The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
o rk m g ra p e r D e n Credit Conditions and the Cyclical Behavior of Inventories Anil K. K ashyap, O w en A. Lamont and Jerem y C . Stein 3 Working P apers Series M acroeconom ic Issues R esearch Department a * CD Federal R eserve Bank of C h icago June 19 9 3 (W P-93-7) FEDERAL RESERVE B A N K O F C H IC A G O CREDIT CONDITIONS AN D THE C Y C L IC A L BEHAVIOR OF INVENTORIES Anil K. Kashyap Owen A. Lamont Jeremy C. Stein First Draft: June 1992 This Revision: June 1993 Abstract This paper examines micro data on U.S. manufacturing firms’ inventory behavior during different macroeconomic episodes. Much of the analysis focuses on the 1981-1982 recession, a recession that was apparently caused in large part by tight monetary policy. We find important cross-sectional effects in this period: "bank-dependent" firms were much more prone to shed inventories than their non-bank-dependent counterparts. A similar pattern emerges during the 1974-1975 recession, in which tight money also appears to have played a role. In contrast, such cross-sectional differences are largely absent during periods o f looser monetary policy in the 1970’s and 1980’ s. Our findings are consistent with the view that: 1) there is a bank lending channel of monetary policy transmission; 2) the lending channel is likely to be particularly important in explaining inventory fluctuations during downturns. *This research is supported by the Federal Reserve Bank of Chicago, the IBM Faculty Research Fund, the National Science Foundation, M IT’s International Financial Services Research Center, and Batterymarch Financial Management. Thanks also to Ben Bemanke, Eugene Fama, Mark Gertler, David Scharfstein, Andrei Shleifer, Robert Vishny, an anonymous referee and seminar participants at numerous institutions for helpful comments and suggestions and to Maureen O ’ Donnell for assistance in the preparation of the manuscript. I. Introduction This paper is motivated by three stylized facts. The first fact is that inventory movements play a major role in business cycle fluctuations. For example, Blinder and Maccini [1991] document that, in postwar U.S. recessions, declines in inventory investment account for an average of 87 percent of the total peak-to-trough movement in GNP. second fact is that recessions usually follow a period of tight credit. The Eckstein and Sinai [1986] argue that each of the six recessions between 1957 and 1982 was preceded by a "credit crunch"— a time of restrictive monetary policy and rising interest rates. At first glance, it would appear that these two facts can be tied together with a simple and obvious story. The story goes as follows: firms’ desired stock o f inventories depends importantly on the cost of carry; as financing becomes more expensive, firms cut back on their inventory holdings. According to this story, one of the most significant effects of restrictive monetary policy is thus its impact on inventory behavior, an impact which is transmitted through a cost-of-financing channel. There is only one problem with this story. Its basic premise—that inventories are sensitive to financing conditions—finds scant support in most empirical work. This is our third stylized fact. As Blinder and Maccini put it in their survey paper, "little influence of real interest rates on inventory investment can be found empirically." (p. 82) So where does this leave the simple "financial" account o f the cyclical behavior of inventories? In our view, it would be premature to dismiss the theory. The failure of empirical models of inventories to find a significant role for financial variables may say more 1 about the inadequacies of the specifications used in these models than about anything else. There are at least two reasons why standard specifications—which typically use security market interest rates such as the commercial paper rate as explanatory variables— might do a poor job of capturing changes in "financial conditions", broadly defined. First, some borrowers may face quantity rationing constraints of the sort described by Stiglitz and Weiss [1981], Jaffee and Russell [1976] and others, and thus may be unable to obtain funds at the observed commercial paper rate. Second, some borrowers may be "bank-dependent", in the sense that they require external financing but do not have easy access to public debt markets. To the extent that there are important variations in the relative cost of bank loans versus commercial paper, the commercial paper rate may again be a poor measure of financing costs for these firms. Kashyap, Stein and Wilcox [1993]-henceforth, KSW-present some aggregate timeseries evidence on the relative costs of bank loans and commercial paper. KSW begin by constructing a quantity financing variable-the "mix"--which they argue captures movements in bank loan supply. The mix is defined as the ratio of corporate bank borrowing to commercial paper borrowing, and the basic intuition is that a decline in the mix is indicative o f a contraction in bank loan supply. Next, KSW demonstrate that tight monetary policy typically leads to a fall in the mix—i.e., tight money causes an inward shift in bank loan supply. Finally, KSW show that when standard inventory models are augmented to include the mix variable, the mix enters in a fashion that is both economically and statistically significant. In other words, information on the state o f bank loan supply does a better job of 2 explaining inventory movements than do open-market interest rates. KSW interpret their results as evidence that: 1) monetary policy has an important effect on bank lending conditions; and 2) a significant number of firms are bank-dependent, and therefore have inventory behavior that is more sensitive to bank lending conditions than to security-market rates. Thus the KSW results at least partially resuscitate the simple financial account of inventory fluctuations given above, while at the same time rationalizing the failure of conventional inventory models to find a significant role for interest rates. However, their focus on aggregate data leaves an important gap still remaining. The "bank lending" theories o f monetary policy transmission stressed by KSW and many previous authors1 make strong cross-sectional predictions that have not yet been tested. In particular, if the lending view is correct, one should expect the inventories of bank-dependent firms to fall more sharply in response to a monetary contraction than the inventories of those firms who have either plenty o f internal funds or access to public debt markets, and who therefore do not need to rely on bank financing. The goal of this paper is to provide an empirical test of this cross-sectional hypothesis. We begin by examining firm-level inventory movements during the calendar year 1982. We focus on this year because it roughly encompasses the five-quarter-long recession of 1981Q3-1982Q4. This recession was preceded by a clear tightening of monetary policy, beginning with the Fed’ s change in operating procedures in October of 1979. In other words, the aggregate data point to this episode as a natural "case study" of a monetary 3 policy-induced contraction, and hence as an obvious place to begin looking for the sorts of cross-sectional effects we are interested in. Our results from the 1982 data confirm the predictions o f the monetary policy/bank lending account of inventory movements. We find that the inventory investment of firms without access to public debt markets is significantly liquidity-constrained. Or said somewhat differently, firms that are bank-dependent—in the sense of having neither bond market access nor large internal cash reserves--do indeed cut their inventories by significantly more during this period than do their non-bank-dependent counterparts. This conclusion appears to be robust to a wide range of variation in econometric specification and estimation technique. While the 1982 results are consistent with the lending view, they leave other possibilities open as well. First, it may be that the liquidity constraints that we document in 1982 are always present, so that there is nothing unique happening in this year. Second, even if the liquidity constraints observed in 1982 are out of the ordinary, they might be attributable to other consequences of the recession, rather than specifically to tight monetary policy. A leading alternative hypothesis is that a recession impairs the value of firms’ collateral. In a world of information and/or moral hazard problems, such a "collateral shock" could increase the costs of external finance, even if banks’ willingness to supply loans (for a fixed amount of collateral) was unchanged.2 Thus the 1982 results do not by themselves prove that the Fed was able to engineer an inward shift in loan supply. Clearly, the lending story and the collateral story are closely related. 