In for a Bumpy Ride? Cash Flow Risk and Dividend Payouts

Transcription

1 In for a Bumpy Ride? Cash Flow Risk and Dividend Payouts Christian Andres, WHU Otto Beisheim School of Management, Vallendar, Germany * Ulrich Hofbaur, WHU Otto Beisheim School of Management, Vallendar, Germany This version: August 2017 ABSTRACT This paper investigates the relation between cash flow risk and dividend policy. Consistent with the notion that shareholders expect stable dividends once a payout has been established, firms with high cash flow risk are more reluctant to initiate dividend payouts. For dividend payers, we find changes in cash flow risk to have an asymmetric effect on dividend changes: While decreases in cash flow risk lower the propensity to cut dividends, firms are not more likely to increase dividends if cash flow risk decreases. Analyzing dividend smoothing, we further document a positive effect of cash flow risk on the speed of dividend adjustments, indicating that higher cash flow risk is partly passed on to shareholders via more volatile dividends. JEL classification: G35; G32 Keywords: Cash flow risk, dividend changes, payout policy * Corresponding author: Christian Andres, WHU Otto Beisheim School of Management, Chair of Empirical Corporate Finance, Burgplatz 2, Vallendar, Germany; tel ; fax ; christian.andres@whu.edu.

2 1. Introduction Developments in cash flow risk have recently gained significant attention in the empirical corporate finance literature (e.g., Han and Qiu, 2007; Bates et al., 2009; Disatnik et al., 2014). Bates et al. (2009) document a substantial increase in cash holdings, with the mean cash-toassets ratio of U.S. firms increasing from 10.5% in 1980 to an average of 23.2% in Analyzing the reasons for the surge in cash holdings, the authors find increased cash flow risk to be the primary driver. Accordingly, firms seem to have increased cash holdings as a precaution to guard against increased cash flow volatility. Bates at al. (2009, p. 1986) also note that their finding is closely related to the disappearing dividend phenomenon, and find significant differences in the development of cash holdings between dividend payers and nonpayers. Literature on the relation between cash flow risk and payout policy mainly focuses on the firm s choice of payout channel, that is, cash dividends versus share repurchases (Jagannathan et al., 2000; Lee and Rui, 2007). Guay and Harford (2000) and Lee and Mauck (2016) investigate whether investors see this choice as a signal of the permanence of the underlying cash flows. Considerably less is known, however, about the question of how a firm s decision to pay dividends and then change these dividends is related to cash flow risk. In addition, it is unclear how changes in risk are related to dividend smoothing, that is, how much of the change in cash flow risk is passed on to shareholders. The latter is of particular interest after the observation of a substantial increase in cash flow volatility since the 1980s (Bates et al., 2009) and given that a relation between risk and dividends can be expected (Hoberg and Prabhala, 2009). We empirically investigate how the riskiness of cash flows affects dividend policy, using data on U.S. firms over a period of 50 years. As indicated above, it is important to point out that term dividend policy comprises several dimensions: i) the decision to pay (or not to pay) 1

3 dividends; ii) given dividend payments, the propensity to adjust these payments; and, finally, iii) the degree of dividend smoothing. We incorporate all of these aspects when investigating the influence of cash flow risk on dividend payouts and thereby provide comprehensive evidence on the link between cash flow risk and corporate dividend policy. Our results suggest that cash flow risk affects the dividend decision across all these dimensions: First, cash flow risk is found to be negatively related to the propensity to pay dividends. This result is well in line with the implications of Myers and Majluf (1984), as ceteris paribus firms with higher cash flow risk are more likely to experience cash flow shortfalls and could thus struggle to fully finance a dividend commitment through internal cash flows. Focusing on the population of dividend payers, we also document that cash flow risk is positively related to dividend reductions and negatively related to dividend increases. Since cash flow risk proxies for the degree of transitory cash flow shocks, this result is consistent with Lintner (1956), who documents that firms adjust their dividends due to non-transitory earnings changes. In addition, we find that changes in cash flow risk affect dividend changes asymmetrically: While decreases in cash flow risk lower the propensity to cut dividends significantly, firms are not more likely to increase dividends after a decrease in cash flow risk. Conversely, higher cash flow risk seems to significantly reduce the propensity to increase dividends and increases the propensity to cut dividends. Finally, we find cash flow risk to be negatively related to dividend smoothing, a result that is in line with Leary and Michaely (2011). This finding reveals that firms do not completely absorb cash flow risk but, rather, pass some of this risk on to shareholders via more volatile dividends. We run a battery of tests to assess the robustness of our results. First, we include important controls as outlined by the prior literature. Second, we run different econometric specifications, including pooled specifications as well as fixed effects specifications, to account for unobserved heterogeneity as a confounding driver for our results (Gormley and Matsa, 2014). Third, we 2

