DIVIDENDS, DEBT, INVESTMENT, AND EARNINGS. Eugene F. Fama and Kenneth R. French * Abstract

Transcription

1 First Draft: March 1997 This Draft: June 1997 Not for Quotation: Comments Welcome DIVIDENDS, DEBT, INVESTMENT, AND EARNINGS Eugene F. Fama and Kenneth R. French * Abstract We study the determinants of dividends and debt, with particular interest in how firms vary dividends and debt in response to investment and earnings. As the Lintner (1956) model predicts, dividends move inexorably toward target payouts. Contradicting the pecking order model of Myers (1984), there is no evidence that dividends accommodate investment. Debt seems to be the residual variable in financing decisions. Investment increases debt, and higher earnings tend to reduce debt. Variation in leverage in response to investment and earnings is, however, temporary. Like the dividend payout, leverage tends to revert to its target. * Graduate School of Business, University of Chicago (Fama) and School of Management, Yale University (French). The comments of Douglas Diamond, Richard Thaler,Vincent Warther, and Luigi Zingales have been helpful.

2 This paper addresses two empirical questions. First, do firms have long-term targets for dividends and debt, and if so, what determines the targets? Second, do earnings and investment produce variation in dividends and debt away from their targets? Our most novel evidence is on the second question, that is, how dividends and debt accommodate variation through time in earnings and investment. But the models that produce these results require estimates of target dividends and debt. To get these estimates, we follow the tracks of earlier research, with largely similar results but also with a few surprises. There is a large literature on the determinants of dividends. [See the review of Allen and Michaely (1995).] The earlier evidence tends to support Lintner s (1956) model in which a firm s dividends move toward a target that is a fixed proportion of earnings. Previous research tests Lintner s model with timeseries regressions for individual firms. This approach has problems. Requiring, say, 20 years of data to estimate the model limits the number of firms in the tests, and even with 20 years of data, time-series regressions produce imprecise parameter estimates. We examine the behavior of dividends with year-by-year cross-section regressions. Our results have the authority of a long time period ( ) and a comprehensive sample of firms (an average of 1,599 per annual regression). In the end, though, we confirm the time-series evidence. Lintner s model, in which dividends are driven by earnings, is a good description of dividend behavior. The literature on the determinants of debt is even larger than the literature on dividends [see the review of Harris and Raviv (1991)], and many models make predictions about target leverage. None get clear support in our tests. The hypothesis that the tax benefits of debt are an important determinant of target leverage [Modigliani and Miller (1963)] is contradicted by evidence that more profitable firms have less leverage. This result also runs counter to Jensen s (1986) agency cost model, which predicts that more profitable firms have more leverage to deter managers from wasting earnings on bad investments. The hypothesis that target leverage is positively related to asset tangibility [e.g., Myers (1977)] is supported by negative relations between R&D expenditures and leverage. But this support is offset by evidence that the relations between target leverage and two common measures of asset tangibility, depreciation and the market-to-book-ratio, are weak for firms that actually have debt. 1

3 The main contribution of the paper is our evidence on the movement of dividends and debt away from their targets in response to earnings and investment. Most striking, we find absolutely no evidence that dividends accommodate investment. Dividends move toward their target proportions of earnings in the manner predicted by Lintner (1956), unperturbed by variation in investment. The insensitivity of dividends to investment seems to contradict a strict version of the pecking-order model of Myers (1984), which says that, because of the asymmetric-information problems that arise when firms issue new securities, retained earnings are first in line to finance investment. Since new issues of equity are relatively rare, the evidence that dividends do not respond to investment suggests that debt is the residual variable in financing decisions. Our tests show that a change in assets indeed tends to push debt in the same direction. Debt also absorbs variation through time in earnings. About $0.25 of every $1 increase in earnings is used to reduce debt. Changes in leverage in response to earnings and investment are, however, transitory. Firms have leverage targets, and in the long term, leverage returns to its target. We begin in section I with a discussion of methodology. Sections II and III present the empirical results. Section IV concludes. I. Methodology Our approach to explaining financing decisions is a bit unusual. First, we attempt to explain changes in dividends and debt over a relatively long horizon of two years, and many of the explanatory variables also cover two-year periods. Second, in the spirit of Fama and MacBeth (1973), we estimate cross-section regressions each year, and use the average slopes and the time-series standard deviations of the slopes to draw inferences. One advantage of this approach is large samples; we typically have more than 1,600 firms per annual regression. Another is that the year-by-year variation in the slopes, which determines the standard errors of the average slopes, includes the effects of estimation error due to the correlation of the residuals across firms. 2