4 Both involve capital market imperfections, and both attribute inventory movements in 1982 to a "cutoff* in the flow of bank credit. The difference is that in the former case the cutoff represents an inward shift in the loan supply schedule, while in the latter it does not. Can the two hypotheses be differentiated? Ideally, we would like to have data for at least: 1) one other recession which was clearly related to tight monetary policy; and 2) one recession which had nothing to do with tight monetary policy (i.e., was caused instead by a supply shock). If we found evidence of liquidity constraints in inventory behavior during the former but not the latter, the lending story would be on firmer ground relative to the collateral story. Unfortunately, aside from the 1982 episode, there are no other such clear-cut natural experiments. As the Eckstein and Sinai [1986] reference above suggests, it is going to be especially difficult to come up with an example of a purely non-monetary recession. Overall, the best we can do is to examine data from both: 1) 1974 (a close approximation to the span of the five-quarter-long recession of 1973Q4-1975Q1); and 2) the 1985-1986 "slowdown". As we will discuss in detail below, the former is a somewhat ambiguous example of another monetary-related recession, while the latter appears to be a period of economic weakness (though not an official recession) that was accompanied by easy monetary policy. And interestingly, liquidity constraints are once again significant in 1974, but completely absent in 1985-1986. Moreover, liquidity constraints are also generally very small in the other years between 1975-1989. Thus the monetary-related recessions of 1974 and 1982 stand out not only relative to 1985-1986, but relative to the rest of the 1970’s and 1980’s as well. This is good news for the lending view. In addition to these intertemporal comparisons, we also examine the cross-sectional 5 determinants o f liquidity constraints during 1982, in another effort to distinguish between the lending story and the collateral story. If the collateral hypothesis explains our 1982 results, we would expect to see liquidity constraints being especially pronounced for that subset of firms whose collateral was most impaired by the recession. in this direction. However, we find little evidence Again, the effect is to put the lending view on somewhat firmer ground. The remainder of the paper is organized as follows. Section II provides some background macroeconomic facts, in an effort to better motivate both: 1) the selection o f 1982 as our primary focus of study; as well as 2) our use of the periods 1974 and 1985-1986 as (admittedly imperfect) comparison episodes. Section III describes our sample and data. In Section IV, we develop and estimate our baseline inventory specifications for 1982, and then perform a number of robustness tests. In Section V, we perform the intertemporal comparisons, re-estimating our baseline model over the 1974 and 1985-1986 periods, as well as over the entire 1974-1989 interval. Section VI analyzes the cross-sectional variation in liquidity constraints during 1982. Section VII concludes. II. Background Macroeconomic Facts A . The 1981-1982 Recession As noted above, our basic empirical strategy is to start by concentrating on an episode for which all the aggregate evidence suggests that tight monetary policy had a real effect on the economy, and on inventories in particular. A p r i o r i , such an episode will be the best place to look for the cross-sectional effects that we are interested in. The 1981-1982 recession would appear ideal for our purposes. 6 In recent history, it stands as the best American example of Milton Friedman’ s [1983] well-known observation th at"... no country has cured substantial inflation without going through a transitional period o f slow growth and high unemployment." (page 202) Indeed, Dombusch and Fischer [1990] argue that "the decision [by the Fed to disinflate] was dramatic because there was little disagreement among economists of widely different macroeconomic persuasions that the move toward tight money would cause a recession along with a reduction in the inflation rate." (page 511) Figure I provides some indicators of the stance of monetary policy over the period 1972-1989. Panel A of the figure plots the Federal funds rate, which Bemanke and Blinder [1992], Goodfriend [1993] and many others believe is a good indicator of the stance of monetary policy. At the time Paul Volcker became Fed chairman in August 1979, the funds rate stood in the neighborhood of 11 percent. Following the October shift in Fed operating procedures, the funds rate began to rise quickly, reaching almost 18 percent by the spring of 1980. With the imposition of the Carter credit controls, short-term interest rates, including the funds rate, dropped precipitously. By late-1980, with the controls lifted, the funds rate again began to climb, reaching 19 percent by year end. During the first eight months of 1981, the funds rate was volatile but on the whole remained high -- averaging around 17.5 percent. In the last quarter of the year, the funds rate began to decline noticeably. Thus on the basis of the funds rate, we are led to conclude that the tightening that started in 1979 was in place at least through the third quarter of 1981. Alternatively, because the funds rate drifted back up during the first part of 1982, before finally retreating to pre-Volcker levels, one might argue that policy was actually tight through the middle of 7 1982. Either way, it seems clear that monetary policy was restrictive at least until the onset o f the recession, which began in the third quarter of 1981. Panel B o f Figure I tells a similar story using another interest rate-based measure of monetary policy, the spread between the prime rate and the commercial paper rate. KSW document that this spread typically rises in the wake of a monetary contraction, and interpret this pattern as evidence that when the Fed tightens, the cost of bank financing increases relative to the cost of open market financing. The prime-CP spread also began to rise shortly after the shift in operating procedures, dipped sharply around the credit control period, and then was high through late 1981. As with the funds rate, there was a final local peak in summer of 1982 before the spread dropped back to its 1979 level. Although the prime-CP spread slightly lags the funds rate, this measure too suggests that tight policy persisted at least through most of 1981. Panel C of Figure I examines a quantity-based indicator, the level of real M2. Here too it appears that monetary policy was tight through most of 1981. The real money supply contracted sharply through 1979 and early 1980, and then was roughly flat until the end of 1981. Indeed, real M2 did not regain its early 1979 levels until the end of 1982. Figure II displays the associated movements in output and inventories. As Panel A shows, real GDP growth was negative in the fourth quarter of 1981, and in three of the next five quarters. Panel B demonstrates that manufacturers cut real inventories for six consecutive quarters, also beginning in the fourth quarter of 1981. In relative terms, inventory behavior in this episode was typical of that seen in recessions: Blinder and Maccini [1991] note that from the third quarter of 1981 through the fourth quarter of 1982, 8 the change in inventory investment represented 90 percent of the output decline. This percentage almost exactly matches their 87 percent average for post-war US recessions. B. The 1974-1975 Recession As noted in the Introduction, it would be desirable if we could replicate our 1982 analysis for another recession in which tight monetary policy also played a clear-cut role. Unfortunately, there does not appear to be another natural experiment quite as clean as 1982. Our next best candidate is the recession of 1974-1975, which according to the NBER, lasted through all four quarters of 1974 and the first quarter of 1975. On the one hand, there seems to be considerable evidence of tight monetary policy leading up to this recession; on the other hand, textbook descriptions of the period (e.g., Dombusch and Fischer [1990], p. 495-496) also attribute the downturn in part to the 1973-1974 oil shock. One piece of evidence for the role of tight monetary policy comes from Romer and Romer [1990]. Their reading of FOM C minutes selects April 1974 as a date when the Fed moved to significantly tighter policy in an effort to combat inflation. Our other indicators confirm this view. The funds rate began the calendar year 1973 at just over 5 percent, and then started a dramatic climb, peaking at almost 13 percent in July 1974. Real M2 declined during the latter part of 1973 and throughout 1974. In addition, there was a pronounced increase in the prime-CP spread, although this spread did not reach the very high levels that were seen in the 1981-1982 episode. As in the 1981-1982 recession, inventory cuts were again substantial, with three consecutive quarters of reductions in manufacturers’ real stocks. The only slight difference 9 from 1981-1982 is the timing of these cuts: in the 1974-1975 recession, the inventory cuts did not begin until the last quarter of the recession, and continued for two quarters after the economy as a whole bottomed out. C. The 1985-1986 "Slowdown" Ideally, we also would like to contrast our results from 1982 to those from a recession that was clearly unrelated to tight monetary policy. exists in the period over which our data is available. Unfortunately, no such episode The closest we can come to this ideal is to examine the two year period 1985-1986. As we will argue below, this seems to have been a period of quite easy monetary policy. And moreover, while it was not classified as an official NBER recession, it was also a time of notable economic weakness. As far as the stance of monetary policy goes, we view 1985-1986 as the decade’s cleanest example of loose policy. First, roughly two years had elapsed since the Fed changed course and started to ease. Thus, even allowing for the famous long and variable lags, one would expect that by 1985 this change would begin to show its real effects (if any). Second, during the 1985-1986 period, all