4 replicate our analysis conducting regressions within deciles of retained earnings, idiosyncratic risk, and firm size. These variables have been shown to significantly affect payout decisions (DeAngelo et al., 2006; Fama and French, 2001; Hoberg et al., 2009). Except for the propensity to pay dividends, our results are not affected when regressing within deciles. Moreover, even when running regressions within size deciles, we find statistically significant results for 60% of the deciles. Fourth, we truncate our observation period covering the years 1997 to 2012 and control for product market fluidity as an alternative explanation for our results. Hoberg et al. (2014) document that firms facing high product market threats have lower propensities to pay out cash to shareholders. Our results with respect to the propensity to pay and adjust dividends are not affected by this combination of truncating the sample and considering product market fluidity as an additional control. Fifth, we account for the evidence of Aivazian et al. (2006), who show that bond ratings affect dividend smoothing. Again, our results are not affected. Finally, we account for the fact that earnings are a major constituent of cash flows and that cash flow variation could therefore be driven by earnings variation. Although earnings and cash flow variation are correlated, our results do not suggest that the documented relation between cash flow risk and dividend smoothing can merely be attributed to earnings variation. Our results contribute to the literature in several respects. First, we provide the first comprehensive empirical analysis on the relation between cash flow risk and dividend policy. In contrast to the prior literature, our analyses jointly include three dimensions of dividend policy: the propensity to pay dividends, the propensity to change dividends, and dividend smoothing. Second, Bates et al. (2009) document that an increase in cash flow risk has fueled the trend in cash hoarding. Moreover, they suggest that the cash holding puzzle and the disappearing dividends puzzle (Fama and French, 2001) are related. We provide supporting evidence for this conjecture and corroborate the findings of Hoberg and Prabhala (2009) by showing that cash flow risk affects the propensity to pay dividends. Third, our results establish 3

5 a direct link between the cash flow risks resulting from a firm s operating business and capital markets in the form of dividends. Hence, the risk of cash flows accruing from the underlying operations significantly impacts whether a firm pays dividends and, given dividends are paid, how it is doing so. In this regard, our study is related to Hoberg et al. (2014). We also add to the understanding of dividend changes over time by showing that cash flow risk is negatively related to dividend smoothing. Less dividend smoothing and hence faster adjustment of dividends to changes in profitability imply that firms with higher cash flow risk only partly smooth their cash flow volatility and pass some of this risk on to shareholders. The remainder of the paper is organized as follows: In the following section we summarize the literature and develop our conjectures. Section 3 describes the data and reports descriptive statistics. Section 4 presents the empirical results. Finally, Section 5 concludes the paper. 2. Literature The academic literature has devoted much attention to the determinants of dividends and time trends in payouts, mainly focusing on three aspects of corporate dividend policy: i) the decision to pay dividends (e.g., Bagwell and Shoven, 1989; Easterbrook, 1984; Fama and French, 2001; Hoberg and Prabhala, 2009; Floyd et al., 2015), ii) the decision to change dividends (e.g., Lang and Litzenberger, 1989; Nissim and Ziv, 2001; Koch and Sun, 2004; Grullon et al., 2005; Li and Lie, 2006), and iii) the propensity to smooth dividends (e.g., Lintner, 1956; Fama and Babiak, 1968; Aivazian et al., 2006; Andres et al., 2009; Leary and Michaely, 2011). In the following, we briefly summarize the most important findings of these strands of the literature with respect to our main research question and form conjectures as to how cash flow risk affects dividend policy in all of the above-mentioned dimensions. Ever since Black s (1976) seminal paper, the question as to why firms pay dividends has not been fully answered. Many theories have been developed and tested, the most popular being 4

6 the cash flow signaling hypothesis (Bhattacharya, 1979; Aharony and Swary, 1980; Miller and Rock, 1985), the free cash flow hypothesis (Easterbrook, 1984; Jensen, 1986), the catering hypothesis (Baker and Wurgler, 2004a, 2004b), as well as explanations around taxes (e.g., Bagwell and Shoven, 1989; Allen et al., 2000; Jacob and Michaely, 2017). Jagannathan et al. (2000) hypothesize that dividends are seen as an ongoing commitment and are therefore used to distribute permanent cash flows. In line with this argument, they find variation in operating income to be negatively related to the probability of paying out cash to shareholders (as either repurchases or dividends). Hoberg and Prabhala (2009) consider idiosyncratic and systematic risk as factors to explain the dividend payout behavior. They document that, after controlling for risk, catering incentives are no longer able to explain the time trends in dividend payouts. Moreover, the authors reveal that risk explains about 40% of the disappearing dividends phenomenon (Fama and French, 2001). Hoberg et al. (2014) show that the dividend payment decision is affected by product market competition. In particular, the authors document that firms facing high product market threats (which are potentially positively related to variation in operating income) are less likely to pay dividends than firms in less competitive industries are. Even though firms increasingly use debt and equity issues to finance dividends (Farre- Mensa et al., 2017), cash flows are the primary source of dividends (see also Myers and Majluf, 1984). With respect to the magnitude of the dividend commitment, Brav et al. (2005) note that maintaining (at least) the current dividend level is a first-order concern for corporations that is as important as investment decisions. Ceteris paribus, firms subject to high cash flow risk will be more likely to experience cash flow shortfalls and consequently struggle more to keep up with the current dividend level. Since dividends are seen as a long-term commitment, which, if broken, is associated with significant costs (Brav et al., 2005), firms should consider the risk of 5

