WHEN DOES THE MARKET MATTER? STOCK PRICES AND THE INVESTMENT … Files/StockPrices... · 2019. 11....

WHEN DOES THE MARKET MATTER?STOCK PRICES AND THE INVESTMENT OF

EQUITY-DEPENDENT FIRMS*

MALCOLM BAKER

JEREMY C. STEIN

JEFFREY WURGLER

We use a simple model to outline the conditions under which corporateinvestment is sensitive to nonfundamental movements in stock prices. The keyprediction is that stock prices have a stronger impact on the investment of“equity-dependent” �rms—�rms that need external equity to �nance marginalinvestments. Using an index of equity dependence based on the work of Kaplanand Zingales, we �nd support for this hypothesis. In particular, �rms that rank inthe top quintile of the KZ index have investment that is almost three times assensitive to stock prices as �rms in the bottom quintile.

I. INTRODUCTION

Corporate investment and the stock market are positivelycorrelated, in both the time series and the cross section. Thetraditional explanation for this relationship is that stock pricesre�ect the marginal product of capital. This is the interpretationgiven to the relationship between investment and Tobin’s Q, forexample, as in Tobin [1969] and von Furstenberg [1977].

Keynes [1936] suggests a very different explanation. He arguesthat stock prices contain an important element of irrationality. As aresult, the effective cost of external equity sometimes diverges fromthe cost of other forms of capital. This affects the pattern of equityissues and in turn corporate investment. This “equity �nancingchannel” has been further developed by Bosworth [1975], Fischerand Merton [1984], Morck, Shleifer, and Vishny [1990], Blanchard,Rhee, and Summers [1993], and Stein [1996].

It has proved dif�cult to determine the relative merits of theseexplanations. This is partly because the equity �nancing channelhas not been articulated in a form that can be sharply distinguished

* We thank Eric Nierenberg for research assistance. We also thank HeitorAlmeida, Nicholas Barberis, Steven Kaplan, Eli Ofek, Lasse Pedersen, DavidScharfstein, Andrei Shleifer, Tuomo Vuolteenaho, seminar participants at Har-vard University, the Kellogg/Zell Center Conference on Risk Perceptions andCapital Markets, New York University, Princeton University, the Tuck School ofBusiness at Dartmouth College, the University of Utah, the World Bank, and theYale School of Management, two anonymous referees, and the editors for helpfulcomments. Financial support from the Division of Research of the Harvard Grad-uate School of Business Administration (Baker) and from the National ScienceFoundation (Stein) is gratefully acknowledged.

© 2003 by the President and Fellows of Harvard College and the Massachusetts Institute ofTechnology.The Quarterly Journal of Economics, August 2003

969

from the traditional view. Tests have had to focus on indirect impli-cations of the two views, or else have had to impose structuralassumptions on the data. For example, researchers taking theformer approach have examined whether the stock market forecastsinvestment over and above other measures of the marginal productof capital, such as pro�tability or cash �ow. If it does not, they argue,then the stock market is probably connected to investment onlyinsofar as it re�ects fundamentals.

This empirical strategy has yielded mixed results. Barro[1990, p. 130] attributes an important independent role to thestock market: “Even in the presence of cash �ow variables, suchas contemporaneous and lagged values of after-tax corporatepro�ts, the stock market variable retains signi�cant predictivepower for investment.” In contrast, Morck, Shleifer, and Vishny[1990, p. 199] conclude from their analysis of �rm-level data that“the market may not be a complete sideshow, but nor is it verycentral,” and Blanchard, Rhee, and Summers [1993, p. 132] sum-marize their study of the aggregate data by stating that “marketvaluation appears to play a limited role, given fundamentals, inthe determination of investment decisions.”

Another empirical strategy is to try to measure inef�ciencydirectly as the difference between market prices and a structuralmodel of ef�cient prices, and then test whether investment is sen-sitive to this measure of inef�ciency. In a study of aggregate Japa-nese data, Chirinko and Schaller [2001] �nd evidence for an equity�nancing channel using this approach. As they point out, however,their conclusions depend on several structural assumptions.

In this paper we take a new approach. We return to the theoryto derive several cross-sectional predictions that are unique to aspeci�c equity �nancing channel. In particular, the model in Stein[1996] implies that those �rms that are in need of external equity�nance will have investment that is especially sensitive to the non-fundamental component of stock prices. Intuitively, a �rm with nodebt and a stockpile of cash can insulate its investment decisionsfrom irrational gyrations in its stock price. But an “equity-depen-dent” �rm that needs equity to fund its marginal investments will beless likely to proceed if it has to issue undervalued shares.1

1. A recent working paper by Polk and Sapienza [2002] also explores the linkbetween stock-market inef�ciency and corporate investment. However, theirmodel assumes that all �rms are �nancially unconstrained, and so they focus ona different set of empirical implications.

970 QUARTERLY JOURNAL OF ECONOMICS

We test several implications of this �nancing channel. To getstarted, we need a proxy for the concept of equity dependence.This concept requires some �nancing friction, or combination offrictions, which makes certain �rms more reliant on outside eq-uity at the margin. Standard corporate-�nance considerationssuggest that equity-dependent �rms will tend to be young, and tohave high leverage, low cash balances, and cash �ows, high cash-�ow volatility (and hence low incremental debt capacity), andstrong investment opportunities.2 One off-the-shelf measurewhich satis�es most of these criteria is an index based on thework of Kaplan and Zingales [1997]. This “KZ index” has alreadybeen adapted for use in large-sample empirical work by Lamont,Polk, and Saa-Requejo [2001], so we can follow their methodology.By taking this approach, as opposed to building our own measureof equity dependence from scratch, we hope to minimize anyconcerns about data mining.

Next, we rank �rms according to this proxy for equity depen-dence, and test in a variety of ways whether those that are classi�edas most likely to be equity dependent have the strongest correlationbetween stock prices (as measured by Q) and subsequent invest-ment. We �nd strong support for this prediction. In our baselinespeci�cation, �rms that rank in the top quintile of the sample interms of the KZ index have a sensitivity of investment to Q that isalmost three times as large as �rms that rank in the bottom quin-tile.3 In fact, in some speci�cations the investment of equity-depen-dent �rms is more sensitive to Q than to cash �ow. This is notewor-thy because it is generally believed that the cash-�ow effect domi-nates the Q effect in investment equations.

While the heightened Q-sensitivity of equity-dependent�rms is consistent with the equity �nancing channel in ourmodel, it also admits other interpretations. The ambiguity arisesbecause the theory predicts that the investment of equity-depen-

2. Morck, Shleifer, and Vishny [1990] look for a �nancing channel using �rmsize as their only proxy for equity dependence. They argue that the “hypothesispredicts that the in�uence of the stock market should be particularly great forsmaller �rms, which rely to a greater extent on external �nancing” [p. 182].However, they �nd little evidence that the investment of smaller �rms is espe-cially sensitive to stock prices, and conclude that there is no support for thehypothesis. As we show, the use of a more fully developed measure of equitydependence leads to quite different conclusions.

3. The result that investment is more sensitive to Q for high-KZ �rmsactually shows up in the small Kaplan and Zingales [1997] sample. However, theirfocus is on a different question—how investment-cash �ow sensitivities vary with�nancial constraints—and they never discuss or interpret this particular �nding.

971WHEN DOES THE MARKET MATTER?

dent �rms should be more sensitive to the nonfundamental com-ponent of stock prices. But Q contains more than just this non-fundamental component. It also embodies information about fu-ture pro�tability—and hence about the quality of investmentopportunities—as well as measurement error arising from ac-counting discrepancies between book capital and economic re-placement costs. As we describe in detail below (see subsectionII.D), one can tell stories in which these other components of Qcovary with our measure of equity dependence in such a way as toinduce the patterns we �nd in the investment-Q regressions.

Ideally, in order to provide a more focused test of the theorythat gets around these problems, we would like to isolate thenonfundamental component of stock prices, and verify that theinvestment of equity-dependent �rms responds more sensitivelyto this component. This is a dif�cult task, but we try to tackle itby using future realized stock returns—speci�cally, returns overthe three years subsequent to the year in which we measureinvestment. The idea is that future realized returns are a noisyestimate of the future returns expected by managers, which inturn include their views about over- or undervaluation. As aresult, in the same way that the investment-Q sensitivity ispredicted to be positive on average and increasing in equitydependence, the investment-future returns sensitivity should benegative on average and increasingly negative in equity depen-dence. We con�rm this prediction.

Finally, we also look brie�y at �nancing patterns. The modelimplies that equity-dependent �rms should have a pronouncedcorrelation between the nonfundamental component of stockprices and the volume of new equity issues. We �nd support forthis hypothesis as well: �rms with high values of the KZ indexalso have equity issuance that responds positively to Q and nega-tively to future returns.

Overall, our results offer support for a speci�c equity �nanc-ing channel in corporate investment. They also complement otherevidence that the cost of external equity has an important, inde-pendent effect on corporate �nancing and investment decisions.For example, Ritter [1991], Ikenberry, Lakonishok, and Ver-maelen [1995], Loughran and Ritter [1995], Speiss and Af�eck-Graves [1995], and Baker and Wurgler [2000] �nd evidence thatequity �nancing patterns depend on the cost of equity, and Bakerand Wurgler [2002] use these results to motivate an alternativeview of capital structure. Shleifer and Vishny [forthcoming] argue


that the cost of equity is a strong determinant of merger activity,explaining the form of �nancing in mergers as well as mergerwaves themselves. Whereas the capital structure �ndings couldbe viewed as �nancial phenomena without signi�cant real effects,however, our results suggest a speci�c mechanism through whichmarket inef�ciency may affect the real economy.

The remainder of the paper is organized as follows. In SectionII we develop several testable hypotheses in the context of asimple model. The model gives some guidance as to how to mea-sure equity dependence in practice, and also provides a frame-work for thinking about competing explanations. In Section IIIwe describe the data, and in Section IV we present the empiricalresults. Section V concludes.

II. HYPOTHESIS DEVELOPMENT

II.A. A Simple Model

We use a simpli�ed version of the model in Stein [1996] todevelop several testable hypotheses about equity dependenceand investment. There is a �rm that can invest K at time 0,which yields a gross return of f(K) at time 1, where f[ is anincreasing, concave function. The ef�cient-market discount rateis r, so the net present value of this investment is f(K)/(1 1 r) 2K. The �rst-best level of investment Kfb is therefore given byf9(Kfb)/(1 1 r) 5 1.

The �rm also has �nancing considerations. Its equity may bemispriced by a percentage d relative to the ef�cient-market value,either overpriced (d . 0) or underpriced (d , 0), while its debt isfairly priced.4 The �rm can issue equity e subject to the con-straints that 0 # e # emax. In other words, it cannot repurchaseequity, and there is an upper bound on how much it can issue.Financing and investment are linked by a leverage constraint,e 1 W 2 K(1 2 D# ) $ 0, where W is the �rm’s preexisting wealth(such as cash on hand, or untapped debt capacity) and D# is the

4. One interpretation, suggested by a referee, is that the debt and equitymarkets are segmented, and have expected-return premiums that vary indepen-dently. This would also fall under our de�nition of mispricing, as it implies thatthe Modigliani-Miller theorem is violated in a way that creates a motive for activemarket timing.


fractional debt capacity of the new assets. This constraint impliesthat the �rm’s debt ratio can fall below D# but cannot exceed it.5

Putting all this together, the �rm’s optimization problem isgiven by

(1) maxe,K

f~K!

