Are the Effects of Monetary Policy Shocks Big or Small?

The paper asks why monetary VARs and the Romer-Romer (2004, “R&R”) narrative approach imply such different real effects of monetary policy shocks, and answers that the gap is largely explained by three identifiable, correctable factors rather than a genuine disagreement about the true policy shock (pp. 1-3). Once these three factors — a different implied path for the federal funds rate (FFR), contamination from the non-borrowed-reserves (NBR) targeting episode, and lag-length sensitivity — are controlled for, the paper argues the competing approaches converge on a “medium range” estimate of the real effects of monetary policy.

At baseline lag specifications, a 100-basis-point policy shock produces wildly different peak effects across methods (Table 2, Panel A, p. 21): a standard recursive VAR (Cholesky-identified, FFR ordered last, 12 lags, following R&R’s variable choices) implies industrial production (IP) falls about 0.7% and unemployment rises 0.16 percentage points; the R&R single-equation approach (a macro variable regressed on 24 autoregressive lags plus 36 lags of the R&R narrative shock, following R&R’s own baseline procedure) implies a much larger 4.3% IP decline and 0.93-point unemployment rise; and a “hybrid VAR” (the baseline VAR with the cumulative R&R shock series substituted for the FFR as the policy variable) lands in between, at a 1.6% IP decline and 0.40-point unemployment rise. Both the sample (monthly US data, 1970:1-1996:12, with R&R shocks derived over 1969:3-1996:12) and prices move similarly: -0.1% (VAR), -4.2% (R&R), -1.8% (hybrid).

A VAR Cholesky shock and an R&R narrative shock are not the same object: an R&R innovation generates a larger and more persistent rise in the FFR than an equivalent VAR shock, and once impulse responses are rescaled onto a common counterfactual FFR path, the difference in estimated macroeconomic effects “largely disappears” (Section II.B, pp. 10-12). Coibion shows this by applying the single-equation approach to all three shock series and normalizing by each series’ implied FFR response (Figure 4); doing so brings the VAR and hybrid-VAR results into line with the R&R results, indicating that part of the apparent disagreement is really a disagreement about how big a “shock” is, not about the transmission mechanism.

R&R shocks during the Federal Reserve’s 1979:10-1982:9 non-borrowed-reserves (NBR) targeting experiment are statistically predictable from lagged macroeconomic variables — F-statistics of 2.28 and 2.45 (both significant at the 5% level) for industrial production and inflation respectively over the full sample (Table 1, pp. 12-16) — which violates the identifying assumption that the shocks are exogenous, and this episode disproportionately drives the large estimated R&R effects. Bernanke-Mihov and Boschen-Mills alternative measures of policy stance show much smaller swings over this period than R&R do, and dropping the NBR period from the sample brings all three approaches’ estimates together (Figure 7); Coibion concludes the original discrepancy is “largely driven by the early Volcker era” (p. 20).

The R&R single-equation results rise nearly monotonically as the number of shock lags increases (Figure 8), while the VAR estimates are comparatively insensitive to lag length; selecting lags by AIC instead of the R&R baseline choice (24 autoregressive, 36 shock lags) reduces the R&R peak IP effect from -4.3% to -3.4%, with model averaging (Kass-Raftery bootstrap weights) giving essentially the same result (Section II.D, pp. 17-20; Table 2, Panel A, p. 21). This is presented as a purely methodological source of disagreement: the R&R two-step procedure’s sensitivity to an atheoretical tuning choice, rather than the VAR’s, generates part of the apparent gap in shock magnitude.

With the NBR-targeting period excluded (1979:10-1983:12 omitted) and lag length chosen by AIC or model averaging, peak industrial production estimates converge across all three approaches to roughly a 1.8% to 2.8% decline, with unemployment rising by about 0.19 to 0.70 percentage points, depending on approach and lag method (Table 2, Panel B, p. 21). The paper states its key conclusion in these terms (p. 20, the paper’s own emphasis): “the real effects of monetary policy shocks appear to be in the medium range, with industrial production dropping by approximately 2-3 percent and unemployment rising by around half a percentage point.”

Three additional shock measures constructed from Taylor-rule residuals — a GARCH(1,1) version of the R&R rule, a time-varying-coefficient (TVC) version re-estimated at each FOMC meeting, and Smets-Wouters (2007) DSGE-filtered shocks — imply 100-basis-point effects of roughly a 2% IP decline and 0.60-point unemployment rise (GARCH); a 4% IP decline, 0.75-point unemployment rise, and 4% price decline (TVC, at baseline lags); and about a 1.5% IP decline, 0.40-point unemployment rise, and 1.5% price decline (Smets-Wouters) (Section III.A-C, pp. 22-27). All three are markedly less sensitive than the original R&R series to lag length, to dropping individual historical episodes, or (for GARCH and Smets-Wouters especially) to excluding the NBR period; correlations with the original R&R shock series are 0.93 (GARCH), 0.82 (TVC), and only 0.40 (Smets-Wouters), and the Smets-Wouters shocks are 69% more volatile than R&R’s.

All four measures (original R&R, GARCH, TVC, and Smets-Wouters) point to a “nontrivial” contribution of monetary policy shocks to historical US fluctuations that is “considerably smaller than implied by the original R&R results but also considerably larger than those implied by the baseline VAR” (Section III.D, p. 28). Exogenous monetary policy is estimated to have contributed approximately 2-3 percentage points of the roughly 5-percentage-point rise in unemployment between 1979 and 1982, and much of the disinflation of the early 1980s is attributed to exogenous policy — a feature the paper notes standard monetary VARs do not share. The contribution of monetary policy shocks to real fluctuations is described as having “drastically declined” since the early 1980s, which the paper interprets as consistent with improved monetary policy conduct (the Great Moderation).

The sample ends in 1996:12, limited by the availability of the Greenbook forecast data the R&R approach requires, so the paper explicitly cannot speak to monetary policy effects after 1996 or in a zero-lower-bound environment (p. 31: measuring policy stance “is rife with difficulties even in calm periods,” and doing so “during extraordinary circumstances such as the Volcker experiment or the current period presents a particularly challenging exercise”). The TVC specification also imposes a narrower information set than the original R&R rule (contemporaneous rather than lagged forecasts, p. 25 fn. 27), and the Smets-Wouters filtered shock uses realized data rather than Greenbook forecasts, so it combines true monetary innovations with rational-expectations forecast errors (p. 26, fn. 28). Online Appendix Monte Carlo simulations show model averaging has a small downward bias relative to the truth, though smaller than the bias from under- or over-fitting lag length.