7 future dividend cuts (and omissions) before deciding to initiate dividend payments. We therefore expect cash flow risk to be negatively related to the propensity to pay dividends. Once dividend payments have been initiated, firms need to decide whether to maintain, cut, or increase these payouts. According to Lintner (1956), firms adjust their dividends due to nontransitory shocks in earnings. However, since the days of Lintner (1956), share repurchases have become an alternative payout mechanism (e.g., Jagannathan et al., 2000; Skinner, 2008). Analyzing both payout channels, Guay and Harford (2000) and Jagannathan et al. (2000) document that the persistence of cash flow changes is related the choice of the payout method. Consistent with Lintner s predictions, the authors find permanent earnings shocks to be distributed to shareholders as dividends, whereas firms prefer to repurchase shares if the underlying (positive) shocks to profitability are transitory. Accordingly, firms facing high cash flow risks could be more conservative with respect to potential dividend increases, since future revisions of this decision (i.e., dividend cuts) will be associated with adverse stock market reactions (e.g., Lintner, 1956; Brav et al., 2005). We therefore conjecture that cash flow risk is negatively related to dividend increases. Regarding the propensity to decrease dividends, our expectation is similar to that regarding the relation between cash flow risk and the propensity to pay dividends: All else being equal, firms subject to higher cash flow risk will be more likely to experience significant cash flow reductions compared to their low-risk counterparts. Since reduced cash flows might not suffice to maintain the current dividend payout commitment, these firms could be forced to reduce dividend payouts. Therefore, we conjecture that cash flow risk is positively related to the propensity to cut dividends. Compared to studies that examine the propensity to pay and change dividends, the literature investigating the determinants of dividend smoothing is less developed. Lintner s (1956) seminal paper documents that, in the event of persistent earnings shocks, firms gradually adjust 6

8 their dividend toward new target payout levels. Accordingly, Lintner models the dividend adjustment pattern as a partial adjustment process. The speed of adjustment describes the time it takes firms to reach new equilibrium payout levels. All else being equal, the higher the speed of adjustment, the faster a new payout level is reached. Therefore, a high speed of adjustment indicates a low degree of dividend smoothing and vice versa. In this respect, Lintner also documents significant cross-sectional variation in the individual speeds of adjustments. Aivazian et al. (2006) were among the first to shed further light on cross-sectional variation in dividend smoothing. They document that the degree of dividend smoothing hinges significantly on whether or not a firm possesses a bond rating. While firms that have rated public debt outstanding smooth dividends significantly, firms that do not tap public debt markets do not seem to smooth their dividends. The authors argue that firms that rely on private debt have less need to use dividends as a device to alleviate agency conflicts and issues related to asymmetric information. Leary and Michaely (2011) examine determinants related to the cross-sectional variation in dividend smoothing in more detail. They document that dividend smoothing behavior is related to agency motives, asymmetric information, and financial constraints. Using earnings volatility as a proxy for asymmetric information, their evidence suggests that the standard deviation of earnings is negatively related to dividend smoothing. 1 In this context, we aim to analyze the influence of cash flow risk on the speed of adjustment. As argued above, we expect high cash flow risk to be associated with low (high) propensities to increase (decrease) dividends. However, these conjectures do not allow us to draw direct inferences on the effect of cash flow risk on dividend smoothing for two reasons: First, focusing on binary dividend events neglects the underlying magnitude of dividend increases and decreases. The magnitude of dividend changes and the underlying earnings shocks, however, 1 With respect to the earnings variable, Leary and Michaely (2011) use different measures of earnings. Some of their analyses use earnings before interest and taxes (EBIT), while other specifications use earnings before interest, taxes, depreciation, and amortization (EBITDA). 7

9 are key parameters of the dividend smoothing process (e.g., Lintner, 1956; Fama and Babiak, 1968). Second, even if there were no systematic difference in the magnitude of dividend changes and earnings shocks between high- and low-cash flow risk firms, lower propensities to increase dividends imply a higher degree of dividend smoothing, whereas high propensities to reduce dividends will have the opposite effect on dividend smoothing. Therefore, the question of whether dividend smoothing is affected by cash flow risk cannot be answered based on our expectations of the propensity to pay and change dividends. Instead, it boils down to the question of whether firms fully absorb cash flow risk or whether this risk is at least partially passed on to shareholders via more volatile dividends. As noted above, Leary and Michaely (2011) show earnings variation to be negatively related to dividend smoothing. Given that earnings are one constituent of cash flows, we it seems plausible that dividend-paying firms only partially absorb cash flow risk. Nevertheless, in the robustness section, we address whether the conjectured relation between cash flow variation and dividend smoothing is driven solely by earnings variation. 3. Data and descriptive statistics 3.1. Data Our sample selection procedure is similar to that of Hoberg and Prabhala (2009). We retrieve information on firms from the Center for Research in Security Prices (CRSP)/Compustat Merged Database. Our sample covers the 50-year period from 1963 to Regarding data availability, we follow the data screening process of Hoberg and Prabhala (2009). In detail, our sample includes those firms with fiscal year-ends in t that have the following data (Compustat data items in parentheses): total assets (6), stock price (199), and shares outstanding (25) at the end of the fiscal year, income before extraordinary items (18), interest expense (15), dividends per share by ex-date (26), preferred dividends (19), and 8

