1) is a tight ex-post decision error guarantee: for any q-cyclically monotone social choice function, the (x,q)-quota mechanism has a Bayes–Nash equilibrium in which the average frequency of incorrect decisions across K problems is bounded by the sum over agents of (|Θ_i| − 1) times the total variation distance between agent i’s quota and the empirical distribution of agent i’s realized type vector. The constants (|Θ_i| − 1) are tight — they cannot be reduced even by arbitrary linking mechanisms without transfers. The core technical challenge is a “cascade of lies”: when an agent’s realized type frequencies depart from his quota, he may misreport in a way that propagates errors across types. The optimal transport reformulation shows this cascade is bounded because, under a cyclically monotone social choice function, an optimal coupling of the empirical and quota distributions can always be chosen whose support contains no nontrivial cycles, so every transport path has length at most |Θ_i| − 1.

Taking expectations (Theorem 2), with quotas set equal to the prior π, the expected decision error is at most (1/√(2K)) times the sum over agents of (|Θ_i| − 1)^(3/2), which is of order 1/√K and tight to within a factor of approximately 1.25. Applied concretely: with three treatment types and K = 200 patients, the expected share receiving the wrong treatment is at most 10%.

Theorem 3 establishes implementation equivalence: a social choice function is (a) one-shot implementable with transfers, (b) π-cyclically monotone, (c) asymptotically implemented by quota mechanisms, and (d) asymptotically implementable by any linking mechanism with transfers, all if and only if each other holds. No linking mechanism, even with transfers, can asymptotically implement social choice functions that quota mechanisms cannot. A quota–transfer duality is identified: the transfer T_i(θ_i’) in the one-shot problem corresponds to the Lagrange multiplier on the quota constraint for type θ_i’, with the two implementations requiring dual pieces of information about the environment.

Theorem 4 bounds the error from misspecified quotas: if the true distribution is π but the quota is set to q, the mechanisms asymptotically implement some social choice function x_π whose expected distance from the target is bounded by Σ_i (|Θ_i| − 1)||q_i − π_i||. With many patients and a quota that underestimates the need for one of three treatments by 1 percentage point, at most 2% of patients receive the wrong treatment. The constants are again tight.

Theorem 5 addresses robustness to agents’ beliefs: in the Bergemann–Morris (2005) rich type-space framework, for any type space satisfying exchangeability and independence, the (x,π)-quota mechanism admits a belief-free equilibrium in which each agent’s strategy depends only on his own payoff type, and the expected average decision error vanishes as K → ∞. The mechanism is belief-robust because each agent knows his opponents must respect the quota, which pins down the marginal distribution of their reports regardless of their beliefs. Extensions treat interdependent values and dynamic settings with sequentially arriving information.

In depth

Q1. What is the fundamental practical problem with quota mechanisms that the paper addresses?

The prior literature showed quota mechanisms work asymptotically when the designer knows the true type distribution and the number of linked decisions is large. In practice, both conditions fail: any finite sample produces an empirical type distribution that deviates from the quota due to sampling variation, and quotas are typically set using imperfect estimates of the population distribution. The paper is the first to quantify the decision errors arising from these two sources of discrepancy.

Q2. What is the decision-error guarantee in Theorem 1 and why are the constants tight?

For a q-cyclically monotone social choice function x and any realization of agents’ private information, the average fraction of incorrect decisions is bounded by the sum over agents i of (|Θ_i| − 1) times ||q_i − marg(θ_i)||. The constants |Θ_i| − 1 are exactly tight: if they were reduced even slightly, the bound would fail for some realization under some linking mechanism. Tightness is demonstrated via a lower bound (Remark 3) that, in the case of a single agent with two types, agrees exactly with the upper bound.

Q3. What is the “cascade of lies” and how does optimal transport resolve it?

When an agent’s empirical type distribution differs from his quota, truthful reporting is infeasible; he must misreport some types, which can propagate further misreporting — a cascade. The key insight is that the agent’s best-response is equivalent to choosing a coupling (joint distribution) of his empirical distribution and his quota that maximizes a linear objective. Because the social choice function is cyclically monotone, Lemma 2 establishes that an optimal coupling exists whose support contains no nontrivial cycles; consequently transport paths visit each type at most once and have length at most |Θ_i| − 1, bounding the total probability moved at (|Θ_i| − 1) times the total variation distance.

Q4. What does the expected error bound (Theorem 2) say quantitatively?

With the quota set equal to the prior π and K problem copies, the expected average fraction of incorrect decisions is at most (1/√(2K)) × Σ_i (|Θ_i| − 1)^(3/2). For a single agent with |Θ| = 3 types and K = 200 problems, the bound evaluates to (1/√400) × (2)^(3/2) ≈ 0.10, so at most 10% of patients receive the wrong treatment. The bound is of order 1/√K and cannot be improved by more than a factor of approximately 1.25.

Q5. What is the implementation equivalence result (Theorem 3) and why is it significant?

Theorem 3 shows that four conditions are mutually equivalent for any social choice function x: being one-shot implementable with transfers (Rochet 1987), being π-cyclically monotone, being asymptotically implemented by (x,π)-quota mechanisms, and being asymptotically implementable by any linking mechanism including those with transfers. The significance is that no richer linking mechanism — even one with monetary transfers — can asymptotically implement anything that quota mechanisms cannot, justifying the focus on quota mechanisms.

Q6. What is the quota–transfer duality identified in Section 5.2?

In the one-shot problem, the transfer T_i(θ_i’) for agent i reporting type θ_i’ corresponds exactly to the Lagrange multiplier on the quota constraint for type θ_i’. The two implementations require dual pieces of information: quota implementation requires knowledge of the type distribution π_i (to set the quota) but not the utility function or cross-agent beliefs; transfer implementation requires knowledge of agent i’s utility function and interim beliefs but not the marginal distribution π_i. A concrete allocation example illustrates that transfers can implement the social choice function without knowing the type distribution, while quotas cannot.

Q7. How does Theorem 4 bound the error from a misspecified quota?

If the quota q is set based on an incorrect estimate but the true distribution is π, the (x,q)-quota mechanisms asymptotically implement some social choice function x_π whose expected total variation distance from the target x is bounded by Σ_i (|Θ_i| − 1)||q_i − π_i||. The constants |Θ_i| − 1 are again tight. Applied to opioid prescription with |Θ| = 3 and a 1 percentage point underestimate (||q − π|| = 0.01) for one treatment, the long-run expected error is at most 2 × 0.01 = 0.02, so at most 2% of patients receive the wrong treatment.

Q8. How is belief robustness (Theorem 5) formalized and what does it require?

The paper adopts the Bergemann–Morris (2005) rich type-space framework, in which each agent has a payoff type and a belief type. Theorem 5 requires the type space to satisfy exchangeability (joint distribution over payoff types is exchangeable across problem copies) and independence (payoff types are independent across agents). Under these conditions, the (x,π)-quota mechanism has a Bayes–Nash equilibrium in which each agent’s strategy depends only on his payoff type vector, not his belief type, and the expected average decision error converges to zero as K → ∞.

Q9. Why is cyclical monotonicity the key structural condition, and what is its relationship to Rochet (1987)?

Cyclical monotonicity requires that no cycle of types would strictly gain, on average, if each type received the allocation intended for the next type in the cycle. Rochet (1987) proved that a social choice function is one-shot implementable with transfers if and only if it is cyclically monotone. Ball and Kattwinkel’s Theorem 3 adds that this same condition characterizes asymptotic implementability by quota mechanisms and by any linking mechanism with transfers, establishing a deep equivalence between the transfer-based and quota-based approaches.

Q10. How does the new quota mechanism formulation differ from Jackson and Sonnenschein (2007) and what are the consequences?

Jackson and Sonnenschein require agents to report a K-vector of types with type frequencies matching the quota, which requires quotas whose components are integer multiples of 1/K and involves additional modifications for general quotas. Ball and Kattwinkel allow each agent to report a type distribution on each problem, with the average of the K distributions constrained to equal the quota. This enables direct application of optimal transport theory; every type gets weakly higher expected utility under the Theorem 1 equilibrium than under the JS equilibrium. Under JS’s definition, Theorem 1 still holds but with an additional error term of order 1/K.

Q11. Does the optimality result in Theorem 1 extend to linking mechanisms with transfers?

Yes. Theorem 1 states that the constants |Θ_i| − 1 cannot be reduced even using arbitrary linking mechanisms — and the text specifies this holds even for mechanisms without transfers. Theorem 3 further establishes that the class of social choice functions asymptotically implementable does not expand when transfers are added, reinforcing the conclusion that quota mechanisms are not dominated by richer mechanisms in the asymptotic sense.

Key concepts

Quota mechanism: A linking mechanism in which each agent’s K reported type distributions must average to a fixed quota profile q; the social choice function is then applied to types independently sampled from each reported distribution. Generalizes mandatory grading curves, prescription quotas, and storable votes procedures.
Cyclical monotonicity (q-cyclical monotonicity): A condition on a social choice function x requiring that no cycle of types would strictly gain, on average, if each type in the cycle received the allocation intended for the next type. With multiple agents, taken in expectation over co-agents’ types drawn from q. Equivalent by Rochet (1987) to one-shot implementability with transfers.
Ex-post decision error: The average, over K problem copies, of the total variation distance between the implemented decision lottery and the socially desired decision lottery, evaluated at a particular realization of private information — not in expectation over types.
Cascade of lies: The phenomenon in which an agent whose empirical type distribution departs from the quota finds it optimal to propagate misreporting across multiple types, amplifying the decision error beyond the minimum necessary to satisfy the quota constraint. Bounded in magnitude by the optimal transport analysis.
Optimal transport reformulation: Each agent’s best-response choice of report vector is recast as selecting a coupling (joint distribution) of his empirical type distribution marg(θ_i) and his quota q_i to maximize a linear objective. The acyclic structure of optimal couplings under cyclical monotonicity yields the tight error bound (|Θ_i| − 1)||q_i − marg(θ_i)||.
Implementation equivalence: The result (Theorem 3) that one-shot implementability with transfers, π-cyclical monotonicity, asymptotic implementation by quota mechanisms, and asymptotic implementability by any linking mechanism with transfers are mutually equivalent conditions on a social choice function.
Belief-free equilibrium: An equilibrium of a quota mechanism in the Bergemann–Morris type-space framework in which each agent’s strategy depends only on his payoff type, not his belief type. Exists under exchangeability and independence, because the quota pins down the marginal distribution of opponents’ reports regardless of beliefs.
Distributional robustness: The property that when the quota q_i is set based on an incorrect estimate of the true distribution π_i, the long-run decision error is bounded by (|Θ_i| − 1)||q_i − π_i||, proportional to the estimation error.

Regulatory Competition in the US Life Insurance Industry

Mon, 01 Jan 0001 00:00:00 +0000

This paper quantitatively assesses the consequences of jurisdictional competition in the US life insurance industry, an $8 trillion market. The central question is whether competition between state regulators over capital requirements for captive reinsurance subsidiaries — a form of regulatory competition — increases or decreases total surplus, and by how much.

US life insurers are regulated at the state level. Since the early 2000s, states have competed to attract captive reinsurance subsidiaries (captives) by setting lower capital requirements on these entities. The externality structure is asymmetric: the captive state earns tax revenues on liabilities transferred to captives and sets their capital requirements, but bears default costs only for policyholders in its own state. Consumer states bear default costs for their own residents even when those policies have been transferred to an out-of-state captive. This mismatch between who sets capital requirements and who bears default costs creates the externality that drives the race-to-the-bottom dynamic studied in the paper.

The empirical setting draws on a novel dataset covering 66 US life insurers from 2005 to 2020, with total liabilities of $1.9 trillion (approximately 25% of the sector). Data sources include NAIC filings via S&P, CompuLife pricing data, A.M. Best ratings, SEC filings, and state legislative records. The author assembles novel data on captives’ capital levels from SEC filings, Iowa Insurance Department captive financial statements, and insurer reinsurance exhibits.

Three motivating empirical findings ground the structural model. First, captives materially reduce insurers’ capital: in 2019, risk-based capital ratios are 23% lower on average after accounting for captives, with the median insurer’s capital declining 24%, and this translates into an increase in 10-year default probability from 1.0% to 2.9%. Second, states’ capital requirements are the primary determinant of where insurers locate captives: a 1 percentage point increase in a state’s captive capital rate is associated with a 1.6 percentage point decrease in the probability an insurer chooses that state (against a 1.1 percentage point unconditional probability), and this holds when insurers switch states over time as capital requirements change. Tax rates, geographic proximity, and amenities are not meaningfully correlated with captive location choice. Third, a difference-in-differences design exploiting Regulation XXX (effective January 1, 2000), which raised capital requirements differentially across product term lengths, shows that 30-year term products — which faced the largest capital requirement increases — experienced price increases averaging 10.3% relative to 10-year term products, with quantities declining monotonically for longer-term products, consistent with an inward supply shift.

The paper develops a structural model of the insurance market with imperfectly competitive insurers, endogenous default following a Leland (1994) framework, discrete choice consumer demand (Berry, 1994), and state regulators who set captive capital rates to maximize a weighted objective over tax revenues, default costs, consumer surplus, and producer surplus. Regulators deviate from a utilitarian social planner in two ways: they are state-based (generating competition and default externalities) and face agency frictions (captured by welfare weights that differ from unity). The demand side implies an average price elasticity of 2.4. The regulator side reveals that state regulators are willing to trade $1 of default costs against $3.5 of tax revenues and $0.59 of consumer surplus — both diverging from the social planner’s equal weighting.

The main counterfactual finding is that eliminating competition by federalizing insurance regulation would cause regulators to raise capital requirements by 19% (3 percentage points), reducing expected default costs by $2.4 billion while lowering consumer surplus by $880 million, for a net total surplus gain of $1.5 billion. Regulator utility would increase by $3.3 billion in equivalent tax revenues. Because regulators over-value consumer surplus relative to default costs, competition exacerbates rather than counteracts their agency frictions, making competition unambiguously welfare-reducing in the baseline. A social planner would set capital requirements even higher than a federal regulator. On distribution, large states such as California and New York gain most from federalization (they bear substantial default costs), while Vermont — the largest captive state by market share — loses because it would forfeit captive tax revenues. Unilateral bans are found to have limited equilibrium consequences: a New York ban on captive use by insurers selling in New York would achieve only 23% of the national default cost reduction that federalization achieves, and a ban on captives domiciled in Vermont would achieve only 10%, as insurers would redirect captives to other states.

Q: What is a captive reinsurance subsidiary and why do states compete to attract them? A: A captive is a wholly-owned subsidiary of a life insurance holding company that reinsures policies written by the operating company, moving liabilities off the operating company’s balance sheet. Captive states earn tax revenues on liabilities transferred to captives and can set their own capital requirements on those entities, which are lower than the uniform NAIC standards applied to operating companies. Because captives are taxed by the state where they are domiciled — not the consumer’s state — captive states can earn tax revenues on policies sold elsewhere, incentivizing competition through lower capital requirements to attract insurers.

Q: What is the default externality at the core of this paper’s argument? A: When an insurer defaults, the shortfall on policies sold to consumers in a given state is borne by that state’s guaranty fund and consumers, regardless of where the captive holding those liabilities is domiciled. So Vermont, as the captive state, sets the capital requirement on liabilities transferred from (for example) Massachusetts policyholders, but does not bear the default cost on those Massachusetts policies. This means Vermont internalizes only the default cost on its own consumers, leading it to set capital requirements lower than it would if it bore the full default cost — a classic externality.

Q: How large is the effect of captives on insurers’ capital levels? A: Using novel data on captives’ actual balance sheets, the author finds that in 2019, the size-weighted average risk-based capital ratio of sample insurers is 23% lower after consolidating captives into the operating company’s balance sheet. The median insurer’s capital ratio decreases by 24%. In terms of default risk, this adjustment corresponds to an increase in the 10-year default probability from 1.0% to 2.9% based on historical insurer default rates.

Q: What is the state of competition among captive domiciles in the data? A: Twenty-two states had passed laws allowing captives as of the sample period, with the set of competing states largely stabilizing after 2013. The market is moderately concentrated: the top five states (Vermont, Arizona, Delaware, Iowa, and South Carolina) account for 80% of all captive liabilities, and the Herfindahl-Hirschman Index is 0.20. Vermont has maintained its position as the largest captive state throughout the period.

Q: What evidence shows that capital requirements — rather than taxes or other factors — drive captive location choice? A: In a linear probability model of captive location with insurer-year fixed effects, a 1 percentage point increase in a state’s captive capital rate is associated with a 1.6 percentage point decrease in the probability that an insurer chooses that state (versus a 1.1 percentage point unconditional probability). Captive tax rates are not meaningfully correlated with location choice, consistent with federal tax laws prohibiting the use of reinsurance to reduce tax liabilities. A changes-on-changes specification confirms that insurers are more likely to shift their captives to states that lower their capital requirements over time.

Q: How does the Regulation XXX natural experiment identify the supply-side effect of capital requirements on insurance prices? A: Regulation XXX, effective January 1, 2000, increased reserve requirements for operating companies on a mechanical basis tied to policy term length, with longer-term products facing larger increases. Using a difference-in-differences design at the insurer-product-month level with insurer-product and month fixed effects, the paper finds that products with larger capital requirement increases experienced larger price increases immediately after the regulation took effect. Thirty-year term products experienced price increases averaging 10.3% relative to 10-year term products (the reference group) within three months. Quantities also declined monotonically for longer-term products, confirming an inward shift of the supply curve rather than a demand shift.

Q: What are the estimated regulator welfare weights, and what do they imply about agency frictions? A: Normalizing the weight on tax revenues to 1, the paper recovers that regulators value $1 of default costs as worth $0.29 (implying $3.5 of tax revenues trades off against $1 of default costs) and value consumer surplus at $0.59 per dollar. Because the social planner sets all weights equal to 1, these estimates show regulators over-weight tax revenues and consumer surplus relative to default costs. The higher weight on consumer surplus is consistent with political backlash from consumers facing high insurance prices.

Q: What is the total surplus effect of eliminating regulatory competition through federalization? A: Federalizing insurance regulation — modeled as a single federal regulator setting a uniform capital rate while holding fixed regulatory frictions — would lead regulators to raise captive capital requirements by 19% (3 percentage points) to internalize the default externality. Expected default costs would fall by $2.4 billion. However, higher capital requirements would raise insurance prices and reduce consumer surplus by $880 million. The net effect is a total surplus increase of $1.5 billion. Regulator utility (in equivalent tax revenues) would increase by $3.3 billion.

Q: Would eliminating both competition and regulatory frictions (i.e., a social planner) produce a different outcome than just federalizing? A: In the baseline estimates, a social planner would set capital requirements even higher than a federal regulator, because regulators’ agency frictions lead them to under-weight default costs relative to consumer surplus, pushing capital requirements below the socially optimal level even absent competition. Competition further exacerbates these frictions by providing an additional incentive to lower capital rates. Thus, in the baseline, competition unambiguously decreases total surplus. The paper also reports results under alternative assumptions, providing a “menu” for policymakers that maps different assumptions about regulators’ frictions to quantitative welfare statements.

Q: What distributional consequences across states explain why federalization has not been adopted? A: Federalization would benefit large states such as California and New York most, because those states bear substantial default costs on large volumes of policies sold to their consumers. States with large captive market shares, primarily Vermont, would be made worse off because they would lose captive tax revenues. These predicted gains and losses align with actual state policy positions: New York has called for a national ban on captives, California forbids insurers from setting up captives there, and Vermont has been the most aggressive state in attracting captive domiciles.