the indicators from Figure I tell a similar story: the funds rate was low and declining, real money growth was healthy and the prime-CP spread was low. Indeed, some observers consider the Fed’ s policy during this period to have been one o f its worst mistakes in recent memory, because on the heels of the favorable oil shock in 1986, the Fed continued with its relatively loose stance, rather than using the opportunity to further cut back inflation. For instance, the Shadow Open Market Committee in its 10 September 1986 Policy Statement wrote (p. 6), "Current Federal Reserve policy is irresponsible. After paying a high price to reduce inflation, the Federal Reserve, urged on by the administration, has returned to the short-sighted policies that produced the inflation of the 1970s."3 In terms o f economic weakness, the brunt of the slowdown seems to have been borne by the manufacturing sector, which was hard-hit by the high foreign exchange value of the dollar. Paul Volcker’s testimony before Congress in July 1985 cites "marked sluggishness in the goods-producing sector." (July 17, 1985, p. 690.) More than a year later, in September 1986, Business Week, noting that industrial production had fallen in five of the previous six months, asserted that "the industrial sector is virtually in a recession." (September 1, 1986, p. 21.) Perhaps most interesting for our purposes, there were significant inventory cuts during this period—as can be seen in Figure II, manufacturers reduced their real stocks during five of eight quarters in 1985-1986. III. Data The sample that we consider is taken from the Compustat data base which tracks publicly traded firms. We restrict our attention to manufacturing companies for which Compustat provides information. Ideally, we would prefer to also examine non-traded firms, since we suspect that these companies are most dependent on bank financing and hence most likely to be susceptible to a credit crunch. Unfortunately, we are unaware of any consistent firm-level data for non-traded companies. Because of this undersampling and because even 11 th e smallest Com pustat firm s (by virtue o f being publicly listed) a re at least m arginally integrated with capital markets, w e conjecture that our analysis if anything understates the m agnitude o f any loan supply effects. In our first set o f tests covering the 1982 episode, w e start w ith the 2,328 U .S . m anufacturing com panies that had com plete (i.e. non-zero and non-m issing) data on assets, sales, inventories and cash holdings for the fiscal years ending in 1980, 1981 and 1982. W e then elim inate the roughly 30 percent o f com panies that reported m ergers o r acquisitions during this period, because these events can induce discontinuities in the balance sheet items that we are studying. Finally, we restrict our attention to the majority o f com panies (approxim ately 60 percent o f the rem aining sample) whose fiscal years end in the fourth quarter, i.e. in either O ctober, N ovem ber or D ecem ber.4 W e do this for tw o reasons. F irst, this tim e period roughly lines up with the recession. Secondly, by using com panies that have sim ilar fiscal years, we ensure that all the firms in our sample are operating under the sam e basic m acroeconom ic conditions. After all these screens, we are left w ith 933 com panies. In our later tests that exam ine other time periods, we reapply the same criteria each y ear to draw the sam ple. For instance, the sample for 1985 was selected by including all U .S . m anufacturing com panies w hose fiscal years ended in the fourth quarter, who had com plete data on assets, sales, inventories and cash holdings for 1983, 1984 and 1985, and w ho w ere not involved in any m erger activity. In applying this set o f screens independently 12 for each year w e are left with an unbalanced panel.5 Table I presents some basic sum m ary statistics for tw o o f the years that w e exam ine, 1982 and 1985. In each year, the table divides our Com pustat sam ple into two sub-samples, representing 1) the firm s that have a bond rating from Standard and Poors at the beginning o f the year in question, and 2) those that do not have such a rating. In addition to data from these two Com pustat sub-samples, the table also includes som e analogous inform ation from the Quarterly Financial Report for Manufacturing Corporations, w hich covers virtually all m anufacturing firm s. The QFR num bers help give som e idea o f the extent to which our Com pustat num bers are representative o f those for all m anufacturing firm s. A couple o f observations stand out. First, the firms with bond ratings are, not surprisingly, much larger than average. These larger firm s also tend to hold somewhat less cash (as a fraction o f assets) than the typical Com pustat com pany. Perhaps a little m ore surprising is the relative behavior o f sales and inventories for larger and sm aller firm s. Both in 1982 and again in 1985, the larger C om pustat firm s (i.e ., those with bond ratings) have m arkedly low er sales and inventory grow th than either the typical Com pustat or Q FR com pany. This phenom enon appears to be part o f a long-run pattern, and is not confined to the two years shown in the table— over the entire 1974-1989 period that we consider, the sm aller non-rated firm s had sales grow th that was on average 1.6 percent faster than that o f the larger, rated firm s. In term s o f the lending view, one suggestive (albeit very crude) com parison can be m ade using the num bers in Table I. Firm s w ithout bond ratings had inventory growth that 13 w as 6.1 percent g reater in 1985 than in 1982. F o r firm s with bond ratings, the com parable 1985-1982 inventory grow th differential is only 2 .4 percent. T hus, the inventory investm ent o f firm s w ithout bond ratings seems to benefit m ore from the shift from tight m oney in 1982 to easy m oney in 1985. IV . Firm -Level D eterm inants o f Inventories D uring 1982 In this section, w e present our em pirical results for 1982. W e start with a "baseline" set o f specifications, which we estim ate using both ordinary least squares (OLS) and instrum ental variables (IV). W e then discuss the econom ic significance o f our param eter estim ates. F inally, w e question some o f the m odeling choices em bodied in our baseline specifications, and exam ine the robustness o f o u r results to a num ber o f alternative specifications. A . Baseline Specifications T able II sum m arizes our baseline regression results for the period 1981:4-1982:4. (i.e ., the calendar year 1982) In each o f the eight regressions, the dependent variable is the change in the log o f firm inventories over the year. T he right hand side variables include a constant term ; the log o f the inventory-sales ratio at the beginning o f the year; the change in. the log o f firm sales over both the current and preceding years; as w ell as 19 dum m y variables corresponding to tw o-digit SIC codes. T hese variables are intended to control fo r the non-fm ancial determ inants o f inventories. In particular, the start-of-period inventorysales ratio and the change-in-log-sales term s can be loosely m otivated by a target adjustm ent 14 model o f the sort seen in Lovell [1961].6 Row 1 in the table represents the simplest possible specification. W e add to the sales and industry controls the variable "LIQ ", which is defined as a firm ’s ratio o f cash and marketable securities to total assets at the beginning o f the period (i.e ., as o f the end o f 1981). The equation is then estimated by OLS. As can be seen, the LIQ variable is strongly significant—it enters with a coefficient o f 0 .3 8 , and a t-statistic o f 3 .5 2 .7 This result is consistent with the notion that liquidity constraints are im portant for inventory behavior, but it is also subject to other interpretations. T he am biguity arises because the LIQ variable may be endogenous, and may be proxying for other factors that should affect inventory behavior. For exam ple, a non-financial explanation might be that LIQ is a proxy for innovations in firm profitability—firm s that have a high value o f LIQ m ight be firms that have recently become more profitable. If this is the case, it would not be surprising to see these firms devoting more resources to inventory investm ent, regardless o f w hether or not they are liquidity-constrained.8 There are two basic ways that one can address this am biguity. The first approach involves using a priori theoretical argum ents to sharpen our predictions relative to the "endogenous L IQ ” hypothesis. The second approach is to estim ate the coefficient on LIQ using an instrum ental variables procedure that should m itigate any endogeneity bias. W e present the results o f both approaches in Table II, and in w hat follows. 