1 1 r2 K 1 de,

subject to

(2) e 1 W 2 K~1 2 D# ! $ 0

and

(3) 0 # e # emax.

Proposition 1 summarizes the solution to this problem.

PROPOSITION 1. Assume that emax . Kfb(1 2 D# ). Then the possibleoutcomes are as follows.

i) If d . 0, then K 5 Kfb and e 5 emax: an overvalued �rminvests at the �rst-best level and issues as much equityas possible.

ii) If d , 0 and W 2 Kfb(1 2 D# ) $ 0, then K 5 Kfb , and e 50: an undervalued �rm with suf�cient wealth W investsat the �rst-best level and avoids issuing equity.

iii) If d , 0 and W 2 Kfb(1 2 D# ) , 0, then K , Kfb : anundervalued �rm with insuf�cient wealth underinvests.This case admits two subcases.a) De�ne Kec by f9(Ke c)/(1 1 r) 5 1 2 d(1 2 D# ). If

W 2 Kec(1 2 D# ) , 0, it follows that K 5 Ke c and e 5Kec(1 2 D# ) 2 W . 0: the �rm issues equity, and bothinvestment and the size of the equity issue are func-tions of the degree of undervaluation d and debt ca-pacity D# .

b) If W 2 Kec(1 2 D# ) $ 0, then K 5 W/(1 2 D# ), and e 50: the �rm does not issue equity and invests as muchas it can subject to its wealth W and the leverageconstraint.

5. These constraints simplify the exposition, but can be dispensed with in afuller model. For example, both repurchases and equity issues could be boundedendogenously by assuming that there are price-pressure effects that increase withthe size of the repurchase or issue. The simple form of the leverage constraint canalso be generalized by having costs of �nancial distress that increase continuouslywhenever the debt ratio exceeds D# . These generalizations are considered in Stein[1996].


The proposition, which is illustrated in Figure I, makes itclear when investment depends on the nonfundamental compo-nent of stock prices d. This happens only in region (iii.a), wheretwo necessary conditions are satis�ed: the stock is undervalued,and available wealth is so low that the �rm would have to issue

FIGURE IGraphical Illustration of Proposition 1

The horizontal axis measures the degree of stock-price misvaluation d. Thevertical axis shows investment K (in Panel A) and equity �nancing e (in Panel B).The square markers show the policy of an equity-dependent �rm. The solid line isthe policy of a nonequity-dependent �rm.


undervalued equity to invest at the �rst-best level.6 We thereforede�ne a �rm as “equity dependent” if W , Kfb(1 2 D# ). The basicmessage of Proposition 1 is that for equity-dependent �rms, mar-ket inef�ciency can act like a �nancial constraint, discouraginginvestment when stock prices are too low.

More precisely, for �rms in region (iii.a) the sensitivity ofinvestment to d is governed by (1 2 D# ). So for a given startingvalue of K, �rms with less debt capacity have investment thatreacts more strongly to stock prices. Intuitively, for a �rm facinga binding leverage constraint, the lower is D# , the more equitymust be issued for each marginal dollar of investment. Hence thestronger is the effect of stock prices on investment.

II.B. Testable Hypotheses

We use Proposition 1 to generate three empirical predictions.The �rst two are about investment behavior, and the third isabout �nancing. We develop these predictions in the context ofthe model, and discuss competing interpretations afterward.

HYPOTHESIS 1. De�ne a �rm as equity dependent if W ,Kfb(1 2 D# ). Equity-dependent �rms display a more positivesensitivity of investment to Q than do nonequity-dependent�rms.

In Hypothesis 1, Q is taken to be an empirical proxy for d, thenonfundamental component of the stock price. Hypothesis 1 istherefore a joint hypothesis about the sources of variation inmeasured Q and the validity of the equity dependence mecha-nism. Note also that Hypothesis 1 does not condition on whether�rms are over- or undervalued. It averages over the overvaluationregion (i), where equity dependence does not matter, and theundervaluation region (iii.a), where it does. (Graphically, thehypothesis averages over both sides of Figure I.) This is an ap-pealing feature of our empirical approach, because we do not have

6. The conclusion that investment is sensitive to stock prices only when d ,0 is a result of the one-sided nature of the leverage constraint. One could alsoimpose the constraint that an equity issue not cause leverage to fall below somecritical value, in which case investment may vary with stock prices even when d .0. The model of Shleifer and Vishny [forthcoming] implicitly embodies such afeature—they assume that overvalued �rms wishing to issue equity cannot simplypark the proceeds in T-bills, and must do something concrete, like acquiringanother �rm, to justify the issue. Hence overvaluation leads to more investmentin the form of mergers.


to take a stand on whether a given �rm is over- or undervalued inabsolute terms in order to test the model.7

Our second hypothesis uses future stock returns, instead ofcurrent Q, as a proxy for the nonfundamental component d instock prices. The intuition is that overpriced stocks have lowerexpected returns going forward, as mispricing is corrected, whileundervalued stocks have higher expected returns. This leads tothe following somewhat sharper prediction.

HYPOTHESIS 2. Equity-dependent �rms display a more negativesensitivity of investment to future stock returns than dononequity-dependent �rms.

If we were to look on average across all types of �rms, thereare a couple of reasons to expect the sensitivity of investment tofuture returns to be negative. Such a relationship might re�ecteither rational variation in the cost of capital, or the sort ofmispricing assumed in our model. Thus, if we want to generatemore speci�c support for the model, we need to focus on thecross-sectional prediction associated with equity dependence.

The main impediment to testing Hypothesis 2 is that realizedreturns are likely to be a very noisy proxy for expected returns,and hence for mispricing. For this reason, and because our maingoal is to understand the relationship between investment and Q,we focus much of our attention on Hypothesis 1. As we discussbelow, however, Hypothesis 2 is particularly useful in rulingout alternative explanations for our �ndings with respect toHypothesis 1.

Note that both hypotheses involve comparisons between�rms that are equity dependent and those that are not—that is,comparisons across the regions in Proposition 1. We can also askwhat happens within region (iii.a) as �rms become “more” equitydependent; i.e., as W falls farther and farther below Kfb(1 2 D# ).It turns out that a globally monotonic relationship between thedegree of equity dependence and the sensitivity of investment tostock prices only obtains if we put certain restrictions on the form

7. In our earlier NBER working paper [Baker, Stein, and Wurgler 2002] weprovide evidence consistent with the more precise prediction that investment ismost sensitive to stock prices for undervalued equity-dependent �rms. We use thelevel of Q as a proxy for over- or undervaluation, and �nd that Hypothesis 1 holdsmore strongly for low-Q than for high-Q �rms as predicted. However, this resultcould also re�ect the fact that there is more measurement error in Q for high-Q�rms, a possibility that is hard to refute. Because of this ambiguity, we focus hereon predictions that do not require us to identify the absolute level of mispricing.


of the production function.8 (This situation is reminiscent of thediscussion by Kaplan and Zingales [1997, 2000] and Fazzari,Hubbard, and Petersen [2000].) Nevertheless, the theory is test-able as long as we can plausibly identify some �rms that are notdependent on equity at all, because the theory unambiguouslypredicts that the sensitivity of investment to stock prices will riseover at least this �rst part of the range of measured equitydependence. Whether the sensitivity continues to increase overthe whole range is an empirical question, however, and not onefor which the theory leaves us with strong priors.

Finally, our model also makes predictions about �nancingbehavior, although they are not as crisp as those for investment.The reason is apparent from Figure I. As the mispricing d goesfrom very negative levels (extreme undervaluation) to positiveones (overvaluation), both types of �rms see their equity issuancee go from 0 to emax. Thus, averaging over the entire range of d’s,both types of �rms have equity issuance that is equally respon-sive to mispricing.9 Consequently, the strongest unconditionalstatement for �nancing patterns is the following.

HYPOTHESIS 3. Equity-dependent �rms have equity issuance thatis positively related to Q and negatively related to futurestock returns.

Hypothesis 3 does not have the same cross-sectional contentas the �rst two hypotheses. It is a prediction of the model, but itshould not carry too much inferential weight on its own. Rather,it should be viewed as a complement to the evidence that wedevelop for investment.10

8. De�ne the “degree” of equity dependence as F 5 K fb (1 2 D# ) 2 W, and thepercentage sensitivity of investment to stock prices as S 5 1/K z dK/dd. (Thismeasure of the sensitivity matches our empirical implementation, where we scaleinvestment by existing assets.) It is straightforward to show that a suf�cientcondition for dS/dF to be positive in region (iii.a)—and hence for the sensitivityS to become ever greater as W declines relative to Kfb(1 2 D# )—is that Kf-(K) 1f0(K) , 0. Among the functions that satisfy this condition are the quadratic, andanything of the form f(K) 5 K/(K 1 A), where A . 2K. Note also that in oursetup, S depends on F within region (iii.a) only through D and not through W,since dK/dW is zero in this region. Nevertheless, the suf�cient condition applieseven in a modi�ed version of the model where dK/dW . 0 in the region of interest.

9. Figure I shows that a sharper prediction emerges if we condition onundervaluation. Given undervaluation, the equity issuance of equity-dependent�rms is especially sensitive to the nonfundamental component of stock prices.Using the level of Q to proxy for over- and undervaluation, our NBER workingpaper �nds some support for this prediction (see also footnote 7).

10. In fact, we should not be surprised if Hypothesis 3 holds, because a largeliterature (cited in the Introduction) has already documented that, on average,


II.C. Competing Explanations

As discussed in the Introduction, differences in the sensitiv-ity of investment to Q can arise for reasons other than thoseemphasized in the model. Broadly speaking, ambiguities arisebecause Q potentially contains three sources of variation: (i)mispricing; (ii) information about the pro�tability of investment;and (iii) measurement error. Our theory is about the �rst of thesecomponents, so the other two can create problems for our infer-ences. Here we discuss these problems in more detail, and de-scribe how they might be addressed in our auxiliary tests.

It is possible that equity-dependent �rms could have agreater sensitivity of investment to Q even in an ef�cient marketwhere all variation in Q comes from variation in the pro�tabilityof investment. Perhaps most notably, the investment of equity-dependent �rms may be constrained by adverse-selection prob-lems in the market for new issues [Myers and Majluf 1984].Several authors have suggested that these problems are worsewhen �rms have weak investment prospects [Lucas and Mc-Donald 1990; Choe, Masulis, and Nanda 1993; Bayless and Chap-linsky 1996]. If so, equity-dependent �rms could have investmentthat is more sensitive to Q simply because a lower value of Qindicates more adverse selection in the new-issues market.

This explanation can be addressed with our tests of Hypothe-sis 2, which focus on the link between investment and futurestock returns. Given that the adverse-selection models are set inan ef�cient market, they do not share our prediction that thesensitivity of investment to future returns will be stronger forequity-dependent �rms.11

Another competing explanation is that different �rms havedifferent technologies, which imply inherently different sensitiv-ities of investment to future pro�tability.12 For example, if in-vestment is concave in Q and equity-dependent �rms tend to havelow Q values, then we would also observe equity-dependent �rmshaving a higher sensitivity of investment to Q. A straightforward

equity issuance is positively correlated with Q and negatively correlated withfuture returns. The hypothesis simply adds that these patterns should hold forequity-dependent �rms in particular.