4 We could also adjust the standard errors of the average slopes for the autocorrelation of the annual slopes. The problem is that we have just 28 time-series observations on the slopes for our sample period. The sample autocorrelations of the slopes are thus imprecise, with standard errors around With such imprecision, the formal cure for autocorrelation can be worse than the disease. (It is worth noting that the popular alternative to Fama-MacBeth standard errors for cross-section regressions, ordinary least squares estimates from regressions that pool the data for all years, typically ignores both the crosscorrelation of the residuals for a given year and the autocorrelation across years.) We use a less formal approach to account for the autocorrelation of the regression slopes. Most of our dependent variables are two-year changes. If one-year changes are serially independent, the overlap of the two-year changes in the year-by-year regressions should induce first-order autocorrelation of 0.5 in the slopes. If this is the only autocorrelation in the slopes, the standard errors of the average slopes, calculated assuming serial independence, are too small by 50%. This suggests that we should require a t-statistic of 4.0, rather than the usual 2.0, to infer reliability. In fact, though not reported in detail, the first-order autocorrelations of the slopes in the regressions cluster between 0.25 and 0.5. Higher-order autocorrelations are more random about zero. Thus, requiring t-statistics of around 4.0 for the average slopes is reasonable. There is also a concern that the regressions are dominated by influential observations. Most of the variables are scaled by assets. This creates influential observations when assets are close to zero. Data errors can also be a problem. To address these issues, each year we drop 0.5% of the observations in each tail of the distribution of each explanatory variable. Because we consider the full sample when we trim each variable, dropping 1% of the observations for K variables causes us to lose fewer than K% of the observations. (See the table legends for details.) Trimming on explanatory variables does not affect the expected values of the regression slopes, but excessive trimming can increase their standard errors. In fact, without showing details, the standard 3

5 errors of the average slopes from the trimmed regressions are typically smaller than the standard errors from regressions that use all the observations. This suggests that the trimming is not excessive. II. Dividends Miller and Modigliani (1961) show that in a perfect capital market, and in the absence of tax effects, dividend policy is irrelevant; it does not affect a firm s value. Previous empirical work [e.g., Fama and Babiak (1968), Choe (1990)] finds, however, that the response of dividends to earnings has a systematic pattern consistent with the target payout model proposed by Lintner (1956). We begin by estimating a cross-section version of Lintner s model. We then turn to the question of primary interest: Do dividends depart from the predictions of the Lintner model to accommodate variation through time in investment? A. The Lintner Model In Lintner s model, a firm s target dividend for year t+2, TD t+2, is proportional to earnings, Y t+2, (1) TD t+2 = ay t+2, but the firm only partially adjusts dividends to the target, (2) D t+2 - D t = ß(TD t+2 - D t ) + e t+2 (3) D t+2 - D t = a 1 Y t+2 + a 2 D t + e t+2. The target payout ratio, a, and the speed of adjustment, ß, are between 0.0 and 1.0, a 1 is aß, a 2 is -ß, and e t+2 is a random error. In effect, then, the Lintner model says that the ratio of dividends to earnings tends to revert to the target a. Equations (1) to (3) are naturally treated as a time-series model for dividends, and that is the way they are usually estimated. As noted earlier, however, the long estimation periods required for time-series tests on individual firms limit the sample to a special group, and even with, say, 20 years of data the parameter estimates are imprecise. We use a different approach. Adding a constant and scaling by total assets, we estimate (3) as the year-by-year cross-section regression, 4

6 (4) (D t+2 - D t )/A t = a + a 1 Y t+2 /A t + a 2 D t /A t + e t+2. To simplify the notation in (4), we omit both the firm subscript that should appear on the regression variables and residuals and the year subscript that should appear on the regression coefficients. Since Lintner s model makes no sense for firms that do not pay dividends, we restrict the estimates of (4) for each year t to firms that pay a dividend in t-2. The results are much the same if the regressions include only firms that pay dividends in t (rather than t-2). The main advantage of the cross-section approach is large samples. The main disadvantage is that the same regression slopes, and thus the same speed of adjustment and target payout, are imposed on all firms. We address this problem below. Average slopes from the year-by-year estimates of (4) are in table 1. Lintner s earnings variable shows up clearly; the positive average slope on Y t+2 /A t is more than 15 standard errors from zero. On the other hand, the estimated speed of adjustment (the negative of the average slope on the lagged dividend, D t /A t ) is small, 0.06, and only 2.31 standard errors from zero. Low speed-of-adjustment coefficients are, however, also found in time-series tests of the Lintner model for recent periods [Choe (1990)]. Finally, estimating the target payout as the negative of the ratio of the average slope on Y t+2 /A t to the average slope on D t /A t produces a number that is too high (above 1.0). We can test the specification of the dividend regressions by adding explanatory variables to (4). The new variables might help explain dividends because they proxy for variation across firms in the Lintner model s target payout and speed of adjustment, which (4) ignores. Or the additional variables might show that dividends are related to more than just measured earnings, suggesting that the Lintner model misses important features of dividend policy. The new variables are the time t+2 market-to-book ratio (V t /A t, the ratio of the firm s total market value to its book value), time t book leverage (I t /A t, the ratio of interest expense to assets) and the lagged (t-2 to t) changes in earnings, dividends, leverage, and assets, dy t /A t dd t /A t, d(i t /A t ), and da t /A t. The expanded regression is (5) dd t+2 /A t = a + a 1 Y t+2 /A t + a 2 D t /A t + a 3 V t+2 /A t+2 + a 4I t /A t + b 1 dy t /A t + b 2 dd t /A t + b 3 d(i t /A t ) + b 4 da t /A t + e t+2. 5