Q: How effective are unilateral state bans as an alternative to federal coordination? A: The paper estimates that a unilateral ban by New York on insurers selling in New York from using captives would achieve only 23% of the national default cost reduction that full federalization would achieve. A unilateral ban on captives domiciled in Vermont — the largest captive state — would achieve only 10% of federalization’s default cost reduction, because insurers would simply relocate their captives to other states that still allow them. This finding underscores the importance of cross-state coordination for meaningful regulatory reform.

Q: What does the model’s demand estimation imply about consumer sensitivity to insurance prices? A: The discrete choice demand model estimated on state-level sales, prices, and product characteristics implies an average price elasticity of demand of 2.4 for life insurance products. This elasticity disciplines the quantitative impact of capital requirements on product markets through their effect on insurance prices.

Q: How does the paper recover regulators’ objective functions? A: The author uses the revealed preferences of state regulators, exploiting regulators’ utility maximization first-order conditions and performing numerical perturbations around those conditions to calibrate the welfare weights (lambdas) on each component of the regulators’ utility function. This approach recovers regulators’ tradeoff weights from their observed policy choices — specifically their captive capital rate decisions — without directly observing regulators’ preferences.

Captive reinsurance subsidiary: A wholly-owned subsidiary of a life insurance holding company that reinsures liabilities from the operating company. Unlike operating companies, captives are regulated by the state in which they are domiciled (the captive state) under that state’s own capital requirements, which are typically lower than the uniform NAIC standards. Captives allow insurers to reduce their overall capital requirements by allocating liabilities to the captive.

Default externality: The mismatch between who sets capital requirements for captives (the captive state) and who bears default costs when an insurer fails (the consumer’s state and its guaranty fund). Because the captive state bears default costs only for its own residents — not for residents of states where the insurer also sells — it has an incentive to set lower capital requirements than it would if it internalized the full default cost, leading to an externality on other states.

Risk-based capital ratio (adjusted for captives): The author’s measure of insurer capitalization after consolidating the captive’s balance sheet with the operating company’s. This adjusted ratio is lower than the statutory risk-based capital ratio that ignores captives, by 23-24% in the 2019 sample, and translates into meaningfully higher default probabilities (from 1.0% to 2.9% over 10 years).

Regulatory agency frictions: Deviations of state regulators’ objective functions from a utilitarian social planner’s, captured by welfare weights (lambdas) on each component of the regulator’s utility. In the paper’s estimates, regulators over-weight tax revenues ($3.5 of tax revenues per $1 of default costs) and consumer surplus ($0.59 per $1 of default costs) relative to the social planner’s equal weighting, consistent with political economy pressures from consumers and revenue incentives.

Captive capital rate: The state-level capital requirement on captives, defined empirically as the sum of capital divided by the sum of liabilities of all captives in the state each year. Higher values represent more stringent requirements. The mean in the sample is 4% with a standard deviation of 3%, and captive capital rates are lower on average than operating company capital rates.

Race to the bottom: The dynamic under which competition between state regulators leads each state to set lower capital requirements than it would absent competition, in order to attract captive tax revenues, resulting in a collectively worse equilibrium with higher default risks. The paper finds this outcome in the baseline: competition lowers capital requirements by 19% (3 percentage points) relative to a federal regulator.

External financing frictions: The costs insurers face in raising equity capital, modeled as a per-dollar cost theta on required capital. These frictions create the supply-side channel through which capital requirements affect insurance prices: higher capital requirements raise insurers’ marginal costs, leading to higher prices and lower quantities, as documented in the Regulation XXX natural experiment.

Rent Guarantee Insurance

Mon, 01 Jan 0001 00:00:00 +0000

Abramson and Van Nieuwerburgh study Rent Guarantee Insurance (RGI), a product in which an insurer pays the landlord on behalf of a tenant who defaults on rent due to a negative income or health expenditure shock, in exchange for a monthly premium proportional to rent. The central question is whether RGI can be designed to be both welfare-improving and financially viable, given the frictions of moral hazard and adverse selection.

The authors develop a dynamic overlapping-generations equilibrium model of the rental market that features endogenous rent default, security deposits, evictions, and homelessness. Households face idiosyncratic persistent and transitory income risk, idiosyncratic medical expenditure risk, and aggregate (cyclical) income risk. Rental contracts are non-contingent, households face borrowing constraints, and housing is indivisible with a minimum quality floor. Landlords set deposits to break even in expectation given observed tenant characteristics. An insurance agency can offer RGI and must also break even in the long run. The model is calibrated to the United States at monthly frequency. Income dynamics are estimated from CPS data (1994–2023) and incorporate transitions among employment, unemployment, out-of-labor-force, and retirement states along with transfer income (unemployment insurance, disability, food stamps) and a progressive tax system. Key moments targeted by Simulated Method of Moments include a delinquency rate of 12.15% (model: 12.69%), average security deposit of $984 (model: $992, from approximately 500,000 Craigslist listings across the 100 largest MSAs), homelessness rate of 1.43% (model: 1.42%), and home-ownership rate of 63.6% (model: 63.2%).

The model’s pre-RGI analysis establishes that persistent income shocks — not transitory shocks or medical shocks — are the primary driver of rent defaults. Default risk remains elevated for 3–6 months following a persistent shock, implying that short-duration RGI coverage is insufficient to prevent eviction; coverage must span multiple months.

The paper’s main policy experiments introduce RGI under different access rules and provider types. Unrestricted RGI (available to all renters) generates large welfare gains through improved risk-sharing and lower security deposits — because insured tenants pose less default risk, landlords lower deposit requirements — but is not financially viable for either a public or private insurer due to moral hazard and adverse selection. Even a public insurer that internalizes the fiscal savings from reduced homelessness cannot break even under unrestricted access.

Restricting access changes the viability calculus sharply. A publicly provided RGI targeted to households at the bottom of the wealth distribution can achieve financial viability: these households are precisely those most prone to homelessness, so the reduction in homelessness expenses — which the public insurer internalizes — offsets the insurance deficit. This restricted public RGI generates substantial welfare gains for the most vulnerable households.

A privately provided RGI must instead target higher-wealth renters to break even, because these households have low default risk (limiting claim payouts) while remaining sufficiently risk averse to pay the premium. The intersection of financial viability and take-up is small, yielding a limited target audience. The private program has minimal impact on housing insecurity, and the most vulnerable households derive little benefit. This pattern matches observed private RGI markets, where providers restrict access to renters in good financial condition.

An RGI mandate — requiring all renters to purchase coverage — mitigates adverse selection by improving the pool of insured tenants, dramatically increasing financial viability and allowing the insurer to reduce the premium substantially while still breaking even. Mandated RGI is highly effective at preventing housing insecurity and generates welfare gains concentrated among the most financially vulnerable households.

Scope conditions: results are calibrated to U.S. income, medical, and housing market parameters as of 2019. The insurer’s borrowing cost matters: the public insurer faces lower, counter-cyclical municipal bond spreads, whereas private insurers face higher, pro-cyclical corporate spreads, which constrains the generosity of private contracts in recessions.

Q: What is Rent Guarantee Insurance and how does it work mechanically in the model? A: RGI is a contract under which a tenant pays a flat monthly premium equal to a fraction kappa of rent. When the insured tenant defaults, the insurer pays the landlord directly and deducts one period from the tenant’s stock of “insurance credit.” The tenant remains housed. Once insurance credit is exhausted, the insurer no longer covers defaults. The insurer sets the premium and the maximum coverage duration to break even in the long run.

Q: Why do most rent defaults arise from persistent rather than transitory shocks? A: The model shows that the renter population is disproportionately exposed to persistent unemployment and labor-force-exit spells, and that negative persistent income shocks are harder to smooth through savings than transitory ones. Default risk remains elevated for 3–6 months after a persistent shock but dissipates quickly after a transitory shock. This implies that RGI coverage periods of only a few months would fail to prevent eviction for the majority of defaulting tenants.

Q: How does RGI affect security deposits in equilibrium? A: Because landlords observe the tenant’s insurance status at lease signing and deposits are set to make landlords break even in expectation, insured tenants pose lower default risk and thus face lower upfront deposit requirements. This deposit reduction is a key welfare channel of RGI, as large deposits tie up a disproportionate share of poor households’ wealth and price the most vulnerable out of housing entirely.

Q: Why is unrestricted RGI financially non-viable even for the public insurer? A: Unrestricted access induces both adverse selection — riskier households self-select into coverage — and moral hazard — insured households alter their default and savings behavior. These effects cause the insurer to run a persistent deficit. Even a public insurer that internalizes the fiscal cost savings from reduced homelessness cannot recoup enough to break even, implying that an unrestricted program would require an ongoing subsidy.

Q: How does publicly provided restricted RGI achieve financial viability? A: By targeting households at the bottom of the wealth distribution — precisely those most prone to homelessness — the public RGI program produces large reductions in homelessness. Because the public insurer internalizes the fiscal expenses associated with shelters, health services, and policing that accompany homelessness, these savings are passed through to the insurer and are sufficient to offset the insurance deficit. No such mechanism is available to a private insurer.

Q: Why must private RGI target higher-wealth renters, and what are the consequences? A: Private insurers must break even using only premium revenue, without access to homelessness cost savings. Higher-wealth renters have lower default probabilities, which limits claim payouts, while remaining sufficiently risk averse to demand coverage and pay the premium. The viable target audience is small given these competing requirements. As a result, private RGI covers few households, has minimal effect on housing insecurity, and provides essentially no benefit to the most vulnerable renters. This pattern is consistent with observed private RGI markets.

Q: What are the two differences between public and private insurers in the model? A: First, the public insurer internalizes the fiscal costs of homelessness (shelters, health services, policing), raising its net benefit from offering coverage. Second, the public insurer borrows at municipal bond spreads — which are lower than corporate spreads and counter-cyclical — whereas the private insurer faces higher, pro-cyclical corporate spreads. Counter-cyclical borrowing costs allow the public insurer to extend more generous coverage precisely when aggregate conditions deteriorate and claims rise.

Q: How does an RGI mandate improve financial viability? A: Mandatory enrollment forces all renters, including low-risk ones, into the insurance pool, which counteracts adverse selection. The expanded and higher-quality pool dramatically reduces per-insured expected claim costs, allowing the insurer to lower the premium substantially while still breaking even. The low-premium mandated policy is then both affordable and effective at preventing housing insecurity, with welfare gains concentrated among the most financially vulnerable renters.

Q: What novel data does the paper use for calibration of security deposits? A: The authors construct a dataset of approximately 500,000 Craigslist rental listings scraped across the 100 largest U.S. metropolitan statistical areas between November 2022 and March 2024 to measure the cross-sectional distribution of security deposits. The average deposit in this dataset is $984, which the model matches closely at $992. The data also reveal that the deposit-to-rent ratio is decreasing in house quality, reflecting the higher default risk of low-income renters in lower-quality units.

Q: What is the paper’s definition of homelessness and what rate does the model match? A: Homelessness is defined broadly to include sheltered homeless, unsheltered homeless (0.6% of households), and doubled-up families (0.83% of households), for a total of 1.43% of U.S. households. The model matches this rate closely at 1.42%.

Q: What is the paper’s key implication for the design of housing policy? A: The central implication is that financial viability and impact on housing insecurity are in tension for private insurers, and cannot both be achieved simultaneously. Only a publicly provided program that internalizes homelessness fiscal costs and faces counter-cyclical borrowing spreads can target the most vulnerable renters, break even, and materially reduce housing insecurity. Private RGI, while viable for a narrow segment, cannot substitute for public provision as a tool against homelessness.

Q: How does RGI relate conceptually to rental assistance programs? A: The paper distinguishes RGI from rental assistance on a structural basis: insurance contracts require tenants to pay premiums, making them potentially self-financing for private providers, whereas rental assistance is a net transfer that can never be self-financing. This conceptual distinction motivates studying whether RGI can be designed to eliminate the need for ongoing fiscal transfers, though the analysis ultimately shows that a public subsidy or mandate is required to serve the most vulnerable renters.

Rent Guarantee Insurance (RGI): A contract under which an insured tenant pays a monthly premium equal to a flat percentage of rent; when the tenant defaults, the insurer pays the landlord directly, preserving tenancy, for a limited number of periods governed by the tenant’s stock of insurance credit.

Insurance Credit: An endowment of periods of RGI coverage that households receive upon entry into the model; each time the insurer pays on behalf of a defaulting tenant, one unit of credit is consumed, and no further coverage is available once credit is exhausted.

Housing Insecurity: In the paper’s framework, the set of outcomes — rent delinquency, eviction, and homelessness — arising from the combination of non-contingent rental contracts, borrowing constraints, and idiosyncratic or aggregate income and medical shocks.

Security Deposit: An upfront payment from tenant to landlord, set by the competitive landlord to break even in expectation given the tenant’s characteristics and insurance status; a key channel through which RGI affects welfare by reducing the upfront cost barrier to obtaining housing.

Moral Hazard (in RGI context): The change in a tenant’s default, savings, and housing choices induced by the presence of insurance coverage, which increases expected claim costs for the insurer relative to a world where behavior is held fixed.

Adverse Selection (in RGI context): The tendency of renters with higher default risk to self-select into RGI when access is unrestricted, worsening the insurer’s risk pool and driving up expected payouts relative to premiums.

Homelessness Externality: The fiscal costs borne by government — for shelters, health services, and policing — that accompany homelessness; the public insurer internalizes these costs, creating a net benefit from RGI that private insurers cannot capture.

Counter-cyclical Borrowing Spread: The feature of public (municipal bond) financing whereby borrowing costs fall during recessions, allowing the public insurer to expand coverage when claims are highest; contrasted with private insurers’ pro-cyclical corporate bond spreads that tighten precisely when aggregate conditions worsen.

Revolutionary Transition: Inheritance Change and Fertility Decline

Mon, 01 Jan 0001 00:00:00 +0000

Gay, Gobbi, and Goñi test Le Play’s (1875) hypothesis that the French Revolution contributed to France’s early fertility decline by abolishing impartible inheritance. In 1793, a series of decrees culminating in the Loi de Nivôse (January 6, 1794) abolished testamentary rights and imposed equal partition of assets among all children — partible inheritance — across France, overriding the mosaic of local customs and written laws that had governed inheritance in the Ancien Régime.

The paper’s central argument is that this reform reduced the economic incentive to have children through indivisibility constraints in agricultural land. Under impartible inheritance, land passed to a single heir undivided, keeping plots above the subsistence productivity threshold even at high fertility. Under partible inheritance, each additional child fragments the land further, potentially pushing plots below the minimum productive size, so households face a strong incentive to limit fertility. A Stone-Geary production function with a minimum land threshold L̄ formalizes this mechanism: when landholdings fall in the binding range (L̄ < L < L̃), fertility is strictly higher under impartible than under partible inheritance.

The authors construct the first complete map of inheritance rules across France’s 435 judicial districts as of 1789, classifying each along two dimensions: partible versus impartible, and whether women were included or excluded. This atlas draws on Brette’s (1904) Atlas des Bailliages and the Nouveau Coutumier Général (Bourdot de Richebourg 1724), covering 141 distinct customs. Treatment is defined as municipalities under impartible inheritance before 1793 whose system was altered by the reforms; control municipalities were already under partible inheritance.

The main identification strategy is a difference-in-differences (DD) design comparing women with varying lengths of remaining fertile years after 1793 — from 0 for women aged 40+ at the reform to 25 for women aged 15 or younger — across treated and untreated municipalities. This is augmented by a regression-discontinuity difference-in-differences (RD-DD) design exploiting sharp discontinuities at judicial district borders. Two independent datasets are used: the Enquête Louis Henry (34,812 women in 39 rural municipalities, family-reconstitution method) and Geni.com crowdsourced genealogies (11,649 women across 2,966 locations after the Blanc 2023 horizontal restriction).

Each additional fertile year of exposure to the 1793 reforms reduced completed fertility by approximately 1 percent. Over the full 25-year fertile cycle, this corresponds to a reduction of roughly 0.7 children, or 24 percent relative to the pre-reform mean of 2.92 surviving children in treated areas. This magnitude equals the entire pre-reform fertility gap between impartible- and partible-inheritance areas (2.9 versus 2.2 children), meaning the reforms closed this gap entirely. DD and RD-DD estimates are similar and not statistically distinguishable from each other, and results replicate across both datasets. Results hold on both the extensive margin (childlessness) and intensive margin (fertility of mothers).

The mechanism is most relevant where smallholder landownership is widespread. France — where 40–80 percent of households owned land at the eve of the Revolution — meets this condition. England and Prussia, with more concentrated landownership, would not be expected to show the same response because the indivisibility constraint would not bind even after partition.

Q: What was France’s inheritance system before the Revolution, and how heterogeneous was it? A: Before 1793, inheritance was governed by 141 distinct customary and written laws applied within 435 judicial districts. The country was broadly divided between the customary-law north (Pays de droit coutumier) and the Roman written-law south (Pays de droit écrit), with substantial local variation within regions. Systems ranged from strictly partible (equal division among all offspring) to impartible (primogeniture, ultimogeniture, or unigeniture). Systems also varied in whether women could inherit or received only a dowry. This geographic variation — rooted in the laws of Germanic peoples after the fall of Rome in 476 CE — is exogenous to late eighteenth-century economic conditions and provides the identifying variation for the paper.

Q: What exactly did the 1793 reforms change, and were they enforced? A: The Loi de Nivôse an II (January 6, 1794) abolished testamentary rights entirely and mandated equal partition of assets among all children, including women, throughout France. The reforms came unexpectedly — only 8 of 571 cahiers de doléances analyzed by Goy (1988) mentioned inheritance — and were motivated by the equality principle, legal unification, and the fear that revolutionary sympathizers would be disinherited (Lataste et al. 1901). Offspring quickly asserted their new rights, and by the late 1790s inheritance disputes were the most common cases before family tribunals (Desan 1997; Poumarède 2011).

Q: What is the model’s core mechanism linking inheritance reform to fertility decline? A: The model uses a Stone-Geary production function with a minimum land threshold L̄ below which output falls to zero. Under impartible inheritance, land passes undivided to a single heir, keeping the farm above L̄ regardless of family size. Under partible inheritance, each child receives an equal share, so adding children risks fragmenting plots below L̄ — a powerful incentive to limit family size. The fertility gap between impartible and partible households is at its maximum when landholdings fall in the intermediate range (L̄ < L < L̃) where the constraint is binding. As land size increases, the constraint becomes less binding but the positive fertility differential persists.

Q: What is the paper’s main quantitative estimate of the reform’s effect on completed fertility? A: Each additional fertile year of exposure to the 1793 reforms reduced completed fertility by approximately 1 percent. Over the full 25-year fertile cycle (ages 15–40), this implies a reduction of roughly 0.7 children, or 24 percent relative to the pre-reform mean of 2.92 surviving children in treated areas. This is nearly identical to the pre-existing fertility gap between impartible- and partible-inheritance areas (0.7 children: 2.9 versus 2.2 surviving children), implying the reforms effectively eliminated the fertility differential.

Q: Are the DD and RD-DD estimates consistent with each other, and do both datasets agree? A: Yes. The DD and RD-DD estimates are similar and not statistically different from each other. The RD-DD design compares women born close to judicial district borders where inheritance rules differed, before and after 1793, exploiting the sharp spatial discontinuity at those borders. Consistency across these two designs — which rely on different identifying assumptions — strengthens causal interpretation. Results are also consistent across the Enquête Louis Henry (family-reconstitution) and Geni.com (crowdsourced genealogies) datasets, which are produced by fundamentally different methodologies.