15 In row 2, w e add another variable to the OLS specification o f row 1. This variable is given by LIQ*B, w here B is a "bond m arket access dum m y" that takes on the value one if the firm in question has a Standard and Poors bond rating as o f the beginning o f the period (i.e ., as o f 1981:4). As noted in Table I, roughly 14 percent o f o u r 1982 sample firm s have such a rating. T he idea behind this interactive term is that a firm should only be bankdependent if two conditions are satisfied: 1) it has a small am ount o f cash on hand; and 2) it is unable to raise money in public m arkets. Thus our bank-dependence hypothesis predicts a positive coefficient on the LIQ term , and a negative coefficient on the LIQ*B term . (That is, the net effect o f internal liquidity for firm s with bond ratings— which is given by the sum o f the LIQ and LIQ*B coefficients—should be sm aller.) In contrast, the "endogenous LIQ" hypothesis makes no such prediction about the coefficient on the LIQ*B term . Continuing with the above exam ple, if LIQ is simply proxying for firm profitability, one might expect that this effect would be sim ilar for all firm s, and hence that LIQ*B would have a coefficient o f roughly zero.9 As can be seen from the table, the coefficient on LIQ*B in row 2 is negative, and, at -0.31, more than three quarters the m agnitude o f the positive coefficient on LIQ. Thus it appears that internal liquidity is much less im portant for firm s w ith access to public debt m arkets, which is consistent with the bank-dependence hypothesis. Rows 3 and 4 make a sim ilar point with a slightly less constrained specification. Rather than using the LIQ*B interaction term , the equations (with ju st the LIQ variable) are run separately for firm s with B = 1 and B = 0 . This allow s the two types o f firm s to have 16 different intercepts and different sensitivities to current and lagged sales. F o r firm s w ithout access to public bond markets, LIQ is again positive, at 0 .4 1 , and strongly significant. F or those with access, LIQ is actually negative at -0.28, though statistically insignificant. The difference between the LIQ coefficients in the two regressions is strongly significant. In Panel B o f Table II, we re-run the specifications in row s 1-4, using (optim al) IV, rather than OLS. In each case, we use a firm ’ lagged value o f LIQ (i.e ., LIQ as o f 1980:4) as an instrum ent for beginning-of-period LIQ. This instrum enting procedure should mitigate any problem s that arise from LIQ proxying for recent innovations in profitability. O f course, even lagged values o f LIQ may contain some inform ation about the perm anent com ponent o f firm profitability. But this should pose less o f a problem , given the rest o f our specification. Recall that we are essentially seeking to explain changes in the stock o f inventories relative to sales. W hile it seems plausible that a secularly m ore profitable firm m ight w ant to m aintain a higher level o f the inventory-to-sales ratio than a less profitable firm , it is harder to imagine why such a firm would want to keep growing its inventories faster relative to sales on a year-in, year-out basis. In spite o f these argum ents, it is still difficult to claim that o u r instrum enting procedure is foolproof. W e recognize that there may be situations in w hich there rem ains som e residual endogeneity bias. But it must be emphasized that the use o f IV is but one o f three lines o f defense against the endogeneity problem . The other tw o, which both derive from our a priori theoretical argum ents, are: 1) the com parison between B = 0 and B = 1 firm s seen ju st above; and 2) the intertem poral com parisons that will be perform ed in Section V below . Even if residual endogeneity problem s could explain the presence o f a positive LIQ 17 coefficient for B = 0 firm s in 1982, it is not at all clear how they could explain the sort o f system atic differences in this coefficient both across types o f firm s and across tim e periods that are predicted by our theory. Thus w hile w e regard the use o f IV as helpful, w e are not resting our entire case on the purity o f our instrum ents. The results in Panel B o f Table II are very sim ilar to those obtained w ith OLS. F or B = 0 firm s, the LIQ coefficient rem ains at 0 .4 1 , w hile for B = 1 firm s it rises slightly, to -0.21. The difference between these two coefficients is still statistically significant at the 5 percent level. A gain, this is consistent with our form ulation o f the bank-dependence hypothesis.10 B. Econom ic Significance o f the Results W hile the LIQ coefficient o f 0.41 for B = 0 firm s may be statistically significant, it is not im m ediately obvious w hether its m agnitude is econom ically im portant. F o r the purposes o f a "back o f the envelope" calculation, we return to the sum m ary statistics in Table I. F irst, the median B = 0 firm in our sam ple cut its real inventories by about 10.2 percent in 1982. Second, the median value o f the LIQ variable for these firm s is 5 .4 percent, with a standard deviation o f 13.6 percent. This means that for a B = 0 firm , a one standard deviation change in LIQ results in an increase in inventories o f 13.6 percent x 0.41 = 5 .7 percent. Loosely speaking, if we start w ith a "typical" B = 0 firm that is cutting its inventories by 10.2 percent, and then increase its cash holdings from , say 5 percent o f assets 18 to 19 percent o f assets, w e elim inate more than h alf o f the inventory reduction." Although w e fully appreciate the potential pitfalls inherent in draw ing a precise structural interpretation from our reduced-form regressions, we nonetheless think these calculations are suggestive. Even if the coefficient used is only half the size (i.e ., 0.2 instead o f 0.41) the effect would still appear to be econom ically m eaningful. Thus although the exact quantitative im portance o f the effect is som ew hat uncertain, it is likely to be non trivial. C. Robustness to A lternative Specifications O ur specifications in Table II embody a num ber o f m odeling choices that could conceivably influence our results. In general, we would like to be able to control for all the fundamental firm -level determ inants o f inventories. (N ote that any industry-w ide or econom y-w ide factors— . interest rates— e.g will be subsumed in the constant term and the industry dum m ies.) At the firm level, sales stand out as the principal driving force behind inventory behavior. T hus, one o f the most im portant open questions is our treatm ent o f the inventory-sales relationship. A lthough w e have controlled for the start-of-period inventory-sales ratio, as well as contem poraneous and lagged changes in sales, we have done so in a fairly unstructured m anner. O ne could im agine appealing to a particular structural m odel o f inventory behavior in an effort to com e up with a more precise specification o f the relationship between inventory and sales. 19 W e take a som ew hat different tack. W e begin by acknow ledging that not only is o u r current treatm ent o f the inventory-sales relationship open to criticism , but; given the large num ber o f alternative models, so is alm ost anything else w e m ight try. Instead, w e focus on a less am bitious objective. R ather than trying to com e up with the single "right" specification o f this relationship, we try to argue that our conclusions about the coefficients on LIQ (for both B = 0 and B = 1 firm s) are relatively insensitive to how w e model the im pact o f sales on inventories. W e do so by trying a variety o f alternative specifications o f the inventory-sales relationship. The results are presented in Table III. First, in Panel A , w e use exactly the sam e IV specification seen in Panel B o f Table II, with one exception— e delete the start-ofw period inventory-to-sales term . This term was actually insignificant in m any o f the regressions o f T able II, so one might question w hether it belongs in the specification. W e chose to keep it in our baseline model because: 1) there are theoretical reasons to believe it should m atter; and 2) as will be seen in Section V below, it does in fact enter significantly, (w ith the predicted negative sign) in all the other years we consider. As can be seen from Panel A o f Table III, how ever, the exclusion o f the inventory-sales ratio has absolutely no effect on our results. In particular, the LIQ coefficient for B = 0 firm s rem ains exactly the sam e, at 0.41. In Panel B, w e m odify the specification in a m ore substantial w ay. W e take as our null model the case w here inventories and sales move in proportional lockstep. T hat is, we use the change in the log o f the inventory-to-sales ratio as our dependent variable, and exclude all sales term s from the right hand side o f the equation. T his is roughly equivalent 20 to constraining the contem poraneous change in sales term to have a coefficient o f unity. A lthough our earlier results suggest that this is probably a very p oor specification o f the nonfinancial aspects o f inventory behavior, it has little effect on the LIQ coefficients— for B = 0 firm s, the coefficient is now 0.33 and still significant, w hile for B = 1 firm s it is still slightly negative and com pletely insignificant. In Panel C , we try instrum enting not only for LIQ, but also for the contem poraneous change in sales. This might be expected to m atter, if, as many theories im ply, anticipated changes in sales influenced inventories differently than unanticipated changes. H ow ever, this m odification has a minimal impact on the LIQ coefficients— for the B = 0 sample, the coefficient drops from 0.41 to 0.37. In Panel D , we add the lagged change in the log o f inventories to the right-hand side o f the baseline specification from Table II. One could m otivate this either: 1) by appealing to gradual adjustm ent in inventories; or 2) in a m ore atheoretical V A R -type fram ew ork, sim ply as an additional control variable. This variation again has alm ost no effect w hatsoever. The lagged inventory term itself has a near-zero coefficient, and the coefficients on LIQ are essentially unchanged. In Panel E, we perform a som ew hat different sort o f robustness check. W e try adding a size control— the log o f assets— our baseline specification. T he summary statistics to in T able