7 A comment on notation. We use d(x t /A t ), with parentheses, to indicate the two-year change in a ratio, whereas dx t /A t denotes the two-year change in the numerator divided by the denominator. For example, d(i t /A t ) = I t /A t - I t-2 /A t-2 is the change in book leverage from t-2 to t, but dd t /A t = (D t - D t-2 )/A t is the change in dividends scaled by A t. The estimates of (5) in table 1 show that lagged investment, da t /A t, and lagged changes in dividends and leverage, dd t /A t and d(i t /A t ), do not add to the explanation of dividends for year t+2. The new variables with explanatory power are the market-to-book ratio, V t+2 /A t+2, lagged leverage, I t /A t, and the lagged change in earnings, dy t /A t. All are positively related to the change in dividends from t to t+2, with average slopes more than 4.9 standard errors from zero. The explanatory power of V t+2 /A t+2, I t /A t, and dy t /A t in (5) might expose shortcomings in Lintner s dividend model. For example, the market-to-book ratio, V t+2 /A t+2, summarizes the expected profitability of the firm s existing and expected future assets. Positive slopes on V t+2 /A t+2 may indicate that dividends respond not only to current earnings, but also to expected future earnings, perhaps in the manner predicted by dividend signaling models [Bhattacharya (1979)]. Similarly, the lagged change in earnings, dy t /A t, may carry information about profitability used to set dividends but missed by current measured earnings. Finally, leverage may have explanatory power because Lintner does not model the relations between dividend and debt policies. The success of Lintner s model in earlier time-series tests suggests an alternative story for the explanatory power of V t+2 /A t+2, I t /A t, and dy t /A t in (5). Perhaps these variables just pick up differences across firms in the model s parameters. To test this possibility, we estimate an expanded version of (5) that includes interaction terms that allow the slopes on the Lintner variables, Y t+2 /A t and D t /A t, to vary as functions of V t+2 /A t+2, I t /A t, and dy t /A t, (6) dd t+2 /A t = a + a 1 (1 + a 1V V t+2 /A t+2 + a 1I I t /A t + a 1Y dy t /A t )Y t+2 /A t + a 2 (1 + a 2V V t+2 /A t+2 + a 2I I t /A t + a 2Y dy t /A t )D t /A t + a 3 V t+2 /A t+2 + a 4I t /A t + b 1 dy t /A t + b 2 dd t /A t + b 3 d(i t /A t ) + b 4 da t /A t + e t+2. 6

8 The estimates of (6) in table 1 say that the interaction terms do not identify reliable variation across firms in the target payout. The interaction slopes that involve Y t+2 /A t are within 3.2 standard errors of zero, and so are below our four-standard-error hurdle. On the other hand, the interaction terms apparently uncover rich variation across firms in the speed of adjustment of dividends to their targets. The slopes on the interaction terms that involve D t /A t are all at least 5.4 standard errors from zero. In the Lintner model of (1) to (3), a faster speed of adjustment implies a more negative slope on D t /A t. All the interaction slopes involving D t /A t are positive. Thus, firms with higher market-to-book ratios, leverage, and lagged earnings growth adjust dividends more slowly to their targets. Offsetting this attenuation of the speed of adjustment, the raw D t /A t slope is much more negative, and more than 7.2 standard errors from zero, when the regressions allow the speed of adjustment to vary across firms. Most interesting, allowing for variation across firms in the speed of adjustment largely absorbs the explanatory power of the market-to-book ratio, the level of leverage, and the lagged change in earnings observed in (5). The slopes on V t+2 /A t+2, I t /A t, and dy t /A t in (6) are only 1.57, 0.52 and 2.62 standard errors from zero. B. Dividends, Investment, and Debt The estimates of (6) suggest that the Lintner model is a reasonable approximation for dividends. We turn next to our main interest and ask whether the variation in dividends left unexplained by the model is in part due to interaction among dividends, investment, and debt. If we take Miller and Modigliani (1961) as our null hypothesis, we can justify using concurrent investment to explain dividends. In the MM world, dividend policy is irrelevant, so investment can affect dividends, but dividends should not affect investment. Thus investment, like earnings, is exogenous with respect to dividends, and we can add concurrent changes in assets [da t+2 /A t = (A t+2 - A t )/A t ] and earnings [dy t+2 /A t = (Y t+2 - Y t )/A t ] to the explanatory variables in (6). 7

9 On the other hand, in the MM world, debt and dividends can be jointly determined, so using concurrent changes in debt to explain changes in dividends can cause an endogeneity problem. Nevertheless, to give perspective on the interaction of the components of net cash flows, we also add di t+2 /A t = (I t+2 - I t )/A t to (6). The full regression is then (7) dd t+2 /A t = a + a 1 (1 + a 1V V t+2 /A t+2 + a 1I I t /A t + a 1Y dy t /A t )Y t+2 /A t + a 2 (1 + a 2V V t+2 /A t+2 + a 2I I t /A t + a 2Y dy t /A t )D t /A t + a 3 V t+2 /A t+2 + a 4I t /A t + b 1 dy t /A t + b 2 dd t /A t + b 3 d(i t /A t ) + b 4 da t /A t + c 1 dy t+2 /A t + c 2 di t+2 /A t + c 3 da t+2 /A t + e t+2. The estimates of (7) in table 1 say that the concurrent change in debt shows no power to explain the change in dividends from t to t+2. The average slope on di t+2 /A t is 0.42 standard errors from zero. Thus, perhaps the endogeneity issue raised by di t+2 /A t is irrelevant because there is no feedback from debt to dividends. In any case, since the average slope on di t+2 /A t is close to zero, including the changes in debt in the regressions does not obscure the effects of other variables which, under our MM null hypothesis, are free of endogeneity problems. The most interesting result from the estimates of (7) is the complete absence of evidence that dividends accommodate investment. The average slope on da t+2 /A t is the wrong sign (positive) and only 0.56 standard errors from zero. Our cross-section evidence that dividends do not respond to concurrent investment confirms Fama s (1974) time-series results for an earlier period and a much smaller sample of firms. The results contradict a strict version of Myers s (1984) pecking order model of financing decisions, which predicts that the asymmetric-information problems that arise in issuing new securities put retained earnings first in line to absorb variation in investment. Aside from the Lintner variables (Y t+2 /A t, D t /A t, and the interaction terms), the only variables that have explanatory power in (7) are the lagged change in earnings, dy t /A t, and perhaps the lagged change in leverage, d(i t /A t ). Although the average slope on dy t /A t is 4.59 standard errors above zero, it has little economic importance. Only 3% of the lagged change in earnings shows up as a future change in dividends. Moreover, we are not inclined to interpret the negative average slope on d(i t /A t ) as evidence 8

