We develop a theory of dividends that relaxes the market efficiency assumption of the dividend irrelevance proof. Our approach is in the spirit of earlier theories of dividends that isolated and relaxed other assumptions of the proof. The essence of catering is that managers give investors what they want. Applied to dividend policy, catering implies that managers will tend to initiate dividends when investors put a relatively high stock price on dividend payers, and tend to omit dividends when investors prefer nonpayers. A simple model formalizes the key tradeoffs between maximizing fundamental value and catering.
Our empirical tests focus on the central prediction of the model, a time series relationship between dividend policy and the relative stock price of current payers and nonpayers. We test this relationship using four stock market measures of the demand for dividend payers. The aggregate propensity to initiate dividends is significantly positively related to each of them, and the propensity to omit dividends is significantly negatively related to some of them. The results are economically substantial: The dividend premium – the difference between the average market-to-book ratios of payers and nonpayers – explains an impressive three-fifths of the time variation in the propensity to initiate.
After an analysis of alternative explanations, we conclude that catering is the most natural explanation for these results. We then ask which set of investors generates time variation in the dividend premium. We do not find strong evidence for tax clienteles, transaction costs, or institutional investment constraints. Instead, the close connection between the closed-end fund discount and the dividend premium variable suggests that investor sentiment may play a significant role in the demand for dividends.
The results suggest several avenues for future research. One is to determine more precisely what psychological and institutional phenomena combine to induce the categorization of payers, and what forces govern the relative demand across payer and nonpayer categories. Another interesting question is to determine more precisely how managers and firms benefit from catering. Managers may trade in their own accounts around catering-motivated decisions, or issue equity at advantageously high prices, in the spirit of Jenter (2001). In light of the model’s suggestion that the net gain to catering includes both an immediate announcement effect and a longer-term recategorization effect, however, and the ambiguities that always arise in measuring abnormal returns (discussed recently in the context of dividend initiations by Boehme and Sorescu (2002)), the ultimate gains to catering will be difficult to pin down.
Catering may also be helpful in understanding recent time series patterns in payout policy. For instance, Fama and French (2001) document that the propensity to pay has been declining over the past few decades. According to the dividend premium variable studied in this paper, investors have favored nonpayers over roughly the same period dividends have been disappearing. To the extent that catering helps to explain why dividends have been disappearing, it may also explain why repurchases have been appearing, as documented by Grullon and Michaely (2002). Accumulating cash has to be paid out somehow. Dividend catering motives may help explain why the switch to repurchases occurred when it did. We are exploring these hypotheses in some work in progress.
This appendix describes the simulations which generate the bias-adjusted coefficients and p-values reported in Table 6. As discussed by Stambaugh (1999), a small-sample bias arises when the explanatory variable is persistent and there is a contemporaneous correlation between innovations in the explanatory variable and stock returns. For example, in the following system
the bias is equal to
where the hats represent OLS estimates. Kendall (1954) shows the OLS estimate of d has a negative bias. The bias for OLS b is therefore of the opposite sign to the sign of the covariance between innovations in dividend policy and returns.
The sign of this covariance is not obvious a priori (unlike when the predictor is a scaled-price variable). To address the potential for bias and conduct inference, we use a bootstrap estimation technique. The approach is identical to Baker and Stein (2002) and is similar to that used in Vuolteenaho (2001), Kothari and Shanken (1997), Stambaugh (1999), and Ang and Bekaert (2001). For each regression in Table 6, we perform two sets of simulations.
The first set generates a bias-adjusted point estimate. We simulate (A1) and (A2) recursively starting with X0, using the OLS coefficient estimates, and drawing with replacement from the empirical distribution of the errors u and v. We throw out the first 100 draws (to draw from the unconditional distribution of X), then draw an additional N observations, where N is the size of the original sample. (For the cumulative three-year regressions, the number of additional draws is one third the size of the original sample, since it contains overlapping returns.) With each simulated sample, we re-estimate (A1). This gives us a set of coefficients b*. The bias-adjusted coefficient BA reported in Table 6 subtracts the bootstrap bias estimate (the mean of b* minus the OLS b) from the OLS b.
In the second set of simulations, we redo everything as above under the null hypothesis of no predictability – that is, imposing b equals zero. This gives us a second set of coefficients b**. With these in hand, we can determine the probability of observing an estimate as large as the OLS b by chance, given the true b = 0. These are the p-values in Table 6.