I raise the possibility that even w ithin, say, the B = 0 category, the larger firm s m ight on average have somewhat different inventory and sales characteristics than the sm aller firm s. As it turns out, how ever, the size control makes absolutely no difference for our results. Indeed, the LIQ coefficient for B = 0 firm s is exactly the sam e as it was in Table II, 21 at 0 .4 1 .1 2 Finally, in o ther regressions not displayed in the table, w e also tested w hether our results w ere dependent on the log-changes form ulation em ployed for inventories and sales. W e used logs in T able II because the raw percentage changes in these variables are highly skew ed-several firm s have extremely low values at the beginning o f the period, and thus show enorm ous percentage changes over the period. Using logs elim inates much o f this skewness, and apparently, produces a better specification. W hen w e re-run the regressions using percentage changes instead, the R2,s are much low er, on the order o f 0.12 rather than 0.36. Again, how ever, we reach a qualitatively sim ilar conclusion: the LIQ coefficient is positive and statistically significant for B = 0 firm s, and close to zero for B = 1 firm s. Although o u r results for the entire sam ple o f m anufacturing firm s look to be robust, there rem ains the question of w hether any particular industries are driving these results. O ne (potentially disturbing) possibility is that the correlation between inventories and LIQ is actually not very broad-based, but rather is due to large effects in ju st one or tw o industries.1 3 Table IV investigates this question. The table docum ents the results o f running separate equations (for B = 0 firms only) for each 2-digit SIC industry for which w e had m ore than 30 observations. Each specification is exactly identical to that in row 7 o f Table II. Although the individual industry estimates are substantially noisier, the overall results suggest that the correlation between inventories and LIQ for B = 0 firm s is indeed quite broad-based. 22 First, 10 o f the 14 point estimates are positive. To put this in perspective, note that under the null hypothesis w here a positive and negative estim ate are equally likely, the probability o f obtaining 10 o r m ore positive values out o f 14 is only 9 .0 percent. Furtherm ore, in spite o f the noise, three o f the positive estim ates are significant at the 10 percent level. In contrast, none o f the negative point estim ates are close to being statistically significant. T he median estim ate is 0 .50, and the unw eighted mean is 0.37. The fact that these values are close to the coefficient o f 0.41 obtained for the entire B = 0 sample suggests that the results in Table II do indeed reflect the behavior o f a "typical" industry, rather than being the consequence of one or two outlier in d u stries.1 4 In sum, we are left with two main conclusions from our analysis o f the 1982 data. First, the IV model used in Panel B o f Table II is probably the single most sensible one, and the param eter estim ates drawn from it are likely the most reliable. Second, the finding that LIQ is a significant determ inant o f inventories for firm s w ithout access to public debt m arkets appears to be both robust and broad-based. V. Comparison of 1982 Versus O ther Periods W e now turn to our analysis of other time periods in the 1970’s and 1980’s. As explained above, o u r aim in looking at these other periods is to have yet another set o f controls against w hich to assess our results for B = 0 firm s in 1982. If the large LIQ coefficient in 1982 stems from contractionary monetary policy and an attendant reduction in bank loan supply, then we should expect to see: 1) a substantial coefficient for 1974 as well, 23 to the extent that this recession also involved tight m onetary policy; 2) sm aller coefficients during the 1985-1986 "slowdown" as well as during other tim es o f easier money. H owever, it should be em phasized that w e also do not necessarily expect a coefficient o f zero during these easy money periods. Even during a tim e o f easy m oney, it is possible that there will still b e som e liquidity effects in inventory investm ent. T hese effects might be due to the usual (tim e-invariant) inform ation and incentive problem s that create a wedge between the costs o f internal and external sources o f finance. Put differently, even when bank lending is relatively "unconstrained" by monetary policy, it may still not be a perfect substitute for internal finance, because an inform ation problem still rem ains between the bank and the firm. Indeed, it is exactly these sorts o f tim e-invariant liquidity effects that have been the subject o f study in the literature on fixed investm ent. Thus our central hypothesis is that tight m onetary policy is likely to intensify the correlation between liquidity and inventory investm ent for B = 0 firm s. T able V investigates this hypothesis. All the specifications in the table are exactly analogous to that in row 7 o f Table II, with the one exception being that, when m ore than one year is included, they allow for a separate intercept term for each year in the regression (i.e ., the regressions contain year dumm ies). Row 1 simply restates our earlier finding from 1982, nam ely a coefficient o f 0.41 on LIQ. Row 2 shows that in 1974, liquidity constraints w ere once again pronounced for B = 0 firm s. T he coefficient on LIQ is 0 .7 1 , and it is highly significant, w ith a t-statistic o f 3 .8 9 .1 This fits with the interpretation o f 1974 as being a tight-m oney-induced recession 5 24 year. Row 3 focuses on our easy money control period, the two years 1985-1986. In these tw o years, the coefficient on LIQ is actually slightly negative, although statistically insignificant. M oreover, the LIQ coefficient in 1985-1986 is significantly different from both that in 1982 and that in 1974. This lends further support to the bank lending account o f inventory m ovem ents. N ot only do the coefficients from 1974 and 1982 stand out relative to 1985-1986, they stand out relative to the other years in the sam ple as well. This is illustrated in row 4. H ere w e com bine all the years between 1974 and 1989, and estimate one giant regression for B = 0 firm s. T he regression contains a LIQ variable, as well as a LIQ *CCR interaction term . H ere CCR is a dum m y variable that takes on the value 1 in the "credit crunch/recession" years 1974 and 1982, and 0 in the other 14 years. The intuition is that the LIQ coefficient in this specification will capture the "average" degree o f liquidity constraint, w hile the LIQ *C C R coefficient will capture the added effect seen in a tight m oney/recession year. T he results in row 4 are striking. T he unconditional LIQ coefficient is only 0 .0 7, w ith a t-statistic o f ju st 1.40. H ow ever, the LIQ*CCR term has a coefficient o f 0 .3 6 , with a t-statistic o f 2.79. Thus while liquidity constraints are substantial in 1974 and 1982, they are alm ost com pletely absent in other y ears.1617 T hese findings— that there is tim e-variation in the im portance o f liquidity constraints -have parallels in the literature on fixed investm ent. G ertler and H ubbard [1988] essentially 25 re-run the original investment-cashflow equations estim ated by F azzari, H ubbard and Petersen [1988], b u t allow liquidity effects to be different in recession years. L ike w e do, they obtain significant positive coefficients on their "recession dum m y". (U nlike us, how ever, they also find that liquidity constraints a re significant for fixed investm ent even in non-recession years.) Sim ilarly, using Japanese data, H oshi, Singleton and Scharfstein [1993] re-run the equations estimated by H oshi, K ashyap and Scharfstein [1991], and find that a large part o f the those results are attributable to periods o f restrictive policy on the part o f the Bank o f Japan. O ur results also help to explain why previous research has failed to uncover significant liquidity effects in inventory investm ent. Given the success o f the (tim e-invariant) F azzari, H ubbard and Petersen [1988] type o f specifications for fixed investm ent, one m ight naturally have expected that analogous results fo r inventories w ould soon follow. But to our know ledge, none have. The results in Table V m ake it clear why: In a typical (non recession) year, liquidity has only a small effect on inventories. Even with an enorm ous am ount o f data, the tim e-invariant com ponent o f LIQ is not statistically significant. Evidently, im portant effects can only be found by narrow ing the search to a period o f tight m onetary policy and econom ic weakness. V I. Cross-Sectional V ariation in the LIQ C oefficient in 1982 W e have m otivated our em pirical w ork prim arily as an investigation o f the "lending view" o f m onetary policy transmission. Although all o u r results to this point are consistent with the predictions o f the lending view, it is still not clear how sharply w e can distinguish 26 between the lending view and other financial accounts o f inventory movem ents also based on capital market im perfections. As noted in the Introduction, one plausible alternative hypothesis is that recessions lead to econom y-wide decline in collateral values. This could— in the presence o f inform ation and/or incentive problem s— increase the cost o f bank finance, even if banks’ w illingness to supply loans (for a fixed am ount o f collateral) was unchanged. The one argum ent w e have made thus far against the "collateral shock" hypothesis involves