10 that leverage deters dividends. The level of leverage, I t /A t, is a better indicator of leverage policy, and the slope on I t /A t in (7) is slightly positive, not negative. Lagged changes in leverage and earnings may capture information about dividends missed by the Lintner model, but regression specification problems are also a possibility. In particular, we doubt that the interaction terms in (6) and (7) capture all variation across firms in the target payout and speed of adjustment coefficients of the Lintner model. On the whole, and like the time-series tests of Fama and Babiak (1968), Fama (1974), and Choe (1990), our cross-section regressions support Lintner s (1956) model for dividends. When we allow for variation across firms in the target payout and speed of adjustment, the Lintner variables (earnings and lagged dividends) show up strongly and other variables lose most of their explanatory power. Most striking, we find no evidence that dividends accommodate investment. In economic terms, partial adjustment toward a target payout (albeit at a snail s pace) seems like a good model for dividends. III. Leverage The dividend tests say that firms do not typically vary dividends in response to investment. Since new issues of equity are relatively rare, the dividend tests suggest that debt is the residual variable in financing decisions. We find that debt is indeed sensitive to investment and earnings. More investment tends to generate more debt, while higher earnings are used to reduce debt. The variation in leverage in response to earnings and investment is, however, temporary. Like others [Jalilvand and Harris (1984), Auerbach (1985), Opler and Titman (1995)], we find that firms have target leverage ratios, and in the long term, leverage returns to its target. We begin by showing that, like the dividend payout, leverage tends to revert to a target value. We then examine how leverage moves away from its target in response to earnings, investment, and dividends. A. Target Leverage and Mean Reversion 9

11 We measure leverage as the ratio of interest expense to book assets, I t /A t. Like Auerbach (1985), we test for mean reversion in leverage with a partial-adjustment model in which the two-year change in leverage, d(i t+2 /A t+2 ) = I t+2 /A t+2 - I t /A t, partially absorbs the difference between a firm s target leverage, E(I t /A t ), and its actual leverage, I t /A t, (8) d(i t+2 /A t+2 ) = a + a 1 [E(I t /A t ) - I t /A t ] + b 1 d(i t /A t ) + e t+2. If firms have leverage targets, and if leverage adjusts only partially toward its target, the parameter a 1 in (8) is positive and less than 1.0. The slope on the lagged change, d(i t /A t ), then measures any additional predictability of leverage due to the time-series properties of leverage that may be missed by the partialadjustment term. Our tests focus on book leverage, I t /A t. One can argue that target debt is a function of the market value of assets, V t, so we should explain market leverage, I t /V t. The literature goes both ways on this issue. We are swayed by the argument of Myers (1977) and others that, because of asymmetric-information problems, intangible assets are not efficiently financed with debt. In this view, leverage targets are probably related more closely to book assets than to total firm value, which includes the value of currently intangible future investments. Moreover, we include V t /A t among the instruments used below to produce a proxy for target book leverage, E(I t /A t ), in (8). If leverage targets are a function of market values, then we should find that target book leverage is positively related to V t /A t. But we shall see that the relation is either negative or non-existent. To estimate (8) we use a two-step cross-section regression approach similar to that in Auerbach (1985) and Opler and Titman (1995). Each year t, we first regress I t /A t on variables assumed to determine target leverage. We then use the fitted values from the first-stage regression for year t as the proxy for E(I t /A t ) in the second-stage estimate of (8) for t. Tables 2 and 3 show the average slopes from the year-by-year first-stage cross-section regressions of I t /A t on potential determinants of target leverage, (9) I t /A t = a + a 1 V t /A t + a 2 YT t /A t + a 3 Dp t /A t + a 4 RD t /A t + a 5 da t /A t + e t. 10