10 (i) preferred stock liquidating value (10), (ii) preferred stock redemption value (56), or (iii) preferred stock carrying value (130). Firms must also have (i) stockholder s equity (216), (ii) liabilities (181), or (iii) common equity (60) and preferred stock par value (130). Total assets must be available in years t and t 1. The other items must be available in year t. (Hoberg and Prabhala 2009, p. 84) We also exclude firms with less than $250,000 in book equity or less than $500,000 in total assets and restrict our sample to firms with CRSP share codes 10 or 11. Finally, due to differences in regulation, we drop utilities (Standard Industrial Classification code 49) and financial firms (Standard Industrial Classification code 6). Moreover, to calculate cash flow risk, we require operating income (13), interest expense (15), income taxes (16), and ordinary dividends (21) to be jointly available for at least three of the last five years. Eventually, we require the first lead of dividends per share (item 26) to be available. Applying these criteria results in 115,854 firm years corresponding to 11,152 firms. Among these observations, 52,340 firm years (4,324 firms) are associated with dividend payments in t + 1. Requiring dividend information to be available for the prior year reduces the number of firm years to 52,062, while the number of firms is unaffected. When investigating dividend smoothing behavior, we additionally apply the sample selection steps of Leary and Michaely (2011), that is, we require at least 10 positive dividend payments per firm and truncate our sample before and after the last dividend payment of each firm. This results in 43,531 firm years (2,126 firms) Descriptive statistics Table 1 presents descriptive statistics of our sample. Since we include a time series of exactly 50 years of data, we split the sample into five periods, each comprising 10 consecutive years. Panel A reports descriptive statistics with respect to the propensity to pay dividends. Consistent with the results of Fama and French (2001), we document that the propensity to pay 9

11 dividends decreases sharply from 1963 to During the first decade of our data, around 74.7% of the firm years are associated with dividend payments. This fraction falls to 29.6% during However, over the most recent 10-year period, the fraction of dividend payers increases to 33.4% again. This result is in line with those of Floyd et al. (2015) and indicates that the propensity to pay dividends has bounced back again in recent years. Corroborating our conjecture, the evidence suggests that cash flow risk (defined as the 5- year rolling standard deviation of cash flows) 2 is significantly (at least at the 1% level) lower for dividend payers. The standard deviation of cash flows is at least twice as high for non-payers during each of our five observation decades. Another striking result displayed in Panel A of Table 1 is the cash flow risk time trend. The average cash flow risk more than doubles, for both payers and non-payers, when moving from the first 10 years of data to the most recent decades. This result is consistent with Bates et al. (2009), who also reveal a significant increase in cash flow risk over the sample period. Panel B of Table 1 reports the propensity to change dividends, broken down into dividend increases and dividend decreases, as well as average cash flow risk for these groups. Compared to the propensity to pay dividends, the fractions of dividend increases and decreases show little variation over the observation period. The minimum for dividend increases amounts to 50.9% during the period and peaks during the subsequent 10 years (70.4%). On the other hand, the minimum and maximum regarding the propensity to decrease dividends amount to 11.5% ( ) and 14.6% ( ), respectively. When investigating whether cash flow risk differs for dividend increases and dividend decreases, we compare dividend increase (decrease) firm years to firm years with no dividend increase (decrease). Consistent with our conjecture, the evidence suggests that cash flow risk is 2 Detailed definitions of the variables can be found in the Appendix. 10

12 negatively related to dividend increases and positively related to dividend decreases. The differences are statistically significant at least at the 1% level for each of the tests conducted. [Insert Table 1 about here] Hoberg and Prabhala (2009) document that idiosyncratic risk is statistically and economically highly relevant for the propensity to pay dividends. In this context, it is important to note that cash flow risk and idiosyncratic risk are correlated. Figure 1 plots the annual means of cash flow risk and idiosyncratic risk over time. To increase the comparability of the two curves, we scale the two graphs so that the corresponding values in 1963, the starting year of our observation period, are set to 100%. As expected, the correlation between the mean annual cash flow risk and the mean annual idiosyncratic risk is relatively high, amounting to 68.8%. Analyzing the time trends displayed in Figure 1, we find the average cash flow risk to have increased fivefold over our observation period. 3 By contrast, idiosyncratic risk peaks at around 400% in 2000 but falls back to below 200% in Moreover, fitting a linear time trend to both curves reveals that the slopes are significantly different (p-value < 0.001; chi-squared, ). Furthermore, when investigating correlations between cash flow risk and idiosyncratic risk on a firm year basis rather than on a mean annual basis, the correlation decreases to 41.1%. Hence, we conclude that cash flow risk and idiosyncratic risk are correlated to some extent. However, the time trend is significantly more pronounced for cash flow risk and the correlation between the two is reduced when moving from annual averages to firm-level data. Nevertheless, we account for the effect of idiosyncratic risk on dividend decisions by controlling for 3 Compared to this increase, the increase within the two groups (dividend payers and non-payers) amounts to slightly more than 100% (Table 1). The larger increase over time documented in Figure 1 is due to the fact that the weights of dividend payers versus non-payers change significantly over time toward non-payers, with nonpayers being subject to higher cash flow risk for each of the five decades summarized in Table 1. 11