the data from 1985-1986. In this period o f easy money but pronounced econom ic weakness, there is no evidence o f liquidity constraints in inventory behavior. H owever, a skeptic might reasonably argue that the 1985-1986 control is not very convincing. Perhaps it takes a severe dow nturn to generate noteworthy collateral effects, and the 1985-1986 period was simply not com parable to 1974 or 1982 in term s o f the m agnitude o f econom ic decline. How else m ight one attem pt to differentiate between the "loan supply shift" hypothesis and the "collateral shock" hypothesis? First, it should be noted that there is other, macro evidence that directly supports the form er. F or exam ple, as was shown in Figure II, the tightening in monetary policy that preceded both the 1974 and 1982 recessions w as accom panied by a widening o f the spread between the prim e rate and the com mercial paper rate. A lso, KSW find that the volum e o f C P issuance was rising relative to bank lending volum e during both episodes. These sorts o f facts are consistent with the notion that the Fed was able to engineer an inward shift in bank loan supply. Returning to the cross-sectional data, the collateral shock hypothesis would seem to m ake a clear prediction: if im paired collateral is responsible for the liquidity constraints we observe in, say, 1982, then the LIQ coefficient should be higher in 1982 for firms with the 27 most im paired levels o f collateralizable assets. The predictions o f the loan supply shift hypothesis on this point are m urkier. On the one hand, an inw ard shift in loan supply could take the form o f banks cutting o ff credit most sharply for their low est-collateral custom ers. On the other hand, this need not necessarily be the case. Thus, if one w ere to find a cross-sectional relationship in 1982 betw een the level o f collateral and the LIQ coefficient, this probably would not differentiate very clearly between the two hypotheses. H ow ever, if no such relationship exists, this would cast doubt on the collateral hypothesis. There are a num ber of ways that one might investigate the cross-sectional relationship between collateral and the LIQ coefficient. Table VI presents one crude approach. W e use as a measure o f collateral the book debt-to-assets ratio, which w e label D R A TIO . First, in row s 1-3 o f the table, we divide our sam ple into thirds, reflecting low , medium and high DRATIO firm s. W e then re-run exactly the same regressions seen in row 7 o f Table II for each o f these sub-sam ples. If the collateral hypothesis is correct, w e should see a higher LIQ coefficient for the high DRATIO firm s, since these are the m ost likely to be collateralim paired. H ow ever, the regressions reveal no support for this proposition. T he high D RA TIO firm s actually have the lowest LIQ coefficient, at 0 .22. The highest coefficient is seen in the middle sam ple, although none o f the differences are close to being statistically significant. The regressions in rows 4 and 5 address the same basic question with a slightly different specification. W e now estim ate a single equation over the entire sam ple, but add an interaction term , LIQ *D RA TIO . This specification effectively allow s the LIQ coefficient to 28 b e a continuous linear function o f a firm 's value o f D RATIO. In row 4, we also add D RA TIO by itself as an additional right-hand-side variable, thereby allow ing for the possibility that collateral might affect both the level o f inventories and its sensitivity to LIQ. (In row 5, we delete the DRATIO term .) T he results in both cases echo those discussed ju st above. In row 4, the coefficient on LIQ *D RATIO is o f the wrong sign, but small and statistically insignificant. T he point estim ates in this equation suggest that a firm with no debt has a sensitivity to LIQ o f 0.52, w hile a firm with a debt ratio o f 40 percent has a sensitivity to LIQ o f 0.52 - (0.45 x 0.40) = 0.34. In row 5, the coefficient on LIQ *D RA TIO changes sign, but again is small and statistically insignificant. Here too, the point estim ates suggest that even firm s with no debt w hatsoever have a sensitivity to LIQ that is very close to the full-sam ple value. In sum, the evidence in Table VI offers no positive support for the collateral hypothesis. H ow ever, one might criticize the tests as being weak ones, for tw o reasons. F irst, the standard errors involved (e .g ., the standard error on the LIQ *D R A TIO coefficient) appear to be fairly large. Thus the tests may simply not be able to uncover econom ically m eaningful effects even when such effects in fact exist. Second, D RATIO is a crude m easure o f collateral, and might suffer from endogeneity problem s. F o r exam ple, those firm s that have the highest values o f "unobserved collateral" and are thus least concerned about losing funding during a recession m ight opt for the highest debt ratios. This would tend to obscure the true relationship between collateral and the m agnitude o f the LIQ coefficient. W e attem pted to address this endogeneity problem in two distinct ways. First, we 29 redid all the tests in Table VI using industry-adjusted, (rather than absolute) values o f DRATIO. T his w ould be expected to help if much o f the endogeneity in capital structure choice resides at the industry level~if, for exam ple, all R & D -intensive firm s choose low debt because their assets do not make good collateral. H ow ever, this adjustm ent had no effect on our results. Second, w e tried to find purer "instrum ents" for shocks to collateral values. H ere w e followed the suggestions o f a referee, who observed that during the period leading up to the 1982 recession, oil prices more than doubled and the value o f the dollar appreciated sharply. If the collateral hypothesis is correct, one would therefore expect to find "oil-sensitive” and "export-sensitive” firm s—whose underlying asset values were presum ably most im paired by these shocks— exhibiting the most pronounced liquidity constraints. Table VII investigates this possibility. In rows 1 and 2, w e split the sam ple into two, roughly equal-sized "oil-sensitive" and "non-oil sensitive" subsam ples. W e assign to the form er category firm s in those industries which had the highest ratios o f 1980 energy cost as a percent o f shipm ents. In rows 3 and 4, we m ake an analogous division o f the sam ple into "export-sensitive" and "non-export-sensitive" subsam ples, this tim e using industry ratios o f exports to shipm ents to assign firm s to the respective categories. As in T able V I, the results yield little support for the collateral shock hypothesis. Indeed, non-oil-sensitive firms show som ew hat greater liquidity constraints than their oilsensitive counterparts, though these differences are not statistically significant. As a check, w e also reproduced the analysis while splitting the sam ple into th ree, rather than tw o, subsam ples based on oil- and export-sensitivity. The results are qualitatively sim ilar to those 30 reported in Table V II. In particular, the largest LIQ coefficient is now found on the least oil-sensitive third o f the sample. VII. Conclusions The one set o f conclusions that em erges most clearly from o u r w ork is com pactly summarized in row 4 o f Table V, and can be stated as follows: C ontrary to much previous research, our results dem onstrate that financial factors do indeed influence inventory movements. M oreover, financial constraints appear to be much m ore binding during tight m oney/recessionary episodes— apart from these episodes, there is little evidence o f inventories being sensitive to financial factors.1 8 A question that is harder to answ er quite as clearly is w hether the financial constraints w e docum ent represent an inward shift in bank loan s u p p ly -i.e ., the bank lending channel o f m onetary policy at w o rk -o r more generally, another type o f recession byproduct, such as a deterioration in collateral values. Our overall reading o f the data here leads us to favor the loan supply shift interpretation. However, we concede that there is probably not very strong evidence against the collateral story. In particular, a skeptic might argue that neither the results from the 1985-1986 slowdown nor from the disaggregated regressions o f Section V I constitute a com pelling rejection o f the collateral hypothesis. Although it is certainly interesting for some purposes to contrast these two hypotheses, it is also worth em phasizing their similarities. 31 U nder either interpretation, the inventory declines seen in recessions are partially due to a cu to ff in bank lending— the tw o hypotheses only disagree about the precise source o f this cutoff. T hus under either interpretation, w e have a financial account o f the cyclical behavior o f inventories, an account that differs sharply from that given in m ost previous w ork on the subject. Finally, w e are left w ith the follow ing question: ju st how econom ically im portant are the financial constraints that w e have identified? T he back-of-the-envelope calculations in Section IV-B suggest that these constraints m ight explain a substantial fraction o f inventory movements during dow nturns. H ow ever, w e recognize that it is difficult to draw precise structural conclusions o f this sort from the reduced-form regressions presented above. Thus w hile we feel quite confident in concluding that financial factors do m atter for inventories, w e are much less confident in assessing exactly how much they m atter.19 Gaining a better understanding o f these econom ic m agnitudes is an im portant topic fo r future research. 