Q: How do the authors verify the parallel trends assumption? A: Figure 6 shows that for cohorts who completed their fertile cycle before 1793, fertility trended downward in parallel across partible- and impartible-inheritance areas: a constant gap of approximately 0.7 children was maintained from women born in the early 1700s (3 versus 2.3 children) through women born in the early 1750s (2.7 versus 2.0 children), the last cohorts to complete fertility before the reforms. The convergence — from 0.7 to 0 children — only begins among cohorts fertile after 1793. The authors also include flexible trend controls interacted with municipality-level religiosity, political support for the Revolution, proximity to administrative centers, and wheat prices, and confirm the main estimate is robust.

Q: What role did the extension of inheritance rights to women play? A: The extension of rights to women was a companion mechanism distinct from abolishing impartible inheritance. Beyond increasing the number of heirs (which directly reduces land per heir), the right to inherit improves a woman’s outside option and postpones entry into marriage, following de Moor and van Zanden (2010). The DD and RD-DD estimates suggest that including women in inheritance and abolishing impartible inheritance had similar effects on fertility. The paper treats these as separate but reinforcing channels.

Q: How do the authors address potential confounders — mortality, migration, and economic conditions? A: On mortality: child mortality did not evolve differently after 1793 across areas with different inheritance rules (Appendix Table A3), and baseline adult mortality (age at death, probability of dying before completing the fertile cycle) was balanced across treated and control areas (Table 1). On migration: the authors explicitly rule out that results are driven by migration. On economic conditions: municipality-specific decade-average wheat prices (Ridolfi 2019) are included as controls for local Malthusian dynamics, and results are robust to their inclusion.

Q: What do the balance tests show? A: Panel A of Table 1 shows that before the reforms, areas with impartible versus partible inheritance were balanced on 9 of 11 individual-level characteristics — including husband and wife age at death, probability of dying before completing the fertile cycle, probability that parents-in-law were alive at marriage, literacy, data accuracy, and age at marriage. The only systematic pre-reform difference was fertility itself (0.7 children). Municipality-level climatic variables, soil suitability, and proxies for mortality uncertainty were also balanced. This is consistent with the origins of these systems in post-Roman Germanic law, which are unrelated to late eighteenth-century economic conditions.

Q: What robustness checks are reported? A: The authors report: (1) permutation tests reshuffling treatment exposure across women and municipalities; (2) non-linear treatment effects across cohorts, showing the heterogeneity required to explain away the baseline estimate is implausibly large per de Chaisemartin and d’Haultfoeuille (2020); (3) exclusion of outlier municipalities; (4) a placebo test for cohorts who completed their fertile cycle before 1793; (5) robustness to alternative sample definitions, treatment definitions, outcome variables, and control groups; (6) Cummins (2020) first-name repetition technique to correct for under-reported child deaths in Henry; (7) terrain characteristics including climatic and soil suitability (Galor and Özak 2016) and ruggedness (Nunn and Puga 2012); (8) for RD-DD: alternative bandwidths, running variable specifications, kernel functions, samples, and border-segment fixed effects. All checks support the main finding.

Q: Why did France experience a fertility decline from inheritance reform while other countries with similar reforms did not? A: The model rationalizes this through landownership structure. The fertility-reducing mechanism operates through indivisibility constraints that bind only when landholdings are small and fragmented — as in France, where 40–80 percent of households owned their land and plots were small. Where landownership is concentrated (England, Prussia), land per heir remains above L̄ even after partible division, so the indivisibility constraint is non-binding and fertility is unaffected by the reform. This provides a structural reason why France’s particular agrarian structure made it uniquely susceptible to this mechanism.

Q: What is the broader historical significance for understanding France’s early demographic transition? A: France’s fertility decline began roughly 50 years before industrialization, making it anomalous relative to standard quantity-quality tradeoff theories linking fertility decline to technological progress and rising returns to human capital. The 1793 reforms provide a legal-institutional explanation for the sharp post-Revolution acceleration visible in Figure 1, which is difficult to attribute to slowly-evolving cultural factors or human capital considerations not yet operative. The estimates imply the reforms brought large impartible-inheritance areas to the low-fertility regime that already characterized partible-inheritance areas, thus sharply accelerating the national transition.

Impartible inheritance: A system under which parents could designate a single heir (through primogeniture, ultimogeniture, or unigeniture) to receive the bulk of the family estate, preventing fragmentation of wealth; in pre-revolutionary France this was associated with extended family households and higher fertility (2.9 surviving children on average) relative to partible areas (2.2).

Partible inheritance: A system under which family wealth was divided equally among all offspring upon death; in the paper’s model this creates an incentive to limit fertility to prevent land fragmentation below the subsistence productivity threshold L̄.

Indivisibility constraint (land threshold L̄): In the Stone-Geary production function, a minimum land input below which agricultural output falls to zero; this is the mechanism through which partible inheritance generates fertility-limiting incentives, since dividing a small plot among many heirs risks crossing L̄ into zero production.

Difference-in-differences (DD) exposure design: The paper’s main identification strategy, using remaining fertile years after 1793 as a continuous treatment-intensity variable (0 for cohorts past fertility at the reform date, up to 25 for cohorts entirely within their fertile years), compared between treated municipalities (impartible → partible) and control municipalities (already partible).

Regression-discontinuity difference-in-differences (RD-DD): An augmented design exploiting the sharp geographic discontinuity at borders between judicial districts with different pre-reform inheritance rules, comparing outcomes on both sides before and after 1793, to address smooth unobserved confounders.

Completed fertility (net): The number of children surviving to age six, preferred over total births because child mortality before 1800 was high (1–1.5 children per mother did not survive to age six per Houdaille 1984), making net fertility the more economically meaningful measure for inheritance and bequest decisions.

Horizontal restriction: A sampling correction applied to crowdsourced genealogical data (Blanc 2023a) that retains an observation only if at least one of the four preceding generations has more than one recorded offspring, correcting for the over-representation of single-child families that arises because Geni users tend to record direct ancestors rather than collateral relatives.

Robust Estimation and Inference in Panels with Interactive Fixed Effects

Mon, 01 Jan 0001 00:00:00 +0000

This paper develops new estimation and inference tools for the coefficient on a covariate of interest in large panel regressions whose unobserved heterogeneity has an interactive fixed effects (factor) structure. The authors demonstrate that standard tools for this model — the least-squares estimator of Bai (2009) and the common correlated effects estimator of Pesaran (2006) — can be heavily biased and severely size-distorted when some of the factors are “weak,” i.e., when factor loadings and factors lack enough variation to be distinguished from noise; in their Monte Carlo designs conventional confidence intervals built on the LS estimator can have almost zero coverage. They propose a debiased estimator together with a bias-aware confidence interval that, given only an upper bound on the number of factors, remains valid uniformly over a class of data-generating processes allowing weak, strong, or nonexistent factors. The construction applies minimax linear estimation to debias a preliminary estimate of the effects matrix, using a nuclear-norm bound on that preliminary estimate’s error, and the estimator attains a faster uniform rate of convergence than existing approaches when weak factors are allowed (reaching the parametric √(NT) rate when N and T grow at the same rate). In 5,000-replication Monte Carlo experiments and an empirical illustration calibrated to the divorce-law studies of Friedberg (1998) and Wolfers (2006), the debiased estimator substantially reduces weak-factor bias without inflating variance and performs comparably to LS when factors are strong, though the bias-aware CIs are often conservative — their oracle length is slightly less than half their actual length, bounding how much the critical value could be tightened. The method requires that the covariate of interest not itself be fully explained by a low-dimensional factor model, which rules out, for example, a treatment indicator that switches on for a subset of units in a single period.

Summary of a forthcoming paper, AI-assisted and human-reviewed. See the linked original for the authoritative claims and full conditions.

In depth

Q1. What problem with interactive fixed effects panels does the paper address?

The paper addresses the failure of conventional estimators and confidence intervals for the regression coefficient β when the factors in an interactive fixed effects model are weak rather than strong. The model is a linear panel regression Yit = Xitβ + Σk Zk,it δk + Γit + Uit in which the unobserved component Γit has a factor structure Γit = Σr λir ftr (factor loadings λir, factors ftr), studied under large-N, large-T asymptotics. The standard least-squares estimator of Bai (2009) is √(NT)-consistent and asymptotically normal under a “strong factor assumption” requiring the loadings and factors to have sufficient variation. When that assumption fails — when factors are present but too weak to separate from the noise term Uit — the estimator cannot recover the true loadings and factors, leaving omitted-variables bias from Γit and producing substantial bias and misleading inference.

Q2. How bad is the problem for conventional methods, and what shows it?

In the authors’ Monte Carlo study, weak factors leave the LS estimator heavily biased and non-normal, and conventional CIs based on it can have almost zero coverage. Their finite-sample distribution plots show the LS estimator centered at the true value when factors are nonexistent or strongly identified, but heavily biased and non-normally distributed at intermediate (“weak”) factor strengths. The simulation tables (5,000 replications) report that the LS estimator is heavily biased and the associated 5%-level tests and 95% CIs are heavily size-distorted unless all factors are strong. The common correlated effects estimator of Pesaran (2006) does not even apply in their designs because the cross-sectional averages of the loadings equal zero.

Q3. What is the paper’s proposed estimator, and how is it constructed?

The paper proposes a debiased estimator that applies the theory of minimax linear estimation to a preliminary estimate of the effects matrix, using a nuclear-norm bound on that estimate’s error. Starting from a preliminary estimate Γ̂pre of the effects matrix Γ together with a bound Ĉ on the nuclear norm ‖Γ − Γ̂pre‖* of its estimation error, the authors form augmented outcomes Ỹit = Yit − Γ̂pre,it and treat the residual effect Γ̃ = Γ − Γ̂pre as a nuisance satisfying the convex constraint ‖Γ̃‖* ≤ Ĉ. They then derive linear weights Ait that optimally use this constraint via minimax linear estimation (Ibragimov and Khas’minskii, 1985; Donoho, 1994; Armstrong and Kolesár, 2018), so that the weights control the remaining omitted-variables bias due to weak factors not captured by Γ̂pre. Bounding the nuclear norm is a convex relaxation of the rank constraint rank(Γ) ≤ R, connecting the approach to the matrix-completion and debiased-LASSO literatures.

Q4. What makes the confidence interval “bias-aware,” and what does it require?

The confidence interval is bias-aware because it uses the nuclear-norm bound Ĉ to explicitly account for the remaining bias in the debiased estimator, and it requires only an upper bound on the number of factors. Rather than assuming the bias is negligible, the CI incorporates the worst-case remaining bias permitted by the bound, which is what allows it to remain valid even under weak factors. The authors show the CI is valid uniformly over a large class of DGPs that allows weak, strong, or nonexistent factors up to the specified upper bound on the number of factors.

Q5. What are the convergence-rate results?

The authors derive rates of convergence that hold uniformly over the DGP class, and show their estimator achieves a faster uniform rate than existing approaches when weak factors are allowed. When N and T grow at the same rate, the estimator attains the parametric √(NT) rate. This improvement is established for the regime that explicitly permits weak factors that cannot be consistently estimated; the authors note their results also apply to the strong and “semi-strong” regimes studied elsewhere, where factors can be consistently estimated and conventional estimators are already asymptotically unbiased and normal.

Q6. What is the key scope condition on the covariate of interest?

An important condition is that the covariate of interest Xit must not itself be entirely explained by a low-dimensional factor model. The method leverages variation in Xit that cannot be explained by a small number of factors. As the authors illustrate, if Xit is the state-year minimum wage, the design requires that states change their minimum-wage laws in different years and often enough to generate such variation. The condition rules out settings where Xit is an indicator for a policy that affects a subset of units starting in a single time period, because then Xit is collinear with the factor model (Xit = λi·ft with λi a treated-unit indicator and ft a post-period indicator) — a fundamental identification problem that other literatures address with additional assumptions.

Q7. Under what error and model conditions do the results hold?

The results hold under conditions similar to Bai (2009) and Moon and Weidner (2015), with the error term mean zero conditional on the regressors and effects but allowed to be heteroskedastic and weakly dependent. Uit is assumed mean zero conditional on X, the controls Z, and Γ, while heteroskedasticity (possibly depending on Xit and Γit) and some weak dependence are permitted. The number of factors R is unknown but assumed small relative to N and T. Unlike some related work, the authors deliberately avoid imposing extra structure such as homoskedasticity or full independence of the errors from the effects and regressor, because such structure would supply additional identifying information and lead to a fundamentally different analysis.

Q8. How does the debiased estimator perform in the Monte Carlo experiments?

In the simulations, the debiased estimator effectively reduces the weak-factor bias without inflating variance, and performs comparably to LS when all factors are strong. Across designs with one and two factors (5,000 replications each), the tables report bias, standard deviation, RMSE, test size, and average CI length. The efficiency gains from debiasing can be very large when a weak factor is present, especially at larger sample sizes, while the cost when factors are strong is minimal. Because LS CIs under weak factors can have zero coverage (being centered on the biased LS estimator and too short), the authors benchmark length against identification-robust “oracle” CIs that invert the LS-based t-statistic using least-favorable critical values; the bias-aware CI’s actual length is at least comparable to, and mostly shorter than, the LS oracle CI length.

Q9. How conservative are the bias-aware confidence intervals?

The bias-aware CIs are often conservative: across most designs their oracle length is slightly less than half their actual length, which bounds how much the critical value could be reduced. This implies the bias-aware critical value cannot be decreased by more than about a factor of two without sacrificing coverage in these Monte Carlos. The authors attribute the conservativeness to two possible sources — the bias bound in their main theorem may be conservative, or there may be additional structure in the initial error or its correlation with the data that the nuclear-norm debiasing does not exploit — and note they cannot rule out that other DGPs would make the critical value non-reducible.

Q10. What does the empirical illustration show?

In an experiment calibrated to the unilateral-divorce-law studies of Friedberg (1998) and Wolfers (2006), a weak factor can severely distort conventional inference, and in the actual data the potential presence of one weak factor is likely to be sufficient to nullify the significance of previously obtained non-robust estimates. Using Kim and Oka (2014) data on a balanced panel of N = 48 states and T = 33 years, with the divorce rate as outcome and a unilateral-divorce-law dummy as the covariate (controlling for state-specific quadratic trends and time effects), the calibrated simulation reproduces the pattern from the abstract design: LS is heavily biased and size-distorted under a weak factor, while the debiased estimator has substantially smaller bias, standard deviation, and RMSE and competitive performance under a strong factor. Applied to the real data, allowing up to one weak factor produces bias-aware CIs substantially wider than the non-robust ones, and the authors find that this potential weak factor is likely sufficient to nullify the significance of the earlier non-robust estimates.

Q11. How does this work relate to existing approaches to weak or rank-deficient factors?

The paper is distinguished by providing inference that remains valid under arbitrary weak factors without assuming the factors can be consistently estimated, which prior approaches generally do not. Robustness results in Bai (2009) and Moon and Weidner (2015) cover the special case where some factors are exactly zero while the rest are strong, but not more general weak factors. Chetverikov and Manresa (2022) also achieve a faster rate under weak factors but assume strong factors when constructing CIs and place a factor structure on the covariate matrix. Lower bounds of Zhu (2019) show no CI can be asymptotically valid under weak factors while matching the performance of Bai’s (2009) CI when factors are strong, underscoring that some cost is unavoidable. The minimax-debiasing strategy parallels debiased-LASSO methods (e.g., Javanmard and Montanari, 2014) for omitted-variable bias in high-dimensional regression.

Key concepts

Interactive fixed effects (factor structure): A model of unobserved panel heterogeneity in which the effect Γit is the sum over factors of a loading times a factor, Γit = Σr λir ftr; equivalently, the matrix of unobserved effects Γ has rank at most R. It generalizes additive fixed effects (αi + γt) and nests the grouped unobserved heterogeneity model as a special case.
Weak factors: Factors whose loadings and/or factors lack sufficient variation across units or over time, so that they cannot be reliably distinguished from the noise term Uit. Under weak factors the strong factor assumption fails and the least-squares estimator’s omitted-variables bias and inference distortions appear; the paper allows arbitrary sequences of such factors, including the nonexistent-factor case.
Strong factor assumption: The condition (as in Bai 2009) that all factor loadings and factors have sufficient variation across i and over t, under which the LS estimator of β is √(NT)-consistent and asymptotically normal. The paper’s contribution is to provide valid estimation and inference without requiring it.
Bias-aware confidence interval: A confidence interval that explicitly incorporates a bound on the estimator’s remaining bias (here via the nuclear-norm bound Ĉ on the preliminary estimate’s error) rather than assuming the bias is asymptotically negligible, enabling uniform validity across factor strengths given an upper bound on the number of factors.
Minimax linear estimation: A method (Ibragimov and Khas’minskii 1985; Donoho 1994; Armstrong and Kolesár 2018) that chooses linear weights minimizing worst-case mean-squared error over a parameter space defined by a convex constraint; here it produces the debiasing weights Ait that optimally use the nuclear-norm constraint ‖Γ̃‖* ≤ Ĉ on the residual effects matrix.
Nuclear norm: The sum of the singular values of a matrix, used as a convex relaxation of the (non-convex) rank constraint rank(Γ) ≤ R; bounding ‖Γ̃‖* constrains the residual effects matrix and is the device through which the bias bound and bias-aware CI are constructed.

Key concepts

Interactive fixed effects (factor structure): A model of unobserved panel heterogeneity in which the effect Γit is the sum over factors of a loading times a factor, Γit = Σr λir ftr; equivalently, the matrix of unobserved effects Γ has rank at most R. It generalizes additive fixed effects (αi + γt) and nests the grouped unobserved heterogeneity model as a special case.
Weak factors: Factors whose loadings and/or factors lack sufficient variation across units or over time, so that they cannot be reliably distinguished from the noise term Uit. Under weak factors the strong factor assumption fails and the least-squares estimator’s omitted-variables bias and inference distortions appear; the paper allows arbitrary sequences of such factors, including the nonexistent-factor case.
Strong factor assumption: The condition (as in Bai 2009) that all factor loadings and factors have sufficient variation across i and over t, under which the LS estimator of β is √(NT)-consistent and asymptotically normal. The paper’s contribution is to provide valid estimation and inference without requiring it.
Bias-aware confidence interval: A confidence interval that explicitly incorporates a bound on the estimator’s remaining bias (here via the nuclear-norm bound Ĉ on the preliminary estimate’s error) rather than assuming the bias is asymptotically negligible, enabling uniform validity across factor strengths given an upper bound on the number of factors.
Minimax linear estimation: A method (Ibragimov and Khas’minskii 1985; Donoho 1994; Armstrong and Kolesár 2018) that chooses linear weights minimizing worst-case mean-squared error over a parameter space defined by a convex constraint; here it produces the debiasing weights Ait that optimally use the nuclear-norm constraint ‖Γ̃‖* ≤ Ĉ on the residual effects matrix.
Nuclear norm: The sum of the singular values of a matrix, used as a convex relaxation of the (non-convex) rank constraint rank(Γ) ≤ R; bounding ‖Γ̃‖* constrains the residual effects matrix and is the device through which the bias bound and bias-aware CI are constructed.