11. In a similar vein, the tests involving future returns can help to separateour theory from one in which the market is in fact ef�cient, but managers believethat it is not, and mistakenly associate high stock prices with overvaluation. SeeJenter [2001] for a discussion of the overlap between these two stories.

12. We thank Nick Barberis for suggesting this possibility.


way to address this possibility is to control directly for nonlin-earities in the investment-Q sensitivity, by including Q2 in theinvestment equation. We can then ask whether our equity depen-dence measure still has a residual connection to this sensitivity.Our tests of Hypothesis 2 are also helpful in discriminatingagainst production-function explanations, since like the adverse-selection stories, they have no implications for future returns.

The effects predicted in Hypothesis 1 might also show up ifmeasurement error in Q were more pronounced for �rms that areless equity dependent. However, it is important to note that sucha pattern in measurement error is precisely the reverse of thatwhich has for many years been discussed in the literature onliquidity constraints. For example, in his discussion of Fazzari,Hubbard, and Petersen [1988], Poterba [1988] points out thattheir results “could be explained on this view because Q is mea-sured with more error for smaller �rms, which tend to be lower-dividend �rms” [p. 202]. Erickson and Whited [2000] and Alti[2003] further develop this point; the latter builds a formal modelto show why measurement error in Q is likely to be greater foryoung, faster-growing, low-dividend �rms. Clearly, if these argu-ments are correct, our approach stands on safe ground. The KZindex scores low-dividend, high-growth �rms as more likely to beequity dependent. If these attributes are associated with moremeasurement error in Q, our tests of Hypothesis 1 will be biasedtoward being excessively conservative.

Nevertheless, it is possible to concoct measurement-errorstories that go in the opposite direction. We address such storiesin two ways. The �rst is to “unpack” the KZ index. The de�nitionof equity dependence leads to speci�c predictions for how each ofthe components of the KZ index should affect the sensitivity ofinvestment to stock prices. If these predictions hold up, advocatesof the measurement-error explanation would need to explain whymeasurement error in Q should be positively correlated withcertain of the KZ components and negatively correlated withcertain others. Our further tests that look at how the sensitivityof investment to stock prices varies with �rm age and debt ca-pacity can be thought of in a similar spirit.

The tests of Hypothesis 2 also address this critique. Measure-ment error in Q is typically thought to arise from an inability toaccurately measure the replacement cost of capital. In our con-text, the concern is that nonequity-dependent �rms do moreintangible investment, which leads to more measurement error


and hence a bias in the sensitivity toward zero. Replacing Q withfuture returns in the investment equations eliminates this type ofmeasurement error.

III. DATA

We study a large, unbalanced panel of Compustat �rms thatcovers 1980 through 1999. The panel excludes �nancial �rms (i.e.,�rms with a one-digit SIC of six), and �rm-years with a bookvalue under $10 million, but otherwise it includes all observa-tions with data on investment, �nancing, equity dependence, andother investment determinants, as described below. The full sam-ple includes 52,101 observations, for an average of 2,605 obser-vations per year.

III.A. Investment

We consider four measures of investment. Our baseline mea-sure is CAPX it/Ait2 1 , the ratio of capital expenditures in year t(Compustat Annual Item 128) to start-of-year book assets (Item6).13 In addition, we look at (CAPX it 1 RDit)/Ait2 1 , which in-cludes research and development expenses (Item 46), and at(CAPXit 1 RDit 1 SGA it)/Ait2 1 , which further includes selling,general, and administrative expenses (Item 189). Finally, we alsoexamine the percentage change in book assets over the year,DA it/Ait2 1 . To reduce the in�uence of outliers, we Winsorize eachof these variables at the �rst and ninety-ninth percentile; i.e., weset all observations beyond these tolerances to the �rst and ninety-ninth percentile values, respectively.14 Panel A of Table I pre-sents summary statistics for these variables.

13. We scale our measures of investment and cash �ow by book assets. Thiscontrasts with some of the literature (e.g., Fazzari, Hubbard, and Petersen [1988]and Kaplan and Zingales [1997]), where the denominator is net plant, property,and equipment (PP&E). Our approach matches our sample, which includessmaller and nonmanufacturing �rms with modest �xed assets, and our alterna-tive measures of investment, which include intangible assets. Nevertheless, weshow in our robustness tests that scaling by PP&E leads to very similar results.

14. We have conducted a variety of tests to determine whether our particulartreatment of outliers makes any difference. As it turns out, all that matters is thatwe do something to tamp down the most extreme realizations of Q, which in theraw data attains a maximum value of 52.5. An alternative to Winsorizing is toreplace the book value of equity in Q with 0.9 times the book value plus 0.1 timesthe market value, thereby bounding the transformed value of Q below 10. Thisprocedure gives virtually identical results to those we report. We thank TuomoVuolteenaho for suggesting this procedure.


III.B. Financing

We consider two measures of external �nancing activity. Forequity issuance we use e it/A it2 1 , the ratio of external equityissues to start-of-year book assets. External equity issues areconstructed as the change in book equity minus the change inretained earnings (DItem 60 1 DItem 74 2 DItem 36). Totalexternal �nance is measured as (e it 1 d it)/Ait2 1, which includesboth equity and debt issues. Debt issues are constructed as thechange in assets minus the change in book equity (DItem 6 2DItem 60 2 DItem 74). These variables are also Winsorized at the�rst and ninety-ninth percentile.

Panel B of Table I presents summary statistics for the �nancingvariables. The mean values are sensitive to major �nancing eventssuch as acquisitions and divestitures, despite the Winsorization andthe restriction on minimum book assets. The medians look morefamiliar, and are stable across the 1980s and 1990s (not reported).

III.C. Equity Dependence

According to our theory, a �rm is more likely to be dependenton equity when W is low (which translates into low pro�tability,cash balances, or previously untapped debt capacity), when Kfb is

TABLE ISUMMARY STATISTICS FOR INVESTMENT AND FINANCING VARIABLES

Full sample Subsample means

N Mean SD Median Min Max 1980–1989 1990–1999

Panel A: InvestmentCAPXt/At2 1% 52,101 8.23 7.85 6.00 0.18 45.22 8.81 7.79

1RDt/At 2 1 52,101 11.41 10.46 8.45 0.24 58.43 10.96 11.761RDt

1 SGAt/At2 1 52,101 40.09 28.93 34.21 1.65 144.04 39.23 40.76DAt /At2 1% 52,101 11.10 28.03 6.41 243.76 153.73 10.30 11.73

Panel B: Financinget/At2 1% 52,101 4.38 14.48 0.64 215.90 92.28 3.00 5.44

1d t/At2 1 52,101 10.51 27.32 4.03 233.70 161.59 8.48 12.08

Source: Compustat, 1980–1999.a. Investment is alternately de�ned as capital expenditures (Item 128) over assets (Item 6); capital

expenditures plus research and development expenses (Item 46) over assets; capital expenditures plusresearch and developmentexpensesplus selling, general, and administrative expenses (Item 189) over assets;and growth in assets.

b. Financing is alternately de�ned as equity issues (DItem 60 1 DItem 74 2 DItem 36) (i.e., the changein book equity minus the change in retained earnings) over assets, and equity issues plus debt issues (DItem6 2 DItem 60 2 DItem 74) over assets.

c. All variables are Winsorized at the �rst and ninety-ninth percentiles.


high (growth opportunities are good), and when the incrementaldebt capacity of new assets D# is low.15 Therefore, a sensibleempirical measure of equity dependence should probably be nega-tively related to operating cash �ow, positively related to proxiesfor growth opportunities, positively related to actual leverage netof cash on hand, and negatively related to the debt capacity ofassets. Firm age may also be a factor, to the extent that young�rms without established reputations may have a harder timeraising bond-market �nance [Diamond 1991].

These observations motivate our interest in the work ofKaplan and Zingales [1997], who undertake an in-depth study ofthe �nancial constraints faced by a sample of 49 low-dividendmanufacturing �rms. Using both subjective and objective criteria,they rank these �rms on an ordinal scale, from least to mostobviously constrained. Most useful for our purposes, they thenestimate an ordered logit regression which relates their qualita-tive ranking to �ve Compustat variables. This regression at-taches positive weight to Q and leverage, and negative weight tooperating cash �ow, cash balances, and dividends. The parame-ters of this regression allow one to create a synthetic “KZ index”of �nancial constraints for a broader sample of �rms, as done inLamont, Polk, and Saa-Requejo [2001]. Following these authors,we construct the �ve-variable KZ index for each �rm-year as thefollowing linear combination:

(4) KZ it ~five-variable! 5 21.002CFit

Ait212 39.368

DIVit

Ait21

2 1.315Cit

Ait211 3.139LEVit 1 0.283Qit,

where CF it/Ait2 1 is cash �ow (Item 14 1 Item 18) over laggedassets; DIVit/A it2 1 is cash dividends (Item 21 1 Item 19) overassets; Cit/A it2 1 is cash balances (Item 1) over assets; LEVit isleverage ((Item 9 1 Item 34)/(Item 9 1 Item 34 1 Item 216)); andQ is the market value of equity (price times shares outstandingfrom CRSP) plus assets minus the book value of equity (Item 60 1Item 74) all over assets. We Winsorize the ingredients of theindex before constructing it.

15. These �rst two factors closely parallel the notion of “�nancial depen-dence” in Rajan and Zingales [1998].


One disadvantage of this index is that the model’s concept ofequity dependence requires a proxy for investment opportunitiesKfb that is distinct from mispricing d. Of the �ve variables in theindex, both low dividends and high values of Q can be thought ofas proxies for strong investment prospects. However, Q will alsocontain information about mispricing d. This dual role for Q isproblematic, since the model has the opposite predictions for theeffects of Kfb and d. In light of this ambiguity, our baselinespeci�cations use a modi�ed four-variable version of the KZ indexthat omits Q. We stress that this is for conceptual cleannessrather than because it has any real effect on the results; as weshow in a robustness check, leaving Q in the index does not alterour basic results. We denote the four-variable version of the indexsimply by KZit:

(5) KZit 5 21.002CFit

Ait212 39.368

DIVit

Ait212 1.315

Cit

Ait211 3.139LEVit.

Henceforth, when we refer to the “KZ index,” we mean thisversion, unless stated otherwise.

Several other aspects of our use of the index deserve fur-ther discussion. First, we do not view the KZ index as a precisemeasure of equity dependence for our sample of �rms, nor thecoef�cients in equation (5) as exactly the “right” weights forthis sample. Rather, we think of Kaplan and Zingales [1997] ashaving nominated several variables that plausibly ought to beindicative of equity dependence, and having shown that each ofthese variables enters with the expected sign. Our use of theindex is simply an effort to restrict ourselves to these previ-ously nominated variables, so as to avoid data mining. Theprecise weights are not really the issue. As we show in arobustness test, very similar results obtain if we reset theweights so that each of the four variables explains one-fourthof the variation in the index.