32 V III. Endnotes 1. Early w ork on the distinction between the "money" and "lending" channels o f m onetary policy transm ission includes M odigliani [1963], Tobin and Brainard [1963], Brainard [1964] and B runner and M eltzer [1964]. M ore recent contributions have com e from Bem anke and Blinder [1988, 1992], King [1986] and R om er and R om er [1990], am ong others. See Kashyap and Stein [1993] for a survey o f much o f this w ork. 2. B em anke and G ertler [1989] develop a model in which shocks to collateral am plify business cycle fluctuations in this fashion. 3. N ote that the Shadow Open M arket Com m ittee is not always critical o f F ed policy. F o r instance, in Septem ber o f 1982 they w rote (p. 3), "W e applaud the Federal R eserve’s com m itm ent and the success o f its policy to reduce in flation." 4. W e also exam ined a sample o f com panies whose fiscal years ended in the third rather than the fourth quarter, i.e. in either July, A ugust and Septem ber. T hese results w ere very sim ilar to the results reported below. W e also used a slightly different definition o f the fourth quarter -- taking firm s whose fiscal years ended in either N ovem ber, D ecem ber o r January. Again the results were essentially identical. 5. Because w e w ere w orried about survivorship bias w e did not use a balanced panel that contained only continuously listed com panies. Indeed, roughly 42 percent o f the com panies in the 1982 sam ple no longer exist and had to be retrieved from the C om pustat research tape. 6. A lternatively, this type o f specification could be motivated by appealing to a cost m inim ization m odel that assumes that firm s face quadratic costs o f producing output and o f deviating from a target inventory-sales ratio. F or instance, K ashyap and W ilcox [1993] show that this sort o f setup gives rise to an error-correction equation fo r inventories that is sim ilar to ours. 7. All standard errors are calculated using W hite’s procedure to correct fo r heteroskedasticity. 8. A lternatively, an endogeneity problem w ith regard to LIQ could arise if firm s planning to increase inventories set aside the cash to do so several months in advance. It should be noted that not all possible endogeneity problem s lead to an upw ard bias in the LIQ coefficient. F o r instance, if those firm s that anticipate having the m ost severe liquidity constraints attem pt to offset them by stockpiling m ore cash, the estim ated LIQ coefficient w ill be pushed tow ard zero. 33 9. A sim ilar logic is invoked by Fazzari, H ubbard and Petersen [1988] and H oshi, Kashyap and Scharfstein [1991] in their analyses o f fixed investm ent. F or exam ple, o u r notion that LIQ should m atter more for the inventories o f firm s w ithout access to public bond markets is analogous to the insight o f H oshi, K ashyap and Scharfstein [1991] that liquidity should be m ore im portant for explaining the investm ent behavior o f Japanese firm s w ithout close ties to an industrial group. H ow ever, there is one key difference between our w ork and these others: w e will also be looking for tim e variation in the LIQ coefficient for firm s w ithout bond m arket access. That is, w e expect this coefficient to be larger during periods o f tight money than at other times. In contrast, the above studies assum e tim e-invariant coefficients. W e discuss these differences in further detail below . 10. To be conservative, we also tried instrum enting with tw ice-lagged LIQ (i.e ., LIQ as o f 1979:4) for the 760 B = 0 firms for which this was available. T he estimated coefficient in this case was 0 .3 6 , with a t-stat o f 1.93. 11. Some care should be taken in interpreting these num bers. Even if inventories w ouldn’t have declined at all without a loan supply effect, this does not im ply that loan supply com pletely "explains" the behavior o f inventories in 1982. In a norm al year, inventories do not stay fla t-ra th e r, they tend to grow . F o r exam ple, in 1981, econom y-w ide inventories rose by $24.6 billion (in 1987 dollars). In 1982 there was a fall o f $17.5 billion. Thus even if loan supply effects can account for a large part o f the decline in 1982, they would still explain less than half o f the abnorm al m ovem ent relative to the previous year. A sim ilar logic im plies that the stylized fact introduced above— that reductions in inventory investment account for 87 percent o f the G NP drop in recessions— should also be interpreted with care. Fluctuations in inventory investm ent are a sm aller part o f the abnorm al m ovem ent o f GNP relative to its normal (increasing) growth path. 12. A distinct question is w hether size might b e a better m easure o f public m arket access than w hether o r not a firm has a bond rating. O bviously, as shown in Table I, large firm s are much m ore likely to have bond ratings, so there is a good deal o f overlap between the two measures. As a practical m atter, both probably capture m arket access in a noisy way, and it may be difficult to disentangle the two. N evertheless, there are a couple o f reasons why we view a bond rating as a slightly better indicator. F irst, it m ore directly addresses the notion o f access to non-bank financing. Second, in the lim ited checking that w e have done, larger firm s w ithout bond ratings seem to behave m ore like sm aller non-rated com panies than like rated com panies. F or instance, even for the larger h alf o f the B = 0 sam ple (i.e ., firm s with assets above the median value o f assets), the coefficient on LIQ in the basic specification (line 7, Table II) is 0.24, with a t-statistic o f 1.82. 13. W e also checked to see whether our results w ere drive by a handful o f outlier firm s. O m itting extrem e values o f the left hand side variable changes the coefficients som ew hat, but does not alter any o f the im portant conclusions. 34 14. We will return to discuss the cross-sectional variation in the LIQ coefficient in much m ore detail in Section VI. 15. Prior to 1982, we were unable to obtain bond rating data from Compustat. Rather than laboriously collecting the data by hand, w e approxim ated the B = 0 subsample in each o f the earlier years as the smallest 85 percent o f the sample by assets. (Recall that in 1982, firm s without bond ratings com prised approxim ately 86 percent o f the sam ple.) F or the years in which we could do it either w ay— ., 1982 and later— i.e using size in this way rather than bond rating to split the sample yielded virtually identical results for the "B = 0 " category. 16. We obtain sim ilar results when we look at how the differential between the LIQ coefficients for B = 0 and B = 1 firm s varies over time. This differential averages -0.03 over the entire 1974-1989 period, but reaches 0 .36 in 1974 and (as seen earlier) 0.62 in 1982. T he differential is also slightly negative in 1985 and 1986. 17. W e also tried re-estim ating Equation 4 o f Table V over the shorter period 1982-1989 using fixed effects and N ewey-W est standard errors, with a balanced panel consisting o f the 196 firms which passed our screens for each year. W e reached a qualitatively similar conclusion: the LIQ* 1982 coefficient is 0.68 with a t-statistic o f 2 .0 1 . 18. These conclusions com plem ent those drawn by G ertler and G ilchrist [1993] from their study o f the QFR data, which allows them to exam ine the com posite balance sheets o f "small" and "large" m anufacturing firms. They find that the inventories o f small firm s decline significantly m ore sharply in response to tight m onetary policy than do the inventories o f large firm s. 19. See Kashyap and Stein [1993] for a m ore detailed discussion o f the issues involved in evaluating the m agnitude o f the bank lending channel o f m onetary policy. 35 IX . References Bem anke, Ben S ., and Alan S. Blinder. "C redit, M oney, and A ggregate D em and," American Econom ic Review. Papers and Proceedings. LX X VIII (1988), 435-39. Bem anke, Ben S ., and Alan S. Blinder. "The Federal Funds Rate and the C hannels o f M onetary T ransm ission," American Econom ic R eview . LXXXII (1992), 901-21. Bemanke,. Ben S ., and M ark G ertler. "Agency Costs, N et W orth, and Business Fluctuations," A m erican Econom ic Review LXXIX (1989), 14-31. Blinder, Alan S ., and Louis J. M accini. "Taking Stock: A C ritical A ssessm ent o f Recent Research on Inventories," Journal o f Econom ic Perspectives. V (1991), 73-96. Brainard, W illiam C. "Financial Interm ediaries and a Theory o f M onetary C ontrol," Yale Econom ic Essays IV (1964), 431-482. Brunner, Karl and Allan H. M eltzer. "The Place o f Financial Interm ediaries in the Transm ission o f M onetary Policy," American Econom ic Review LIX (1964), 372-382. D om busch, Rudiger and Stanley Fischer. 1990). M acroeconom ics (New York: M cG raw -H ill, Eckstein, Otto, and Allen Sinai. "The M echanisms o f the Business C ycle in the Postw ar E ra," in The A m erican Business Cycle: Continuity and C hange. R obert J. G ordon, ed. (Chicago: U niversity o f Chicago Press, 1986), 39-122. Fazzari, Steven M ., R. Glenn H ubbard, and Bruce C . Petersen. "Financing C onstraints and Corporate Investm ent," Brookings Papers on Economic Activity (1988), 141-195. Friedm an, M ilton. Bright Prom ises. Dismal Perform ance. New Y ork: H arcourt, Brace, Jovanovich (1983). G ertler, M ark, and Simon G ilchrist. "M onetary Policy, Business C ycles and the Behavior o f Small M anufacturing F irm s," mimeo (1993). G ertler, M ark, and R. Glenn H ubbard. "Financial Factors in Business Fluctuations," Financial M arket V olatility." Kansas City: Federal Reserve Bank o f Kansas C ity (1988). G oodfriend, M arvin. "Interest Rate Policy and the Inflation Scare Problem : 1979-1992," Econom ic Q uarterly. LXXIX, Federal Reserve Bank o f Richmond (1993), 1-24. 