Self-Fulfilling Debt Crises with Long Stagnations

Mon, 01 Jan 0001 00:00:00 +0000

Layer 1: Overview

This paper asks whether sovereign debt crises can be self-fulfilling — triggered by lenders’ expectations of default rather than by weak fiscal fundamentals alone — and whether such crises are empirically plausible. Following the mechanism of Calvo (1988), high expected default probabilities require high interest rates to compensate lenders, but high interest rates in turn raise the cost of debt service and the probability of default, making the pessimistic expectations self-confirming. The key theoretical contribution is to show that this multiplicity of equilibria is state-dependent: it arises only in periods of stagnation, when the endowment process is in a persistent low-growth regime. The paper modifies a standard infinite-horizon sovereign default model (in the spirit of Eaton-Gersovitz and Arellano 2008) by introducing a two-state Markov regime-switching process for trend growth and by having the borrower choose current debt rather than debt at maturity — a timing assumption that is essential for multiplicity. Calibrating the output process to Argentina, Brazil, Italy, Portugal, and Spain using 1980–2017 data, the paper finds that for intermediate levels of debt and in low-growth states, interest rates can be either low (around 4%) or high (around 46%) depending on the coordination of lenders’ beliefs — a self-fulfilling crisis range that reproduces the qualitative features of the European sovereign debt crisis of 2010–2012 and the Argentine crisis of 1998–2002. In high-growth states, the multiplicity region is negligibly small or absent entirely.

In depth

Q1. What is the Calvo (1988) mechanism, and why does it require a bimodal endowment process?

The Calvo mechanism generates multiple equilibrium interest rates for a given level of debt: lenders’ expectation that the borrower will default in the low-output state forces them to charge a high interest rate to break even, but the high rate raises the debt service burden and makes default more likely, validating the pessimistic expectation. For this self-confirming loop to sustain multiple stable equilibria, the interest rate correspondence — mapping debt levels to possible interest rates — must have an upward-sloping region at both the low and high rate. A unimodal endowment distribution generates a correspondence with a downward-sloping high-rate segment (higher debt → lower high interest rate), which is inadmissible and eliminates multiplicity. A bimodal distribution with well-separated high and low growth states, as observed empirically in crisis-prone countries, generates an upward-sloping correspondence at both rates, creating a region of intermediate debt levels where either rate is an equilibrium.

The second key model feature is the timing of moves: the borrower chooses current debt (amount borrowed today) rather than debt at maturity (the repayment obligation). When the borrower chooses debt at maturity, it implicitly pins down the default probability and therefore the interest rate, eliminating multiplicity. When the borrower chooses current debt, the interest rate is determined by lenders and can take either the high or the low value consistent with break-even pricing, given the chosen debt level.

Q2. How does the paper calibrate the endowment process and what does the estimation reveal?

The paper estimates a two-state Markov regime-switching model for annual GDP per capita growth for Argentina, Brazil, Italy, Portugal, and Spain using data from 1980 to 2017, and finds clear evidence of a bimodal distribution with persistent high- and low-growth regimes across all five countries. Estimated using a Bayesian MCMC algorithm with the Kim (1994) filter, the posterior means for the benchmark cross-country calibration are: low-growth rate gL = −1.0% per year, high-growth rate gH = 3.0% per year, persistence of low-growth state pL = 0.60, persistence of high-growth state pH = 0.80, and standard deviation of transitory shocks σ = 0.015. The average gap between gL and gH across countries is approximately 6 percentage points, more than three times the standard deviation of the transitory shock — confirming the bimodal structure that is essential for multiplicity. Both growth regimes are persistent, with the low-growth state having 60–80% persistence across countries.

The quantitative model uses these estimates together with standard parameters: risk-free rate R* = 3.5%, recovery rate κ = 75%, discount factor β = 0.75, and risk aversion γ = 3. The sunspot process governing equilibrium selection is i.i.d. with a 5% probability of the bad (high-rate) sunspot in each period.

Q3. What does the calibrated model predict for interest rates and when do self-fulfilling crises occur?

In the calibrated quantitative model, the multiplicity region is present only in the low-growth state and for intermediate debt levels; in the high-growth state, the multiplicity region is either empty or negligibly small. In the low-growth state with intermediate debt, the interest rate schedule features two admissible equilibria: a low-rate equilibrium consistent with a low probability of default (1.7% in the benchmark simulation) and a high-rate equilibrium consistent with a high probability of default (60.1%). Both are sustained by self-confirming expectations. The scenario simulation illustrates this starkly: two economies starting from identical wealth and facing identical growth-shock sequences but different sunspot realizations in period t=2 (when they are in the low-growth state) face interest rates of 4.0% versus 46.4% and next-period default probabilities of 1.7% versus 60.1%, respectively, with no difference in fundamentals.

The model also generates endogenous austerity: borrowers optimally refrain from increasing debt to avoid discrete jumps in interest rates, both at the fundamental threshold (driven by the growth regime) and at the expectations-driven threshold (driven by the sunspot). In low-growth states facing the bad sunspot, the borrower either bunches at a low debt level below the multiplicity region or makes a discrete jump above it, echoing the binary fiscal-adjustment dynamics observed in crisis episodes.

Q4. How does the model interpret the European debt crisis and the role of the ECB?

The model provides a direct interpretation of the European sovereign debt crisis: the southern European economies (Italy, Spain, Portugal) entered a low-growth state around 2009–2010, which created conditions for the Calvo mechanism to operate; spreads jumped to high-rate equilibria driven by expectations rather than by fundamentals alone. The ECB’s announcement of the Outright Monetary Transactions (OMT) program in September 2012 — the commitment to purchase sovereign bonds in secondary markets — shifted lenders’ beliefs from the bad-sunspot equilibrium to the good-sunspot equilibrium, collapsing spreads substantially even without actual intervention. In the model’s language, a credible lender of last resort can eliminate the bad equilibrium by committing to lend at the low-rate schedule, rendering the high-rate self-fulfilling expectations non-viable. The Argentine crisis of 1998–2002 fits the model as an alternative trajectory: Argentina entered the low-growth state with a 7% spread on 35% debt-to-GDP and, without a lender of last resort intervention, eventually defaulted in 2002 — consistent with the bad-sunspot equilibrium path.

Q5. What role does persistence of the low-growth state play?

The persistence of the low-growth state (pL) is the key parameter governing the severity of self-fulfilling crises: higher pL generates higher equilibrium interest rates in the bad-sunspot equilibrium and a larger multiplicity region. Intuitively, if the economy is likely to remain in the low-growth state for a long time, the probability of default conditional on entry into the bad equilibrium is very high, requiring lenders to charge very high interest rates to break even. The higher the interest rate, the more debt service costs compress fiscal space, making default even more likely and potentially sustainable at even lower debt levels. The paper shows in robustness exercises that the multiplicity result is robust to reasonable perturbations in pL, κ (recovery rate), σ (transitory shock standard deviation), gL, and gH, with the key ingredient being the bimodal structure of the endowment process rather than any single parameter value.

Q6. What are the policy implications for lenders of last resort?

The central policy implication is that a lender of last resort — such as the ECB or the IMF — is justified precisely when fundamentals are weak, not because fundamentals alone cause the crisis but because weak fundamentals create conditions in which expectations can trigger a self-fulfilling crisis. Intervening in the bad-sunspot equilibrium by committing to supply funds at low-rate terms makes the high-rate equilibrium infeasible: lenders cannot expect default because the lender of last resort ensures the borrower can always roll over at low rates. The model thus rationalizes the design of the OMT: a credible commitment with no limit on size is sufficient to rule out the bad equilibrium without necessarily requiring actual asset purchases. The paper notes that such interventions will also have effects on the economy outside the period of crisis, since the availability of backstop financing may affect the equilibrium path more broadly.

Key Concepts

Calvo (1988) mechanism: the self-fulfilling loop in sovereign debt markets in which high lender expectations of default require high interest rates for break-even pricing, which raise the actual default probability and thereby confirm the initial pessimistic expectations; generates multiple equilibrium interest rates for a given debt level.
state-dependent multiplicity: the feature of the model in which multiple interest rate equilibria arise only in periods of low and persistent growth (stagnation), not in high-growth regimes; the central quantitative finding that self-fulfilling crises are empirically plausible only when growth fundamentals are weak.
endogenous austerity: the borrower’s optimal choice to hold debt below the multiplicity region to avoid discrete jumps in interest rates triggered by either fundamentals or expectations; reflected in the flat portions of the debt policy function and consistent with fiscal consolidation patterns observed in crisis episodes.
sunspot variable: the exogenous coordination device that selects among the multiple equilibrium interest rate schedules; takes a bad or good realization each period according to an i.i.d. process, with the bad sunspot selecting the high-rate schedule and the good sunspot selecting the low-rate schedule in the low-growth state.

Silence to Solidarity: How Communication About a Minority Affects Discrimination

Mon, 01 Jan 0001 00:00:00 +0000

This paper examines how two types of communication about a minority group affect discriminatory behavior: (i) horizontal communication between majority-group members, and (ii) top-down communication from agents of authority such as the legal system. The setting is urban Chennai, India, where the paper measures discrimination against thirunangai — a community of transgender women who are India’s most visible LGBTQ+ group — in a field experiment with 3,397 participants.

Discrimination is measured using incentivized hiring choices. Participants are offered a free grocery delivery and make 10 binary choices over which worker will carry out the delivery, with worker gender (cisgender male, cisgender female, or transgender) varying across options. The stakes are real: one choice is randomly selected and implemented 2–9 weeks later. Participants in the control condition are highly discriminatory: they are 19 percentage points (32%) less likely to hire a transgender worker than a non-transgender worker (p<0.001), and are willing to sacrifice grocery items worth 1.9 times their median daily per capita food expenditure to avoid a 15-minute interaction with a transgender worker.

The first main treatment involves randomly assigning participants to a 3-person group discussion with two neighbors, in which they discuss and make collective hiring choices over the same options. The key outcome is participants’ subsequent private, individual hiring choices. The discussion eliminates anti-transgender discrimination on average: participants in the discussion arm are 17 percentage points (42%) more likely to select a transgender worker in their private post-discussion choices relative to the control group (p<0.001), so that discrimination is no longer statistically distinguishable from zero (p=0.30). The discussion’s effect is partially persistent: approximately one month later, discussion participants are still 4 percentage points more likely to select transgender workers in hypothetical hiring choices (p=0.03), representing roughly 25% of the short-run effect.

The second main treatment cross-randomizes a video shown before hiring choices. The legal rights video informs participants of a Supreme Court ruling affirming that transgender people hold the same fundamental constitutional rights as other citizens. This reduces discrimination by 10.3 percentage points (p<0.001). A rights messaging video — which argues that transgender people should have equal rights without invoking legal authority — reduces discrimination by a smaller 5.8 percentage points (p=0.001), and there is some evidence the legal-authority version is more effective (p of difference in [0.01, 0.12]). However, the legal rights video’s effect is only 59% as large as the discussion’s effect (p of difference in [0.002, 0.04]), and it does not persist at the one-month follow-up (p in [0.12, 0.51]).

The paper rules out two candidate mechanisms for the discussion’s effects and supports a third. First, the discussion does not work primarily through correcting misperceived norms: while control-group participants do overestimate peer discrimination by 5 percentage points, the discussion reduces predicted discrimination by 24 percentage points — far more than a corrected misperception could explain (at most 21% of the effect under generous assumptions). Second, the discussion does not work through virtue signaling alone: a “No discussion (public)” arm in which participants make individually-visible choices shows no reduction in discrimination on average (p=0.83). Third, the paper provides affirmative evidence for a persuasion channel: participants in a “listener” arm, who silently observe a 2-person discussion without participating, discriminate 13 percentage points less than the control group (p<0.001), an effect that is highly persistent at the 2–9 week follow-up (11 percentage points, p<0.001). The persuasion mechanism is further supported by the finding that pro-trans participants are more vocal: each additional transgender worker chosen in post-discussion private choices is associated with a 32% higher probability of speaking first (p=0.03) and a 27% higher probability of dominating the discussion (p=0.02). Statements about transgender workers during discussions were 5.7 times more likely to be positive than negative. Listeners who heard moral argumentation about equality, rights, and giving opportunities subsequently discriminated less (p<0.001).

Scope conditions: the study is conducted among urban Chennai residents (85% female), where transgender identity is visually recognizable and socially salient, awareness of the 2014 Supreme Court ruling is low (36% could not identify a single legal right transgender people hold), and a wedge exists between descriptive norms (high actual discrimination) and prescriptive norms (93% of the control group rate explicit discrimination as wrong). The model’s “sweet spot” logic implies these effects may not generalize to settings where discrimination is either near-universal (no privately pro-trans individuals to be vocal) or already minimal (no incentive to persuade).

Q: How is anti-transgender discrimination measured in the experiment? A: Participants make 10 incentive-compatible binary hiring choices over grocery delivery workers, with one choice randomly selected and implemented 2–9 weeks later. Discrimination is defined as the reduction in the probability of selecting the alternative worker when that worker is transgender versus non-transgender, conditional on other option characteristics such as items offered and reliability score. Participants are told they will have a 15-minute conversation with the selected worker, ensuring anticipated social contact. The design is framed as market research to obfuscate the study’s purpose; only 8% correctly guessed the true focus.

Q: How large is baseline discrimination in the control group? A: In the No discussion (private) control condition, participants are 19 percentage points (32%) less likely to hire a transgender worker than a non-transgender worker (p<0.001). In willingness-to-pay terms, participants sacrifice grocery items worth 1.9 times their median daily per capita food expenditure (Rs. 127 on a base of Rs. 67) to avoid selecting a transgender worker. Even when a transgender worker dominates on both items and reliability score, participants in the control group still select the non-transgender worker 47% of the time.

Q: What is the main effect of the 3-person group discussion on subsequent discrimination? A: Participants who engage in a group discussion with two neighbors are 17 percentage points more likely to select a transgender worker in their subsequent private individual choices (p<0.001). This eliminates average discrimination entirely: in the discussion arm, the probability of selecting a transgender worker is not statistically distinguishable from the probability of selecting a non-transgender worker (p=0.30). The willingness-to-pay to avoid a transgender worker falls from Rs. 127 to Rs. 13 (p of difference < 0.001), and is no longer significantly different from zero (p=0.265).

Q: How persistent are the effects of the group discussion? A: At the 2–9 week follow-up survey (mean 35 days), discussion participants are approximately 4 percentage points more likely to select transgender workers in hypothetical hiring choices (p=0.03). This represents approximately 25% of the short-run 17 percentage point effect, a decay rate comparable to the persistence of US political advertising effects in the political science literature (Hill et al., 2013, estimate 10–15% remaining after 30 days).

Q: What is the effect of the legal rights video, and how does it compare to the discussion? A: The legal rights video — informing participants of the Supreme Court ruling affirming transgender people’s fundamental constitutional rights — increases the probability of selecting a transgender worker by 10.3 percentage points (p<0.001). The rights messaging video, which argues that transgender people should have equal rights without invoking legal authority, increases it by 5.8 percentage points (p=0.001). The legal rights video’s effect is only 59% as large as the discussion’s 17 percentage point effect (p of difference in [0.002, 0.04]), and unlike the discussion, neither video’s effect is detectable at the one-month follow-up (p in [0.12, 0.51]).

Q: Does the legal rights video work through a different channel than the rights messaging video? A: There is evidence that the legal authority of the Supreme Court matters beyond the content of the rights message. The legal rights video is more effective than the rights messaging video at reducing discrimination (p of difference in [0.01, 0.12]), and the legal rights video (but not the rights messaging) affects participants’ beliefs about the legal status of transgender people (as measured by a summary index). Both videos shift perceived descriptive norms — participants predict others will select transgender workers more, by 2–6 percentage points — but neither significantly affects attitudes as measured by a list experiment or disapproval questions.

Q: Does the discussion work through correcting misperceived norms? A: This channel can account for at most a small fraction of the effect. Control-group participants do overestimate peer discrimination by 5 percentage points in incentivized predictions (p<0.001, as measured by predicted probability of selecting a transgender worker). However, the discussion reduces predicted discrimination by 24 percentage points (p<0.001), far exceeding the initial misperception. Even under generous assumptions in which the misperception is precisely corrected, this mechanism could account for no more than 21% of the discussion’s treatment effect (95% CI: [8.9%, 32.5%]).

Q: Does the discussion work through virtue signaling? A: The evidence rules out virtue signaling as the primary channel. The “No discussion (public)” treatment arm makes participants’ individual hiring choices visible to their group members, exogenously increasing social image concerns in the absence of a discussion. This has no detectable average effect on discrimination (p=0.83), indicating that social image concerns alone — without the persuasive content of an actual discussion — do not explain the reduction in discrimination generated by the group discussion.

Q: What is the evidence for the persuasion mechanism? A: The “listener” treatment arm provides direct evidence. In this arm, one participant silently observes a 2-person discussion without speaking, then makes private individual choices. Listeners discriminate 13 percentage points less than the control group (p<0.001), an effect statistically indistinguishable from full discussion participants. Since listeners changed their behavior based solely on what they heard and saw, this constitutes evidence of persuasion. The listener effect is highly persistent at the 2–9 week follow-up (11 percentage points, p<0.001) and holds on a robustness outcome designed to be completely private. The implied persuasion rate is 29%, described as high relative to values in the literature (DellaVigna & Gentzkow, 2010).

Q: Why do pro-trans participants persuade others — what drives the discussion’s content? A: Pro-trans participants are disproportionately vocal. Each additional transgender worker chosen in post-discussion private choices (a proxy for pro-trans private attitudes) is associated with a 32% higher probability of speaking first (p=0.03) and a 27% higher probability of dominating the discussion (p=0.02), but only when discussing a choice involving a transgender worker. The overall tone of discussions is strongly pro-trans: statements about transgender workers are 5.7 times more likely to be positive than negative. Participants who hear moral argumentation about equality, rights, and giving opportunities subsequently discriminate significantly less (p<0.001).

Q: Does the discussion work by changing statistical (belief-based) discrimination? A: Partially, baseline discrimination in the control group is partly statistical: despite transgender workers having the same average reliability scores as others, participants rate them as less likely to complete a delivery, and revealing the true reliability score makes participants 2.9 percentage points more likely to select a transgender worker (an effect unique to transgender workers). However, the discussion does not significantly affect beliefs about transgender workers’ reliability, and there is no detected reduction in the belief-based component of discrimination in the discussion arm (though the test is underpowered).

Q: Are the effects of the discussion and the legal rights video additive? A: The two interventions appear to combine approximately linearly for the legal rights video: there are no detected interaction effects (p in [0.83, 0.96]). By contrast, there is weak evidence of a negative interaction between the rights messaging video and the discussion, suggesting these two may be substitutes — consistent with the rights messaging video’s content being similar to the pro-trans moral argumentation already present in discussions.

Q: What alternative explanations are ruled out? A: The paper tests and finds no support for: (i) photo characteristics such as perceived caste driving results; (ii) social image concerns affecting even post-discussion private choices (the “extra private” robustness outcome designed to be unobservable by neighbors yields similar results); (iii) increased contemplation or deliberation about choices; (iv) experimenter demand effects or social desirability bias (treatment effects do not differ for the 8% who guessed the study’s purpose); (v) increased salience of the transgender category; and (vi) cheap talk from low stakes (choices were incentive-compatible and implemented).

Q: What is the study’s theoretical model for why pro-trans participants speak out? A: The paper develops a model combining social signaling (people want to fit in with their group; Bénabou & Tirole, 2006) with direct persuasion (participants can change each other’s preferences through messages). Under the right conditions, only pro-trans participants send persuasive pro-trans messages. This occurs in a “sweet spot” range: when average discrimination is not so strong that no one is privately pro-trans, and not so weak that pro-trans participants lack an incentive to persuade (since they are already in the majority). The context in Chennai — high actual discrimination but strong social norms against it — satisfies this sweet spot condition.

Q: What are the policy implications regarding horizontal versus top-down communication? A: In this context, facilitating horizontal communication between neighbors is a more effective tool for reducing discrimination than top-down communication about legal rights: the discussion’s effect is 1.7 times larger than the legal rights video (17 p.p. vs. 10.3 p.p.) and partially persists at one month, whereas the legal rights video’s effect does not persist. However, the legal rights video does reduce discrimination relative to the rights messaging video, suggesting that communicating the legal authority of the Supreme Court carries independent weight beyond rights advocacy messaging. Both interventions are complementary when combined.

