What Do a Million Observations on Banks Say About the Transmission of Monetary Policy?

The paper frames the lending channel as resting on the claim that Federal Reserve open-market operations can directly shift bank loan supply schedules, which in turn requires that a version of the Modigliani-Miller theorem fail for banks: “implicit in such banks-don’t-matter formulations is the assumption that a version of the Modigliani-Miller (M-M) theorem holds for banks,” under which “open-market operations have no meaningful effect on banks’ activities” because banks are indifferent between financing themselves with reservable transactions deposits and non-reservable instruments like large CDs. The authors argue M-M is likely to break down because “many of the major classes of bank liabilities which escape reserve requirements are not covered by deposit insurance,” so that if a bank tries to replace lost insured deposits with uninsured CDs, adverse-selection problems in the CD market mean investors will not fully offset the loss, giving open-market operations “real consequences for bank lending behavior” whenever “there is a link between the reservability and insurability of bank liabilities.”

The paper draws on Stein’s (1995) adverse-selection-based model of banking, which combines three features: a bank lending channel in which open-market operations affect loan supply and the loan-Treasury spread; a financing-friction-driven “buffer-stock” demand for liquid assets, with banks facing the worst adverse-selection problems (plausibly, small banks) holding the highest ratios of liquid securities to loans; and a general-equilibrium role for bank-level imperfections in shaping both loan-Treasury spreads and the response of Treasury rates themselves to monetary shocks. From this the authors derive two formal hypotheses: first, d²L_it/dB_it dM_t < 0 – for banks without perfect access to uninsured finance, the sensitivity of lending to balance-sheet strength intensifies when monetary policy tightens; second, d³L_it/dB_it dM_t dSIZE_it > 0 – this effect should be most pronounced for the smallest banks, which are least able to frictionlessly raise uninsured finance.

The paper explains that earlier evidence – aggregate declines in bank lending after a monetary contraction (Bernanke and Blinder 1992), or evidence that small non-financial firms are more liquidity-constrained after a tightening – “admits another interpretation” in which declining loan demand, not loan supply, drives the results, and that even cross-firm commercial-paper-versus-bank-loan comparisons (Kashyap, Stein and Wilcox 1993) face alternative compositional explanations. The authors’ key methodological move is to hold bank size fixed and exploit within-size-class differences in balance-sheet liquidity, which they argue is harder to explain away: to attribute their results to a standard interest-rate channel, one would need “those banks whose customers’ loan demand is most sensitive to monetary policy systematically opt to hold less in the way of liquid assets” – a “heterogeneous risk aversion” story the authors call “somewhat implausible from a theoretical perspective” and separately test against the data (Q6).

The paper’s sample is drawn from the Call Reports (Consolidated Reports of Condition and Income) that every insured U.S. bank must file, covering 1976:Q1 to 1993:Q2 and yielding 961,530 bank-quarters, with regression samples of 746,179 observations (C&I loans) and 836,885 (total loans) after excluding merger-affected quarters, outliers, and (for C&I) banks where C&I lending is a trivial share of the loan book. The authors implement a two-step regression: in the first step, for each of six bank-size classes and each quarter, they regress bank-level loan growth on four lags of itself, lagged balance-sheet liquidity B_it-1, and Federal Reserve district dummies, yielding a time series of estimated liquidity-sensitivity coefficients β_t for each size class; in the second step, they regress these β_t series on the contemporaneous value and four lags of a monetary-policy indicator (plus a time trend), in both a “univariate” specification and a “bivariate” specification that also controls for real GDP growth. The authors prefer this two-step approach over a more tightly parameterized one-step interaction regression because it “imposes no a priori structure on the time-series properties of lending volume,” instead allowing a different macro shock in each period for each Federal Reserve district.

The paper uses the federal funds rate, the discrete Boschen-Mills index, and a continuous indicator built from Bernanke and Mihov’s (1995) structural VAR, which “nests previous structural VAR’s based on more specific assumptions about Fed operating procedures” such as funds-rate targeting or non-borrowed-reserves targeting. The authors use all three because “no single one of our measures of monetary policy is likely to be perfect”: the funds rate (and, they find, the Bernanke-Mihov measure, which inherits much of the funds rate’s high-frequency volatility even during the Volcker period) may be a poor indicator of the stance of policy in the early-1980s non-borrowed-reserves-targeting period, while the Boschen-Mills index’s discreteness “effectively introduces a substantial degree of measurement error” at higher frequencies.