12 The explanatory variables in (9) are motivated by theoretical predictions (outlined below) about how target leverage is related to the profitability and tangibility of a firm s assets. Besides the market-to-book ratio, V t /A t, the variables include earnings before interest and taxes (YT t /A t ), depreciation (Dp t /A t ), research and development expenditures (RD t /A t ), and investment [da t /A t = (A t - A t-2 )/A t ], all measured relative to book assets. Since the partial-adjustment model (8) makes no sense for firms that have no debt, our proxy for target leverage in the second-stage regressions is from the estimates of (9) that exclude unlevered firms (specifically, firms that do not pay interest in year t-2). Since the literature seems unaware that inferences about target leverage are sensitive to whether the tests include unlevered firms, we also report estimates of (9) that use all firms. We show average slopes for two time periods, (table 2) and (table 3). Chan, Jegadeesh, and Lakonishok (1995) find that the Compustat sample of firms becomes essentially complete in The regressions thus have better coverage than those for earlier years. Without showing the details, we can report that the dividend regressions for are similar to those for in table 1. For leverage, the results are a bit more sensitive to time period. Much of the theoretical literature on leverage focuses on predictions about how the profitability and tangibility of a firm's assets determine its target leverage, E(I t /A t ). We outline these predictions and then examine the evidence from the estimates of (9). Tangibility: Asymmetric Information and Agency Problems -- Myers (1977), among others, argues that tangible assets are efficiently financed with debt, but because of asymmetric-information problems, intangible assets like research and development (R&D) and future growth opportunities are financed more efficiently with equity. We interpret Myers (1977) as predicting that target leverage is positively related to the ratio of depreciation to assets, Dp t /A t (a proxy for the fraction of existing assets that are tangible). One might also interpret Myers (1977) as predicting a negative relation between target leverage and the ratio of R&D to assets, RD t /A t, or the market-to-book ratio, V t /A t (a proxy for future growth opportunities). We attempt to explain book leverage, I t /A t, however. Since book assets, A t, do not include R&D or future 11

13 growth opportunities, the predictions of Myers (1977) about the relations between book leverage and V t /A t or RD t /A t are unclear. The predictions of Myers and Majluf (1984) about the relations between target leverage and V t /A t or RD t /A t seem clearer. In their model, asymmetric-information problems make managers reluctant to issue outside equity or risky debt, and managers sometimes forgo profitable investments if they must be financed with risky securities. To avoid this cost, firms likely to have strong growth opportunities choose lower target leverage. If V t /A t and RD t /A t are proxies for growth opportunities, leverage should be negatively related to these variables. Many empirical studies conclude that leverage is positively related to asset tangibility [Marsh (1982), Bradley, Jarrell, and Kim (1984), Auerbach (1985), Long and Malitz (1985), Titman and Wessels (1988), Opler and Titman (1995), Rajan and Zingales (1995)]. Our results on this issue are mixed. The average slopes on RD t /A t in the estimates of (9) in tables 2 and 3 are strongly negative. Thus, consistent with the asset tangibility hypothesis, firms that spend more on R&D have less leverage. For the market-to-book ratio, V t /A t, and depreciation, Dp t /A t, however, the evidence for a relation between asset tangibility and leverage is fragile; it hinges on whether the tests include firms with no debt. The average slope on V t /A t in tables 2 and 3 is reliably negative when the estimates of (9) include all firms. When we drop unlevered firms, the V t /A t slopes for and are only 0.57 and 2.19 standard errors below zero. Both are far from our four-standard-error benchmark. We infer that unlevered firms have high market-to-book ratios and thus larger fractions of intangible assets. But among firms that have debt (87% of Compustat firms), there is no reliable relation between V t /A t and leverage. Similarly, the average slope on depreciation, Dp t /A t, is reliably positive in the estimates of (9) that include firms with no debt. When we exclude unlevered firms, the average Dp t /A t slope for shrinks by 90% and the t-statistic falls to There is a similar 80% decline in the Dp t /A t slope when unlevered firms are dropped from the tests, but the slope is reliably positive (t = 5.57). This statistical reliability is testimony to the power of the test, not the economic importance of the result. Skipping the details, the average Dp t /A t slope for implies that a firm with depreciation two 12

14 standard deviations above the mean has only 6.7% more leverage than average. We conclude that unlevered firms indeed have low ratios of depreciation to book assets and thus low fractions of tangible assets, but among levered firms any relation between Dp t /A t and target leverage is weak. As noted above, the empirical literature commonly reports that leverage is positively related to asset tangibility. Our finding that (R&D aside) the positive relation between asset tangibility and leverage is largely driven by firms that have no debt is, however, apparently new. Profitability: Taxes, Agency Problems, the Pecking Order, and Non-Debt Tax Shields -- If there is a tax advantage to corporate debt [Modigliani and Miller (1963)], then, to save taxes, more profitable firms should have higher target leverage. Jensen s (1986) agency cost model also predicts a positive relation between target leverage and profitability. To deter managers from wasting free cash flows on bad investments, firms with more profitable assets should have more leverage. Myers s (1984) pecking order model does not imply that firms have target leverage ratios. But given that firms have such targets, one might interpret the pecking order model as predicting that more profitable firms have lower leverage targets because they finance more of their investment with retained earnings. This is opposite to the predictions of the tax hypothesis and Jensen s (1986) agency cost model. DeAngelo and Masulis (1980) argue that, controlling for profitability, target leverage is negatively related to non-debt tax shields such as R&D expenditures and depreciation because they reduce the likelihood that the future tax benefits of debt are realized. The negative slopes on RD t /A t in the estimates of (9) in tables 2 and 3 support the DeAngelo-Masulis (1980) hypothesis, but the slopes on Dp t /A t, which tend to be positive, contradict their prediction. Pre-tax profitability is the only variable besides R&D that shows consistent explanatory power in the estimates of (9). The average slopes for YT t /A t are all strongly negative (t-statistics below -7.4). Investment, da t /A t = (A t - A t-2 )/A t, is also a potential proxy for profitability since more profitable firms are likely to grow more rapidly. The slopes on investment are also negative, but only the average slopes clear our four-standard-error hurdle. Rajan and Zingales (1995) also report a negative relation 13