36 Hoshi, Takeo, A nil Kashyap, and David Scharfstein, "Corporate Structure, Liquidity and Investment: Evidence from Japanese Industrial Groups," Quarterly Journal o f Economics. C V I (1991), 33-60. H oshi, Takeo, D avid Scharfstein and Kenneth J. Singleton. "Japanese C orporate Investm ent and Bank o f Japan G uidance o f Com m ercial B ank L ending," in Japanese M onetary Policy edited by Kenneth Singleton. Chicago: U niversity o f Chicago Press (1993), 63-94. Jaffee, D w ight, and Thom as Russell. "Im perfect Inform ation, U ncertainty, and C redit R ationing," Q uarterly Journal o f E conom ics. X C (1976), 651-66. K ashyap, A nil K ., Jerem y C. Stein and D avid W . W ilcox. "M onetary Policy and C redit Conditions: Evidence from the Com position o f External F inance", A m erican Econom ic R eview . LX X XIII (1993), 78-98. Kashyap, Anil K. and Jerem y C. Stein. "M onetary Policy and Bank Lending" in M onetary P olicy, edited by N. Gregory M ankiw . Chicago: U niversity o f Chicago Press (1993). K ashyap, Anil K. and David W . W ilcox, "Production and Inventory Control at the General M otors Corporation D uring the 1920’s and 1930’s," A m erican Econom ic Review LX X XIII (1993), 383-401. K ing, Stephen R. "M onetary Transm ission: T hrough Bank Loans o r Bank Liabilities?," Journal o f M oney. Credit and Banking XVIII (1986), 290-303. Lovell, M icheal C. "M anufacturers’ Inventories, Sales Expectations, and the Acceleration P rinciple," E conom etrica. XXIX (1961), 293-314. M odigliani, F ranco. "The M onetary M echanism and Its Interaction with Real Phenom ena," Review o f Econom ics and Statistics. (1963), 79-107. R om er, C hristina D ., and David H . R om er. "N ew Evidence on the M onetary Transm ission M echanism ," Brookings Papers on Econom ic A ctivity. (1990), 149-213. Shadow Open M arket Com m ittee, "Policy Statem ent and Position Papers: Septem ber 12-13, 1982," G raduate School o f M anagem ent, U niversity o f R ochester, W orking Paper PPS-82-2. Shadow Open M arket Com m ittee, "Policy Statem ent and Position Papers: Septem ber 21-22, 1986," G raduate School o f M anagem ent, U niversity o f R ochester, W orking Paper PPS-86-6. Stiglitz, Joseph, and Andrew W eiss. "C redit Rationing in M arkets with Im perfect Inform ation," American Econom ic R eview . LX X I (1981), 393-410. 37 Tobin, Jam es, and B rainard, W illiam . "Financial Interm ediaries and the Effectiveness o f M onetary C ontrol," American Econom ic R eview . LIII (1963), 383-400. 38 Table I Summary Statistics for Selected Periods 1982 Sam ple 1985 Sample N um ber o f Firm s Compustat sam ple w ithout bond rating with bond rating 933 802 131 841 698 143 M edian Assets beginning o f period (nominal $) Compustat sam ple w ithout bond rating with bond rating $44.2 M $31.8 M $1573.2 M $44.2 M $29.6 M $1325.2 M M edian Liquid Assets to Total Assets 5.00% 5.37% 3.51% 6.84% 7.65% 4.65% 5.05% 3.08% Com pustat sam ple w ithout bond rating with bond rating -8.06% -6.88% -11.84% -3.03% -1.67% -6.33% Q FR sam ple % change in sales -9.21% -3.42% -10.91% -10.20% -14.32% -4.98% -4.11% -11.88% -7.92% -4.98% Com pustat sam ple w ithout bond rating with bond rating Q FR sample average liquid assets to total assets M edian % Change in Sales (real $) M edian % Change in Inventories (real $) Com pustat sam ple w ithout bond rating with bond rating Q FR sample % change in inventories Notes: The Compustat sample was s l c e following t e procedure described i the t x . eetd h n et The "QFR" sample comes from the Quarterly Financial Report f r Manufacturing Corporations. o Real s l s and inventory changes c l ae a culated using the consumer price index. Table I I Baseline Spec f c t o s 1981:4- 1982:4: iiain, ALOG(INV) v . LOG(INV/SALES); ALOG(SALES); ALOG(SALES).,; LIQ; LIQ*B, Industry Controlss ( - t t s i s i parentheses) tsaitc n SAMPLE LOGAN V/SALES'! A. ALOG(SALES) ALOG(SALES). LIQ L1Q*B S! N OLS Regressions 1 All Firms . -0 .1 (•5 -4) .59 (7.88) .13 (2.57) .38 (3.52) _ .36 933 2 All Firms . -0 .1 (.5 -4) .58 (7.81) .13 (2.56) .39 (3.57) -3 .1 (-1. 9 6) .36 933 3 B = 0 . -0 .1 (.6 -3) .58 (7.56) .14 (2.58) .1 4 (3.63) .36 802 4 B = 1 . -1 .1 (2 - .80) .76 (4.90) -.18 (.2 -8) -.28 (-1.62) .46 131 .36 933 .36 933 .36 802 .46 131 B. _ IV Reeressionsfwith LIO,as instrument f r LIO) o 5. All Firms -0 .1 (.3 -4) .58 (7.85) .13 (2.57) .39 (2.67) All Firms -0 .1 (.4 -4) .58 (7.79) .13 (2.56) .39 (2.70) 7. B = 0 -0 .1 (.4 -3) .58 (7.55) .14 (2.59) .1 4 (2.70) 8 B = 1 . -1 .1 (-2.75) .76 (4.88) -.18 (.1 -8) -2 .1 (-1.06) 6. _ -.28 (1 - .45) — *A11 regressions use White’ robust e r r . Industry controls are dummy variables corresponding t 2 d g t SIC codes. s ros o -ii Table III Alternative Specifications of Inventory Equations, 1981:4-1982:4 ( - t t s i s i parentheses; a l equations estimated using IV) tsaitc n l SAMPLE LOGfINV/SALES’ ! ALOGfSALESl ALQGCSALESi. LIO ALOG(INV). LOGfASSETSl Rf N A. No Logflnv/Salesl on Right-Hand Side 1 . 2 . B = 1 .57 (7.22) B = 0 — .14 (2.88) .1 4 (2.75) .36 802 .76 (4.49) -.14 (.6 -6) -.09 (.3 -4) .42 131 ... .33 (2.36) B. Dependent Variable i ALocflnv/Sales) s 3 . B = 0 4. B = 1 ... ... — .04 802 .17 131 -1 .1 (.0 -5) C. Instrumenting for ALop(Sales) with ALoe(Inv), 5 . B = 0 -.02 (.0 -7) .67 (3.03) 6 . B = 1 -.12 (-2.58) 1.23 (2.52) .13 (2.59) -.34 (1 - .05) .37 (2.42) — — .35 802 .34 131 -.19 (.0 -9) D. ALogdnvl,on Right-Hand Side .40 (2.65) .03 (4) .7 -.25 (-1.01) -.17 (.3 -8) .13 (9) .7 7 . B = 0 -.02 (.4 ‘ -5) .57 (7.62) .12 (2.14) 8 . B = 1 -.12 (-2.75) .74 (5.23) — .36 802 .46 131 E. LocfAssetsl on Right Hand Side 9 . B = 0 -0 .1 (•4 -3) .58 (7.57) .14 (2.61) .1 4 (2.61) -- -.00 (.5 -2) .36 802 1. 0 B = 1 -1 .1 (-2.80) .78 (5.05) -.19 (.6 -8) -.24 (-1.22) — -0 .1 (.5 -7) .46 131 Table IV Disaggregated Results, B=0 firms, 1981:4-1982:4 ( - t t s i s i parentheses; a l equations estimated using I.V.)” tsaitc n l * SIC # and C a s f c t o lsiiain LIO R? M 35 machinery, ex e e t i a . lcrcl .33 (1.25) .45 126 36 e e t i a and elect o i lcrcl rnc equip. .45 (1.26) .57 117 38 instruments and r l t d eae equip. 1.07 (2.89) .32 77 28 chemicals and a l e lid prod. -8 .1 (1 - .18) .48 73 34 f b i a e metal arctd prod. -. 7 3 (.0 -5) .29 62 37 transportation equip. 11 .1 (7) .9 .16 50 33 primary metal i d s r e nutis .24 (7) .7 .22 50 30 rubber and misc. p a t c lsi prod. .54 (1.18) .26 48 20 food and kindred prod. -0.11 (0 - .30) .22 43 26 paper and a l e prod. lid -.07 (.6 -3) .42 40 27 printing and publishing .84 (1.94) .56 40 23 apparel and r l t d prod. eae .56 (1.77) .43 35 32 s o e clay and g ass tn, l .70 (5) .9 .25 35 22 t x i e m prod. e t l ill .74 (6) .0 .42 34 *Specification i i e t c l t t a used i Table I, row 7 s dnia o ht n I . Table V Comparison of Results for 1982 v . Other Years s ( - t t s i s i parentheses, a l equations estimated using I V ) tsaitc n l ..* SAMPLE LOGflNV/SALESl ALOGfSALESl ALOGCSALES1. LIO 1 1982 . (B=0 firms only) -0 .1 (.4 -3) .58 (7.55) .14 (2.57) .1 4 (2.70) 2 1974 . (B=0 firms only) -1 .1 (4 - .43) .58 (10.62) .00 (0) .7 -.20 (-6.64) .57 (6.69) .14 (3.02) 4. 1974-1989 (B=0 firms only) -.14 (-14.01) .61 (19.34) .09 (4.04) _ -.02 (.8 -1) .07 (1.42) R? N .36 802 • .32 .1 7 (3.89) 3 1985-1986 . (B=0 firms only) LIO*CCR 877 .34 .36 (2.78) 1364 .32 13,203 ’Specifications are analogous t those i Table I, row 7 except t a they allow for d f e e t i t r e t terms fo each y a . • ‘ o n I , ht ifrn necp r er CCR = 1 f r 1982 and 1974, 0 otherwise. o Table VI Relationship Between LIQ Coefficient and Debt-to-Assets Ratio for B=0 Firms i 1982 n ( - t t s i s i parentheses, a l equations estimated using I V ) tsaitc n l ..* SAMPLE LOG IINV/SALES) ALOG SALES ALOGfSALES). LIO LIO*DRATIO DRATIO E! N Sample Mediar. D/TC A. Separate Regressions fo Low, Medium and High DRATIO Firms r 1 Firms with DRATIO . i Bottom Third of n Sample -.02 (.4 -4) .76 (6.14) .19 (1.34) .26 (1.26) 2 Firms with DRATIO . i Middle Third of n Sample -0 .1 (.3 -1) .50 (5.22) .17 (1.58) .62 (1.71) 3 Firms with DRATIO . i Upper Third of n Sample -0 .1 (.8 -1) .56 (4.87) .47 253 „ .22 (5) .6 .14 (2.62) .52 (2.45) -.45 (.3 -4) .1 4 (2.61) .17 (0.27) .33 254 .22 .39 252 •• .12 (1.81) .05 .42 .37 759 .22 .37 759 .22 B. Continuous I t r c i n between LIO and DRATIO neato 4 Full . Sample -0 .1 (.3 -4) .59 (7.55) 5. Full Sample -0 .1 (.4 -3) .59 (7.08) .14 (2.55) .12 (9) .4 Table VII Relationship Between LIQ Coefficient and O and Export- S n i i i y ilestvt f r B=0 Firms i 1982 o n ( - t t s i s i parentheses, a l equations estimated using I V ) tsaitc n l ..* SAMPLE LOGdNV/SALES^ 1 . Oil-Sensitive 2 . -.04 (.8 -7) ALOG(SALES) ALOGfSALESl, LIQ E! N .55 (5.07) .32 (3.13) .25 (1.08) .35 468 Non-Oil-Sensitive - 0 .1 (.2 -2) .9 5 (7.67) .08 (2.68) .47 (2.98) .40 334 3 . Export-Sensitive -.04 (.6 -7) .66 (7.27) .12 (2.07) .50 (2.33) .45 424 4 . Non-ExportSensitive .1 0 (2) .2 .42 (3.02) .25 (1.45) .19 378 .3 1 (1.57) Note: Shaded areas denote recessions as i e t f e by dniid Figure I Monetary Policy I d c t r , 1972-1989 niaos A Federal F u n d s Rate C. R e a l M 2 t e NBER. h Figure I I Movements i G D P and Inventories, 1972-1989 n A. C h a n g e in R e a l G D P Percent B illio n s o f 1 9 8 7 D o lla r s B. C h a n g e in M a n u f a c t u r e r s I n v e n t o r i e s B illio n s o f 1 9 8 7 D o lla r s Percent 50