Horizontal communication: Communication between members of the majority group about a minority, as distinct from contact between majority and minority groups or top-down communication from authority. In this paper, operationalized as a group discussion among three neighbors who make collective hiring choices.

Top-down communication: Communication from agents of authority — here, the legal system — about a minority group’s rights. Measured via a video informing participants of a Supreme Court ruling affirming transgender people’s constitutional rights.

Anti-transgender discrimination: In the paper’s own measurement, the reduction in the probability that a worker is chosen because they are transgender (relative to being non-transgender), conditional on other delivery option characteristics. Measured in incentivized, privately-elicited binary hiring choices.

Expressive law hypothesis: The theory that changes in the law affect behavior by changing people’s perception of the prevailing social norm, not (only) through deterrence. The paper tests this by comparing a legal rights video (invoking Supreme Court authority) to a rights messaging video with identical content but no legal backing, finding the legal-authority version more effective.

Persuasion channel: The mechanism by which discussion participants change each other’s preferences through persuasive messages, particularly moral arguments about equality and rights. Distinguished in the paper from virtue signaling (publicly visible pro-trans behavior) and norm correction (updating misperceived beliefs about peer behavior).

Pluralistic ignorance: A setting in which people misperceive how common discriminatory attitudes are among their peers, potentially hiding genuine minority support for the discriminated group. The paper tests this as a candidate mechanism and finds it can account for at most 21% of the discussion effect.

Sweet spot condition: The range of average group discrimination levels in which pro-trans participants have both the motivation and opportunity to speak out persuasively — discrimination is not so universal that no one is privately pro-trans, and not so minimal that the pro-trans participants feel no need to persuade others. The paper argues the Chennai context satisfies this condition.

The Effect of Education Policy on Crime: An Intergenerational Perspective

Mon, 01 Jan 0001 00:00:00 +0000

This paper studies the intergenerational effects of education policy on crime, asking whether a compulsory schooling reform that reduced crime among those directly exposed also reduced crime among their children. The authors exploit the staggered municipal rollout of Sweden’s comprehensive school reform, implemented gradually between 1949 and 1962 across more than 1,000 municipalities, which increased compulsory schooling by one to two years, abolished tracking into academic and vocational streams after 6th grade, and introduced a uniform national curriculum. The parent generation consists of all individuals born in Sweden between 1945 and 1955 (approximately 447,000 men and 450,000 women), and their children form the child generation (426,721 sons observed from age 15 to 29). Crime is measured by administrative conviction records from the Swedish National Council for Crime Prevention covering 1973–2010.

The empirical strategy is difference-in-differences, comparing changes in conviction rates across cohorts in municipalities that implemented the reform at different times, with treatment assigned based on the parent’s birth municipality to avoid endogenous sorting bias. Standard errors are clustered at the municipality level. Parallel trends validity is supported by three tests: results are unchanged when municipality-specific linear trends are included, placebo tests using incorrect reform dates yield effects indistinguishable from zero, and residuals from crime regressions show no correlation with municipality-specific trends.

The main finding is a significant 0.79 percentage point (pp) decline in conviction rates among sons of fathers exposed to the reform (p-value < 0.002), representing a 3.4 percent reduction relative to baseline. The decline spans multiple crime types: violent crime fell by 0.27 pp, traffic-related crime by 0.45 pp, fraud by 0.22 pp, and other offenses by 0.41 pp — percentage reductions of three to six percent across categories. Multiple convictions fell by 0.43 pp (5.8 percent). These second-generation effects are driven entirely by paternal exposure: the impact of maternal reform exposure is an order of magnitude smaller and statistically insignificant, and the difference between paternal and maternal effects is itself significant (p-value 0.048 for any conviction, 0.009 for multiple convictions). Effects on daughters in the child generation are much smaller, with only the residual “other crime” category showing a significant 0.129 pp (15.5 percent) decline.

The asymmetry between paternal and maternal transmission is explained by the first-generation effects of the reform. For men, the reform increased schooling by 0.32 years, earnings by approximately 1 percent, the probability of white-collar employment by 1.2 percent, cognitive skills by 0.14 standard deviations, noncognitive skills by 0.17 standard deviations, spousal earnings by 1,022 SEK per year, and overall household income by approximately 1 percent. For women, the reform increased education by 0.21 years but did not raise earnings, household income, or white-collar employment, and did not reduce their already low crime rates. Only 13 percent of women in the 1945–55 cohorts were at or below the compulsory schooling threshold, versus 20 percent of men, substantially limiting the reform’s bite for women.

A mediation analysis decomposes the intergenerational transmission through three channels: fathers’ education accounts for 64.8 percent of the indirect effect, the decline in paternal crime accounts for 18.5 percent, and the increase in household disposable income accounts for 16.7 percent. The direct effect (unexplained by these mediators) accounts for 48 percent of the total effect. The paper also documents that children of treated fathers attended schools with lower peer crime rates and lived in neighborhoods with lower youth crime rates, supporting a neighborhood and peer effects channel alongside human capital and role-model channels.

Scope conditions: the study covers male children observed to age 29 in Sweden; results apply to a context of near-universal administrative records, a specific postwar schooling reform, and cohorts born 1945–1955 in a Nordic welfare state.

Q: What is the magnitude of the intergenerational crime reduction caused by the reform?

A: Sons of fathers exposed to the reform experienced a 0.79 pp decline in conviction rates (p-value < 0.002), corresponding to a 3.4 percent reduction relative to the baseline conviction rate of approximately 24 percent for the child generation by age 29. Multiple convictions fell by 0.43 pp, a 5.8 percent reduction. These magnitudes are similar in percentage terms to the direct crime reduction the reform caused among fathers themselves.

Q: Does the reform’s intergenerational effect on crime differ by the sex of the treated parent?

A: Yes. The intergenerational effect is driven entirely by paternal exposure to the reform: the effect of maternal exposure is an order of magnitude smaller and insignificant at any conventional significance level. The difference between paternal and maternal effects is statistically significant, with p-values of 0.048 for any conviction and 0.009 for multiple convictions. The paper attributes this asymmetry to the much weaker first-generation effects of the reform on women’s earnings, household income, crime rates, and neighborhood sorting.

Q: Which crime types declined significantly among sons of treated fathers?

A: Significant declines were found in violent crime (−0.27 pp, Romano-Wolf p-value 0.09), traffic-related crime (−0.45 pp, RW p-value 0.057), fraud (−0.22 pp, RW p-value 0.09), and other offenses (−0.41 pp, RW p-value 0.047), each representing a three-to-six percent reduction relative to the mean incidence of that crime type. Property crime and drug-related crime did not show significant declines.

Q: What were the direct effects of the reform on the parent generation’s human capital?

A: For men, the reform increased schooling by 0.32 years, earnings by approximately 1 percent, the probability of white-collar employment by 1.2 percent, cognitive skills by 0.14 standard deviations, and noncognitive skills by 0.17 standard deviations, all measured at military enlistment. Spousal earnings increased by 1,022 SEK per year and overall household income rose by approximately 1 percent. For women, education increased by 0.21 years and marriage market matches improved, but earnings, household income, and white-collar employment probability did not increase significantly.

Q: Why did the reform have stronger first-generation effects on men than on women?

A: The average share of individuals at or below the compulsory schooling threshold — the margin at which the reform was binding — was 20 percent for men but only 13 percent for women in the 1945–55 cohorts. Because fewer women were constrained by the old compulsory schooling limit, the reform increased their education by less and produced smaller downstream effects on earnings and labor market outcomes.

Q: What are the three channels through which the reform reduces child crime, and what is the relative contribution of each?

A: The paper identifies three channels: (1) the human capital channel, whereby increased parental education raises household income and child human capital; (2) the role model channel, whereby reduced paternal crime participation directly reduces son’s crime; and (3) the neighborhood and peer effects channel, whereby higher income enables sorting into lower-crime neighborhoods and better schools. The mediation analysis attributes 64.8 percent of the indirect effect to fathers’ increased education, 18.5 percent to the decline in paternal crime, and 16.7 percent to the increase in household disposable income. The direct effect unexplained by these three mediators accounts for 48 percent of the total effect.

Q: What is the role model effect, and how strong is it in the parent generation?

A: The role model channel operates through the strong intergenerational persistence in crime participation: sons are 2.06 times more likely to participate in crime if their fathers have been convicted (Hjalmarsson and Lindquist, 2012). The reform reduced the incidence of any conviction among treated men by 1.5 pp and repeat convictions by 1.5 pp — the latter representing an approximately 8 percent decline from a lower base. For women, the reform produced no reduction in crime, providing no analogous role model improvement through the maternal channel.

Q: How does neighborhood and school peer quality change for children of treated fathers versus treated mothers?

A: Sons of fathers exposed to the reform moved to neighborhoods with lower youth crime rates (−0.087 pp) and attended schools with lower peer crime rates (−0.077 pp). In contrast, sons of mothers exposed to the reform experienced higher neighborhood crime rates (p-value 0.06) and higher school peer crime rates (p-value 0.01), the opposite direction. This asymmetry helps explain why only paternal treatment generates significant second-generation crime reductions.

Q: What happens to other outcomes for children of treated fathers beyond crime?

A: Sons experienced a 1.2 percentile increase in school GPA (RW p-value 0.05), a 2.3 pp increase in employment (RW p-value 0.04), a matching 2.3 pp decline in unemployment benefit receipt, a reduction in hospitalization of 2.4 days (17 percent, RW p-value 0.02), and a decline in prescribed drugs of 31 doses (2.8 percent, RW p-value 0.09). The decline in prescribed drugs for sons is driven by nervous system drugs and painkillers, pointing to improved mental health. Daughters of treated fathers show a significant reduction in welfare dependency but no other significant improvements.

Q: How does the paper validate the parallel trends assumption?

A: Three tests are reported. First, including municipality-specific linear trends leaves the main coefficient unchanged (p-value 0.85 for the trend terms themselves). Second, placebo contrasts using incorrect reform implementation dates produce effects indistinguishable from zero for all tested dates. Third, graphical inspection of regression residuals shows no correlation with municipality-specific trends. Together these provide strong support for the identifying assumption.

Q: Are the results sensitive to using a linear probability model instead of a nonlinear model?

A: A Monte Carlo experiment was conducted replicating observed crime rates across municipalities and imposing the estimated average treatment effect. Assuming the true data-generating process is a probit model, the linear probability model biases the estimated average effect upward by only 5 percent — a difference that is statistically indistinguishable from zero in the actual data — validating the OLS approach.

Q: What is the broader policy implication of the findings?

A: The results show that well-designed education policies can reduce crime not only among the directly treated generation but also among their children, amplifying the social benefits of reform across generations. The authors interpret this as consistent with the theoretical framework of Becker and Tomes (1979) on intergenerational transmission of human capital, and suggest that education policy evaluations that focus only on the treated generation substantially understate total social returns.

Intergenerational transmission of education reform effects: the phenomenon whereby an education policy that raises parental human capital produces improvements in children’s outcomes — including crime — through multiple channels including resource increases, parental role modeling, and neighborhood sorting, beyond any direct policy exposure of the child generation.

Comprehensive school reform (Sweden, 1949–1962): a nationally mandated restructuring of compulsory schooling that extended required attendance by one to two years, abolished selection into academic and vocational tracks after 6th grade, and introduced a uniform national curriculum, rolled out staggered across 1,055 Swedish municipalities.

Human capital channel: the mechanism by which increased parental education raises earnings and household income, enabling greater investments in children’s development and exploiting complementarity between parental and child human capital in the skill production function, thereby raising children’s opportunity cost of crime.

Role model channel: the mechanism by which reduced parental crime participation directly reduces children’s crime, operating through the transmission of norms and information across generations; identified empirically by the strong intergenerational correlation in convictions (sons with convicted fathers are 2.06 times more likely to be convicted themselves).

Neighborhood and peer effects channel: the mechanism by which increased parental income from the reform enables sorting into residential neighborhoods and schools with lower youth crime rates, exposing children to peers less involved in illegal activities and thereby reducing their own crime participation.

Mediation analysis: a decomposition method following Heckman, Pinto, and Savelyev (2013) that quantifies the share of a total treatment effect accounted for by specific intermediate variables (here: fathers’ education, fathers’ crime participation, and household disposable income) versus the direct unexplained effect.

Conviction rate: the proportion of individuals in a given generation and observation window who received at least one criminal conviction in Swedish administrative records; used as the primary outcome measure because it captures offenses that led to a court appearance, excluding minor infractions resolved by direct fine.

The Impact of Unions on Nonunion Wage Setting: Threats and Bargaining

Mon, 01 Jan 0001 00:00:00 +0000

This paper estimates the impact of unions on nonunion wage setting in the United States over the period 1980–2010, distinguishing two channels through which unions affect nonunion wages: (1) a traditional threat channel, in which nonunion firms raise wages to preempt unionization by making workers indifferent between forming a union and remaining nonunion (an “emulation wage”); and (2) a bargaining channel, in which nonunion workers use the availability of high-paying union jobs as part of their outside option when bargaining individually with their employer, so that a decline in union job prevalence or the union wage premium erodes nonunion bargained wages even at firms that face no direct unionization threat.

The authors build a search-and-bargaining model grounded in Nash bargaining, with endogenous union formation and, in the most complete version, the possibility of nonunion firm responses to the threat of unionization. Workers in this model can be employed at simple nonunion firms, union firms, or union-emulating firms. The model is embedded in a multi-industry, multi-city framework following Beaudry, Green, and Sand (2012), which formalizes the mechanism by which higher-rent jobs in a city raise outside options and therefore wages for workers in all other jobs throughout that local labor market. This cross-city, within-industry variation is the primary source of identification.

The empirical implementation uses Current Population Survey Merged Outgoing Rotation Groups (1983–2020) and CPS May extracts (1978–1982), pooling observations around 1980, 1990, 2000, 2010, and 2020 across 43 cities and 51 industries. To address endogeneity of outside option variables — which may be correlated with unobserved local productivity shocks — the authors construct Bartik-style instruments based on start-of-period local industry and union employment composition interacted with national changes in industry growth, industry wage premia, and union job transition probabilities. The threat channel is identified by the interaction of the probability a firm in a given industry-city cell faces a union election (proxied using NLRB data) with the outside option value of union workers. The authors derive a model-based overidentifying restriction, test it, and cannot reject it, providing support for their identification strategy.

The central quantitative finding is that de-unionization accounts for approximately 38% of the 16% decline in the mean real (composition-constant) wage in a typical US city between 1980 and 2010. One-third of that de-unionization effect arises from a standard shift-share component — workers moving from higher-paying union jobs to lower-paying nonunion jobs — while two-thirds arises from spillover channels affecting nonunion wage setting. The spillover effects are almost entirely attributable to the bargaining channel rather than the traditional threat channel; the threat probability was too low, even in 1980, to generate large emulation effects in the aggregate. The total impact of a one-dollar increase in the outside option value for the mean wage in industry i is estimated at 1.78 dollars once within-industry feedback loops are included.

The paper finds no evidence of bargaining spillovers in the 1980s specifically, the decade of the sharpest unionization declines. The offsetting forces were declining probabilities of finding union jobs and simultaneously rising union wage premia — with the model explaining the premium increase as a consequence of nonunion firms no longer needing to emulate union wages once the threat of their shop being organized receded substantially. After 1990 the threat stabilized at a low level, the premium declined, and the outside-option effect of declining unionization became the dominant force.

Heterogeneity results show that spillover effects are larger for women than men, and that de-unionization accounts for 43% of the real wage decline for women versus 27% for men. For workers without post-secondary education, de-unionization accounts for 43% of their real wage decline. The traditional threat effect is statistically insignificant in states with Right-to-Work laws, consistent with the interpretation that identification captures emulation responses to unionization threat.

Q: What are the two channels through which unions affect nonunion wages in this model? A: The traditional threat channel operates when nonunion firms raise wages to make workers indifferent between unionizing and remaining nonunion, thereby forestalling a costly union election. The bargaining channel operates because nonunion workers can credibly point to available union jobs when bargaining individually; a decline in union job prevalence or the union wage premium therefore weakens nonunion workers’ outside options and lowers their bargained wages even at firms that face no direct unionization threat.

Q: How large is the overall contribution of de-unionization to the US wage decline between 1980 and 2010? A: The paper estimates that de-unionization accounts for 38% of the approximately 16% decline in the mean composition-constant real wage in a typical US city between 1980 and 2010. One-third of that 38% arises from the direct shift-share effect of workers moving from higher-paying union to lower-paying nonunion employment; the remaining two-thirds arises from spillover effects on nonunion wages.

Q: Which spillover channel dominates in the decomposition, and why? A: The bargaining channel dominates almost entirely. The traditional threat channel is statistically significant but quantitatively small because the probability that any given nonunion firm faced a union election was low even in 1980, so the scope for emulation to affect aggregate wages was limited. The bargaining channel, by contrast, operates through the outside options of all nonunion workers searching across many industries and cities, giving it broader aggregate reach.

Q: Why was there no measurable bargaining spillover in the 1980s despite the decade’s large drop in union density? A: During the 1980s, two forces offset each other: the probability of a nonunion worker finding a union job fell sharply, but the union wage premium rose substantially over the same period, so the expected value of the union outside option changed little. The paper explains the rising premium as a consequence of nonunion firms reducing their emulation wages as the threat of unionization receded, causing nonunion wages to fall faster than union wages and thus mechanically widening the premium. After 1990, when the threat stabilized at a low level, the premium declined and the net outside-option effect of continued de-unionization became the dominant spillover force.

Q: What is the estimated multiplier effect of an improvement in outside options on nonunion wages? A: The total impact of a one-dollar increase in the outside option value on the mean wage in a given industry is estimated at 1.78 dollars once within-industry feedback loops — in which an improved outside option raises wages, which in turn improves outside options elsewhere — are accounted for.

Q: How do the authors address endogeneity of the outside option variables? A: They construct Bartik-style instruments based on start-of-period local industry and union employment composition interacted with national-level changes in industry growth, industry wage premia, and the probability of transitioning to a union job. This strategy isolates variation in local outside options that is driven by predetermined compositional exposure rather than contemporaneous local shocks. They derive a model-based overidentifying restriction, test it in the data, and cannot reject it, supporting the validity of the instrument.

Q: How do the authors address selection bias arising from the changing composition of union and nonunion workers as unionization declines? A: They implement a generalized Heckman two-step approach, including a quartic in the change in the proportion unionized to control for selectivity. After this correction, they cannot reject the null of no selectivity effects, and the main estimated coefficients change very little, indicating that compositional selection is not the primary driver of their results.

Q: What heterogeneity is found across gender groups? A: Both the bargaining and traditional threat effects are larger for women than for men. Men experienced a decline in mean real wages between 1980 and 2010 more than double that experienced by women, but spillover effects are of identical size, so de-unionization accounts for a larger share of women’s wage decline (43%) than men’s (27%).

Q: What heterogeneity is found by education level? A: For workers with a high school education or less, the traditional threat effect estimate is twice as large as the bargaining effect, while the reverse holds for workers with post-secondary education. Workers without post-secondary education experienced real wage declines nearly triple those of the more educated group, and de-unionization accounts for 43% of the lower-educated group’s wage decline.

Q: How do the authors validate that they are identifying the threat channel rather than some other effect? A: The traditional threat effect is estimated to be statistically insignificant in states with Right-to-Work (RTW) laws, where the legal environment substantially reduces the ability of workers to organize and therefore reduces the credible threat of unionization that would induce nonunion firms to emulate union wages. This pattern is consistent with the interpretation that the identified effect captures firm emulation responses to a genuine unionization threat.

