Measuring Monetary Policy

Bernanke and Mihov argue that the standard recursive VAR indicators of monetary policy – orderings built around the federal funds rate, nonborrowed reserves, or the nonborrowed/total reserves ratio (Strongin 1995) – are Cholesky-style restrictions that do not correspond to an explicit model of how the Federal Reserve actually operates in the market for bank reserves, so empirical results can be sensitive to which indicator is chosen. Instead of picking one ordering on priors, they build a “semi-structural” model of the reserves market with parameters governing how much the Fed accommodates non-policy disturbances, and show that the standard indicators are special cases of this model corresponding to particular parameter restrictions – which lets the indicators be tested against each other rather than one being assumed correct.

The system is a six-variable monthly VAR for the United States, 1965:1-1996:12, split into a “policy block” of total reserves (TR), nonborrowed reserves (NBR), and the federal funds rate (FFR), and a “macro block” of real GDP, the GDP deflator, and a Dow-Jones spot commodity price index, with the macro block assumed block-exogenous within the month (policy-block shocks cannot affect the macro variables contemporaneously) – a timing assumption the authors say “seems consistent with… monthly data” (Section II, p. 874). Lag length is chosen by a general-to-specific procedure starting from 15 lags, with 13 retained for the full monthly sample; the model is also re-estimated at biweekly frequency and across four subsamples (1965:1-1979:9, 1979:10-1996:12, 1984:2-1996:12, 1988:9-1996:12) as robustness checks.

Within the policy block, BM specify three structural equations – reserves demand (elastic in the funds rate, slope alpha), borrowed-reserves supply (elastic in the funds-rate/discount-rate spread, slope beta), and a nonborrowed-reserves/policy equation in which the Fed’s accommodation of demand and borrowing shocks is governed by parameters phi^d and phi^b – and invert these to recover the exogenous policy shock v^s as a linear combination of the three VAR residuals. Setting (phi^d, phi^b) to particular values nests five identifications: FFR (phi^d=1, phi^b=-1: the Fed fully accommodates all shocks to hold the funds rate fixed), NBR (phi^d=0, phi^b=0: pure nonborrowed-reserves targeting), NBR/TR a la Strongin (alpha=0, phi^b=0), BR (phi^d=1, phi^b=alpha/beta: borrowed-reserves targeting), and a just-identified (JI) model that imposes only alpha=0 and estimates phi^d and phi^b freely.

Estimation proceeds in two steps: the reduced-form VAR is estimated by OLS, and its residuals are then used as data to estimate the structural parameters (alpha, beta, phi^d, phi^b) by efficient GMM, with the overidentifying restrictions of each of the four restricted models (FFR, NBR, NBR/TR, BR) tested against the data. Because the JI model imposes only one restriction (alpha=0), it has no overidentifying restrictions of its own to test and instead serves as the benchmark the other four are checked against.

The NBR model’s overidentifying restrictions are rejected at conventional significance levels in every subsample except 1979:10-1982:10 – the conventional dates of the Volcker nonborrowed-reserves-targeting experiment – while the NBR/TR (Strongin) specification has the best overall fit across subsamples and the FFR model performs well before 1979 and after 1988. Consistent with this, a Hamilton (1989) two-state regime-switching model fit directly to the structural accommodation parameters (phi^d, phi^b), without imposing subsample break dates by assumption, independently locates two breaks – one in late 1979, one in 1982 – bracketing the same Volcker episode: phi^d and phi^b are estimated at roughly 0.863 and -0.778 outside that window (an interest-rate-focused regime that largely accommodates reserves-demand shocks) versus roughly 0.183 and -0.216 within it (an NBR-targeting regime with little demand accommodation) (Figure I, p. 891).

All five identifications generate impulse responses that are qualitatively “reasonable” – an expansionary shock raises output relatively quickly, peaking around 12-18 months, and raises prices more slowly and persistently – but the magnitudes diverge sharply: the cumulative output response under NBR/TR is more than twice as large as under NBR, and at a four-year horizon the price response under the NBR identification is about four times the response under FFR. The authors conclude “these differences are certainly large enough to have important effects on the inferences one draws about the effects of monetary policy” (p. 894); the exercise itself is explicitly “meant only to be illustrative” (p. 896), since it applies full-sample JI parameter estimates to a VAR that pools across the regime changes documented in Q5.

Algebraically, each overidentified indicator (FFR, NBR, NBR/TR, BR) can be written as a linear combination of the true policy shock v^s plus reserves-demand shocks v^d and borrowing shocks v^b, weighted by phi^d and phi^b (Eqs. 13-16); when the Fed accommodates demand shocks heavily (phi^d close to 1, as estimated outside the Volcker window), the NBR indicator in particular loads heavily on v^d rather than v^s, biasing its implied impulse responses. Formalizing this as a bias term in the NBR-based liquidity-effect estimate (Eq. 17), the authors evaluate it at the JI model’s GMM estimates and find it negative and mostly growing in magnitude – -0.777 in the full sample, -0.904 in 1988-1996, and -1.035 in 1984-1996 (a sign reversal, i.e. the NBR indicator implies the wrong sign for the liquidity effect in that subperiod) – leading them to conclude that “the ‘vanishing liquidity effect’ may well be the result of using a biased indicator of monetary policy, rather than of a change in the economy” (p. 896).

Beyond the five discrete indicators, BM construct a continuous overall policy-stance index from the full vector of structural innovations (A^{-1}(I-G)P), in which one element’s VAR innovations correspond exactly to the JI model’s policy shocks; because the funds rate and nonborrowed-reserves growth are not measured in comparable units across regimes, the raw series is normalized by subtracting a trailing 36-month moving average, with zero defined as “normal” policy. This index correlates 0.71 with the Boschen-Mills (1991) qualitative index at monthly frequency, and the Romer and Romer (1989) narrative dates line up with peaks of maximum policy tightness in BM’s index “rather than points at which policy changed from expansionary to contractionary” (p. 898).

The authors note that the block-exogeneity timing assumption is shared with the standard Cholesky approach and is not independently tested beyond its plausibility for monthly data; that the discount rate is set to zero in estimation because discount-rate changes are infrequent, a simplification that could matter in periods of active discount-rate adjustment; and that the model’s institutional structure is calibrated to the U.S. reserves market, so applying it elsewhere requires adapting the reserves-market equations to local institutions (p. 900). They also flag, as noted in Q6, that the illustrative impulse responses pool across regimes their own regime-switching results show are structurally unstable – though a robustness check using regime-switching rather than full-sample parameter estimates gives similar results (fn. 24, p. 892).