Beyond noting the theoretical implausibility of heterogeneous risk aversion, the paper offers three empirical checks: first, under the bivariate specification, “if the heterogeneous risk-aversion story is true, the coefficients on GDP growth… should be negative,” whereas “if the rational buffer-stocking story is true, the coefficients… should be positive” – and the paper reports the GDP coefficients are “for the most part positive,” consistent with the lending-channel interpretation; second, since the largest banks face negligible capital-market frictions, “any non-zero coefficients on M_t in the second-step regression [for the largest banks] must purely reflect the direction of the cyclicality bias” – and the paper finds these coefficients are consistently positive (the opposite sign needed for the risk-aversion story), which “cuts in favor of the rational buffer-stocking story” and implies the small-bank estimates “might be understating the true effects of monetary policy.” The paper distinguishes its results from Bernanke and Gertler’s (1995) “balance sheet channel” (in which tight policy weakens small non-financial firms’ creditworthiness and so reduces their access to any external finance, not specifically bank credit) by noting that its identification strategy is based on bank-level, not borrower-level, balance-sheet characteristics.

For C&I loans, the paper reports that across all six combinations of the three monetary measures and the univariate/bivariate specifications, “the point estimates are negative” for small banks, “consistent with the theoretical prediction,” with five of six significant at the 2.4% level or better and four of six below 0.6%; the only exception is the bivariate Boschen-Mills specification, which the authors attribute to that index’s measurement error. The small-bank/big-bank differential is positive and “always larger in absolute value than the corresponding figures obtained for the small banks in isolation,” statistically significant at the 8.0% level or better in all six specifications and below 2.2% in five of six – which the authors summarize as “overwhelming support for both of our key hypotheses.”

The results for total loans are “qualitatively similar, but quite a bit weaker”: all six small-bank point estimates are negative and all six big-bank estimates positive (directionally consistent), but the small-bank coefficient magnitudes are “typically only about one-third to one-fifth that of the corresponding estimates” for C&I loans, and none of the six small-bank test statistics is significant at conventional levels (though two have p-values below 12.1%). The paper offers two explanations, not mutually exclusive: first, if the “rational buffer-stocking” bias (Q6) is real, it should be worse for total loans, where aggregating across loan categories of different cyclicality “exacerbates the bias,” making the C&I-loan estimates the more conservative (and more reliable) ones; second, a simpler explanation “that has nothing to do with endogeneity biases” is that because of their shorter maturity, banks “can adjust the volume of C&I loans outstanding more readily than they can adjust their lending in other major categories, such as long-term mortgages.”

Focusing on the Bernanke-Mihov/C&I-loan combination the authors describe as their most precisely estimated, the paper reports that “for the small-bank category, every single one of the individual monetary-policy coefficients is negative in the univariate specification, and all but one are negative in the bivariate specification,” and that “the implied response of β_t to a monetary shock has an economically plausible hump shape for the small banks, with the coefficients increasing over the first couple of lags and then gradually dying down” – a dynamic pattern consistent with a genuine, gradually-building supply response rather than an artifact of a single noisy coefficient.

The paper’s stated conclusion is that “it now seems pretty hard to deny the existence of a lending channel of monetary policy transmission, at least for the U.S. in our sample period,” based on results that are “strongly statistically significant,” “impervious to a wide range of variations in specification and estimation technique,” and difficult to rationalize with a standard interest-rate channel without invoking the empirically-rejected heterogeneous-risk-aversion story. The authors are explicit, however, that this leaves open “the next logical question”: quantitatively, how important is the lending channel for aggregate investment and GDP? Answering that, they note, requires knowing both what fraction of the banking sector is “small” in the relevant sense (roughly 25% of system assets, using the bottom-95%-of-size-distribution cutoff, which they call “modest, but certainly non-trivial”) and “the elasticity with which borrowers can substitute between bank and non-bank forms of credit on short notice” – a borrower-side question the paper states “will not be easy to answer satisfactorily” but that “must eventually be addressed” for a full picture of banks’ role in monetary transmission.