Q: What are the policy implications of the distinction between the two channels? A: The traditional threat effect can only be activated by increasing union power directly, since it depends on a credible risk of a firm’s workforce voting to unionize. The bargaining channel, however, is not union-specific: any policy that raises workers’ outside option values — such as eliminating non-compete agreements or expanding access to higher-paying jobs in a local labor market — can generate similar wage spillovers. Unions are one powerful mechanism for doing this, but not the only one.

Q: What is the theoretical model structure, and what distinguishes it from Taschereau-Dumouchel (2020)? A: The model is built on TD’s search-and-bargaining framework with endogenous union formation, in which unions can threaten to withdraw the entire workforce from production whereas individual nonunion workers can only threaten to withdraw their own labor. The key modifications are: (1) the hiring-channel mechanism of TD (firms skew toward skilled workers who dislike unions) is replaced with a direct wage-emulation mechanism; (2) the BGS multi-industry, multi-city framework is incorporated to allow outside options to vary with the composition of jobs across industries in a locality; and (3) a single skill level with multiple industries is used, keeping the model tractable for empirical implementation.

Q: What data sources are used and over what period? A: The primary dataset is the Current Population Survey Merged Outgoing Rotation Groups for 1983–2020 combined with CPS May extracts for 1978–1982, covering workers aged 25–65 not enrolled in school. The sample is organized into 93 geographic areas (43 cities), 51 industries based on 1980 Census classification, and analyzed at 10-year intervals (1980, 1990, 2000, 2010, 2020) with three-year pooling windows to reduce noise. NLRB case data on union elections proxies for unionization threat probabilities, and County Business Patterns data are used in constructing emulation probabilities.

Traditional threat effect: The mechanism by which nonunion firms raise wages to an “emulation wage” — the level that makes workers indifferent between unionizing and remaining nonunion — in order to preempt the costs of a union election, thereby reducing the net benefit of unionization below the threshold required for workers to vote for a union.

Bargaining channel (bargaining spillover effect): The mechanism by which the availability of union jobs in a local labor market raises the outside option of nonunion workers during individual Nash bargaining, so that declines in union job prevalence or the union wage premium lower nonunion bargained wages even at firms not directly facing a unionization threat.

Outside option: In the model’s Nash bargaining framework, the value a worker (or firm) obtains if negotiations break down — for nonunion workers, this is the expected value of searching across both nonunion and union jobs weighted by transition probabilities and wage rents in each sector.

Emulation wage: The wage a nonunion firm sets that is just high enough to make workers indifferent between unionizing and remaining nonunion, determined by the firm’s calculation of the threshold below which workers would prefer to bear the costs of unionization.

Union formation (endogenous): In the model, unionization occurs when the surplus workers gain from collective bargaining exceeds the costs of organizing; firms can influence this calculus through wage emulation or direct anti-union actions, making union formation an equilibrium outcome rather than an exogenous event.

Bartik-style instrument (outside option instrument): An instrument for local outside option values constructed by interacting start-of-period local employment composition across industries with national-level changes in industry growth, industry wage premia, and union job transition probabilities, isolating variation in outside options driven by predetermined exposure to national trends rather than local demand shocks.

Shift-share (between) component: The portion of the aggregate wage effect of de-unionization attributable to the direct reallocation of workers from higher-paying union jobs to lower-paying nonunion jobs, distinct from spillover effects on nonunion wage setting itself.

The Macroeconomic Consequences of Early Childhood Development Programs

Mon, 01 Jan 0001 00:00:00 +0000

Layer 1: Overview

This paper embeds early childhood development (ECD) investment into a general-equilibrium (GE), heterogeneous-agent, overlapping-generations model calibrated to U.S. data in order to quantify the aggregate and distributional consequences of large-scale, universal government ECD programs. The central finding is that a universal program spending $13,500 per child-year on children aged 0–3 — the same level as a well-studied North Carolina randomized controlled trial — generates long-run welfare gains of 12.7% in consumption-equivalent units for newborns under the veil of ignorance, income growth of 10.6%, an intergenerational mobility increase of 28.2% (roughly half the US–Canada gap), and a lifetime-earnings inequality reduction of 2.0% (roughly half the US–Germany gap). The key mechanism is dynastic: investing in a child today not only raises that child’s own skills and income but creates a better parental background — in terms of skills, assets, and education — for the next generation, so that more than two-thirds of the welfare gains accrue through this intergenerational channel rather than from the direct effect on the intervened generation. General equilibrium compresses the college wage premium and reduces welfare gains by approximately one-third relative to partial-equilibrium projections, but the policy remains self-financing in the long run. The model is validated against the first- and second-generation experimental evidence from Garcia et al. (2020, 2024), replicating both the 15 p.p. college graduation rate increase and the 1.54 lifetime income return per dollar spent that those RCTs documented.

Summary of a published paper, AI-assisted and human-reviewed. See the linked original for the authoritative claims and full conditions.

In depth

Q1. What is the core market failure that motivates government ECD investment?

The paper identifies two inter-related reasons why private early childhood investments fall below the social optimum: parents cannot borrow against their children’s future income (no child-to-parent compensation contract), and borrowing constraints together with uninsurable idiosyncratic return risk further depress parental investment below even that constrained optimum. Under complete markets with compensating contracts, a poor parent who invests in a high-skilled child could smooth lifetime consumption intergenerationally. Without such contracts, the entire cost of investment falls on the parent in the early life-cycle when assets and income are low, reducing investment incentives sharply. Government ECD spending financed by future taxation on the child’s higher income imperfectly replicates this missing insurance-borrowing mechanism. The model finds that uncertain returns to skill investments — which the government can spread across the population but parents cannot insure privately — and incomplete credit markets are quantitatively more important than imperfect altruism in driving underinvestment.

Q2. How is the model structured, and what is the role of the dynastic framework?

The model is a dynastic overlapping-generations Aiyagari life-cycle economy with four stages (childhood, college, work/parenthood, retirement), each of four-year periods, in which children’s cognitive and non-cognitive skills are determined by CES-aggregated parental time and money investments in ages 0–3, calibrated using the Cunha et al. (2010) skill-formation function. At age 28 (period j=8), the working agent becomes a parent and chooses parental time τ and money m to invest in the child’s skill development across two periods; these decisions interact because time and money are estimated to be imperfect complements (CES exponent γ estimated from data). College attendance is endogenous — it depends on assets, skills, and a school-taste shock — and can be financed by parental transfers (constrained to be non-negative), work, or subsidized student loans. The dynastic structure means that changes in the distribution of parental skills, assets, and education feed forward into the next generation’s initial conditions, which is the source of the large long-run intergenerational amplification.

Q3. What are the main quantitative results of the benchmark universal ECD policy?

A universal permanent policy investing $13,500 per child-year (ages 0–3), financed by an endogenous labor income tax, produces in the long-run steady state: a welfare gain of 12.7% in consumption-equivalent units (for a newborn under the veil of ignorance), a labor income increase of 10.6% (driven by an 11.7% rise in labor productivity), an intergenerational mobility increase of 28.2% as measured by minus the rank-rank coefficient, and a lifetime-earnings variance reduction of 2.0%. The level of $13,500 is both the historically implemented per-child cost in the Garcia et al. (2020) RCT and close to the welfare-maximizing amount in the model (welfare peaks at 13.1% at a slightly higher spending level). Children of low-skilled, non-college parents gain the most (welfare gain 9.1%) versus children of college-educated, high-skilled parents (welfare gain 4.1%). Taxes in the long run are approximately unchanged from baseline because the expanded tax base offsets the direct program cost; in partial equilibrium without GE forces, taxes fall by 2.5 p.p. but GE compression of wages leaves only a negligible tax reduction in the benchmark.

Q4. How do the decomposition exercises isolate the relative importance of long-run dynamics versus GE and taxation?

Decomposing the 12.7% benchmark gain across four counterfactual implementations reveals that long-run intergenerational dynamics account for over two-thirds of total welfare gains, while GE forces reduce gains by roughly one-third and taxation costs are approximately offset by higher revenues in the long run. Specifically: (i) a one-generation, partial-equilibrium version of the policy generates only 5.2% welfare gain; (ii) adding long-run intergenerational effects (permanent policy) raises this to 14.5%, an increase of 9.3 p.p.; (iii) further allowing for balanced-budget taxation in PE reduces gains from 17.3% to a net 12.5%; and (iv) incorporating GE effects — which compress the college wage premium — reduces labor productivity gains and welfare gains by about one-third. More than two-thirds of the aggregate welfare gain comes from changes in the distribution of initial conditions (newborns born into higher-skilled, better-resourced families) rather than from higher utility at a fixed initial-state distribution.

Q5. How does the model validate against experimental evidence, and what does this imply for RCT estimates?

The model replicates two key experimental benchmarks: the Garcia et al. (2020) RCT finding of a 15 p.p. college graduation rate increase and a per-dollar lifetime income return of 1.55 (the model generates 1.54 for children of median-income parents), and the Garcia et al. (2024) estimate that second-generation income gains are 29% of first-generation gains (the model generates 20%, at or below that empirical estimate). The validation is run as a small-scale, partial-equilibrium, one-generation exercise — exactly matching the RCT design — so the comparison is clean. The fact that the model replicates both generations’ effects provides confidence in the intergenerational amplification mechanism. The paper interprets this as evidence that RCT evaluations of short-run, small-scale programs systematically underestimate the long-run benefits of universal programs, with the ratio of long-run GE gains to short-run PE estimates falling between 2 and 3 across a range of alternative education policies.

Q6. How do general equilibrium forces shape distributional outcomes, and why do they cut in opposite directions?

GE forces create a tension: they simultaneously generate most of the inequality reduction (by compressing the college wage premium) and eliminate most of the labor-productivity and welfare gains from partial-equilibrium projections. As the universal program raises the share of college graduates, the relative wage of college workers falls. This wage compression is the primary driver of the 2.0% reduction in lifetime-earnings variance — approximately equal to half the US–Germany inequality gap — but it also lowers the productivity return on human capital investment, reducing GDP gains from 17.2% in partial equilibrium to 10.6% in general equilibrium. Because wages of college graduates fall, the government’s long-run tax savings from higher aggregate earnings are nearly eliminated compared to the partial-equilibrium case (from 2.5 p.p. reduction to negligible reduction). GE forces thus explain both the largest distributional benefit of the policy and its largest welfare cost.

Q7. What do robustness and extension exercises reveal about scalability and alternative policies?

The main results are robust across alternative elasticity-of-substitution parameter values for parental time and money investments, and the welfare-maximizing spending level closely tracks the benchmark $13,500 program. A scale-up extension — in which the early childhood input requires college-graduate labor — finds nearly identical long-run welfare gains but smaller first-generation gains because program costs initially rise as college labor becomes scarcer. A comparison of alternative policies (investments in older children, investment subsidies rather than direct investments, college subsidies, parenting education programs) shows that policies investing directly in young children’s skills consistently achieve larger long-run GE welfare gains relative to their short-run PE estimates than alternative designs, because the intergenerational “better-parents” mechanism is most pronounced for early childhood. Among these alternatives, using equivalent resources as a lump-sum transfer at age 16 yields only 4.1% welfare gain — less than one-third of the 12.7% from early childhood investment — confirming that in-kind investments in childhood are more efficient than cash at the same fiscal cost when parents cannot be compensated by their children.

Q8. What do the transition dynamics imply for the political economy of ECD investment?

If the policy is introduced permanently, every new cohort born after the introduction is better off (each successive generation benefits more as the parental background improves), but older generations alive at the time of introduction face net welfare losses of approximately 1–3% on average because they bear higher taxes while receiving only indirect benefits through their children. The paper calculates that if the government uses debt to smooth the financing cost over time — shifting part of the burden to future generations who will be richer — losses to initial cohorts are reduced to the point where a majority of adults would vote in favor of the policy. More than three-quarters of the long-run welfare gains are achieved within one generation of permanent policy implementation, alleviating concerns that benefits require an implausibly long time to materialize.

Key Concepts

early childhood development (ECD) investment: publicly provided direct monetary investments in children aged 0–3, modeled as a perfect substitute for parental money investments m in the CES skill-formation technology; the paper’s primary policy instrument, calibrated at $13,500 per child-year to match the Garcia et al. (2020) North Carolina RCT.
dynastic intergenerational amplification: the mechanism by which a government ECD investment raises not only the directly intervened child’s skills and income but also improves the distribution of parental skills, assets, and education for the next generation, amplifying aggregate welfare gains so that more than two-thirds of the total gain accumulates via this channel rather than the direct first-generation effect.
child-to-parent compensation constraint: the assumption that parents cannot borrow against their child’s future income or receive direct compensation from the child for parental investments; identified as the primary source of underinvestment in the model alongside borrowing constraints and return uncertainty.
skill-formation technology: a nested CES function adapted from Cunha et al. (2010) in which child skills θ’ depend on current child skills (cognitive θc and non-cognitive θnc), parental skills θ, and a CES aggregate of parental time τ and money investments m; the calibrated complementarity between time and money implies that government money investments also crowd in parental time investment.
general equilibrium skill-premium compression: the decline in the relative wage of college graduates that occurs in a GE model when the policy raises the share of college workers; the mechanism that simultaneously generates the policy’s distributional gains (reduced wage inequality) and reduces its aggregate productivity gains by approximately one-third relative to partial equilibrium.
consumption-equivalent welfare gain under the veil of ignorance: the percentage increase in steady-state consumption that makes a newborn — who does not yet know her parental background — indifferent between being born in the baseline and the policy steady state; the paper’s primary welfare measure, equal to 12.7% under the benchmark universal program.

The Nature of Long-Term Unemployment: Predictability, Heterogeneity, and Selection

Mon, 01 Jan 0001 00:00:00 +0000

Layer 1: Overview

This paper studies the sources of long-term unemployment (LTU, defined as failing to find a job within six months) using administrative data on the universe of unemployment spells in Sweden from 1992 to 2016, merged with exceptionally rich individual characteristics including income and employment histories, employer records, asset portfolios, and IQ scores. The central question is how much of the observed decline in aggregate job-finding rates with unemployment duration reflects genuine state dependence — structural deterioration of individual prospects from being unemployed — versus dynamic selection, the mechanical compositional shift as higher-job-finding workers exit first. Using machine-learning prediction models and a complementary multiple-spell identification strategy, the paper finds that observable heterogeneity in LTU risk is substantial: the hold-out R-squared of the baseline prediction model is 15%, and rises to more than twice that value when rich administrative variables are added relative to a model using only standard socio-demographics. Applying the prediction model across durations, the paper shows that persistent heterogeneity can account for approximately 49% of the observed decline in aggregate job-finding rates over the spell (from 70% to 55% between 0 and 6 months), and potentially as much as 88% under a proportionality assumption on selection based on unobservables. In contrast, the same rich data reject the hypothesis that compositional changes in the pool of unemployed workers explain the cyclical rise in LTU risk in recessions.

In depth

Q1. What is the conceptual framework for decomposing duration dependence?

The paper develops a statistical framework that decomposes the observed change in average job-finding rates over the spell into true duration dependence — the individual-level within-spell decline — and dynamic selection, the change in pool composition as higher-job-finding workers exit first. The key insight is that dynamic selection is identified by the persistent covariance in job-finding probabilities across durations: if a worker who has a high job-finding probability early in a spell also has a high probability later, selection of such workers out of unemployment will lower the average for the remaining pool even if no individual’s rate changes. The framework shows that the hold-out R-squared of a prediction model provides a lower bound for the share of variance in outcomes that is ex ante determined, and that the cross-duration covariance of predictions identifies the persistent heterogeneity component that drives selection.

The paper further shows that combining prediction-based identification (which recovers observable plus transitory heterogeneity) with multiple-spell identification (which recovers persistent observable and unobservable heterogeneity) yields a tighter lower bound on overall heterogeneity. Observable and unobservable approaches are complementary: the multiple-spell approach identifies persistent unobservable heterogeneity that observables miss, while the observables approach captures transitory heterogeneity that changes across spells.

Q2. What is the predictive power of observable characteristics?

The baseline prediction model using standard administrative variables achieves a hold-out R-squared of 15%, measuring the share of variance in job-finding outcomes that is predictable from characteristics determined before the spell begins. This estimate is more than twice as large as a model using only basic socio-demographics (age, gender, education, marital status, citizenship, number and age of children). Prior employment history — even if available for only one or two years — is the most powerful predictor, potentially proxying for unobservable worker characteristics. Additional variables available for limited samples or years (occupation, assets, IQ, union membership) add only modest further predictive power beyond the baseline, suggesting saturation of the observable signal. The predictive power of a linear model is nearly as high as the ensemble of LASSO, gradient-boosted trees, and random forests, indicating that the gains come from data richness rather than nonlinearities exploited by machine learning algorithms.

The unobserved heterogeneity estimated using repeated unemployment spells corresponds to roughly half of the estimated observable heterogeneity. The combined lower bound on the variance that is ex ante determined is at least 19% of total variation in job-finding outcomes.

Q3. How much of duration dependence is explained by dynamic selection?

Applying the prediction model at multiple durations within the spell, the paper finds that nearly three-quarters of predictable heterogeneity is persistent over the spell of unemployment, implying that dynamic selection accounts for at least 49% of the observed aggregate decline in job-finding rates from the start to 6 months into the spell. In 2006, the aggregate 6-month job-finding rate fell from 70% at spell start to 55% at 6 months of ongoing unemployment — a 15 percentage point decline. The lower bound from persistent observable heterogeneity accounts for a decline from 70% to 62.7%, or 49% of the total observed decline. Under a proportionality assumption that unobservable selection mirrors observable selection, the paper estimates that dynamic selection can explain as much as 88% of the observed decline.

The paper also finds substantial heterogeneity in the individual-level dynamics across workers with different observable characteristics: the individual-level decline in job-finding over the spell is strongly negatively correlated with the job-finding probability at the start of the spell, meaning that workers who start with lower job-finding chances also experience stronger within-spell declines, further compressing the heterogeneity in job-finding over time.

Q4. What does the analysis imply about the proportional hazard assumption?

The paper tests and rejects the key assumption in proportional hazard models that job-finding rates decline at the same proportional rate across all workers. The individual predictions at different durations reveal significant heterogeneity in the dynamics across workers with different observable characteristics: workers do not all experience the same proportional decline with duration. This rejection is robust to corrections for sampling error and to non-parametric tests. The heterogeneity in dynamics is empirically distinguishable from the level heterogeneity in job-finding probabilities, and exists over and above what dynamic selection alone would generate.

Q5. Can compositional changes in the pool of unemployed explain LTU risk in recessions?

Despite the paper’s finding that rich observable characteristics explain substantial duration dependence over the spell, the same data reject the heterogeneity hypothesis for cyclicality: compositional changes in the observable characteristics of workers who become unemployed in recessions do not translate into higher predicted LTU risk for the average unemployed worker. The distribution of predicted job-finding risk changes over the business cycle, but the direction is not consistent with the hypothesis that recessions selectively pull in high-LTU-risk workers. Instead, unemployed workers are exposed to substantial changes in LTU risk over the business cycle that operate through within-individual declines in job-finding chances, not through composition. Recessions disproportionately hurt the job-finding prospects of workers with lower education and income, but this is a general worsening of individual prospects rather than a compositional shift in the pool.

Q6. What are the policy implications?

