Monetary policy shocks: What have we learned and to what end?

Questions & answers

Q1. What is the paper’s organizing framework for comparing different monetary policy shock identification schemes?

The chapter writes the model as a structural VAR, A₀Z_t = A(L)Z_{t-1} + ε_t, where ε_t contains mutually orthogonal structural shocks and A₀ captures contemporaneous relationships among variables; the central identification problem is recovering A₀ from the reduced-form covariance matrix, and the chapter’s “recursiveness assumption” posits that a monetary policy shock is orthogonal to all variables in the Federal Reserve’s contemporaneous information set — operationalized as zero restrictions placing certain variables in a “predetermined block” that cannot respond within the period to the policy shock. A formal result (their Proposition 4.1) shows that under this assumption, the policy shock is uniquely identified regardless of how variables are ordered within the predetermined or non-predetermined blocks — a technical point that removes ordering as a genuine source of ambiguity within the recursive approach, in contrast to how Cholesky-ordering choices are sometimes portrayed.

Q2. What are the three benchmark recursive identification schemes, and how are they distinguished?

The three schemes differ only in which variable serves as the policy instrument: the “FF” scheme uses the federal funds rate; the “NBR” scheme uses nonborrowed reserves (plus extended credit); and the “NBR/TR” (Strongin) scheme uses the ratio of nonborrowed to total reserves, exploiting the assumption that total reserve demand is essentially interest-inelastic in the short run so that a policy shock reallocates reserves between nonborrowed and borrowed components. The standard deviation of the estimated shock differs substantially across schemes (0.71 percentage points at an annual rate for the FF shock versus an implied 0.39 for the NBR shock), and pairwise correlations among the three shock series are only moderate (0.51 to 0.82), which the authors note implies “at least two must be confounded by nonpolicy shocks as well” — yet the qualitative dynamic responses remain consistent across all three.

Q3. What are the main documented effects of a contractionary monetary policy shock?

Following a contractionary shock under any of the three benchmark schemes, the federal funds rate rises persistently, nonborrowed reserves fall persistently (a strong liquidity effect), total reserves are initially insulated but eventually fall by roughly 0.3%, M2 falls immediately and persistently, commodity prices decline with an initial delay, the GDP deflator remains essentially flat for about a year and a half before declining, and real GDP falls in a hump-shaped pattern with the maximal decline occurring roughly one to one-and-a-half years after the shock. The authors summarize this as their central qualitative finding, common to “all three policy shock measures,” even though the underlying shock series are not highly correlated with one another.

Q4. What is the “price puzzle,” and how does the chapter show it can be resolved?

The price puzzle — a term the chapter attributes to Eichenbaum (1992) — refers to the anomalous finding that some recursive identification schemes imply a contractionary monetary policy shock leads to a sustained rise in the price level, evident historically around events like the 1974 oil-price shock; the chapter reports that when the federal funds rate is used as the policy indicator without including commodity prices in the Fed’s information set, the puzzle is statistically significant (a bootstrap test finds the price response positive in 96-98% of simulations at 2-, 4-, and 6-quarter horizons). Following Sims’s (1992) conjecture that the puzzle arises because the omitted information set fails to capture inflation information the Fed was already reacting to, the chapter confirms (citing Christiano et al. 1996a and Sims and Zha 1998) that including current and lagged commodity prices in the Fed’s information set makes the puzzle largely disappear for the benchmark FF and NBR schemes, which the authors describe as having become “standard practice.”

Q5. What do variance decompositions reveal about the importance of monetary policy shocks, and why do the authors caution against overinterpreting them?

The FF benchmark shock accounts for an estimated 21% of the 4-quarter-ahead forecast error variance in output (with a wide 95% confidence interval of 7-41%), rising to 44% at 8 quarters, while the NBR benchmark shock accounts for only about 7% at 4 quarters and 10% at 8 quarters — leading the authors to conclude that “inference about the importance of monetary policy shocks depends sensitively on which policy shock measure is used.” None of the three schemes attribute much of the price level’s volatility to policy shocks even at a three-year horizon, underscoring that the qualitative dynamic responses are far more robust across identification schemes than the quantitative variance-decomposition estimates.

Q6. How robust are the results to using a fully simultaneous (non-recursive) identification, and to narrative measures of policy shocks?

Implementing a version of the Sims-Zha (1998) non-recursive model — in which the Fed’s reaction function contemporaneously sees only commodity prices and a monetary aggregate, not current output or the aggregate price level — the chapter finds the qualitative response of the system “quite similar” to the benchmark FF and NBR recursive results, and shows that whether goods-market variables are treated as predetermined relative to the policy shock is “not important” for the resulting inference. Comparing to Romer and Romer’s (1989) narrative approach — dates when the Fed is judged, from historical records, to have deliberately tightened policy to fight inflation — the chapter reports the qualitative responses (funds rate up, commodity prices down, monetary aggregates down, employment down with a delay, prices little affected initially) are again “quite similar” to the benchmark recursive results, though the narrative episodes imply roughly double the maximal impact on the federal funds rate (about 100 basis points versus about 60 for the FF shock).

Q7. What happens when the recursiveness/exogeneity assumptions are altered or tested directly?

Excluding current real GDP from the Fed’s contemporaneous information set produces an implausible result — output initially rising before eventually falling, which is inconsistent with any known monetary transmission model — leading the authors to reject this specification, whereas excluding the current price level leaves the benchmark results “virtually unaffected.” Testing Bernanke and Mihov’s (1998) overidentifying restriction that a particular reserve-market parameter equals zero, the chapter reports this restriction is “strongly rejected” by the data (bootstrap p-value below 0.001%) for the NBR-based model, concluding that “BM’s claim to have rejected the benchmark NBR model is unwarranted,” though the FF model’s estimated parameter has a sign that Bernanke and Mihov’s own theory would rule out.

Q8. What does the chapter conclude about the overall robustness of monetary policy shock analysis, and what does it leave open?

The chapter’s central conclusion is that “qualitative inference about the effects of a monetary policy shock is quite robust to the different shock measures discussed” — spanning recursive VARs with different policy instruments, a non-recursive simultaneous model, and narrative-based shock measures — even as quantitative magnitudes (shock size, variance shares, precise timing) vary considerably across approaches. The authors note, as a methodological caution largely separate from their main results, that estimated parameters of the Fed’s reaction function are not directly interpretable as behavioral parameters, since they convolve true behavioral responses with projections onto the econometrician’s necessarily limited information set — a limitation of using any single VAR specification to recover the Fed’s “true” reaction function.