Second, Kaplan and Zingales [1997] focus on a small sampleof low-dividend manufacturing �rms. Our sample is muchbroader, so one might worry about the appropriateness of apply-ing the KZ index to it. However, in our robustness checks, weshow that we get very similar results when the index is appliedeither to a subsample similar to that originally studied by Kaplan


and Zingales—i.e., low-dividend manufacturing �rms—or to itsexact complement.16

Third, when we modify the KZ index by dropping Q, we donot adjust the coef�cients on the other four variables. We havereestimated the four-variable version on the original KZ sampleusing the data provided in the appendix of Kaplan and Zingales[1997], and found that the coef�cients on the other four variablesare virtually identical whether or not Q is included in the regres-sion. In other words, Q is approximately orthogonal to the linearcombination of the other four variables implied by the index, so asa practical matter there is no need to adjust those coef�cientswhen excluding Q.

Fourth, the index does not include every characteristic thatseems likely to be associated with equity dependence. In auxiliarytests, we therefore supplement the index with two other vari-ables. AGEit is the number of years since the �rm’s IPO, de�nedas the current year minus the �rst year Compustat reports anonmissing market value of equity. As noted above, this variablemay serve as a proxy for reputation and access to lending mar-kets. Industry s(CF/A)i is the industry average standard devia-tion of cash �ows, a proxy for industry debt capacity. In combi-nation with the raw level of leverage in equation (5), this variablehelps us to get closer to the notion of debt capacity suggested bythe model, which is the �rm’s current leverage relative to its debtcapacity. We calculate the standard deviation of cash �ows acrossthe subset of �rm-year observations for each industry using theindustry de�nitions in Fama and French [1997].

Panel A of Table II shows summary statistics for the KZindex, its four ingredients, and the other two equity-dependenceproxies. By multiplying the coef�cients in equation (5) by thestandard deviation of the components, one can see that the indexis especially sensitive to variation in dividends and leverage.

III.D. Other Investment Determinants

Following Fazzari, Hubbard, and Petersen [1988] and others,our baseline investment equation includes year �xed effects, �rm

16. Relatedly, our baseline approach also differs from that of Kaplan andZingales [1997] in that we scale the components of the KZ index by assets ratherthan by net plant, property, and equipment (Item 8). Again, this is because oursample includes small and nonmanufacturing �rms. Although scaling by assetsproduces fewer outliers in the raw KZ index, the results are similar with eitherdenominator, as we show in our robustness tests.


�xed effects, start-of-year Q, and contemporaneous cash �ow. Qand cash �ow are as de�ned above. Our tests of Hypothesis 2 alsorequire future stock returns. We use the raw three-year cumula-tive return from CRSP, beginning at the end of the period inwhich investment is measured. The choice of three years is basedon evidence which suggests that equity-issuance-related mispric-ing tends to unwind over roughly this horizon (e.g., Loughran andRitter [1995] and Baker and Wurgler [2000]).

IV. EMPIRICAL RESULTS

IV.A. Hypothesis 1: Investment and Stock Prices

We begin with a simple test of Hypothesis 1, which predictsthat the investment of equity-dependent �rms is more sensitiveto stock prices than that of nonequity-dependent �rms. We assign

TABLE IISUMMARY STATISTICS FOR EQUITY DEPENDENCE AND OTHER

INVESTMENT DETERMINANTS

Full sample Subsample means

N Mean SD Median Min Max 1980–1989 1990–1999

Panel A: Equity dependence variables (t 2 2)KZ Index 52,101 20.04 2.60 0.28 218.43 3.68 20.12 0.03CFt2 2 /At2 3% 52,101 12.57 28.75 8.97 239.81 225.92 13.31 11.99DIV t2 2/At2 3% 52,101 1.98 4.67 0.60 0.00 37.65 2.34 1.71Ct2 2/At2 3% 52,101 14.95 30.13 5.01 0.01 223.79 12.80 16.61LEVt2 2% 52,101 35.33 25.62 34.29 0.00 124.13 35.76 35.00AGEt2 2 52,101 10.72 6.08 9.00 1.00 31.00 8.86 12.17Industry s(CF/A) 52,101 144.49 223.89 64.89 11.34 914.11 142.44 146.08

Panel B: Other investment determinantsQt2 1 52,101 1.47 0.93 1.17 0.53 6.15 1.28 1.62Rit ,t1 3% 41,819 58.37 111.70 36.96 292.32 565.24 54.79 62.63CFt/At2 1% 52,101 8.20 11.68 9.19 242.43 36.52 9.25 7.39

Source: Compustat and CRSP, 1980–1999.a. Equity dependence is de�ned using the Kaplan and Zingales [1997] index of �nancial constraints

(excluding Q from the index). This modi�ed version of the index has four components: cash �ow (Item 14 1Item 18) over assets; cash dividends (Item 21 1 Item 19) over assets; cash balances (Item 1) over assets; andleverage ((Item 9 1 Item 34)/(Item 9 1 Item 34 1 Item 216)). We also consider two additional measures ofequity dependence,�rm age and the industry standard deviation of cash �ow over assets between 1980 and1999. Industry de�nitions follow Fama and French [1997]. Age is equal to the current year minus the IPOyear.

b. Q is de�ned as the market value of equity (price times shares outstanding from CRSP) plus assetsminus the book value of equity (Item 60 1 Item 74) over assets. Rit, t13 is the cumulative stock return in yearst 1 1, t 1 2 and t 1 3 from CRSP.

c. All variables are Winsorized at the �rst and ninety-ninth percentiles except for �rm age and theindustry standard deviation of cash �ow.


each �rm to a quintile according to its median value of KZit overthe full sample period.17 For each KZ quintile we then estimatethe following investment equation:

(6)CAPX it

Ait215 ai 1 at 1 bQit21 1 c

CF it

A it211 uit.

Hypothesis 1 predicts that the coef�cient b will generally increaseas KZ increases.

Table III shows that there is indeed a strong relationshipbetween KZ and the effect of stock prices on investment. Thecoef�cient b rises from 0.012 in the �rst quintile to 0.033 in the�fth quintile. Thus, the �rms that are most likely to be equitydependent according to the KZ index have a sensitivity of invest-ment to stock prices that is almost three times as large as �rmsthat are unlikely to be equity dependent.

This pattern in the b coef�cients is our main result, but as anaside it is also interesting to look at the pattern of c coef�cients inTable III. Consistent with the small-sample results of Kaplan and

17. Our results are not sensitive to the technique used to classify �rm-yearobservations. In our robustness checks below, we experiment with two alterna-tives. In the �rst, we allow a �rm’s KZ quintile to vary from year to year, so, forexample, if its leverage increases, it may move to a higher quintile. In the second,we assign �rms to quintiles based on their median values of the index over�ve-year subperiods, rather than over the full sample period. Each of thesealternatives lead to very similar results.

TABLE IIIEQUITY DEPENDENCE AND THE LINK BETWEEN INVESTMENT AND STOCK PRICES

KZ index

Q t2 1 CFt/At 2 1

R2N b (se) [t-stat] c (se) [t-stat]

Quintile 1 10,427 0.012 (0.0013) — 0.110 (0.0146) — 0.622 10,415 0.018 (0.0015) [2.86] 0.124 (0.0133) [0.74] 0.613 10,421 0.024 (0.0024) [4.45] 0.117 (0.0140) [0.38] 0.584 10,418 0.032 (0.0030) [6.17] 0.137 (0.0130) [1.42] 0.625 10,420 0.033 (0.0042) [4.80] 0.145 (0.0134) [1.80] 0.62

a. Regressions of investment on Q and cash �ow by equity dependence quintile. We sort �rms into �vequintiles according to the �rm median Kaplan and Zingales [1997] index of �nancial constraints (excludingQ from the index), performing separate regressions for each group. Year and �rm �xed effects are included.

b. Investment is de�ned as capital expenditures over assets. Q is de�ned as the market value of equityplus assets minus the book value of equity overassets. Cash �ow is de�ned as operating cash �ow over assets.

c. t-statistics test the hypothesis of no difference between the coef�cient in each quintile and quintile 1.d. Standard errors and t-statistics are heteroskedasticity-robust, clustered by �rm, with all �ve regres-

sions estimated simultaneously.


Zingales [1997], we �nd no discernible pattern in this coef�cientacross the KZ quintiles. It is almost the same in quintile 1 (0.110)as in quintile 5 (0.145), and bounces around nonmonotonicallyin between. This particular result, however, is sensitive to thenature of the speci�cation. If we keep everything else thesame and lag the cash-�ow term one year (so that we are usingCF it2 1 /A it2 2 instead of CF it/A it2 1), the cash-�ow coef�cient cnow shows a pronounced increasing pattern, going from 0.037 inquintile 1 to 0.134 in quintile 5 (not reported). The pattern of bcoef�cients, on the other hand, is essentially unaffected by thisvariation: it now goes from 0.014 to 0.034 across the �ve quintiles(not reported).18

To get a better understanding of economic magnitudes, notefrom Table II that the standard deviation of Q in our sample is0.93. Thus, in the highest KZ quintile the impact of a one-stan-dard-deviation shock to Q is to alter the ratio of capital expendi-tures to assets by 0.031 (0.033 3 0.93 5 0.031). When comparedwith either the median or the standard deviation of this invest-ment measure (0.060 and 0.079, respectively), this effect is sub-stantial. As another benchmark, note that the standard deviationof the cash-�ow-to-assets ratio is 0.12, so that in the highest KZquintile, a one-standard-deviation shock to cash �ow moves theinvestment ratio by 0.017 (0.145 3 0.12 5 0.017). Thus, amongthe �rms most likely to be equity dependent, stock prices have alarger effect on investment than does cash �ow. This stands incontrast to the general belief that the effect of cash �ow domi-nates that of Q in investment equations.

Turning to statistical, as opposed to economic signi�cance,there are a couple of ways to evaluate the precision of our results.First, and most simply, we report a t-test of the difference be-tween our quintile 1 and quintile 5 b coef�cients in Table III.When we estimate these �ve regressions simultaneously (withresiduals clustered at the �rm level in an effort to deal withpotential serial correlation and with heteroskedasticity-robust

18. We use the speci�cation with a contemporaneous cash-�ow term as ourbaseline for two reasons. First, this seems to be the convention in the literature.Second, we are interested in having the cash-�ow term be as good a control aspossible for fundamentals, not in making structural inferences with respect to it.This suggests that we should use the most recent cash-�ow information available.In contrast, if the goal were to test whether cash �ow has a causal impact oninvestment controlling for investment prospects, this approach might be lesspalatable, precisely because of contemporaneous cash �ow’s informativenessabout future pro�tability.


standard errors), the t-statistic for this difference is a signi�cant4.80.