13 idiosyncratic risk in the subsequent analyses. Additionally, we address issues related to the correlation of cash flow risk and idiosyncratic risk in the robustness section. 4 [Insert Figure 1 about here] 4. Multivariate results 4.1. Propensity to pay dividends Table 2 reports the results for regressions investigating the determinants of the propensity to pay dividends. Specifications (1) to (3) are estimated using a pooled logistic model and specifications (4) to (6) report the results of conditional firm fixed-effects logistic estimations. Hence, the latter three specifications include only firms that switch their dividend payment status at least once. Consistent with Fama and French (2001), we include controls for the market to book ratio (MTB), asset growth, size, and profitability. 5 Moreover, we control for systematic and idiosyncratic risk (Hoberg and Prabhala, 2009), cash holdings (Bates et al., 2009), the earned/contributed capital mix, and persistence of the decision to pay dividends (DeAngelo et al., 2006). All explanatory variables are lagged by one year. Consistent with our conjecture, we find the coefficient of cash flow risk to be negative and highly statistically significant (at least at the 1% level) across all six regression specifications. Nevertheless, including controls attenuates the magnitude of the cash flow risk coefficient. Regarding the control variables, the results in Table 2 are consistent with the literature. The signs of the variables Size (+), MTB ( ), Asset Growth ( ), and ROA (+) are identical to the findings of Fama and French (2001), as well as Hoberg and Prabhala (2009). Moreover, idiosyncratic ( ) and systematic risk ( ) also corroborate the results of Hoberg and Prabhala. 4 Babenko et al. (2016) document that beyond idiosyncratic risk cash flow risk is also related to systematic risk. The correlation between cash flow risk and systematic risk amounts to 0.06 in our sample. 5 We follow related papers such as Fama and French (2001) in using NYSE-percentiles as a proxy for size. Our results are not affected when substituting the NYSE-percentile by the natural logarithm of total assets. 12

14 However, the effect of systematic risk on the probability of paying dividends is not statistically significant at conventional levels. Furthermore, the results with respect to the sign and significance of Earned Equity (+) are in line with the findings of DeAngelo et al. (2006). Finally, corroborating DeAngelo et al., we find the decision to pay dividends to be persistent (Div. Prior Year). [Insert Table 2 about here] To gain deeper insights into the relation between cash flow risk and the dividend payment decision, we report the economic relevance of the propensity to pay dividends in Table 3. Specifically, Table 3 reports estimated (i.e., predicted) probabilities based on model specifications (1) to (3) in Table 2. To calculate the probabilities, all explanatory variables, except for cash flow risk, are kept at their means. We set cash flow risk to different values ranging between the 10th and 90th percentiles. The coefficients used to calculate the probabilities correspond to those of specifications (1) to (3) of Table 2. We are thus able to calculate the (estimated) probabilities of paying dividends by only varying cash flow risk and keeping everything else constant. The difference in the propensity to pay dividends when moving the cash flow risk variable from the lowest to the highest decile ranges from (specification (1)) to (specifications (3)) percentage points. Accordingly, all else being equal, firms located in the 90th percentile of cash flow risk are between 29.9% and 86.7% less likely to pay dividends than their counterparts in the first percentile. Therefore, in line with our conjecture, we conclude that cash flow risk has a pronounced economic impact on the decision to pay dividends. [Insert Table 3 about here] 13

15 4.2. Propensity to Change Dividends Having documented that cash flow risk significantly affects a firm s propensity to pay dividends, we turn to the question of whether cash flow risk also impacts the decision to adjust dividends. To answer this question, we limit our sample to dividend payers only. Table 4 reports the results of regressions using the decision to increase dividends as the dependent variable. Again, all explanatory variables are lagged by one year. Consistent with the specifications displayed in Table 2, we include controls for the market to book ratio, asset growth, profitability, size, retained earnings, cash holdings, and risk. Moreover, we control for changes in profitability (Lintner, 1956) and the current dividend yield. As Andres and Hofbaur (2017) suggest, dividend increases tend to be persistent, which is why we also include the lagged dependent variable (in specifications (4) and (8)). Doing so, we also account for the partial adjustment of dividends (Lintner, 1956). Specifications (1) to (4) report the results of pooled logistic estimations, whereas specifications (5) to (8) display the coefficient estimates of conditional fixed effects estimations. Standard errors are clustered at the firm level. The coefficient of cash flow risk is negative and statistically significant at least at the 1% level across all model specifications. Again, adding controls diminishes the magnitude of the cash flow risk coefficient from to in the case of pooled estimations and from -8.5 to -5.9 in the case of conditional fixed effects logit estimations. Nevertheless, the statistical significance of the cash flow coefficient remains unchanged. Focusing on the control variables, we document results consistent with the literature. The variables Size, MTB, Asset Growth, and ROA (with a positive sign for dividend increases and a negative sign for dividend decreases) are consistent with the results of Hoberg et al. (2014). Moreover, the significantly positive (negative) effect of the change in return on assets (D. ROA) on dividend increases (decreases) confirms Lintner s (1956) finding that changes in dividends are driven by changes in earnings. The results further suggest that dividend increases are 14

