Monetary Policy Shocks: What Have We Learned and to What End?

Summary

A comprehensive Handbook of Macroeconomics chapter (written 1998) reviewing empirical strategies for identifying monetary policy shocks using vector autoregressions (VARs). The paper motivates the recursiveness assumption — the Federal Reserve (Fed) observes slow-moving variables (output, prices, commodity prices) contemporaneously but not same-period financial quantities — and benchmarks three policy instruments: federal funds rate (FF), non-borrowed reserves (NBR), and their ratio to total reserves (NBR/TR). Qualitative agreement across all three recursive schemes, the non-recursive Sims-Zha model, and the narrative Romer-Romer approach is documented: a contractionary shock produces a persistent output decline, and the price puzzle is resolved by including commodity prices in the information set. The paper frames the entire exercise through the Lucas program: VAR impulse responses serve as stylized facts that structural dynamic stochastic general equilibrium (DSGE) models must match.

Key Claims

• Lucas program as organizing principle: Impulse responses from identified VARs are stylized facts for model-building, not ends in themselves; structural models must be calibrated to reproduce them.
• Recursiveness assumption: The Fed observes the "slow" block $X_t$ (output, prices, commodity prices PCOM) contemporaneously but not the "fast" block $X_{2t}$; policy instrument $S_t$ separates the two. Cholesky identification on $Z_t = (X_t, S_t, X_{2t})$.
• Identification invariance: The dynamic response of $Z_t$ to the monetary policy shock is identical across a whole family of $A_0$ matrices differing by an orthogonal rotation $W$ of the lower-right block; Cholesky is one member of this equivalence class.
• Three benchmark instruments (FF, NBR, NBR/TR): Federal funds rate (Bernanke-Blinder 1992), non-borrowed reserves (CEE 1992), and ratio NBR/TR (Strongin 1995) give qualitatively similar impulse responses.
• Price puzzle: Without commodity prices $\mathrm{PCOM}_t$ in the slow block, the price level rises after a contractionary shock; including $\mathrm{PCOM}_t$ (Sims 1992) resolves this.
• Liquidity effect: M1 and M2 initially rise, then persistently fall after a contractionary shock — consistent with a short-run liquidity effect.
• Sims-Zha non-recursive scheme: Allowing the Fed to observe contemporaneous prices and output in a block-simultaneous $A_0$ gives qualitatively similar results to the benchmark.
• Romer-Romer narrative approach: Federal Open Market Committee (FOMC)-minutes-based contractionary episodes are approximately the innovation to NBR orthogonal to the FF rate; impulse response functions (IRFs) are qualitatively similar to recursive VAR results.
• Variance decomposition: FF shocks account for 21%, 44%, and 38% of output forecast-error variance at 4, 8, and 12 quarter horizons; NBR shocks account for only 7–10%. Both measures explain a negligible fraction of price-level forecast variance. Policy shocks dominate forecast-error variance of the policy instruments themselves at short horizons but are not the dominant source of business cycle fluctuations.
• Policy rule pitfall: Estimated feedback rule coefficients are a convolution of the true policy rule and the data collection/revision process; a backward-looking rule may appear forward-looking due to real-time data gaps, not genuine forward-looking behavior.
• Structural identification theory (Section 3): The identified set for $A_0$ is $\mathcal{Q}_V \cap \mathcal{Q}_\tau \cap \mathcal{Q}_S$, where $\mathcal{Q}_V$ constrains $A_0$ to be consistent with the reduced-form covariance ($A_0 A_0' = \Sigma^{-1}$), $\mathcal{Q}_\tau$ encodes the theoretical zero restrictions (recursiveness and block structure), and $\mathcal{Q}_S$ requires statistical identifiability. The order condition ($l \geq k(k-1)/2$ restrictions for a $k$-variable system) is necessary but not sufficient; the rank condition of Rothenberg (1971) is sufficient for local identification. Under the recursiveness assumption, only the monetary policy shock column of $A_0^{-1}$ is identified — the remaining columns are not.
• Bernanke-Mihov critique (Section 4.5.2): Bernanke and Mihov (1998) propose a structural model of the federal funds market and impose the overidentifying restriction $\gamma = 0$ (borrowed reserves insensitive to NBR) to discriminate among the benchmark schemes via a likelihood-ratio test. CEE demonstrate this restriction is empirically rejected ($F = 3.48$, $p < 0.001$) and theoretically unjustified — Goodfriend (1983) shows a positive co-movement between NBR and borrowed reserves is expected — invalidating Bernanke-Mihov's (BM) discriminating test.
• Exogenous policy rule (CEE 1997b): Standard VAR experiments recover the economy's response to an identified policy shock; DSGE model experiments must specify the entire path of the policy instrument as exogenous. CEE propose that structural models be evaluated by matching the VAR-estimated impulse response function of money growth to the policy shock, removing endogenous feedback and ensuring model and data experiments are conducted under identical interventions.
• Subsample stability: A modified subsample stability hypothesis — stable impulse responses across subsamples but potentially different shock variances — is not rejected. The Volcker disinflation period (early 1980s) exhibits markedly larger policy shock variances without a detectable change in propagation dynamics.

Concepts Introduced or Extended

• Monetary Policy Shocks
• Structural Identification
• Impulse Response Function
• Vector Autoregression
• Variance Decomposition
• Price Puzzle

Entities Mentioned

• Lawrence J. Christiano
• Martin Eichenbaum
• Charles L. Evans
• Christopher Sims
• Thomas Sargent
• Ben S. Bernanke
• Ilian Mihov

Quotes

"The central idea, which we borrow from Robert Lucas, is to use model economies as laboratories in which to conduct experiments that cannot be performed in actual economies."

My Take

The paper's lasting contribution is methodological: it establishes the discipline of using VAR impulse responses as a benchmark that structural models must match, rather than treating the VAR as a structural model itself. The identification invariance result (§4) is underappreciated — it shows that a whole family of $A_0$ matrices all generate the same policy IRF, validating Cholesky identification as a member of a larger equivalence class. The price puzzle as a specification diagnostic (resolve by including $\mathrm{PCOM}_t$ or question the identifying assumptions) has become standard practice. The qualitative robustness across recursive, non-recursive, and narrative strategies is the paper's key empirical finding.

The critique of Bernanke-Mihov is particularly sharp: BM's $\gamma = 0$ restriction, proposed as an overidentification test of the benchmark schemes, is itself rejected empirically ($F = 3.48$, $p < 0.001$) and lacks theoretical justification, leaving the benchmark schemes intact. The CEE (1997b) exogenous policy rule proposal — evaluate DSGE models against the VAR-estimated money-growth IRF rather than interest rate paths — is underappreciated relative to the identification results but provides the most operationally useful bridge between reduced-form VARs and structural model evaluation.