An alternative approach, in the spirit of Fama and MacBeth[1973], can be used if one is very concerned about serial correla-tion at the �rm level and does not trust the cluster adjustment todeliver proper standard errors. This approach is illustrated inFigure II. Using the same methodology as before, we now divide�rms into twenty KZ-index groups, instead of �ve. For each groupwe estimate the b coef�cient of investment on Q. We then regressthese twenty b estimates against their respective ordinal KZ-index rankings. In other words, we treat each b coef�cient simplyas a data point, without making any assumptions about theprecision with which it is estimated. This twenty-data-point re-gression yields a point estimate of 0.0014 and a t-statistic of 8.44.Intuitively, this methodology infers that the b coef�cients areprecisely estimated by virtue of the fact that most of them clusterclose to the �tted regression line in Figure II. One can show thatthe point estimate obtained with this approach is roughly consis-

FIGURE IIEquity Dependence and the Link between Investment and Stock Prices

The �gure shows the sensitivity of investment to Q by equity dependence group.We sort �rms into twenty groups according to the �rm median Kaplan andZingales [1997] index of �nancial constraints (excluding Q from the index) overthe period from 1980 to 1999, performing separate regressions for each group.


tent in magnitude with the pattern of coef�cients in the quintileregressions.

IV.B. Robustness of Hypothesis 1 Results

In Tables IV and V we explore the robustness of our basicresult. We alter various aspects of the speci�cation such as thevariables included in the KZ index, the scaling of these indexcomponents, the horizons over which �rms are classi�ed by theindex, the other variables in the investment equations, the com-position of the sample, and the de�nition of investment.

In Row 1 of Table IV we reproduce our baseline speci�cationfor reference: recall that the coef�cient b rises from 0.012 to 0.033as we move from the bottom to the top quintile of the KZ index. InRow 2 we revert to the original �ve-variable version of the indexused in Lamont, Polk, and Saa-Requejo [2001], which includes Q.In this case, the b coef�cient goes from 0.011 in quintile 1 to 0.027in quintile 5, which differs little from the baseline result.

In Row 3, we go back to the four-variable version of the index,but classify �rms based on their �ve-year median value of KZ,rather than their median value over the entire sample period. InRow 4 we push this time-varying classi�cation further, and re-classify �rms every year. Speci�cally, we assign a �rm to a KZquintile in any given year t based on its value of the index in yeart 2 2. These changes in the classi�cation horizon do not affect theresults.

In Row 5 we scale the elements of the KZ index by property,plant, and equipment (PP&E), as opposed to by assets.19 Thisstrengthens the results somewhat, with b now almost quadru-pling from the �rst quintile to the �fth quintile. In Row 6 we addthe lagged value of the capital-expenditure ratio to the right-handside of the regressions. This brings down the b coef�cients a bit,but does not change their relative proportions. In Row 7 we deletethe cash-�ow term from the regression entirely, so that Q is theonly explanatory variable for investment. Again, the results arevery similar to those in the base case.

In Row 8 we restrict ourselves to a subsample similar to thelow-dividend manufacturing sample used by Kaplan and Zingales[1997] to �t the index weights. More precisely, we focus on manu-

19. In Row 5, capital expenditures and cash �ow are still scaled by assets.When we scale the variables in the investment equation by PP&E as well, b risesfrom 0.063 to 0.113 from the �rst to the �fth quintile. The difference of 0.050 hasa t-statistic of 4.05.


TA

BL

EIV

EQ

UIT

YD

EP

EN

DE

NC

EA

ND

TH

EL

INK

BE

TW

EE

NIN

VE

ST

ME

NT

AN

DS

TO

CK

PR

ICE

S :R

OB

UST

NE

SS

Bot

tom

quin

tile

23

4T

opqu

inti

leT

op2

bott

om

b(s

e)b

(se)

b(s

e)b

(se)

b(s

e)D

b[t

-sta

t]

1.K

Z(b

ase

case

)0.

012

(0.0

013)

0.01

8(0

.001

5)0.

024

(0.0

024)

0.03

2(0

.003

0)0.

033

(0.0

042)

0.02

1[4

.80]

2.K

Z(�

ve-v

aria

ble

inde

x)0.

011

(0.0

015)

0.01

9(0

.001

7)0.

020

(0.0

020)

0.02

8(0

.002

4)0.

027

(0.0

030)

0.01

6[4

.74]

3.K

Z(�

ve-y

ear

med

ian

)0.

011

(0.0

014)

0.01

9(0

.001

7)0.

024

(0.0

022)

0.03

0(0

.002

9)0.

034

(0.0

046)

0.02

3[4

.72]

4.K

Z(a

nnu

al)

0.01

0(0

.001

6)0.

019

(0.0

019)

0.02

4(0

.002

9)0.

033

(0.0

036)

0.03

5(0

.004

5)0.

026

[5.4

0]5.

KZ

(PP

&E

scal

ing)

0.01

0(0

.001

2)0.

015

(0.0

014)

0.02

5(0

.002

1)0.

032

(0.0

030)

0.03

9(0

.003

9)0.

030

[7.2

8]6.

Lag

ged

CA

PX

/Ain

clud

ed0.

009

(0.0

012)

0.01

2(0

.001

4)0.

019

(0.0

021)

0.02

5(0

.002

7)0.

028

(0.0

037)

0.01

9[4

.96]

7.C

F/A

excl

uded

0.01

6(0

.001

3)0.

022

(0.0

016)

0.02

8(0

.002

5)0.

037

(0.0

033)

0.03

8(0

.004

8)0.

022

[4.4

7]8.

Low

-div

.mfg

.,P

P&

Esc

alin

g0.

008

(0.0

012)

0.01

3(0

.001

9)0.

015

(0.0

023)

0.02

5(0

.003

3)0.

024

(0.0

040)

0.01

6[3

.69]

9.C

ompl

emen

tsa

mpl

eto

#80.

018

(0.0

026)

0.02

4(0

.003

1)0.

028

(0.0

043)

0.04

2(0

.004

5)0.

047

(0.0

068)

0.03

0[4

.12]

10.K

Z(r

eset

toeq

ual

wei

ghts

)0.

014

(0.0

012)

0.02

0(0

.001

8)0.

027

(0.0

027)

0.03

5(0

.003

5)0.

031

(0.0

047)

0.01

7[3

.60]

a.R

egre

ssio

ns

ofin

vest

men

ton

Qan

dca

sh�o

wby

equi

tyde

pend

ence

quin

tile

.We

sort

�rm

sin

to�

vequ

inti

les

acco

rdin

gto

vari

ous

mod

i�ca

tion

sof

the

Kap

lan

and

Zing

ales

[199

7]in

dex

of�n

anci

alco

nst

rain

ts,

perf

orm

ing

sepa

rate

regr

essi

ons

for

each

grou

p.T

heca

sh-�

owco

ef�c

ient

sar

en

otre

port

ed.

Yea

ran

d�r

m�x

edef

fect

sar

ein

clud

ed.

b.In

vest

men

tis

de�n

edas

capi

tale

xpen

ditu

res

over

asse

ts.Q

isde

�ned

asth

em

arke

tval

ueof

equ

ity

plus

asse

tsm

inus

the

book

valu

eof

equi

tyov

eras

sets

.Cas

h�o

wis

de�n

edas

oper

atin

gca

sh�

owov

eras

sets

.c.

The

�rs

tro

wis

our

base

line

spec

i�ca

tion

,w

hich

clas

si�e

s�r

ms

acco

rdin

gto

thei

rm

edia

nfo

ur-v

aria

ble

KZ

inde

x(e

xclu

ding

Q).

The

seco

ndro

wu

ses

the

�rm

med

ian

�ve

-var

iabl

eK

Zin

dex

(inc

ludi

ng

Q).

The

thir

dro

wus

esa

�ve-

year

med

ian

KZ

inde

x(e

xclu

ding

Q).

The

fou

rth

row

uses

anan

nual

KZ

inde

xm

easu

red

atti

me

t2

2(e

xclu

ding

Q).

The

�fth

row

scal

esth

eK

Z-i

ndex

com

pone

nts

by�x

edas

sets

inst

ead

ofto

tala

sset

s.T

hesi

xth

row

incl

udes

lagg

edin

vest

men

tas

anin

depe

nden

tva

riab

le.T

hese

vent

hro

wex

clu

des

cash

�ow

.Th

eei

ghth

row

rest

rict

sth

esa

mpl

eto

man

ufa

ctur

ing

�rm

s(S

ICco

des

2000

thro

ugh

3999

)wit

ha

rati

oof

divi

den

dsto

net

inco

me

ofle

ssth

an10

perc

ent,

and

also

scal

esth

eK

Z-in

dex

com

pone

nts

by�x

edas

sets

.The

nint

hro

wst

udie

sth

eco

mpl

emen

tsa

mpl

eto

that

stud

ied

inth

eei

ghth

row

,aga

insc

alin

gby

�xed

asse

ts.

Th

ete

nth

row

rese

tsth

eK

Z-in

dex

wei

ghts

soth

atea

chva

riab

leex

plai

nsth

esa

me

port

ion

ofth

ein

dex

vari

atio

n.d.

Stan

dard

erro

rsan

dt-

stat

isti

csar

ehe

tero

sked

asti

city

-rob

ust,

clus

tere

dby

�rm

,wit

hal

l�v

ere

gres

sion

sin

aro

wes

tim

ated

sim

ult

aneo

usl

y.


facturing �rms in SIC codes 2000 to 3999, and further requirethat these �rms have a ratio of dividends to net income of lessthan 10 percent (the dividend criterion Kaplan and Zingalesapply to approximate a subsample studied in Fazzari, Hubbard,and Peterson [1988]); this screen reduces the number of observa-tions from 52,101 to 26,725. We also follow Kaplan and Zingalesmore literally by using PP&E scaling in the components of the KZindex.20 Row 8 shows that the b coef�cient in this subsampleagain just about triples from the bottom to the top quintile. Row9 further shows that the complementary subsample—i.e., �rmsthat are not low-dividend manufacturers—also generates similarresults. Taken together, these results suggest that our previous�ndings are not an artifact of having applied the KZ index outsideof its original setting.

Finally, in Row 10 we show that the results are not overlysensitive to the exact weights that the KZ index gives to its fourcomponents. We reset the weights so that each component ex-plains an equal amount of the variation in the index. This againgives similar results.

20. In Row 8, capital expenditures and cash �ow are still scaled by assets.When we scale the variables in the investment equation by PP&E as well, b risesfrom 0.054 to 0.090 from the �rst to the �fth quintile. The difference of 0.037 hasa t-statistic of 2.90.

TABLE VEQUITY DEPENDENCE AND THE LINK BETWEEN INVESTMENT AND STOCK PRICES:

INTERACTIVE SPECIFICATIONS

N

Q t2 1 Qt2 1 z KZ CFt/At2 1

R2b [t-stat] c [t-stat] d [t-stat]

CAPX/A 52,101 0.021 [22.04] 0.011 [6.87] 0.129 [20.96] 0.611RD/A 52,101 0.028 [21.63] 0.013 [6.71] 0.109 [12.89] 0.691RD 1 SGA/A 52,101 0.049 [18.01] 0.021 [4.19] 0.258 [14.64] 0.84

DA /A 52,101 0.080 [20.43] 0.058 [10.16] 1.117 [38.00] 0.41

a. Regressions of investment on Q, Q interacted with equity dependence, and cash �ow. Year and �rm�xed effects are included.

b. Investment is alternately de�ned as capital expenditures over assets; capital expenditures plusresearch and development expenses over assets; capital expenditures plus research and development ex-penses plus selling, general, and administrative expenses over assets; and growth in assets. Q is de�ned asthe market value of equity plus assets minus the book value of equity over assets.

c. The measure of equity dependence is the �rm median Kaplan and Zingales [1997] index of �nancialconstraints (excluding Q from the index), standardized to have unit variance.

d. Cash �ow is de�ned as operating cash �ow over assets.e. t-statistics use heteroskedasticity-robust standard errors, clustered by �rm.