16 persistent (Increase Prior Year), in line with the partial adjustment of dividends as described by Lintner. [Insert Table 4 about here] Table 5 investigates the economic relevance of cash flow risk on the decision to increase dividends. The procedure to obtain the estimates in Table 5 is similar to that in Table 3. Changing cash flow risk from the lowest to the highest decile results in a difference in the propensity to increase dividends ranging between (specification (1)) to -8.4 (specification (4)) percentage points. This corresponds to a relative difference between the lowest and the highest cash flow risk deciles ranging between -12.7% and -20.2%. [Insert Table 5 about here] Table 6 is structured similarly to Table 4 but analyzes the effect of cash flow risk on dividend decreases. Model specifications thus differ only in switching the dependent (indicator) variable from dividend increases to decreases and controlling for the lagged dependent variable accordingly. Across all eight specifications, the results of Table 6 suggest that dividend payers facing high cash flow risks are more likely to reduce their dividend payouts. In our investigation of the economic relevance of cash flow risk for dividend decreases, Table 7 reveals a wedge that ranges between 1.7 to 6.1 percentage points in absolute terms and from 25.1% to 62.6% in relative terms when changing the cash flow risk variable from the 10th to the 90th percentile. Therefore, consistent with our conjectures, we conclude that cash flow risk has an economically and statistically significant effect on dividend changes. Firms subject to high cash flow risk are less (more) likely to increase (decrease) their dividend payout than their counterparts subject to low cash flow risk. [Insert Tables 6 and 7 about here] 15

17 The results presented thus far investigate the relation between different levels of cash flow risk and the propensity to increase (decrease) dividends. To assuage endogeneity concerns, we lag all independent variables and measure cash flow risk over the past five years (i.e., the cash flow risk from t = -5 to t = -1 explains dividend changes in t = 0). In addition, the conditional logit regressions control for unobserved heterogeneity between firms. Based on these coefficient estimates, Tables 5 and 7 use predicted values to quantify the effect of different levels of cash flow risk on the propensity to increase dividends (Table 5) or decrease dividends (Table 7). In other words, these estimates assume that the relation between the independent variables and increases/decreases is stable across firms (i.e., the explanatory variables are set at their mean values). In the following, we analyze the effect of changes in cash flow risk on the propensity to change dividends more directly, focusing on the degree to which cash flow risk changes. Specifically, for each year of our sample period, we classify firms into 10 risk classes (i.e., deciles) based on their cash flow risk. We then create a transition matrix that captures changes in cash flow risk classes over time and use this information to define substantial increases (decreases) in risk (i.e., cash flow shocks) as firm years in which a firm moved up (down) a minimum number of cash flow risk deciles. Table 8 displays the (logit) estimation results (and z-statistics) of the relation between varying degrees of changes in cash flow risk and the propensity to increase/decrease dividends. We only report the coefficient estimates for the (cash flow) risk change variable, but all models contain the same set of independent variables as the corresponding specifications in Tables 4 and 6. Panel A of Table 8 shows the effect of increases in risk on the propensity to increase/decrease dividends and Panel B the effect of decreases in risk. We vary the definition of risk increases and decreases according to the magnitude of the risk change. Increases (decreases) are defined as such if the firm moves up (down) at least three, two, or one decile(s) of cash flow risk. Since most firm years constitute an unchanged risk decile (a histogram of 16

18 the change distribution is roughly normally shaped with excess kurtosis), these changes capture substantial cash flow changes (in decreasing order of magnitude). The results in Panel A show that increases in cash flow risk have a highly statistically significant effect on the propensity to change dividends (at least at the 1% level in all model specifications). Sharp increases in cash flow risk (i.e., at least a three-decile increase) lower the propensity to increase dividends and increase the propensity to cut dividend payouts. In addition, the coefficient estimates of the cash flow change variable decrease monotonically from large (³ 3 deciles) to small (³1 decile) changes in cash flow risk for both dividend increases and decreases. In contrast to risk increases, decreases in cash flow risk seem to have very different effects on the propensities to increase and decrease dividends. As the results in Panel B of Table 8 show, even large decreases in risk do not lead to a higher propensity to raise dividends. The coefficient estimates of the conditional fixed effects logit estimation are positive but fall short of statistical significance at conventional levels (z-values range from 1.43 to 1.59). On the other hand, decreases in risk lower the propensity to cut dividends significantly (only two out of 18 specifications do not yield a coefficient estimate that is significant at the 1% level). Firms that experience decreases in cash flow risk thus seem to avoid (additional) dividend cuts but seem reluctant to raise dividends following a period of high(er) risk. [Insert Table 8 about here] In sum, the results above indicate an asymmetric adjustment of dividend payments to changes in cash flow risk. While increases (decreases) in cash flow risk positively (negatively) affect the propensity to cut dividends, the effect of changing cash flow risk on dividend increases depends on the direction of the risk change. When cash flow risk increases, firms indeed seem to be reluctant to increase dividends. On the other hand, reductions in cash flow risk do not seem to increase the propensity to increase dividends. This result is consistent with 17