The policy implications differ sharply between genuine state dependence and dynamic selection as sources of LTU. If duration dependence at the individual level is large — skill atrophy, scarring, signaling stigma — early intervention is warranted to interrupt the deterioration process before it becomes irreversible. The paper’s evidence that selection is quantitatively dominant over the spell of unemployment suggests that much of the observed fall in aggregate job-finding with duration does not reflect structural deterioration of individual prospects, limiting the case for early intervention aimed at preventing skill erosion. However, the paper also finds that individual-level declines are heterogeneous and concentrated among workers with already-low job-finding chances, suggesting that targeted risk-profiling — already used by Public Employment Services in many countries including Sweden — is well motivated. The separate finding that recessions generate genuine within-individual worsening of prospects (rather than compositional shift) implies that countercyclical support is warranted on different grounds.

Key Concepts

dynamic selection: the mechanical change in the composition of the unemployment pool as workers with higher job-finding probabilities exit unemployment first, leaving an increasingly low-job-finding-probability remainder; identified in the paper as the dominant driver of the observed aggregate decline in job-finding rates with unemployment duration, accounting for at least 49% and potentially up to 88% of the observed decline.
genuine state dependence (true duration dependence): the within-individual structural deterioration of a worker’s job-finding rate from the experience of unemployment itself — skill erosion, employer stigma signaling — as distinct from compositional selection; the paper finds this is a minor contributor to the observed aggregate duration dependence pattern in Sweden.
persistent heterogeneity: the covariance of individual job-finding probabilities across different durations (or spell cohorts); the component of overall heterogeneity that drives dynamic selection, identified empirically by the cross-duration covariance of prediction model outputs in hold-out samples.
LTU risk profiling: the practice of predicting workers’ probability of becoming long-term unemployed at spell entry using observable characteristics, to target active labor market programs; the paper provides the statistical foundation and quantifies the gains from richer administrative data relative to standard survey-based characteristics.

The Zero-Beta Interest Rate

Mon, 01 Jan 0001 00:00:00 +0000

Layer 1: Overview

This paper proposes and measures the zero-beta rate — the expected return on a portfolio of stocks with zero market beta, constructed to be orthogonal to the SDF innovations spanned by standard factors — as the correct intertemporal price of consumption, and argues that safe interest rates (Treasury bill yields) are not. Using 130 stock portfolios (81 sorted on combinations of beta, size, value, investment, and profitability; 49 industry portfolios) and GMM estimation with five macro instruments (T-bill yield, inflation, term spread, excess bond premium, and the U6 unemployment rate) over January 1973 to December 2020, the paper estimates the zero-beta rate to average 8.3% per year in real terms with a standard deviation of 9.3%, producing a spread of roughly 7.6% per year over the expected real Treasury bill yield. The paper then shows that this zero-beta rate fits the aggregate consumption Euler equation remarkably well: the macro instruments that best predict the real return of the zero-beta portfolio are nearly proportional to those that predict real consumption growth, a non-mechanical result that survives when the sample is restricted to exclude COVID. Statistical Euler equation tests (Stock-Wright [2000] weak-instrument-robust GMM) reject the Euler equation for all IES values when applied to the Treasury bill, fail to reject it for any IES value when applied to the volatile market return (weak identification), but fail to reject it only for IES below 0.5 (risk aversion above 2) when applied to the zero-beta rate — providing identification from the intermediate predictability of the zero-beta portfolio. Monetary policy shock regressions using Romer-Romer and Nakamura-Steinsson shocks further show that an unexpected monetary tightening raises the real Treasury bill yield but lowers the real zero-beta rate, consistent with the Euler equation’s prediction that the intertemporal price should fall when expected consumption growth falls. Finally, the high level and volatility of the zero-beta rate implies that the entire variation of the price-dividend ratio of a consumption claim can be attributed to variation in the zero-beta rate without requiring time-varying equity risk premia — resolving the equity premium puzzle and Campbell’s [1991] excess volatility puzzle simultaneously, at the cost of an unexplained convenience spread on safe assets.

In depth

Q1. What is the zero-beta rate and how does the paper construct it?

The zero-beta rate is the expected return on a portfolio of stocks that is constructed to have zero covariance with all included asset-pricing factors; in a model where the SDF innovations are spanned by those factors, the expected return of this portfolio equals the intertemporal marginal rate of substitution — the correct price for rearranging consumption over time. The paper follows a three-step procedure. First, it estimates the betas of 130 CRSP stock portfolios with respect to seven factors (Fama-French 5-factor model augmented with a bond excess return factor and a default spread factor) using the Ledoit-Wolf (2017) shrinkage estimator for the factor covariance matrix, to mitigate the rank problem arising from 130 portfolios and 574 monthly observations. Second, it uses the betas to construct the minimum-variance zero-beta portfolio — the portfolio that minimizes return variance subject to having zero exposure to each factor, exploiting all 130 portfolios. Third, it regresses the return of this portfolio on five macro instruments (T-bill yield, lagged inflation, term spread, excess bond premium [EBP], and U6 unemployment) using GMM, in an exactly-identified system in which the same instruments used to predict the portfolio return are used as moment conditions. The fitted value of this regression is the zero-beta rate — the predictable component of the zero-beta portfolio return.

The key econometric property is that the GMM procedure simultaneously estimates the factor loadings and the predictive regression in a way that accounts for the estimation error in both. The standard errors for the predictive coefficients (γ) account for the fact that the betas used to construct the zero-beta portfolio are themselves estimated. Comparatively, an infeasible OLS regression predicting the zero-beta portfolio return has nearly identical point estimates and only slightly smaller standard errors, showing that the main estimation uncertainty comes from the predictive regression rather than from the beta estimation step.

Q2. What are the key properties of the estimated zero-beta rate?

The estimated zero-beta rate has three properties that distinguish it from safe rates and make it a plausible intertemporal price: it is predictable by macro instruments (with all predictors except the unemployment rate individually significant at 5%), it is high and volatile (8.3% annually on average, std dev 9.3%, compared to a low and stable Treasury bill yield), and it co-moves in the direction economic theory predicts with macro conditions. The spread between the zero-beta rate and the expected real Treasury bill yield averages roughly 7.6% per year. The average real zero-beta rate is similar to the average real return of the CRSP market index (which averages 11.8% annualized nominal, or about 8.1% real), consistent with earlier estimates of the average zero-beta return by Hong and Sraer (2016) and Bali et al. (2017). The standard deviation of the zero-beta portfolio’s excess return over its expected value is about 2.7% per month (9.4% annualized), substantially below the standard deviation of the market return, which is why the paper can reject predictability for the zero-beta portfolio even though predicting market returns is notoriously difficult.

In terms of time-series patterns: (1) the zero-beta rate increases more than one-for-one with the Treasury bill yield (consistent with Treasury bills having money-like qualities per Nagel [2016]); (2) it is decreasing in lagged inflation; (3) it falls when a recession is likely — specifically, the unemployment rate, term spread, and EBP collectively predict the zero-beta rate so that it is particularly low when U6 unemployment is low, the yield curve is inverted, and the EBP is high, conditions associated with elevated recession risk (Kiley [2022]).

Q3. How does the zero-beta rate fit the aggregate consumption Euler equation?

The central empirical finding is that the macro instruments that best predict the real return of the zero-beta portfolio are nearly proportional to the macro instruments that best predict real consumption growth — a non-mechanical result because the two regressions are estimated entirely separately, on different data series, with no consumption data used to construct the zero-beta rate. In the linearized Euler equation, the real zero-beta rate should predict real consumption growth in proportion to 1/σ (the IES). If the prediction coefficients on the instruments for the zero-beta rate are γ₀, and the corresponding coefficients for consumption growth are γc, then the vector Δ = γ₀ − σ × γc should be close to zero. Graphically, the expected real zero-beta rate and expected real consumption growth track each other closely over the full sample (Figure 1 in the paper), while the expected real Treasury bill return bears essentially no resemblance to expected consumption growth. With five instruments (L=5), the result is not mechanical: one would need L=1 to always find a σ that makes the result hold. The result is even stronger when the sample ends in December 2019, excluding the COVID episode, because the COVID consumption collapse introduces four-to-seventeen standard deviation consumption growth realizations that attenuate the fit.

Robustness: ridge-penalized estimation (using cross-validation to minimize out-of-sample squared forecast error) substantially attenuates both expected consumption growth and the zero-beta rate toward zero, but they remain approximately proportional. The result holds across alternative specifications (different factor models, different instruments) tested in Appendix Section G.

Q4. What do the statistical Euler equation tests find?

Using Stock-Wright (2000) weak-instrument-robust GMM tests of the non-linear consumption Euler equation, the paper finds: (a) the T-bill return fails the Euler equation test for all IES values (rejected); (b) the market return fails to reject the Euler equation for almost any IES values (weak-instruments problem, consistent with Yogo [2004]); and (c) the zero-beta rate fails to reject the Euler equation for IES values below 0.5 (σ above 2, i.e. risk aversion above 2 under CRRA), and rejects for IES above 0.5. The procedure tests the instrumented non-linear Euler equation: for a conjectured value of σ, it estimates δ (the discount factor) from the unconditional Euler moment, then tests the instrumented Euler moments for each of the five instruments. The test statistic is chi-square with 5 degrees of freedom; the confidence set is the set of σ values that cannot be rejected. The zero-beta portfolio’s intermediate predictability (between the easily-predicted T-bill and the hard-to-predict market return) provides the identification that gives this test meaningful power. The paper obtains its preferred IES estimate of approximately 0.2 (σ ≈ 5) by noting that scaling the zero-beta rate down by a factor of five makes it match expected consumption growth most closely.

The test faces a boundary problem: when σ is very large (above 10), the April 2020 consumption collapse creates a very large SDF realization that dwarfs all others, making the variance-covariance matrix nearly singular and the test uninformative. For this reason, the analysis is restricted to σ ≤ 10.

Q5. What happens to the zero-beta rate after a monetary policy shock?

Using the Romer-Romer (2004) and Nakamura-Steinsson (2018) identified monetary policy shocks, the paper finds that a surprise tightening raises the nominal and real Treasury bill yield but lowers the real zero-beta rate — opposite to what the standard Euler equation with the Treasury bill predicts. Both shocks are normalized to a 100-basis-point increase in the federal funds rate on impact. The estimated effect on the real Treasury bill yield is an immediate increase of roughly the same magnitude (slightly more transitory for the Romer-Romer shock). In contrast, the real zero-beta rate falls following the shock, and the effect is larger and more persistent for the Nakamura-Steinsson shock.

This result is consistent with the consumption Euler equation applied to the zero-beta rate, because the monetary shock also lowers expected consumption growth (well-established in the impulse-response literature). The decomposition in Appendix Section E shows the mechanism: a higher Treasury bill yield raises the zero-beta rate, but the monetary tightening also flattens the yield curve and widens credit spreads (raises the EBP); since both the term spread and EBP are predictors of the zero-beta rate with large coefficients, the indirect effects through these variables dominate and lower the zero-beta rate overall. This finding resolves a tension in structural macro models: while the standard New Keynesian model uses the Euler equation with the safe rate (as in Smets and Wouters [2003, 2007]) and requires habits or wedges to match the hump-shaped consumption response to monetary shocks, the Euler equation with the zero-beta rate is satisfied without additional mechanisms. A stylized three-period New Keynesian model in Appendix Section F shows a monetary tightening can simultaneously raise the safe rate and lower the zero-beta rate through endogenous changes in the convenience spread.

Q6. Can the zero-beta rate explain valuation ratio variation without time-varying risk premia?

The paper shows that the zero-beta rate is sufficiently high, volatile, and persistent to generate the observed variation in the price-dividend ratio of a consumption claim under the assumption of a constant and small equity risk premium, reproducing the finding of Campbell (1991) that discount rates must vary — but attributing the variation to the zero-beta rate rather than to time-varying excess returns. The Campbell-Shiller decomposition of the log price-dividend ratio implies that the expected price-dividend ratio must predict either future real zero-beta rates, future excess returns, or a combination. Under the hypothesis of constant expected excess returns, the price-dividend ratio variation is driven entirely by zero-beta rate variation. In a VAR that includes the five macro instruments plus the CAPE ratio, the implied standard deviation of expected log price-dividend ratio of a consumption claim is approximately 28% — comparable to the 27% standard deviation in the Campbell-Cochrane (1999) model calibration, which achieves this variation through habit formation and time-varying risk premia. The equivalent calculation using the Treasury bill yield rather than the zero-beta rate produces a standard deviation of only 9%, consistent with Campbell’s (1991) original finding that the risk-free rate variation is insufficient.

The paper interprets this as a resolution of both the equity premium puzzle (the average zero-beta return is approximately equal to the average market return, implying a roughly zero equity premium over the zero-beta rate) and the excess volatility puzzle (the zero-beta rate generates sufficient variation in the discount rate). The trade-off is that the unexplained spread between the zero-beta rate and the Treasury bill yield — averaging 7.6% annually — is instead left as a “convenience puzzle.” The paper argues this reframing is progress because the type of models required to explain a large convenience spread on safe assets (frictions, segmented markets, money-like demand for liquid assets) are quite different from those designed to explain large time-varying risk premia.

Key Concepts

zero-beta rate: the expected return on the minimum-variance portfolio of stocks that has zero covariance with each of the included asset-pricing factors; in a model where the SDF innovations are spanned by those factors, equals the conditional expectation of the intertemporal marginal rate of substitution, making it the correct intertemporal price of consumption; measured in this paper at 8.3% annually on average using GMM on 130 CRSP stock portfolios, January 1973–December 2020.
convenience spread: the gap between the zero-beta rate and the expected real Treasury bill yield, averaging approximately 7.6% per year in this paper’s estimates; interpreted as the non-pecuniary value that holders of safe assets (Treasury bills and equivalents) receive from liquidity, collateral, and money-like services — not an expected excess return relative to consumption risk but a departure from the risk-return tradeoff for investors who value safety and liquidity independently of consumption hedging.
SDF-orthogonal equity portfolio: the minimum-variance portfolio constructed by the paper to have zero covariance with all seven asset-pricing factors (Fama-French 5 plus bond and default factors); the portfolio whose expected return equals the zero-beta rate because, by construction, no factor risk premium enters its expected return; estimated using the Ledoit-Wolf (2017) shrinkage estimator applied to 130 stock portfolios to address the rank problem in beta estimation.
Euler equation failure with safe rates: the empirical finding that the expected real Treasury bill return does not co-move with expected real consumption growth — the standard failure documented by Hansen-Singleton (1983), Dunn-Singleton (1986), and Yogo (2004) — which the paper reinterprets not as a structural failure of the representative agent model but as a consequence of using the wrong interest rate; the same Euler equation holds when applied to the zero-beta rate, which the paper argues is the correct intertemporal price.

Trade with Nominal Rigidities: Understanding the Unemployment and Welfare Effects

Mon, 01 Jan 0001 00:00:00 +0000

Standard international trade models assume perfectly flexible prices and full employment. This paper introduces nominal rigidities (downward wage rigidity) into a quantitative trade model and asks how this changes the welfare gains from trade liberalization. The central finding is that standard flexible-price estimates overstate the welfare gains by approximately one-third: trade liberalization can generate unemployment in import-competing sectors when wages cannot fall, and the forgone output from these workers is a welfare cost that flexible-price models miss entirely. The paper calibrates the degree of downward wage rigidity to cross-country data on unemployment volatility and shows that the magnitude of the overstatement is robust across a range of calibrations. An analytical decomposition separates the allocative efficiency gains (which standard models capture) from the employment losses (which they miss), clarifying when the overstatement is large versus small.

Summary of a forthcoming paper, AI-assisted and human-reviewed. See the linked original for the authoritative claims and full conditions.

In depth

Q1. How does downward wage rigidity create unemployment from trade liberalization?

When import competition reduces demand for domestic labor in an affected sector, wages cannot fall fast enough to clear the labor market under downward rigidity; workers are priced out of re-employment in the short run, generating cyclical unemployment in the import-competing sector that persists until real wages adjust through inflation erosion. The unemployment is involuntary and represents forgone production — a social cost that flexible-price models attribute to zero by assumption.

Q2. Why is the overstatement approximately one-third?

The one-third figure comes from the ratio of the employment-loss welfare cost to the total flexible-price welfare gain in the paper’s benchmark calibration; the rigidity-driven employment loss is large enough relative to the allocative efficiency gain to reduce net welfare gains substantially, but not so large as to eliminate them. This ratio is not universal — it depends on the degree of wage rigidity, the sectoral composition of trade exposure, and the speed of labor reallocation — but the paper shows it is robust across plausible parameter ranges.

Q3. Does trade liberalization still generate net welfare gains?

Yes, on net the welfare gains from trade remain positive even with downward wage rigidity — the overstatement of one-third means the true gains are positive but smaller than flexible-price models predict, not negative. The paper does not argue against trade liberalization but against using flexible-price welfare estimates without adjustment for unemployment costs.

Key concepts

downward wage rigidity : the empirical constraint that nominal wages adjust slowly downward; the key friction this paper adds to the quantitative trade model, generating unemployment in sectors hit by import competition.

welfare overstatement : the gap between the flexible-price welfare gain from trade liberalization (the standard model’s prediction) and the true gain once unemployment costs from nominal rigidity are accounted for; approximately one-third in the paper’s benchmark calibration.

US Public Debt and Safe Asset Market Power

Mon, 01 Jan 0001 00:00:00 +0000

Layer 1: Overview

This paper asks whether the U.S. government exploits its market power as the dominant global supplier of safe assets when setting the quantity of public debt, and quantifies the macroeconomic consequences of this strategic behavior. The paper develops a two-country general equilibrium model in which U.S. public debt provides a non-pecuniary benefit to foreign holders (capturing liquidity, collateral, and safety value) and the U.S. is the monopoly provider of this asset — facing a downward-sloping demand curve for Treasuries, so that issuing more debt reduces the convenience yield. The paper then tests empirically whether the data favor this monopoly model over a price-taking benchmark, exploiting the industrial organization insight that rotations in the demand curve (changes in elasticities during high- versus low-volatility regimes) can distinguish strategic from competitive behavior. Using quarterly data from 1935 to 2020, the paper finds that the data reject price-taking behavior in favor of the monopoly model across a wide range of specifications. Quantitatively, the monopoly calibration implies that U.S. market power generates approximately 45% of the observed convenience yield as a markup (about 30 basis points out of 68 basis points on average), causes safe asset supply to be roughly half what it would be under price-taking, and generates welfare gains to the U.S. of 0.21% in permanent consumption equivalents — almost half of which is attributable to market power rather than to the non-pecuniary value itself.

In depth

Q1. What is the theoretical framework for U.S. market power in safe assets?

The paper develops a deterministic, infinite-horizon, two-country model in which the U.S. is the sole provider of an asset with a non-pecuniary benefit to foreign (Rest of World) households, so the U.S. faces a downward-sloping demand curve for its public debt and acts as a monopolist in equilibrium. In the model, purchasing U.S. public debt yields a non-pecuniary flow benefit captured by an increasing, concave function f(b*). Because of this benefit, the equilibrium return on U.S. debt is lower than the return on capital — the gap being the convenience yield, defined as the spread between the U.S. capital return and the return on U.S. public debt. The U.S. Ramsey government internalizes the inverse demand function for its debt when solving its optimal fiscal problem, creating a standard monopoly markup: the equilibrium markup equals the inverse of the demand elasticity, µ = 1/ε_D, where ε_D is the price elasticity of foreign demand for U.S. Treasuries. Under price-taking, the markup is zero and the convenience yield reflects only the non-pecuniary value; under the monopoly model, the convenience yield is inflated by the markup, reducing debt supply below the competitive level.

Q2. How does the paper test monopoly versus price-taking behavior empirically?