In Table V we consider how our basic results carry over to thethree other measures of investment. To do this in a compactfashion, we pool the observations and run a series of interactivespeci�cations of the form,

(7)Iit

A it215 ai 1 at 1 bQ it21 1 cQ it21 z KZ i 1 d

CF it

Ait211 u it,

where Iit denotes one of the four measures of investment, and KZi

is the sample-median value of the KZ index for �rm i.21 Thecoef�cient of interest in this case is c. As predicted, c is positivefor each of the four investment measures, with strong statisticalsigni�cance in each case.

Equation (7) also provides a convenient speci�cation withinwhich to examine one of the competing explanations for ourresults discussed earlier. That is, different types of �rms mayhave different production technologies, which imply different sen-sitivities of investment to Q in the absence of mispricing. Morespeci�cally, investment may be naturally concave in Q, and so ifequity-dependent �rms tend to have lower Q values, they wouldbe expected to have higher sensitivities. A simple way to examinethis explanation is to add a Q2 term to the speci�cation andexamine whether c changes. In unreported results we �nd that,while the coef�cient on Q2 does turn out to be signi�cantly nega-tive, the c coef�cient is barely affected. This is true for eachmeasure of investment considered in Table V.22

IV.C. Decomposing KZ and the Effect of OtherEquity-Dependence Indicators

The KZ index results may appear to be something of a blackbox. We know the sign that each variable takes in the index, andwe have established the overall effect of the index on the sensi-tivity of investment to stock prices. Here we examine how eachcomponent of the index individually affects this sensitivity. Wealso examine the effect of other variables, not in the KZ index,that may help to identify equity dependence. As argued above,

21. Including the interaction between cash �ow and KZ in equation (7) doesnot affect the results. The interaction term is statistically insigni�cant in threeout of four regressions in Table V, and c is virtually unchanged in all four cases.In the second row, c drops from 0.012 to 0.011. In the fourth row, c drops from0.058 to 0.056. In the �rst and third rows, c is unchanged.

22. The �rst c coef�cient in Table V drops from 0.011 to 0.08; the second from0.013 to 0.010; the third from 0.021 to 0.018; and, the fourth from 0.058 to 0.053.All four remain statistically signi�cant at the 1 percent level.


demonstrating that the sensitivity responds in the expected di-rection to each of a number of variables forces an advocate ofcompeting explanations—based on, for example, measurementerror in Q, or differences in production technologies across�rms—to tell a somewhat more convoluted story.

To perform the KZ index decomposition, it is again conve-nient to pool all observations, and run the following interactivespeci�cation:

(8)Iit

A it215 ai 1 at 1 bQit21

1 Q it21 z F c1

CFA 1 c2

DIVA 1 c3

CA 1 c4LEVG 1 d

CF it

Ait211 u it,

where the unsubscripted versions of the variables CF/A, DIV/A,C/A, and LEV refer to sample-median values for �rm i. We runthis regression for each of four measures of investment. Thede�nition of equity dependence makes predictions for the signs ofthe interactions: c1 should be negative, as should c2 and c3 . Incontrast, c4 should be positive.

Panel A of Table VI shows that the predictions for c1 , c2, c3 ,and c4 are largely borne out. The sharpest results are for c4 , theinteraction on the leverage term. Across all four de�nitions ofinvestment, c4 is always signi�cantly positive, indicating a strongtendency for levered �rms to have investment that is more sen-sitive to stock prices. The results for c1 , c2 , and c3 are somewhatweaker, but generally consistent with the theory—they take onthe predicted sign in eight of twelve cases altogether, and inseven of the eight cases where they are statistically signi�cant.

In Panel B we consider two other proxies for equity depen-dence that are not included in the KZ index: �rm age and indus-try cash-�ow volatility. We reestimate (8) augmented with theinteractions of these two variables with Q, denoting the interac-tion coef�cients as c5 and c6 , respectively. The level of �rm age isalso included, but its coef�cient is not reported. We predict that c5will be negative, on the premise that younger �rms are morelikely to be dependent on equity, and c6 will be positive, sincehigher cash �ow volatility implies lower debt capacity, whichagain contributes to equity dependence. The results in Panel Bstrongly support these predictions. Both c5 and c6 have the ex-pected sign for all four measures of investment, and are signi�-cant in all but one case.


TA

BL

EV

IE

QU

ITY

DE

PE

ND

EN

CE

DE

CO

MP

OS

ITIO

N

CF

/AD

IV/A

C/A

LE

Vlo

g(A

GE

t)In

dust

rys

(CF

/A)

c 1[t

-sta

t]c 2

[t-s

tat]

c 3[t

-sta

t]c 4

[t-s

tat]

c 5[t

-sta

t]c 6

[t-s

tat]

KZ

inde

x2

22

1pr

edic

tion

22

21

21

Pan

elA

:K

Zin

dex

vari

able

sC

AP

X/A

0.00

6[3

.68]

20.

006

[24.

07]

20.

006

[25.

66]

0.00

7[4

.79]

1R

D/A

20.

000

[20.

03]

20.

006

[23.

36]

20.

004

[22.

47]

0.00

8[4

.14]

1R

D1

SG

A/A

20.

020

[23.

69]

0.00

3[0

.59]

0.00

2[0

.68]

0.02

2[5

.92]

DA

/A2

0.02

5[2

2.92

]2

0.03

0[2

6.17

]0.

002

[0.3

4]0.

027

[4.9

3]P

anel

B:

Add

itio

nal

mea

sure

sof

equi

tyde

pen

denc

eC

AP

X/A

0.00

6[4

.16]

20.

004

[23.

31]

20.

006

[26.

26]

0.00

5[3

.43]

20.

004

[25.

67]

0.00

7[4

.93]

1R

D/A

0.00

0[0

.17]

20.

005

[22.

73]

20.

004

[22.

85]

0.00

6[3

.08]

20.

004

[23.

98]

0.00

7[4

.14]

1R

D1

SG

A/A

20.

019

[23.

47]

0.00

6[1

.22]

0.00

0[0

.10]

0.02

2[5

.34]

20.

008

[23.

65]

0.00

4[1

.67]

DA

/A2

0.02

4[2

2.84

]2

0.02

6[2

5.35

]0.

000

[0.0

0]0.

025

[4.2

3]2

0.01

1[2

4.02

]0.

010

[2.7

7]

a.R

egre

ssio

ns

ofin

vest

men

ton

Q,

Qin

tera

cted

wit

hth

eco

mpo

nent

sof

equ

ity

depe

nde

nce

,and

cash

�ow

.Onl

yth

eco

ef�

cien

tson

the

Qin

tera

ctio

nte

rms

are

repo

rted

.Yea

ran

d�

rm�

xed

effe

cts

are

incl

ude

d.b.

Inve

stm

ent

isal

tern

atel

yde

�ned

asca

pita

lex

pend

itur

esov

eras

sets

;ca

pita

lex

pend

itu

res

plus

rese

arch

and

deve

lopm

ent

expe

nses

over

asse

ts;

capi

tal

expe

ndit

ures

plu

sre

sear

chan

dde

velo

pmen

tex

pens

espl

usse

llin

g,ge

nera

l,an

dad

min

istr

ativ

eex

pens

esov

eras

sets

;an

dgr

owth

inas

sets

.Qis

de�n

edas

the

mar

ket

valu

eof

equi

typl

usas

sets

min

us

the

book

valu

eof

equi

tyov

eras

sets

.c.

The

�rs

tpa

neld

ecom

pose

sth

eef

fect

ofth

eK

apla

nan

dZi

ngal

es[1

997]

inde

x(e

xclu

ding

Q)

into

its

four

com

pone

nts:

�rm

med

ian

oper

atin

gca

sh�

owov

eras

sets

;�rm

med

ian

divi

dend

sov

eras

sets

;�

rmm

edia

nca

shba

lanc

eov

eras

sets

;an

d�

rmm

edia

nle

vera

ge.

The

seco

ndpa

nel

adds

two

addi

tion

alm

easu

res

ofeq

uity

depe

nden

ce,

�rm

age,

and

the

indu

stry

stan

dard

devi

atio

nof

cash

�ow

over

asse

tsbe

twee

n19

80an

d19

99.T

hele

vel

of�r

mag

eis

incl

uded

alon

gw

ith

the

inte

ract

ion

betw

een

�rm

age

and

Q.

d.A

llco

mpo

nen

tsof

equi

tyde

pend

ence

are

stan

dard

ized

toha

veu

nit

vari

ance

.e.

t-st

atis

tics

use

hete

rosk

edas

tici

ty-r

obus

tst

anda

rder

rors

,cl

uste

red

by�r

m.


Overall, these results paint a detailed picture of the type of�rm that is likely to have a high sensitivity of investment to stockprices: a young, nondividend-paying �rm, with low cash �ow andcash balances, and with high leverage relative to the debt capac-ity of its assets. This picture �ts the theoretical de�nition ofequity dependence rather well. And while it does not de�nitivelyrule out competing explanations for our results, it does suggestthat such explanations would have to be quite intricate andmultidimensional.

IV.D. Hypothesis 2: Investment and Future Stock Returns

We now turn to Hypothesis 2. The basic strategy for testingthis hypothesis is to replace Q in the investment equations withfuture stock returns. More precisely, we now run regressions ofthe following form:

(9)CAPXit

Ait215 at 1 bRit,t13 1 c

CF it

Ait211 u it,

where Rit , t1 3 is the return on �rm i’s stock over the three-yearperiod from the end of year t to the end of year t 1 3. As discussedabove, one might expect b to be negative on average whethervariation in Rit ,t1 3 re�ects either mispricing or differences in therational cost of capital. The unique prediction of our theory isagain a cross-sectional prediction—b should be more negative forequity-dependent �rms. Note that such a prediction does notfollow from any of the alternative explanations of our Hypothesis1 results, such as differences in measurement error in Q across�rms, differences in production technologies, or adverse selectionthat varies with Q.

Nevertheless, it is worth pointing out another type of mea-surement error problem that might arise in this setting. Therealized return Rit ,t1 3 is a noisy proxy for what we want tocapture, which is the expected return. If the signal-to-noise ratioin this proxy—i.e., the ratio of forecastable return variance tounexpected return variance—varies systematically with the de-gree of equity dependence, this could bias our results. Vuol-teenaho [2002] provides a variance decomposition of stock returnswhich is very helpful in addressing this issue. It turns out thatthe signal-to-noise ratio is in fact greater for smaller stocks—inother words, a greater fraction of small stocks’ returns are pre-dictable based on the set of variables in his vector autoregression.


