On Measuring the Effects of Fiscal Policy in Recessions

Parker’s diagnosis is that “almost all DSGE and VAR analyses do not answer the questions asked in the introduction” because both methodologies are built, implicitly or explicitly, around linear dynamics, so that “a Keynesian-style multiplier — one that is large in recessions and small at other times — is ruled out in the dynamics of any DSGE model linearized around a single point and in any standard VAR because it does not allow state dependence” (Sec. 1–2). He opens by noting that in January 2009 the Obama Administration turned to economists for the size of a stimulus package and got “a wide range of answers” (Krugman’s call for much more stimulus citing a multiplier of 1.5 versus Barro’s citation of a WWII-implied multiplier of 0.8), reflecting that policymakers were relying on multi-equation forecasting models “inconsistent with the best practices of modern macroeconomics.”

Parker argues that existing methods “almost entirely ignore the state of the economy and estimate ’the’ government multiplier, which is presumably a weighted average of the one we care about — the multiplier in a recession — and one we care less about — the multiplier in an expansion” (Abstract). The “old Keynesian” view he takes as a maintained possibility — not an established fact — is that fiscal stimulus is effective only when there is slack (high unemployment, low capacity utilization); in a boom, by contrast, expansionary fiscal policy would “primarily raise interest rates and crowd out private consumption and investment, as in the neoclassical model.”

A linearized solution to a DSGE model implies “that the partial derivative of output or consumption… with respect to government spending or taxes is the same following a large positive shock to the economy as following a large negative shock,” so the effect of spending is identical in “a roaring boom” and “the depths of a recession” (Sec. 2.1). Parker notes this is not merely a solution-method artifact for most cases: “most models — certainly the textbook neoclassical model and textbook New Keynesian model — are close to linear in structure,” so the linear approximation is already a very good description of their true nonlinear dynamics. He does point to one documented exception: Christiano, Eichenbaum, and Rebelo (2011) show the multiplier “can be very large — 3 or more” when the nominal interest rate is held fixed at the zero lower bound versus allowed to vary, but he stresses this nonlinearity “depends on the fixity of the interest rate and not on the degree of slack in the economy” — a different channel from the Keynesian idle-resources story.

Because a VAR is a linear system, “the impulse response to an unexpected or exogenous increase in government spending is constrained to be the same independent of the state of the business cycle” (Sec. 2.2), so “the majority of the VAR literature on the efficacy of government spending in recessions suffers a similar problem to that of its DSGE sibling.” He singles out Auerbach and Gorodnichenko (2010) as “the most significant work to date” using a nonlinear smooth-transition structural VAR, which finds multipliers “quite different” across states — cumulative 5-year multipliers of 0 to 0.5 in a boom versus 1 to 1.5 in a recession — but cautions this “raises many issues about how best to statistically model nonlinear dynamics,” since the method smooths impulse responses so that normal-time data inform both boom and recession dynamics.

Parker argues that “even research that seeks to estimate the multiplier in recessions actually estimates the marginal multiplier not the total multiplier,” and that “if the marginal multiplier declines with the size of the stimulus, then government actions… will tend to make the marginal multiplier significantly lower than the average multiplier in recessions.” He sketches the logic with a government optimizing continuously: if policy is at its optimum, a small further deviation from that optimum has approximately zero marginal effect on expected discounted utility by a standard first-order condition, even though the total effect of the whole countercyclical program could be large — so “the measured, marginal multiplier could be below one while the total multiplier is well above one.” This is explicitly presented as a theoretical possibility he flags, not a measured magnitude.

Parker states plainly that “we have few severe recessions from which to learn… and each (potential) recession has a different set of causal factors, a different set of contemporaneous confounding factors, and different policies themselves,” so that “there is little information in macroeconomic data alone to inform policy.” He surveys several partial remedies the literature has tried — using asset-market data or survey expectations to sharpen counterfactuals (Fisher and Peters 2009; Ramey 2011; Auerbach and Gorodnichenko 2010), or pooling across countries (Perotti 1999; Ilzetzki, Mendoza, and Vegh 2010) — but notes the latter is limited because “few countries have aggregate data of the quality and length of that in the U.S.,” and because multipliers likely differ across countries’ monetary regimes, tax systems, and labor-market institutions.

Parker’s proposed remedy is to build nonlinear DSGE models and discipline them not just by fitting aggregate correlations but by matching “estimates from microeconomic studies of the causal effects of policies on economic outcomes taking prices as given” — i.e., well-identified partial-equilibrium causal estimates, which he argues are desirable because they combine “independence” (estimated on data separate from the aggregate dynamics being explained, guarding against overfitting) and “relevance” (closely tied to the specific causal channel the model needs to get right), following the terminology of Prescott (1986). This stands in contrast to “most of the current literature which derives quantitative discipline from best-fitting (linear) aggregate dynamics and/or from external estimates of structural parameters.”

Two main examples are discussed: tax-rebate studies and a cross-state military-spending study. On tax rebates, Parker cites Johnson, Parker, and Souleles (2006) and Parker, Souleles, Johnson, and McClelland (2011), which use the randomized timing of 2001 and 2008 rebate disbursements to estimate that households increased nondurable consumption by “a quarter to a third” of the rebate amount — “much larger than that consistent with the models of household behavior under rational expectations” in standard neoclassical or New Keynesian models — and that this response was “stronger among low-income or low-asset households,” pointing to liquidity constraints as a source of state-dependence. He contrasts two ways researchers have used this evidence: Kaplan and Violante (2011) build a heterogeneous-household model in which the consumption response to rebates “varies by a factor of two across different economic scenarios,” while Huntley and Michelangeli (2010) and Oh and Reis (2011) fit linear-dynamics models to the same micro moments and find a multiplier for tax rebates “very small (less than 0.2)” — illustrating that matching micro evidence does not automatically generate state-dependence unless the model also has a channel for it. On military spending, Nakamura and Steinsson (2011) estimate that the output and employment effects of state-level military spending “are roughly twice as large… when the national unemployment rate is above its median as when it is below its median,” a cross-sectional statistic Parker notes is independent of aggregate dynamics, relevant to the general-equilibrium multiplier, and sharp enough to distinguish among competing DSGE specifications (a New Keynesian model with nonseparable consumption-leisure preferences matches it; baseline New Keynesian and neoclassical variants do not).

Parker is explicit that such evidence “only provide[s] information about microeconomic effects and leaves the rest of the economy for specification by the researcher” — it disciplines one causal link in the model’s chain but cannot, by itself, pin down the full general-equilibrium response. He also notes that to the extent aggregate nonlinearity comes from state-dependence in individual-level behavior (rather than from the changing composition of a heterogeneous population), cross-sectional micro evidence collected specifically during deep recessions, contrasted with normal-time evidence, would be especially informative — implying the current evidence base, collected mostly in a single recessionary episode (2008-09) or from a single policy (the 2001/2008 rebates), is not yet sufficient to establish state-dependence in general.

The stated conclusion is that “a fruitful avenue for ongoing and future research is to build (possibly highly) nonlinear models and match their implications not just to correlations in aggregate or microeconomic data, but also to estimates from microeconomic studies of the causal effects of policies on economic outcomes.” Parker frames this as a genuinely open research agenda rather than a solved problem: theoretical tools for building state-dependent nonlinear DSGE models already exist in principle but “may require refinement” for practical use, and it is “not obvious which frictions are the right ones” to generate the needed curvature, though he points to informational frictions in financial markets — relevant to both the Great Depression and 2008-09 — as one promising candidate.