15 between leverage and profitability. Titman and Wessels (1988) find no relation between book leverage and profitability, but in a more restricted sample of firms. Our evidence that more profitable firms have less leverage is bad news for the tax hypothesis and for Jensen s (1986) agency cost hypothesis, but it seems consistent with Myers s (1984) pecking order model. On the whole, though, the pecking order model does not capture the essence of our results. The simple version of this model predicts that firms finance investment first with retained earnings, then with debt, and only as a last resort with stock. In such a world, firms do not have target leverage ratios, and the evidence that leverage is slowly mean-reverting can be read as embarrassing. Target dividend payouts also are not in the spirit of the pecking order model and some other model, like Lintner s (1956), is needed to explain them. Finally, the evidence that investment does not produce even temporary variation in dividends away from the predictions of the Lintner model seems rather damning for what we take to be the central prediction of the pecking order model -- that retained earnings are first in line to absorb variation in investment. We are a bit worried that the estimates of (9) over-state the negative relation between target leverage and profitability. We shall see that variation in earnings generates transitory variation in leverage. This suggests that leverage is likely to be temporarily high when profitability is low. Thus, the negative profitability slopes in (9) may be due in part to transitory variation in leverage in response to profitability, rather than to variation in target leverage. We are reluctant to drop YT t /A t from (9) since the only other variable with consistent explanatory power for firms with debt is R&D. Instead, we have tried the simple alternative of dropping target leverage E(I t /A t ) from (8) and from (10), the expanded version of (8) tested below. Dropping E(I t /A t ) amounts to assuming that target leverage is the same for all firms. Without showing the details, we can report that our main results on the mean reversion of leverage and the transitory variation in leverage in response to earnings and investment show up in these simplified regressions. Tables 2 and 3 summarize the second-stage estimates of the partial-adjustment model (8). The estimates confirm that leverage is mean-reverting and the mean reversion is captured rather well by the 14

16 partial-adjustment model. As the model predicts, the average slopes on the E(I t /A t ) proxy in the estimates of (8) are positive. Moreover, the large t-statistics (above 6.5) on the E(I t /A t ) slopes confirm that the fitted values from the first-stage estimates of (9) capture meaningful differences in target leverage across firms. As predicted by the partial-adjustment model, the average slopes on I t /A t in the estimates of (8) are negative (t-statistics below -9.0), and the slopes on E(I t /A t ) and I t /A t are roughly equal in absolute value. The two-year rate of mean reversion is, however, apparently rather slow, about 17% for and 19% for B. Earnings, Dividends, Investment, and Debt Consider an expanded version of the partial-adjustment model for leverage, (10) d(i t+2 /A t+2 ) = a + a 1 [E(I t /A t ) - I t /A t ] + Z B + e t+2. Z is a vector of variables that produce changes in leverage other than those due to mean reversion. Given our interest in the interplay between investment and financing decisions, our prime candidates for the Z variables are investment, earnings, and dividends. We lean on Modigliani and Miller (1958) to argue that investment and earnings can cause concurrent changes in leverage, but the reverse should not be true. On the other hand, in the MM world, changes in leverage and dividends from t to t+2 may be jointly determined, resulting in an endogeneity problem. Nevertheless, to give perspective on the interaction of the components of cash flows, we include the concurrent change in dividends, dd t+2 /A t, as well as the changes in assets and earnings, da t+2 /A t and dyi t+2 /A t, among the Z variables in (10). (We use dyi t+2 /A t, earnings before interest but after taxes, because we want to measure the response of leverage to earnings, not tax effects.) Finally, to allow for lags in the response of leverage to earnings, dividends, and investment, the Z variables also include predetermined versions of these variables. The estimates of (10) are in tables 2 and 3. One interesting result is that the additional explanatory variables in (10) have little effect on the estimated rate of mean reversion of leverage. The average slopes 15

17 on E(I t /A t ) and especially I t /A t in (10) are similar to those in (8). Thus, the mean reversion of leverage is slow, but apparently it is relentless and not affected much by other variables. Adding other variables reinforces the evidence that the change in leverage from t to t+2 tends to reverse the most recent two-year change. The negative average slopes on d(i t /A t ) in (10) are further from zero than those in (8). In practical terms, however, the estimates of (10) suggest that the two-year change in leverage only reverses 11% to 14% of the lagged change. The statistical reliability of these estimates (tstatistics below -9.8) again confirms that the cross-section regressions have power. Except for lagged investment (discussed below), the predetermined Z variables in (10) show no reliable explanatory power. Thus, leverage does not seem to respond to the lagged market-to-book ratio or to lagged levels or changes in dividends or earnings. The most interesting results from the estimates of (10) are the strong responses of debt to concurrent changes in earnings and to concurrent and lagged investment. The average slopes on dyi t+2 /A t are negative and more than 8.0 standard errors below zero. When earnings growth is high, leverage tends to fall. For better perspective on the dollar response of debt to earnings, tables 2 and 3 include estimates of (10) in which the dependent variable is the change in interest, di t+2 /A t = (I t+2 - I t )/A t, rather than the change in leverage, d(i t+2 /A t+2 ) = I t+2 /A t+2 - I t /A t. In the regressions for the change in interest, the average slopes on the concurrent change in earnings are ( ) and ( ). With interest rates of around 10%, these slopes imply that 20-30% of the variation in earnings is absorbed by concurrent variation in debt. Note, though, that the variation in leverage in response to earnings is temporary accommodation. The mean reversion of leverage, apparent in the average slopes on E(I t /A t ) and I t /A t in the regressions for the change in leverage, implies that leverage eventually returns to its target. The estimates of (10) do not produce a reliable concurrent relation between the change in leverage, d(i t+2 /A t+2 ), and investment, da t+2 /A t. Thus, debt issued to finance concurrent investment does not typically change a firm s leverage; interest and assets move in roughly the same proportion. This inference is reinforced by the regressions to explain changes in interest. The average slopes on da t+2 /A t,