This could potentially pose a problem for us, since there is a weakconnection between the KZ index and �rm size. Fortunately,however, we are able to use Vuolteenaho’s results to explicitlycalibrate the signal-to-noise ratio for each of our KZ categories; asit turns out, the differences across categories are much too smallto account for the results we report below.23

Equation (9) drops the �rm �xed effect that we included inour previous Q speci�cations. The results to follow are actuallystronger if we keep the �xed effect in, but using a �xed effect thatis effectively �tted over the �rm’s entire sample period makes theinterpretation of the future-returns term problematic.24 By leav-ing out the �xed effect, we have all the variables set up in such away that if we were to put future returns on the left-hand sideinstead of the right, we would be in a position to conduct alegitimate no-look-ahead return-forecasting exercise. Indeed, thefull-sample results that we have are isomorphic to such a fore-casting exercise, implying that, controlling for cash �ow, higherinvestment forecasts lower returns.25 If one wants to preserve thespirit of a �xed effect, but avoid the “look-ahead” problems that itentails, another approach is to measure the investment and cash-�ow terms relative to their contemporaneous industry averages.Demeaning in this way leads to estimates similar to those re-ported here.

Table VII presents the results for investment and futurereturns. (The requirement of returns data reduces the number ofobservations from 52,101 to 41,819.) The table shows that thebasic prediction of Hypothesis 2 is supported. The coef�cients onfuture returns are negative and signi�cant in each KZ quintile,and they go from 20.004 in the �rst quintile to 20.007 in the �fth

23. Vuolteenaho’s Table IV [p. 247] allows us to calculate the ratio of fore-castable return variance to unexpected return variance separately for each size(market capitalization) decile. For each year in our sample, we assign every �rmto its closest size decile. We can then compute the average value of this signal-to-noise ratio over the entire sample period for each KZ category, taking intoaccount the fact that stocks in different KZ categories have different size distri-butions. The ratios are 0.193, 0.203, 0.209, 0.226, and 0.236 for KZ quintiles 1through 5, respectively.

24. For example, if returns are negative in the past, a �xed effect will makesubsequent returns look high even if they are in fact zero.

25. In a recent paper, Titman, Wei, and Xie [2001] do this kind of forecastingexercise explicitly, demonstrating that one can earn high benchmark-adjustedreturns by buying the stocks of low-investment �rms and shorting the stocks ofhigh-investment �rms. They are interested in a different set of issues than we are,however, and so they do not explore the cross-sectional predictions that we focuson. See also Polk and Sapienza [2002].


quintile. The t-statistic on the difference of these estimates is22.40.

As before, another simple and robust way to assess the sta-tistical signi�cance of the coef�cient pattern across KZ categoriesis with the type of scatter plot that we introduced in Figure II. Werepeat this exercise for future returns in Figure III. A regressionof the twenty individual b coef�cients against their ordinal KZ-index rankings yields a slope of 20.0002, with a t-statistic of23.23. In spite of this signi�cant t-statistic, however, a visualcomparison of Figures II and III suggests that our estimates of bare markedly less precise in the future-returns regressions thanthey were in the Q-based regressions. This is not surprising, sincefuture realized returns are bound to be a very noisy proxy formispricing.

In unreported regressions, we have also replicated the ro-bustness analysis of Tables IV and V, as well as the KZ-indexdecomposition of Table VI, with future returns inserted every-where in place of Q. The results generally hold up to differentmeasures of investment, and to variations in scaling techniquesand other details. The most noteworthy wrinkle is that because ofthe greater noise inherent in working with future returns, we nolonger have the power to do much with the KZ-index decomposi-tion—when we interact future returns with each of the index

TABLE VIIEQUITY DEPENDENCE AND THE LINK BETWEEN INVESTMENT AND FUTURE RETURNS

KZ index N

R it , t1 3 CFt/At2 1

R2b (se) [t-stat] c (se) [t-stat]

Quintile 1 8307 20.004 (0.0009) — 0.228 (0.0169) — 0.142 8292 20.004 (0.0011) [20.33] 0.215 (0.0165) [20.57] 0.073 8449 20.006 (0.0010) [21.66] 0.250 (0.0207) [0.82] 0.074 8407 20.008 (0.0009) [23.08] 0.311 (0.0204) [3.11] 0.085 8364 20.007 (0.0009) [22.40] 0.290 (0.0186) [2.44] 0.08

a. Regressions of investment on future returns and cash �ow by equity dependence quintile. We sort�rms into �ve quintiles according to the �rm median Kaplan and Zingales [1997] index of �nancialconstraints (excluding Q from the index), performing separate regressions for each group. Year �xedeffects are included.

b. Investment is de�ned as capital expenditures over assets. Rit ,t13 is the cumulative stock return inyears t 1 1, t 1 2, and t 1 3 from CRSP. Cash �ow is de�ned as operating cash �ow over assets.

c. t-statistics test the hypothesis of no difference between the coef�cient in each quintile and quintile 1.d. Standard errors and t-statistics are heteroskedasticity-robust, clustered by �rm, with all �ve regres-

sions in a panel estimated simultaneously.


components separately, the individual interaction coef�cients areoften statistically insigni�cant.26

Finally, we have tried using both Q and future returns to-gether in the same regression (along with cash �ow) to explaininvestment. This is arguably a too stringent test, in the sense thatthe future-return term will now only attract a signi�cant coef�-cient to the extent that it contains information about expectedreturns above and beyond that in Q. Nevertheless, the patternsacross KZ categories are similar—albeit slightly muted—to those

26. We have also tried a variant of the analysis mentioned in footnote 7,splitting the sample based on whether future returns are positive or negative, andrunning the Q-based regressions separately for each subsample. Again, the ideahere is to explore a subtler prediction of our model—that Hypothesis 1 should holdmore strongly for undervalued �rms, i.e., those with higher future returns. Thisappears to be the case: for �rms with positive future returns, the coef�cient on Qgoes from 0.011 to 0.032 as we move from KZ quintile 1 to 5, a 191 percentincrease. For �rms with negative future returns, the coef�cient on Q goes from0.016 to 0.030, an 88 percent increase. However, given the imprecision associatedwith splitting the sample on future returns, this difference-in-difference is notsigni�cant.

FIGURE IIIEquity Dependence and the Link between Investment and Future ReturnsThe �gure shows the sensitivity of investment to future stock returns by equity

dependence group. We sort �rms into twenty groups according to the �rm medianKaplan and Zingales [1997] index of �nancial constraints (excluding Q from theindex) over the period from 1980 to 1999, performing separate regressions for eachgroup.


we obtain when we enter Q and future returns separately. Inparticular, the coef�cients on Q continue to be positive and in-creasing in the KZ index, while the coef�cients on future returnscontinue to be negative and decreasing.27 This is consistent withthe view that both Q and future realized returns contain inde-pendent noisy information about future expected returns, andhence about mispricing.

IV.E. Hypothesis 3: Financing, Stock Prices, and Stock Returns

Our last hypothesis is that the equity issuance of equity-dependent �rms is positively related to Q and negatively relatedto future returns. Verifying this hypothesis would lend furthersupport to the model’s premise that �nancing considerations areat the heart of the link between investment and stock prices. Itwould also further justify our assumption that the KZ index is auseful indicator of equity dependence; i.e., it would establish thathigh-KZ �rms do in fact issue meaningfully more equity whentheir stock prices are high. We test Hypothesis 3 with the sameregression framework used for the previous hypotheses, simplychanging the dependent variable from investment to �nancing.

Panel A of Table VIII looks at the relationship between�nancing and Q. In the �rst row the �nancing variable is equityissues over assets, e it/A it2 1 . The sensitivity of equity issuance toQ is strongly signi�cant across all KZ quintiles, and rises from0.021 in the �rst quintile to 0.064 in the �fth quintile. Thus, �rmsclassi�ed as most likely to be equity dependent have equity issu-ance that is strongly tied to their stock prices, as predicted inHypothesis 3.28

The model does not imply that equity-dependent �rms �-nance 100 percent of their marginal investment with equity is-sues, of course. According to Proposition 1, in the relevant regionof the parameter space (iii.a) the leverage constraint is binding,so at the margin new equity and debt are raised in proportions(1 2 D# ) and D# , respectively. This suggests that equity-dependent

27. The coef�cients on future returns are now 20.003, 20.003, 20.004,20.006, and 20.005 for KZ quintiles 1 through 5, respectively. The differencebetween quintiles 1 and 5 has a t-statistic of 1.53; that between quintiles 1 and 4has a t-statistic of 1.94.

28. As discussed earlier, the theory makes a clear-cut prediction for therelative strength of the effect across different KZ classes only if one conditions onundervaluation. See the NBER working paper version for empirical tests alongthese lines.


TA

BL

EV

III

EQ

UIT

YD

EP

EN

DE

NC

EA

ND

TH

EL

INK

BE

TW

EE

NE

XT

ER

NA

LF

INA

NC

E,S

TO

CK

PR

ICE

S,A

ND

RE

TU

RN

S

Bot

tom

quin

tile

23

4T

opqu

inti

leT

op2

bott

om

b(s

e)b

(se)

b(s

e)b

(se)

b(s

e)D

b[t

-sta

t]

Pan

elA

:Fin

anci

ngan

dQ

Ext

erna

leq

uity

—al

l�

rms

0.02

1(0

.004

0)0.

051

(0.0

050)

0.04

0(0

.005

6)0.

056

(0.0

069)

0.06

4(0

.008

5)0.

043

[4.5

9]E

xter

nal

equi

typl

usde

bt—

all

0.03

6(0

.006

5)0.

077

(0.0

080)

0.08

7(0

.009

0)0.

104

(0.0

115)

0.13

6(0

.016

1)0.

100

[5.7

5]P

anel

B:

Fin

anci

ngan

dfu

ture

retu

rns

Ext

erna

leq

uity

—al

l�

rms

20.

007

(0.0

018)

20.

013

(0.0

021)

20.

012

(0.0

020)

20.

012

(0.0

012)

20.

008

(0.0

010)

20.

001

[20.

71]

Ext

erna

leq

uity

plus

debt

—al

l2

0.00

8(0

.003

0)2

0.01

9(0

.003

1)2

0.02

0(0

.003

3)2

0.02

6(0

.002

5)2

0.02

0(0

.002

6)2

0.01

2[2

2.94

]

a.R

egre

ssio

ns

ofex

tern

al�n

ance

onQ

and

cash

�ow

,by

equ

ity

depe

nden

cequ

inti

le(P

anel

A),

and

regr

essi

ons

ofex

tern

al�n

ance

onfu

ture

retu

rns

and

cash

�ow

,by

equi

tyde

pend

ence

quin

tile

(Pan

elB

).W

eso

rt�r

ms

into

�ve

quin

tile

sac

cord

ing

toth

e�

rmm

edia

nK

apla

nan

dZ

inga

les

[199

7]in

dex

of�

nanc

ial

cons

trai

nts

(exc

ludi

ngQ

),pe

rfor

min

gse

para

tere

gres

sion

sfo

rea

chgr

oup.

b.E

xter

nal�

nanc

ing

isal

tern

atel

yde

�ned

aseq

uity

issu

esov

eras

sets

and

aseq

uit

ypl

us

debt

issu

esov

eras

sets

.Qis

de�n

edas

the

mar

ket

valu

eof

equi

typl

us

asse

tsm

inus

the

book

valu

eof

equi

tyov

eras

sets

.Rit

,t1

3is

the

cum

ulat

ive

stoc

kre

turn

inye

ars

t1

1,t

12,

and

t1

3fr

omC

RSP

.C

ash

�ow

isde

�ned

asop

erat

ing

cash

�ow

over

asse

ts.

c.Y

ear

�xed

effe

cts

are

incl

uded

inbo

thpa

nel

s.F

irm

�xed

effe

cts

are

incl

uded

inth

e�

rst

pane

lon

ly.

d.St

anda

rder

rors

and

t-st

atis

tics

are

hete

rosk

edas

tici

ty-r

obus

t,cl

uste

red

by�r

m,w

ith

all

�ve

regr

essi

ons

ina

row

esti

mat

edsi

mu

ltan

eou

sly.