19 the managerial reluctance to (in particular) positive changes to dividends if these changes would need to be reversed in the near future, as documented in the literature (e.g., Lintner, 1956; Brav et al., These findings also have implications for the workhorse of empirical studies of corporate dividend policy, Lintner s (1956) partial adjustment model. In his partial adjustment model, Lintner implicitly assumes symmetric adjustment of dividends in reaction to changes in earnings Dividend smoothing The evidence above suggests that cash flow risk affects dividend payouts in various dimensions. Firms subject to high cash flow risk are less likely to pay dividends, and, given they are dividend payers, less (more) likely to increase (cut) dividends than otherwise similar firms with lower cash flow risk. In this section, we investigate whether and, if so, how cash flow risk affects dividend smoothing. The most common approach to measure dividend smoothing is to estimate Lintner s (1956) partial adjustment model. However, this measure has been shown to induce upward-biased estimates, particularly for low speeds of adjustment (Leary and Michaely, 2011). We therefore apply the two-step procedure developed by Leary and Michaely (2011) to estimate the speed of adjustment. The procedure is as follows: In the first step, we calculate the deviation from the target payout ratio (Dev) as the difference between the target payout ratio (which is measured as the firm s median payout ratio over the sample period) and the previous year s dividend. In a second step, we regress the change in dividends on the deviation variable calculated in the first step. The coefficient of Dev then denotes the speed of adjustment. We start by splitting our sample into two groups based on the level of cash flow risk. Accordingly, we use the median cash flow risk to split the sample. Since both cash flow risk and dividend smoothing (Leary and Michaely, 2011) change significantly over time, we conduct the median split on a within-industry year basis using the Fama and French 12-industry 18

20 classification. 6 Table 9 reports the results. For the below-median cash flow risk group, we document a smoothing coefficient of 9.4%. The smoothing coefficient for the above-median cash flow risk group is two percentage points higher, indicating a higher speed of adjustment and, therefore, less dividend smoothing. This finding corresponds to a relative difference of 21.3%. Hence, consistent with our conjecture, the results suggest that cash flow risk is negatively related to dividend smoothing. [Insert Table 9 about here] We supplement the univariate analysis with fixed effects panel regressions that control for various other variables that have been shown to affect dividend smoothing. These variables include market-to-book ratios, time trends, and size (Leary and Michaely, 2011), as well as the four-quarter structure in dividend increases documented by Andres and Hofbaur (2017). 7 To investigate whether dividend smoothing is affected by cash flow risk, we include an interaction term Dev CF-Risk. The results are displayed in Table 10. Across all model specifications, we find the interaction term Dev CF-Risk to be positive and statistically significant at least at the 1% level, which indicates faster adjustments to new dividend levels for firms with high cash flow risk. Adding controls does not profoundly affect the magnitude of the coefficient, which varies between 0.75 (specification (3)) and 0.87 (specification (6)). Moreover, consistent with Andres and Hofbaur (2017), we document that smoothing is more pronounced for firms that raise their dividends in predictable four-quarter cycles. In addition, in line with Leary and Michaely (2011), the results suggest that market-to-book ratios are negatively related to dividend smoothing. 6 Since we focus on dividend payers only, we use Fama and French s 12-industry classification rather than the 49-industry classification to obtain a fair number of observations for each industry year. 7 We account for the relation between bond ratings and dividend smoothing in the robustness section. 19

21 In total, the evidence displayed in Tables 9 and 10 corroborates our conjecture that cash flow risk is negatively related to dividend smoothing. Less dividend smoothing and hence faster adjustment of dividends to changes in profitability imply that firms with higher cash flow risk smooth their cash flow volatility less than firms with lower cash flow risk and pass more of this risk on to shareholders. [Insert Table 10 about here] 4.4. Robustness So far, we have documented that firms facing high cash flow risks are less likely to pay dividends, are more (less) likely to cut (increase) dividends, and tend to smooth dividends less than their counterparts subject to low cash flow risk. We have noted above that cash flow risk and other explanatory variables such as idiosyncratic risk could be correlated. Therefore, disentangling the effects of cash flow risk and idiosyncratic risk is a challenge. We elaborate on this issue in this section. As pointed out before, prior literature has outlined at least three important factors that may influence dividend policy but are also correlated with cash flow variation: retained earnings (De Angelo et al., 2006), idiosyncratic risk (Hoberg and Prabhala, 2009), and firm size (Fama and French, 2001). If our interpretation of a causal relation between cash flow variation and dividend policy holds, matching observations based on retained earnings, idiosyncratic risk or size should not affect our results. To address this question, we form deciles based on earned equity, idiosyncratic risk, and size. The first (last) deciles of earned equity, idiosyncratic risk, and size contain the lowest (highest) values of the respective variable. In a next step, in model specifications similar to those reported above, we investigate whether cash flow risk affects the propensity to pay, increase, and decrease dividends by running regressions within each of the generated deciles. Table 11 reports the results. In Panel A, the regression specification is equivalent to specification (3) of Table 2; in Panels B and C, it corresponds to specification (4) of Tables 4 20