18 ( ) and ( ), are similar to the average values of I t /A t, ( ) and ( ), and more than 8.4 standard errors from zero. The concurrent response of debt to investment is not, however, the total response. In the regressions to explain the change in debt, di t+2 /A t, the slopes on lagged investment, da t /A t, are similar to the slopes on concurrent investment (with t-statistics above 9.8). Similarly, in the regressions to explain the change in leverage, d(i t+2 /A t+2 ), the slopes on lagged investment, da t /A t, are strongly positive. We speculate that the lagged response of debt to investment is driven by the behavior of earnings and leverage in the years before investment. The average of the year-by-year correlations between investment, da t+2 /A t, and the lagged change in leverage, d(i t /A t ), is negative, (Because the data for individual firms are noisy, a correlation of is relatively large.) Thus, firms seem to reduce leverage before periods of strong asset growth. An interesting (but clearly speculative) story for our overall results is that firms reduce leverage when earnings are high and they anticipate strong investment. They then finance the investment concurrently with about the average amount of leverage. Finally, they reverse the earlier reduction in leverage when the newly acquired assets begin to pay off. The average slopes on the concurrent change in dividends, dd t+2 /A t, in the estimates of (10) for are negative and close to our four-standard-error hurdle for reliability. But this is one of the few cases where the results are sensitive to the sample period. The dd t+2 /A t slopes for (table 3) are 30% and 50% smaller than the averages, and they are less than two standard errors from zero. Moreover, the dividend regressions in table 1 say that after controlling for the variation in dividends predicted by the Lintner model, there is no marginal relation between concurrent changes in dividends and debt. There is a hint in table 2 that dividend policy affects debt decisions. In the estimates of (10) for the change in debt, the average slope on the lagged level of dividends, D t /A t, for is positive and 4.01 standard errors from zero. Thus, controlling for the many other influences on debt (earnings, investment, and the mean reversion of leverage), firms that pay more dividends are more likely to issue more debt. This is consistent with Lintner s (1956) hypothesis that firms are reluctant to cut dividends and 17

19 will issue debt to maintain dividends. The same value (0.07) of the average slope on D t /A t for is, however, only 2.60 standard errors from zero, so the evidence is shaky. Finally, confirming Jung, Kim, and Stulz (1996) and Opler and Titman (1995), our regressions to explain the change in debt, di t+2 /A t, show that firms with higher market-to-book ratios, V t /A t, are less likely to issue debt. Jung et al. and Opler and Titman interpret this result in terms of agency cost and asymmetric-information stories in which firms with growth opportunities have less leverage [e.g., Myers (1977), Myers and Majluf (1984)]. This view is contradicted, however, by the evidence in tables 2 and 3 that for firms with debt in their capital structures, target leverage and changes in leverage are not reliably related to V t /A t. Our guess is that the reliable negative relation between changes in debt and V t /A t is more about the short-term dynamics of financing decisions than about long-term targets. In the near term, firms with higher market-to-book ratios (strong firms) have more flexibility in their financing decisions and so are less likely to issue debt to finance investment. But in the longer term, there is no reliable relation between target leverage and V t /A t for firms that have debt. 18

20 V. Conclusions We study the determinants of dividends and debt, with particular interest in how firms vary dividends and debt in response to investment and earnings. Lintner s (1956) model, in which firms adjust dividends toward a target proportion of earnings, continues its long reign as a good description of dividends. Most interesting, we find no evidence that firms adjust dividends to accommodate investment. Instead, debt seems to be the residual variable in financing decisions. Investment tends to increase debt, and debt tends to absorb variation in earnings. Changes in leverage in response to earnings and investment are, however, transitory. Like others, we find that firms have target leverage ratios, and that leverage is slowly mean-reverting. 19