�rms should also be raising a signi�cant amount of debt on themargin.

We examine this prediction using a total external �nancevariable that includes both equity and debt issues, (eit 1 d it)/Ait2 1 . The results are in the second row of Panel A. They showthat for �rms in KZ quintile 5, the sensitivity of total �nance to Qis 0.136, whereas for equity issues alone it is 0.064. In otherwords, for high-KZ �rms, a marginal increase in the stock priceleads to roughly equal increases in equity and debt �nance. Thisseems consistent with the model’s prediction that equity-depen-dent �rms issue equity and debt in lockstep. Moreover, the quan-titative response of equity issues to stock prices is substantialrelative to the response of total �nance, consistent with the spiritof the equity �nancing channel.

Panel B changes the proxy for mispricing from Q to futurereturns. The conclusions are much the same as in Panel A. Equityissuance is signi�cantly negatively related to future returns forhigh-KZ �rms, and the response of equity issues to future returnsin this group again accounts for roughly half of the response oftotal external �nance to future returns.

Finally, because of the dif�culty of measuring equity anddebt issues from Compustat data, we have examined whether theresults in Table VIII are robust to alternative de�nitions of �-nancing variables. Those used in Table VIII are backed out frombalance sheet data. In unreported regressions we �nd very simi-lar results using equity and debt issues taken from the cash �owstatement.29

V. CONCLUSION

Rather than restating our results, we close with a �nal ca-veat regarding interpretation. If one takes our model literally, itmight be tempting to conclude that the investment behavior ofequity-dependent �rms must necessarily be less ef�cient thanthat of nonequity-dependent �rms. After all, according to themodel, nonfundamental movements in stock prices introduce vola-tility into the investment of equity-dependent �rms, thereby mov-

29. Speci�cally, we remeasure net equity issues as Compustat Item 108minus Item 115, and debt issues as the change in long-term debt (Item 111 2 Item114). With these de�nitions, the sensitivity of equity issuance to Q goes from 0.017in the �rst KZ quintile to 0.042 in the �fth quintile, and sensitivity of totalexternal �nance to Q from 0.026 in the �rst quintile to 0.107 in the �fth quintile.


ing them away from the �rst best. But we caution readers againstjumping to this sort of welfare conclusion, because it is quitesensitive to the perhaps unrealistic assumption that, absent �-nancial constraints, managers always act in the interest of theirstockholders.

Consider the following embellishment of the model. Every-thing is as before, except that unconstrained managers are sub-ject to an agency problem that leads them to prefer excessivelysmooth investment in the face of changes in fundamentals. If onelayers on top of this agency problem our equity �nancing channel,the same basic positive implications emerge: all else equal, theinvestment of equity-dependent �rms will respond more to stockprices than that of nonequity-dependent �rms. But the normativeimplications of the model will be very different. Since the noneq-uity-dependent �rms are now underreacting to stock prices, theinvestment of the equity-dependent �rms may actually be closerto ef�cient on average.30 In other words, starting from a second-best situation, the distortion inherent in the equity �nancingchannel may help to alleviate the distortion coming from manag-ers’ tendency to smooth investment.

GRADUATE SCHOOL OF BUSINESS, HARVARD UNIVERSITY, AND NATIONAL BUREAU OF

ECONOMIC RESEARCH

DEPARTMENT OF ECONOMICS, HARVARD UNIVERSITY, AND NATIONAL BUREAU OF

ECONOMIC RESEARCH

STERN SCHOOL OF BUSINESS, NEW YORK UNIVERSITY

REFERENCES

Alti, Aydogan, “How Sensitive Is Investment to Cash Flow When Financing IsFrictionless?” Journal of Finance, LVII (2003), 707–722.

Baker, Malcolm, Jeremy Stein, and Jeffrey Wurgler, “When Does the MarketMatter? Stock Prices and the Investment of Equity-Dependent Firms,” NBERWorking Paper No. 8750, 2002.

Baker, Malcolm, and Jeffrey Wurgler, “The Equity Share in New Issues andAggregate Stock Returns,” Journal of Finance, LV (2000), 2219–2257.

——, “Market Timing and Capital Structure,” Journal of Finance, LVII (2002),1–32.

Barro, Robert J., “The Stock Market and Investment,” Review of Financial Stud-ies, III (1990), 115–132.

Bayless, Mark, and Susan Chaplinsky, “Is there a Window of Opportunity forSeasoned Equity Issuance?” Journal of Finance, LI (1996), 253–278.

Blanchard, Olivier, Chanyong Rhee, and Lawrence Summers, “The Stock Market,

30. Indeed, researchers commonly associate a greater sensitivity of invest-ment to stock prices with a higher degree of ef�ciency. See, e.g., Scharfstein [1998]and Rajan, Servaes, and Zingales [2000] for examples in the context of internalcapital markets.


Pro�t, and Investment,” Quarterly Journal of Economics, CVIII (1993),115–136.

Bosworth, Barry, “The Stock Market and the Economy,” Brookings Papers onEconomic Activity (1:1975), 257–290.

Chirinko, Robert S., and Huntley Schaller, “Business Fixed Investment and‘Bubbles’: The Japanese Case,” American Economic Review, XCI (2001),663–680.

Choe, Hyuk, Ronald W. Masulis, and Vikram Nanda, “Common Stock Offeringsacross the Business Cycle: Theory and Evidence,” Journal of Empirical Fi-nance, I (1993), 3–31.

Diamond, Douglas W., “Monitoring and Reputation: The Choice between BankLoans and Directly Placed Debt,” Journal of Political Economy, XCIX (1991),689–721.

Erickson, Timothy, and Toni M. Whited, “Measurement Error and the Relation-ship between Investment and Q,” Journal of Political Economy, CVIII (2000),1027–1057.

Fama, Eugene F., and Kenneth R. French, “Industry Costs of Equity,” Journal ofFinancial Economics, XLIII (1997), 153–193.

Fama, Eugene F., and James D. MacBeth, “Risk, Return and Equilibrium: Em-pirical Tests,” Journal of Political Economy, LXXXI (1973), 607–636.

Fazzari, Steven M., R. Glenn Hubbard, and Bruce C. Petersen, “Financing Con-straints and Corporate Investment,” Brookings Papers on Economic Activity(1:1988), 141–195.

Fazzari, Steven M., R. Glenn Hubbard, and Bruce C. Petersen, “Investment-CashFlow Sensitivities Are Useful: A Comment on Kaplan and Zingales,” Quar-terly Journal of Economics, CXV (2000), 707–712.

Fischer, Stanley, and Robert C. Merton, “Macroeconomics and Finance: The Roleof the Stock Market,” Carnegie-Rochester Conference Series on Public Policy,XXI (1984), 57–108.

Ikenberry, David, Josef Lakonishok, and Theo Vermaelen, “Market Underreac-tion to Open Market Share Repurchases,” Journal of Financial Economics,XXXIX (1995), 181–208.

Jenter, Dirk, “Market Timing and Managerial Portfolio Decisions,” Harvard Uni-versity, working paper, 2001.

Kaplan, Steven N., and Luigi Zingales, “Do Investment-Cash Flow SensitivitiesProvide Useful Measures of Financing Constraints?” Quarterly Journal ofEconomics, CXII (1997), 169–215.

Kaplan, Steven N., and Luigi Zingales, “Investment-Cash Flow Sensitivities AreNot Valid Measures of Financing Constraints,” Quarterly Journal of Econom-ics, CXV (2000), 695–705.

Keynes, John Maynard, The General Theory of Employment, Interest, and Money(London: Macmillan, 1936).

Lamont, Owen, Christopher Polk, and Jesus Saa-Requejo, “Financial Constraintsand Stock Returns,” Review of Financial Studies, XIV (2001), 529–554.

Loughran, Tim, and Jay Ritter, “The New Issues Puzzle,” Journal of Finance, L(1995), 23–51.

Lucas, Deborah, and Robert McDonald, “Equity Issues and Stock Price Dynam-ics,” Journal of Finance, XLV (1990), 1019–1043.

Morck, Randall, Andrei Shleifer, and Robert Vishny, “The Stock Market andInvestment: Is the Market a Sideshow?” Brookings Papers on EconomicActivity (2:1990), 157–215.

Myers, Stewart C., and Nicholas S. Majluf, “Corporate Financing and InvestmentDecisions When Firms Have Information That Investors Do Not Have,”Journal of Financial Economics, XIII (1984), 187–221.

Polk, Christopher, and Paola Sapienza, “The Real Effects of Investor Sentiment,”Northwestern University, working paper, 2002.

Poterba, James, “Comments on Fazzari, Hubbard, and Petersen,” Brookings Pa-pers on Economic Activity (1:1988), 200–204.

Rajan, Raghuram G., Henri Servaes, and Luigi Zingales, “The Cost of Diversity:The Diversi�cation Discount and Inef�cient Investment,” Journal of Finance,LV (2000), 35–80.


Rajan, Raghuram G., and Luigi Zingales, “Financial Dependence and Growth,”American Economic Review, LXXXVIII (1998), 559–586.

Ritter, Jay, “The Long-Run Performance of Initial Public Offerings,” Journal ofFinance, XLII (1991), 365–394.

Scharfstein, David S., “The Dark Side of Internal Capital Markets II: Evidencefrom Diversi�ed Conglomerates,” NBER Working Paper No. 6352, 1998.

Shleifer, Andrei, and Robert Vishny, “Stock Market Driven Acquisitions,” Journalof Financial Economics, forthcoming.

Speiss, D. Katherine, and John Af�eck-Graves, “Underperformance in Long-RunStock Returns Following Seasoned Equity Offerings,” Journal of FinancialEconomics, XXXVIII (1995), 243–267.

Stein, Jeremy C., “Rational Capital Budgeting in an Irrational World,” Journal ofBusiness, LXIX (1996), 429–455.

Titman, Sheridan, K. C. John Wei, and Feixue Xie, “Capital Investments andStock Returns,” University of Texas, working paper, 2001.

Tobin, James, “A General Equilibrium Approach to Monetary Theory,” Journal ofMoney, Credit and Banking, I (1969), 15–29.

von Furstenberg, George M., “Corporate investment: Does Market ValuationMatter in the Aggregate?” Brookings Papers on Economic Activity, (2:1977),347–397.

Vuolteenaho, Tuomo, “What Drives Firm-Level Stock Returns?” Journal of Fi-nance, LVII (2002), 233–264.


WHEN DOES THE MARKET MATTER? STOCK PRICES AND THE INVESTMENT … Files/StockPrices... · 2019. 11....

Documents

Transcript of WHEN DOES THE MARKET MATTER? STOCK PRICES AND THE INVESTMENT … Files/StockPrices... · 2019. 11....