The paper applies the conduct-testing approach of Bresnahan (1982) and the model selection test of Rivers and Vuong (2002): since rotations in the demand curve (changes in the elasticity, holding the level fixed) shift prices only if the firm exploits market power, a finding that convenience yields increase in high-elasticity regimes while quantities decrease is evidence of strategic behavior. The paper uses a regime indicator for periods of high global volatility (measured by the rolling standard deviation of MSCI UK Index returns over 1935–2020) as the demand rotator: during high-volatility periods, investors’ demand for safe assets is more inelastic (flight-to-safety), causing the demand curve to both shift outward and rotate (become steeper). The monopoly model predicts that the U.S. responds to the more inelastic demand by raising the convenience yield through higher markups and restricting supply, whereas the price-taking model attributes any convenience yield increase purely to shifts in marginal cost.

Empirically, the data show that convenience yields are higher and debt-to-GDP ratios lower during high-volatility periods — inconsistent with the price-taking model’s prediction that both prices and quantities should rise in a demand shift, and consistent with the monopoly model’s prediction of reduced supply. The estimated demand semi-elasticities are −0.20% per log-point in low volatility and −0.59% per log-point in high volatility (OLS), implying demand elasticities of 1.07 in low volatility and 3.18 in high volatility. The Rivers-Vuong test statistics reject price-taking in favor of the monopoly model at the 1% significance level under both OLS and IV specifications and across a wide range of assumed cost elasticities.

Q3. What is the quantitative magnitude of safe asset underprovision?

Using a demand elasticity of 2.2 (the average of the OLS and IV estimates from specifications without the demand rotator, consistent with the prior literature), the paper’s calibrated monopoly model implies that the U.S. safe asset supply is approximately half as large as it would be if the U.S. acted as a price taker: the steady-state total safe assets-to-GDP ratio is 0.39 in the monopoly equilibrium versus 0.59 in the competitive equilibrium. The markup accounts for approximately 45% of the average convenience yield of 68 basis points, implying a markup of about 30 basis points. The interest rate on U.S. public debt is 0.97% in the monopoly equilibrium versus 1.09% in the competitive equilibrium — a difference of 12 basis points — reflecting both the lower debt level and the higher convenience yield that the monopoly generates. These results hold across alternative parameterizations of the cost and demand elasticities.

Q4. What are the welfare implications of safe asset market power?

Market power generates significant welfare gains to the U.S. and welfare losses to the Rest of World: transitioning from the monopoly steady state to an economy with no special role for U.S. assets costs the U.S. 0.21% in permanent consumption equivalents and benefits the Rest of World by 0.34%; transitioning to a competitive equilibrium (price-taking but maintaining the special role) costs the U.S. 0.08% and benefits the Rest of World by 0.10%. This decomposition implies that roughly 60% of the U.S. welfare gain from its safe asset status is attributable to the non-pecuniary value (the benefit function f), and approximately 40% is attributable to market power per se. The interpretation is that the U.S. captures surplus from global safe asset demand through both the intrinsic value of its debt and through monopoly rents from restricting supply. The paper interprets these welfare gains as a quantification of “exorbitant privilege” arising from the supply side rather than from risk premium considerations.

Q5. What happens when safe asset competition increases?

The paper analyzes the effects of introducing Cournot competition among multiple sovereign safe asset suppliers, finding that while the aggregate supply of global safe assets increases substantially with more competitors, the U.S. public debt level itself is fairly stable, borrowing costs for the U.S. increase, and the Rest of World welfare improves. With N=2 symmetric Cournot competitors, the aggregate safe asset supply approximately doubles relative to the monopoly baseline, but each supplier’s equilibrium quantity is roughly unchanged. As N increases further, aggregate supply continues to grow, convenience yields fall, and interest rates on U.S. debt rise. A domestic financial fringe competing with U.S. government debt is modeled differently: because the U.S. government internalizes domestic fringe profits, domestic competition results in less competitive pressure, higher markups, and smaller welfare losses for the U.S. than the same amount of competition from foreign suppliers. These results quantify the macroeconomic stakes of initiatives to create alternative safe assets, such as euro area supranational safe bonds or Chinese reserve currency aspirations.

Q6. What identifies strategic versus competitive behavior using debt holder composition?

As a complementary identification strategy, the paper exploits time variation in the composition of U.S. Treasury holders: foreign investors (primarily official sector) have more inelastic demand than domestic investors (primarily financial institutions and mutual funds), and the increasing share of foreign investors since the 1970s implies a secular decline in the average demand elasticity. The paper estimates demand elasticities separately for the two groups, finds the foreign investor curve is more inelastic, and uses the implied time-varying average elasticity as a second demand rotator. The conduct test under this alternative approach also rejects price-taking in favor of the monopoly model. The monopoly model explains the observed increase in long-term convenience yields since the 1970s through rising markups driven by the shift toward less elastic foreign investors, rather than through rising marginal costs of debt issuance.

Key Concepts

safe asset market power: the U.S. government’s ability to internalize the downward-sloping foreign demand curve for U.S. Treasuries and restrict supply to maintain a high convenience yield; the paper provides the first formal empirical test and quantification of this strategic behavior in the global safe asset market.
convenience yield markup: the component of the observed convenience yield on U.S. Treasuries attributable to monopoly pricing rather than to the intrinsic non-pecuniary value of the assets; estimated at approximately 45% of the total convenience yield (about 30 out of 68 basis points) under the paper’s baseline demand elasticity of 2.2.
demand rotator (Bresnahan identification): a variable that changes the elasticity of demand without shifting its level, enabling identification of strategic conduct: observing that prices rise and quantities fall when demand becomes more inelastic (as during high-volatility regimes) is evidence of monopoly pricing, since a price taker would not respond to an elasticity change alone.
safe asset underprovision: the quantity distortion from monopoly pricing in the global safe asset market; the paper estimates the steady-state safe-asset-to-GDP ratio is approximately 50% lower in the monopoly equilibrium than in the competitive benchmark, reflecting the standard monopoly restriction of output to exploit the downward-sloping demand curve.

Why Is Intermediating Houses So Difficult? Evidence from iBuyers

Mon, 01 Jan 0001 00:00:00 +0000

This paper examines frictions in dealer intermediation in durable consumer goods markets, using iBuyers — technology-driven real estate companies such as Opendoor and Offerpad — as a lens. The central research question is why dealer intermediation, which provides immediate liquidity by purchasing assets onto a balance sheet and reselling, is so limited in the U.S. housing market (valued at $50 trillion and representing roughly 70% of the median household’s net worth) relative to other durable goods markets such as automobiles.

The authors use CoreLogic deed transaction data and MLS listing data from five markets with substantial iBuyer presence (Phoenix, Las Vegas, Dallas, Orlando, and Gwinnett County, Georgia) over 2013–2018, covering arm’s-length, non-foreclosure single-family home and condominium transactions. They supplement this with Redfin ZIP-level data on listing speed and American Community Survey demographics. iBuyers are identified as Opendoor, Offerpad, Knock, Zillow, and Redfin.

The empirical analysis documents that iBuyers grew from roughly 1% market share in Phoenix in 2015 to about 6% by 2018, acting as balance-sheet intermediaries who hold properties for a median of 105 days. iBuyers purchase homes at a 3.1 percentage point (pp) discount relative to comparable homes sold in the same ZIP-quarter, and sell at a 2.2 pp premium relative to other institutional sellers, for a combined gross spread of approximately 5.3 pp (reported in the abstract and body as ~5%). Sellers to iBuyers show a 6.8 pp higher rate of market exit post-sale and a 4.0 pp higher probability of purchasing before selling, consistent with demand for immediacy from impatient, relocating households.

Two key frictions constrain intermediation. First, adverse selection: iBuyers rely on algorithmic valuation models (AVMs) that explain over 80% of price variation in iBuyer transactions versus only 68% in non-iBuyer transactions, leaving a residual of soft information (odor, neighbor quality) that sellers know but algorithms cannot capture. iBuyer presence is over three times greater in the lowest pricing-uncertainty tercile versus the highest, and a one standard deviation increase in pricing uncertainty reduces iBuyer presence by 1.23 pp within a ZIP and reduces gross spread per transaction by 1.5 pp. Second, underlying illiquidity: iBuyers are almost entirely absent in market segments where the probability of sale within three months (PSALE) falls below 50%, despite strong seller demand.

To quantify these frictions, the authors build and calibrate a continuous-time directed search equilibrium model with a dealer intermediary subject to adverse selection. Six parameters are calibrated to match empirical moments: iBuyer market share (5%), purchase discount (3.1 pp), sale premium (2.2 pp), iBuyer concentration in the most versus least liquid PSALE quartiles, impatient seller fraction, and median iBuyer holding time. The calibrated adverse selection parameter (α = 0.35) means the intermediary correctly identifies 35% of low-quality homes as such; the impatient seller share (μ = 0.18) means 18% of unmatched sellers are highly impatient; and the vacancy depreciation rate (d = 0.02) means 2% per period for unoccupied homes. External validation via a difference-in-differences comparison of Phoenix against other markets yields model-consistent predictions of a 0.5 pp reduction in time on market and a 0.8 pp increase in house prices.

Counterfactual experiments reveal that introducing a 30-day acquisition delay (rather than near-instantaneous) reduces iBuyer market share from 5% to below 2%; eliminating the signal entirely (α = 0) drops market share to just above 1%; and enabling iBuyers to rent vacant properties during the holding period could raise market share above 7.5 pp. A 50% reduction in PSALE reduces iBuyer market share roughly proportionally.

The calibrated model is then applied to other durable goods markets by varying informational asymmetry, liquidity, and depreciation parameters. Cars — more homogeneous (year/make/model/mileage fully characterizes value), mobile (transportable across markets), and depreciating primarily through use — are predicted to support dealer intermediary market shares of 40–55%, consistent with observed U.S. car dealer market share of ~50%. Reducing the depreciation rate from the housing level (d = 0.02) to a car-like level (d = 0.005) alone increases intermediary market share by about 5 pp. Houses — heterogeneous, immobile, and depreciating through time rather than use — are predicted to support near-zero intermediation under pre-iBuyer technology. The authors also explain COVID-19 iBuyer suspensions (reduced market liquidity made resale untenable) and Zillow’s November 2021 exit (very liquid markets eroded the iBuyer speed premium, worsening adverse selection while rapid price appreciation degraded AVM accuracy).

Q: What discount do iBuyers pay when purchasing homes, and what premium do they earn when selling? A: iBuyers purchase homes at a 3.1 pp discount relative to comparable homes sold in the same ZIP code and quarter, with a t-statistic of 8.55. They sell at a 2.2 pp premium relative to other institutional sellers. The combined gross spread is approximately 5.3 pp (referred to throughout the paper as roughly 5%).

Q: How large is the iBuyer market share, and in which markets did they operate? A: iBuyer market share grew from approximately 1% in Phoenix in 2015 to roughly 6% by 2018. In Gwinnett County, Las Vegas, and Dallas/Orlando, shares reached approximately 4%, 4%, and 2% respectively by 2018. The analysis covers five markets: Phoenix, Las Vegas, Dallas, Orlando, and Gwinnett County (suburban Atlanta).

Q: What is the evidence that iBuyer sellers are impatient rather than simply lower-quality-house owners? A: Sellers to iBuyers exhibit a 6.8 pp higher rate of market exit (defined as purchasing a home outside the county or making no subsequent real estate purchase within 12 months), consistent with relocation-driven impatience. They also have a 4.0 pp higher probability of purchasing a new home before completing the sale of their current home, which is enabled by the iBuyer transaction’s speed facilitating mortgage approval conditional on the existing property’s sale.

Q: How do the authors measure adverse selection risk and what is its relationship to iBuyer presence? A: Adverse selection is proxied by the squared residual from a hedonic pricing regression — the variation in transaction prices unexplained by observable characteristics — computed at the ZIP-year level for non-iBuyer transactions. iBuyer presence is over three times greater in the lowest pricing-uncertainty tercile than in the highest. A one standard deviation increase in pricing uncertainty reduces iBuyer presence by 1.23 pp within a ZIP (controlling for ZIP fixed effects, local prices, house age, and square footage), and reduces gross spread per transaction by 1.5 pp.

Q: What role does underlying asset liquidity play in constraining iBuyer intermediation? A: iBuyers concentrate almost entirely in market segments where the ex ante probability of selling within three months (PSALE) exceeds 50%, and are essentially absent where PSALE falls below 50%. This holds even though sellers in low-PSALE segments have strong demand for immediacy, implying that illiquidity raises intermediation costs above the demand-side willingness to pay a discount.

Q: What does the model’s calibration reveal about the share of impatient sellers and the accuracy of iBuyer signals? A: The calibrated adverse selection parameter α = 0.35 means the intermediary correctly identifies 35% of low-quality homes as low quality (the signal is moderately but imperfectly informative). The calibrated impatient seller share μ = 0.18 means approximately 18% of unmatched sellers are highly impatient and willing to accept a significant price discount for immediacy. The vacancy depreciation rate d = 0.02 implies a 2% per period cost for unoccupied properties.

Q: How important is transaction speed to the iBuyer model? A: Introducing a 30-day acquisition delay (rather than near-instantaneous purchase) reduces iBuyer market share from 5% to below 2% — a reduction of more than 60%. The model mechanism is that the primary iBuyer customers are highly impatient sellers who place extreme value on immediate transactions; even a moderate delay substantially reduces their willingness to accept a price discount.

Q: What happens if iBuyers lose their ability to distinguish between high- and low-quality homes? A: Setting the signal accuracy to zero (α = 0, the “naive intermediary” case) causes iBuyer market share to fall from 5% to just above 1%. Without any quality signal, severe adverse selection forces the intermediary to offer substantially lower prices to break even, which in turn reduces the number of sellers willing to transact.

Q: How much would enabling iBuyers to rent vacant properties during the holding period affect market share? A: The rental-enabled iBuyer counterfactual shows that market share could increase above 7.5 pp from the baseline 5%, because rental income would allow iBuyers to offer higher purchase prices while offsetting carrying costs. This suggests that rental infrastructure or policy changes permitting temporary rentals would substantially expand the scope of dealer intermediation in housing.

Q: How does the model validate itself externally? A: The authors use a difference-in-differences design comparing Phoenix (earlier and larger iBuyer entry) to the other four markets. The model predicts iBuyer entry should reduce average time on market and increase house prices; the DiD results show a 0.5 pp reduction in time on market and a 0.8 pp increase in house prices in Phoenix relative to comparison markets post-entry, consistent with model predictions.

Q: Why did iBuyers suspend operations during the COVID-19 pandemic despite having a contactless technological advantage? A: The model explains the suspension through the liquidity channel: iBuyers’ value proposition depends on quickly reselling acquired properties, not merely on contactless buying. When market liquidity collapsed during lockdowns (transaction volumes fell sharply), iBuyers could not resell properties quickly, making intermediation unprofitable regardless of their purchasing-side technological advantage. As liquidity recovered, iBuyers resumed operations.

Q: What does the model say about Zillow’s exit from iBuying in November 2021? A: In very liquid markets, the iBuyer speed advantage shrinks because homeowners can sell quickly in the traditional market anyway, reducing the discount sellers accept when selling to an iBuyer. With a smaller discount, adverse selection worsens because only sellers with unfavorable private information (knowing their house has problems the algorithm overvalued) choose the iBuyer route. The pandemic-era housing market also featured rapid price appreciation that degraded AVM accuracy trained on historical data, compounding adverse selection. Zillow reported having significantly overpaid for homes, consistent with this mechanism.

Q: Why is dealer intermediation approximately 50% in car markets but near-zero historically in housing? A: The model, applied to car-market parameters, predicts 40–55% dealer intermediation, consistent with observed U.S. car market shares. Three structural differences explain the gap: (i) cars are more homogeneous (year/make/model/mileage sufficiently characterizes value), reducing adverse selection; (ii) cars are mobile and can be transported across markets, increasing effective liquidity; and (iii) cars depreciate primarily through use, so holding a car on a dealer lot incurs lower value loss than leaving a house vacant. Reducing the depreciation rate from the housing calibration (d = 0.02) to a car-like level (d = 0.005) alone raises predicted intermediary market share by about 5 pp.

Q: Does subjective value dispersion (heterogeneity in buyer preferences) play a large role in limiting intermediation? A: While subjective value dispersion plays a significant role in shaping search market equilibrium (affecting match quality and the gains from household-to-household search), the model finds its effect on the overall level of intermediation is comparatively less pronounced than informational asymmetry, market liquidity, or the opportunity cost of vacancy.

Q: What evidence supports the claim that iBuyers use algorithmic pricing? A: Observable property characteristics and ZIP-quarter fixed effects explain over 80% of price variation in iBuyer transactions, compared to only 68% in non-iBuyer transactions. The higher R-squared for iBuyer transactions is consistent with iBuyers relying on measurable, formalizable characteristics rather than soft information (such as odors or neighbor property conditions) that traditional buyers gather through physical visits.

Q: What are the structural limits on iBuyer expansion even with improved technology? A: Even with enhanced pricing technology (lower α), the scope for dealer intermediation remains narrow because strong incentives persist for iBuyers to avoid markets where algorithmic valuation is difficult, such as older and less homogeneous housing stock. The fundamental barriers — heterogeneity, immobility, and high vacancy opportunity cost — cannot be overcome by technology alone, meaning iBuyers are unlikely to reach the ~50% market share seen in automobile dealer markets.

iBuyers: Technology-driven real estate companies (principally Opendoor and Offerpad) that use automated valuation models and online platforms to make near-instantaneous cash offers on homes, functioning as dealer intermediaries who purchase properties onto their balance sheet and resell after a short holding period, thereby providing immediate liquidity to sellers who would otherwise wait 90+ days in the traditional listing process.

Dealer (Balance Sheet) Intermediation: A form of market-making in which an intermediary purchases an asset outright and holds it on its own balance sheet while finding a subsequent buyer, as distinct from matchmaking intermediaries (brokers) who connect buyers and sellers without taking ownership. The intermediary earns a gross spread between purchase and sale prices.

Adverse Selection (in iBuyer context): The problem arising because sellers possess soft private information about their property (odors, hidden defects, neighbor quality) that algorithmic valuation models cannot capture, while traditional buyers can acquire this information through physical visits. Because iBuyers price quickly without visits, they disproportionately attract sellers of unobservably lower-quality homes, as measured in the paper by the calibrated parameter α = 0.35 (the fraction of low-quality homes the intermediary correctly identifies).

Algorithmic Valuation Model (AVM): The pricing technology used by iBuyers to value homes near-instantaneously using observable property characteristics. The paper measures AVM performance by the R-squared of a hedonic regression: over 80% for iBuyer transactions versus 68% for non-iBuyer transactions, with the residual representing information the algorithm misses and traditional buyers discover through visits.

PSALE (Probability of Sale within 3 Months): An ex ante measure of a property’s underlying liquidity, estimated from a probit model on non-iBuyer listings, capturing the probability that a given home sells within three months of listing. The paper uses PSALE as the key liquidity variable; iBuyers are almost entirely absent where PSALE falls below 50%.

Occupancy Cost: The value loss incurred when a house is held vacant on an intermediary’s balance sheet — encompassing both foregone housing service flows (which continue to benefit occupants under traditional listing but are lost under iBuyer ownership) and ongoing maintenance and depreciation costs (calibrated at d = 0.02 per period). This cost distinguishes housing from goods like cars that depreciate primarily through use rather than time.

Gross Spread: The difference between the price at which an iBuyer sells a property and the price at which it purchased that property, expressed as a percentage of the acquisition price. The paper documents a gross spread of approximately 5% (combining the 3.1 pp purchase discount and the 2.2 pp sale premium), which is persistently positive over the sample period.