21 References Allen, Franklin, and Roni Michaely, 1995, Dividend policy, in R. A. Jarrow, V. Maksimovic, and W. T. Ziemba (eds), Handbooks in Operations Research and Management Science: Finance, Auerbach, Alan J., 1985, Real determinants of corporate leverage, in Benjamin J. Friedman (ed.), Corporate Capital Structures in the United States, Chicago, University of Chicago Press. Bhattacharya, Sudipto, 1979, Imperfect information, dividend policy, and the bird in the hand fallacy, Bell Journal of Economics and Management Science 10, Bradley, Michael, Gregg A. Jarrell, and E. Han Kim, 1984, On the existence of an optimal capital structure: Theory and evidence, Journal of Finance 39, Chan, Louis K.C., Narasimhan Jegadeesh, and Josef Lakonishok, 1995, Evaluating the performance of value versus glamour stocks: The impact of selection bias, Journal of Financial Economics 38, Choe, Hyuk, 1990, Intertemporal and cross-sectional variation in corporate dividend policy, Ph.D. thesis, Graduate School of Business, University of Chicago. DeAngelo, Harry, and Ronald Masulis, 1980, Optimal capital structure under corporate and personal taxation, Journal of Financial Economics 8, Fama, Eugene F., 1974, The empirical relationships between the dividend and investment decisions of firms, American Economic Review 64, Fama, Eugene F., and Harvey Babiak, 1968, Dividend policy of individual firms: An empirical analysis, Journal of the American Statistical Association 63, Fama, Eugene F., and James D. MacBeth, 1973, Risk, return, and equilibrium: Empirical tests, Journal of Political Economy 81, Harris, Milton, and Artur Raviv, 1991, The theory of capital structure, Journal of Finance 46, Jalilvand, Abholhassan, and Robert S. Harris, 1984, Corporate behavior in adjusting to capital structure and dividend Targets: An Econometric Study, Journal of Finance 39, Jensen, Michael C., 1986, Agency costs of free-cash-flow, corporate finance, and takeovers, American Economic Review 76, Jung, Kooyul, Yong Cheol Kim, and René M. Stulz, 1996, Timing, investment opportunities, managerial discretion, and the security issue decision, Journal of Financial Economics, 42, Lintner, John, 1956, The distribution of incomes of corporations among dividends, retained earnings, and taxes, American Economic Review 46, Long, Michael S., and Ileen B. Malitz, 1985, Investment patterns and financial leverage, in Benjamin J. Friedman (ed.), Corporate Capital Structures in the United States, Chicago, University of Chicago Press. 20

22 Marsh, Paul, 1982, The choice between equity and debt: An empirical study, Journal of Finance 37, Miller, Merton H., and Franco Modigliani, 1961, Dividend policy, growth, and the valuation of shares, Journal of Business 34, Modigliani, Franco, and Merton H. Miller, 1958, The cost of capital, corporation finance, and the theory of investment, American Economic Review 48, Modigliani, Franco, and Merton H. Miller, 1963, Corporate income taxes and the cost of capital: A correction, American Economic Review 53, Myers, Stewart C., 1977, Determinants of corporate borrowing, Journal of Financial Economics 5, Myers, Stewart C., 1984, The capital structure puzzle, Journal of Finance 39, Myers, Stewart C., and Nicholas S. Majluf, 1984, Corporate financing and investment decisions when firms have information the investors do not have, Journal of Financial Economics 13, Opler, Tim, and Sheridan Titman, 1995, The debt-equity choice, working paper, Boston College. Rajan, Raghuram G, and Luigi Zingales, 1995, What do we know about capital structure? Some evidence from international data, Journal of Finance 50, Titman, Sheridan, and Roberto Wessels, 1988, The determinants of capital structure choice, Journal of Finance 43,

23 Table 1 -- Regressions to Explain Changes in Dividends, (D t+2 -D t )/A t : , 28 Years The variables (Compustat data item numbers in parentheses) are as follows. A t is total book assets (6). D t is total dividends paid duri interest expense (15). V t, the total value of the firm, is its common stock price (199) times shares outstanding at the end of fiscal year [taken to be, in order and as available, redemption value (56), liquidating value (10), or par value (130)], plus total book liabilities (18 balance sheet deferred taxes and investment tax credit (35). Y t is equity income (237) plus (when available) income statement deferre credit (51). The regressions are run for each year t of the period using NYSE, AMEX, and NASDAQ firms on Compustat with data for t regression, here or in other tables. Part A of the table show means (across years) of the regression intercepts (Int) and slopes, and t-sta defined as the mean divided by its standard error [the times-series standard deviation of the regression coefficient divided by (27) ½ ]. years) of the means and standard deviations (Std) of the regression variables. The regressions require that firms pay dividends in t-2. This causes the average number of firms in the year-by-year regressions for 19 from 2612 to The regression for each year drops the extreme 0.5% of the firms in each tail of the cross-section distributions of e variables, Y t+2 /A t, D t /A t,v t+2 /A t+2, I t /A t, dy t /A t, dd t /A t, d(i t /A t ), and da t /A t. This causes the average number of firms to drop by about If the change in leverage, d(i t /A t ) = I t /A t -I t-2 /A t-2, is replaced by the change in debt, di t /A t = (I t -I t-2 )/A t, the average slopes for di t /A t are d(i t /A t ). The slopes on other variables are largely unaffected. Part A: Average Regression Coefficients and t-statistics for the Averages (V t+2 /A t+2 ) * (I t /A t ) * (dy t /A t ) * Int Y t+2 /A t D t /A t Y t+2 /A t D t /A t Y t+2 /A t D t /A t Y t+2 /A t D t /A t V t+2 /A t+2 I t /A t dy t /A t dd t /A t d(i t /A t ) da Mean t(mn) Mean t(mn) Mean t(mn) Mean t(mn) Part B: Means and Standard Deviations of the Regression Variables (V t+2 /A t+2 ) * (I t /A t ) * (dy t /A t ) * Int Y t+2 /A t D t /A t Y t+2 /A t D t /A t Y t+2 /A t D t /A t Y t+2 /A t D t /A t V t+2 /A t+2 I t /A t dy t /A t dd t /A t d(i t /A t ) da

24 Mean Std