The long and variable lags of monetary policy: Evidence from disaggregated price indices

The paper asks why the aggregate U.S. price level takes so long to respond to a monetary policy shock, and argues that Friedman’s (1960, 1961) “long and variable lags” doctrine applies across the cross-section of individual prices, not only to the aggregate time series. Using highly disaggregated Personal Consumption Expenditures Price Index (PCEPI) data, the authors document “stark differences in the timing and magnitude of the responses across price categories, including some prices that show an initially positive response” to a monetary tightening, and use this heterogeneity to explain why the aggregate response looks slow and muted even though individual prices are often moving substantially.

The analysis uses monthly PCEPI subcomponents for the United States from 1982 to 2008 — a sample chosen to exclude the zero lower bound, non-standard monetary policy, and the COVID period — disaggregated into as many as five levels of granularity, down to 136 individual price categories at the finest level (level 5), which form a balanced panel. The identified shock is the external instrument constructed in the authors’ companion paper, Aruoba and Drechsel (2023): a natural-language-processing/machine-learning measure applied to the Federal Reserve’s Greenbook (Tealbook) forecasting documents, following the narrative logic of Romer and Romer (2004), designed to back out exogenous variation in the federal funds rate target rather than relying on market interest-rate surprises around FOMC announcements — the latter avoided because such surprises can reflect “information effects” or responses to non-standard policy rather than a clean policy shock.

Each price component i is analyzed with a separate local projection (following Jordà 2005) at every horizon h from 0 up to 60 months, regressing the log price h periods ahead on the contemporaneous monetary policy shock, a control vector, and the lagged log price level; the coefficient on the shock at each horizon is the impulse response, scaled to a 100-basis-point policy tightening. The control vector for each series is chosen by a combinatorial search over macroeconomic variables and their first 10 lags — selecting up to 7 controls that maximize fit at the 24-month horizon, a procedure the authors note runs on the order of 2.4 billion candidate regressions per price series — and standard errors are computed with Newey-West HAC correction using a bandwidth equal to h+1.

Because the aggregate PCEPI is approximately an expenditure-share-weighted sum of the subcomponent price levels, the re-aggregated impulse response at each horizon is simply the same weighted sum of the individual-series coefficients — but because price responses are correlated across categories, the authors estimate a Seemingly Unrelated Regressions (SUR) system for each horizon and use a feasible generalized least squares (FGLS, via GMM with a Cholesky factorization) estimator to obtain the correct joint covariance matrix. The correctly computed SUR standard error of the aggregate response includes cross-series covariance terms that a naive standard error (built only from the individual OLS variances) ignores; since price movements are mostly positively correlated, the naive approach understates the true standard error and “would reject a zero response early on” at the 90% confidence level where the SUR-corrected bands do not — meaning the correction changes the substantive conclusion about when the aggregate response becomes significant, not just its precision.

Following a 100-basis-point contractionary shock, the point estimate for headline PCEPI is essentially flat (even slightly positive, though not statistically distinguishable from zero at the 90% confidence interval) for roughly the first three years, turning significantly negative only after about 43 months and reaching a peak reduction of roughly 4% around 54 months; core PCEPI (excluding food and energy) follows a similar timing but with a smaller peak reduction of roughly 2.5%, estimated more precisely. This pattern reproduces Christiano, Eichenbaum, and Evans’s (1999) finding that “the aggregate price level does not fall to a statistically significant extent for 4 years after an FFR increase”; the authors themselves flag that these specific magnitudes should ultimately be checked against their Figure 1 rather than treated as exact to the decimal.

Disaggregating further reveals sharply different response profiles: at the 2-category level, durable goods prices decline fastest (significant after about 2.5 years, peaking around a 5% reduction near 4 years) while services move more slowly and NPISH (nonprofit institution) prices show no significant response at all; at the finest 136-series level, roughly 13% of series are never significant in either direction, and fully half of the series show a significantly positive response to the tightening at some horizon. The authors’ central interpretive claim is that “the reason why the headline and core response is flat and fall after 4 years is not because all prices are unchanged for 4 years” — instead, many individual prices move substantially in both directions, and the aggregate only turns clearly negative once enough series’ declines outweigh the persistently positive ones; used motor vehicle prices, for example, rise initially because higher rates suppress new-vehicle purchases and push substitution demand toward used vehicles.

A contribution-level decomposition — which weights each subcomponent’s price change by its expenditure share — shows that a small number of categories account for most of the eventual aggregate movement, with motor vehicle fuels/lubricants/fluids as the single largest contributor, alongside food and nonalcoholic beverages purchased off-premises, hospital and nursing-home services, financial services, housing, and household utilities. Consistent with the aggregate pattern, almost all of these contribution-level impulse responses are flat for the first few years and only turn strongly negative after roughly 40 months.

The data show no meaningful negative relationship between how quickly a price category’s response to the shock peaks and how frequently that category’s prices adjust empirically — if anything the relationship is weakly positive, the wrong sign for a Calvo-style sticky-price model, which predicts more frequently adjusting prices should respond faster. Menu-cost models fare no better, the authors argue, since in that framework a firm that does adjust must move its price in the direction implied by the money-supply change, which cannot generate the positive responses observed in about half the series; instead, the authors favor a heterogeneous cost-channel mechanism (building on Barth and Ramey 2001 and Ravenna and Walsh 2006) combined with demand substitution across goods, in which a tightening acts as a demand shock for some sectors but raises marginal costs — and hence prices — for financially constrained sectors in the short run, with Baumeister, Mumtaz, and Peersman’s (2013) multi-sector sticky-price-with-cost-channel model described as coming “closest to what we envision as a suitable framework.”

Re-aggregating the disaggregated impulse responses using 1959 expenditure shares instead of 2023 shares yields a similar aggregate response in both magnitude and timing, even though the underlying consumption basket shifted substantially over that period (the food-and-beverages share fell by about 12 percentage points while the health-care share rose by about 11 points) — indicating that decades of structural change in what households buy have not accelerated the economy’s aggregate inflation lag. The authors are explicit about scope limits: the 1982-2008 sample excludes the zero lower bound, COVID, and the 2022-2023 tightening cycle, so the results may not generalize to those episodes; the underlying shock series itself comes from a companion paper’s NLP/ML methodology that treats Fed decisions as exogenous, even though the authors note elsewhere that the 2022-2023 decisions were largely systematic rather than exogenous; and a stylized calculation of the price reduction implied by applying their estimates to the 2022-2023 hiking cycle is explicitly described as “intentionally provocative” and “not a serious policy counterfactual.”