22 and 6. Since matching on deciles does not completely remove variation for the variable on which the deciles have been formed, we also control for the corresponding variable when running the within-decile regressions. Panel A of Table 11 investigates the propensity to pay dividends. We document a statistically significant coefficient of cash flow risk for deciles 3 to 10. In other words, cash flow risk does not seem to explain variation in the propensity to pay dividends in the first quintile. This result could be due to the fact that only 4.26% of the observations in the lowest two deciles of Earned Equity pay dividends. Hence, the absence of retained earnings could already cause these firms to not pay dividends. Consequently, cash flow risk possesses little explanatory power. When splitting the sample according to deciles of idiosyncratic risk, we find a statistically significant influence of cash flow risk for the first nine deciles. Within the 10th decile of idiosyncratic risk, the propensity to pay dividends amounts to only 3.9%. Therefore, the high exposure toward idiosyncratic risk could predominantly shape the dividend payment decision for these firms. When running regressions by size deciles, we find the evidence to be slightly more mixed: We document a significant coefficient of cash flow risk for six of the 10 deciles. Panels B and C focus on the propensity to increase and decrease dividends, respectively. The results reveal a statistically significant coefficient for all decile regressions, the exception being the ninth size decile in Panel C. Hence, our results suggest that the relation between cash flow risk and dividend policy holds when matching firms based on deciles of earned equity and idiosyncratic risk. Regarding size deciles, the results are somewhat more mixed with respect to the propensity to pay dividends. Nevertheless, we document significant results for the vast majority of the deciles. [Insert Table 11 about here] A recent paper by Hoberg et al. (2014) shows that product market threats have an economically and statistically significant influence on dividend policy. In detail, firms facing 21

23 high product market threats are found to be less likely to pay dividends. Moreover, given dividend payments, firms seem to be more (less) likely to cut (increase) dividends than their counterparts exposed to lower product market threats. Hence, product market threats are an important factor of influence on dividend policy that we did not account for so far. Table 12 addresses this issue. In regressions similar to those reported in Tables 2, 4, and 6, we control for product market threats when investigating the propensity to pay dividends (specifications (1) and (2)), increase dividends (specifications (3) and (4)), and decrease dividends (specifications (5) and (6)). Since data on product market fluidity are not available before 1997 and after 2011, we truncate the time series accordingly. The evidence in Table 12 suggests that our results are robust to the inclusion of product market threats. After controlling for product market fluidity, firms facing high cash flow risk are still less likely to pay dividends, more likely to cut dividends, and less likely to increase dividends. [Insert Table 12 about here] Aivazian et al. (2006) document that bond ratings are a key determinant of dividend smoothing. Firms relying on public debt markets implement dividend smoothing consistent with the Lintner model, whereas firms using private debt seem to smooth their dividends to a lesser extent. Since data on bond ratings are only available from 1985 onward, our previous analysis did not account for a potentially confounding effect of bond ratings. Table 13 reports the results for the relation between cash flow risk and dividend smoothing and controls for bond ratings. Accordingly, the time series comprises the years The speed of adjustment coefficient in specification (1) (8.4%) is lower than that of specification (1) in Table 9, which covers the period from 1973 to This finding is consistent with the notion of a positive time trend in dividend smoothing (Leary and Michaely, 2011). The results with respect to the relation between cash flow risk and dividend smoothing remain unchanged. Moreover, consistent with Aivazian et al., we find bond ratings to be positively related to dividend 22

24 smoothing (specifications (3) and (4)). However, when controlling for other factors that affect bond ratings, we do not find a statistically significant relation between bond ratings and dividend smoothing. 8 [Insert Table 13 about here] Regarding dividend smoothing, Leary and Michaely (2011) demonstrate that earnings variation is negatively related to dividend smoothing, a result that is consistent with their asymmetric information hypothesis. We provide evidence that cash flow risk is negatively related to dividend smoothing. However, since earnings are an important constituent of cash flows, earnings and cash flows are correlated. The same should hold when moving from levels to the variations of the two variables. Therefore, our results with respect to the effect of cash flow risk on dividend smoothing could simply proxy for the effects of earnings variation on dividend smoothing. To elaborate on this issue, we separately and jointly account for the effects of cash flow risk and earnings risk on dividend smoothing. We define earnings in four different ways: EBITDA (variable ROA (1)), income before extraordinary items (ROA (2)), income before extraordinary items, available for common stockholders (ROA (3)), and EBIT (ROA (4)). Table 14 displays the corresponding results. When moving from levels to standard deviations over the recent five-year interval, the correlations between cash flow risk and ROA risk amount to 0.67 (ROA (1)), 0.59 (ROA (2)), 0.59 (ROA (3)), and 0.64 (ROA (4)). For the earnings risk variables based on EBIT and EBITDA, we document a negative and significant relation to dividend smoothing. However, this result no longer holds when cash flow risk is included among the independent variables (specifications (2) and (8)). The results with respect to the two alternative earnings definitions based on income do not reveal any significant results. Moreover, when we jointly include cash flow risk and earnings risk, the prior results with respect to the 8 When we only include Dev, Bond Rating, and the corresponding interaction term, the coefficient of the interaction term is negative and significant at the 5% level (results not tabulated for brevity). 23