Very preliminary. Comments welcome. Contributed capital versus retained earnings: tax differences and value implications.

Transcription

1 Very preliminary. Comments welcome. Contributed capal versus retained earnings: tax differences and value implications October, 2002 by Siyi Li Douglas A. Shackelford and Jacob K. Thomas Columbia Business School, New York, NY Universy of North Carolina, Chapel Hill, NC We thank Deen Kemsley, Ed Maydew and Terry Shevlin and workshop participants at UNC- Chapel Hill and Rutgers Universy for helpful comments.

2 Contributed capal versus retained earnings: tax differences and value implications Abstract The investor-level tax rate on dividends that is capalized in share prices is an important issue that remains unresolved. Recent research has examined whether the tax code feature that exempts dividend taxes on return of contributed capal can be used to infer the extent of dividend tax capalization. Intuive arguments have been made by Harrris and Kemsley (1999) to generate predictions for how the coefficients estimated from regressions of stock prices on book value of equy and trailing earnings should vary wh the proportion of equy represented by contributed capal. While the observed results suggest that share prices capalize dividend taxes at a fairly high rate, some have questioned the regression specification used, since these predictions are not formally derived, and also raised concerns about empirical issues relating to the results. We return to the primive stream of taxable and tax-exempt dividends and derive an alternative specification. Our results do not support the hypothesis that stock prices distinguish between contributed capal and retained earnings. We are, however, unable to offer evidence relating to the underlying question regarding the extent of dividend tax capalization. 1

3 Contributed capal versus retained earnings: tax differences and value implications 1. Introduction Determining the extent to which shareholder taxes affect equy prices is currently the most active, important, and controversial area in tax research in accounting. Arguments have been offered for both the full capalization of dividend taxes (e.g. King, 1977) as well as the opposing view that ltle or no dividend taxes are capalized in share prices (e.g. Miller and Scholes, 1978). A difficulty in resolving this question revolves around the inabily to observe directly the price effects of taxes on dividends and capal gains. This occurs in part because the amount of taxes that shareholders expect to pay on future dividends is unobservable. As a result, a host of creative investigations have arisen. All share the same goal of providing a window onto the relation between share prices and shareholder taxes. They include tests of share price movement around ex-dividend dates (e.g., Guenther (2000), changes in tax legislation (e.g., Guenther and Willenborg, 1999), and changes in corporate control (e.g., Landsman and Shackelford, 1995; Erickson, 1998). Others explo instutional distinctions, such as the dividends-received deduction (Erickson and Maydew, 1998), the ESOP interest exclusion (e.g., Shackelford 1991) or the kink in capal tax rates when shares qualify for favorable longterm treatment (e.g., Blouin, et al, 2002). Another set of papers explos the fact that shareholders are only taxed on a portion of the distributions they receive from corporations. When a company distributes property to shareholders, the distributions are taxable to the extent eher accumulated or current earnings and profs (E&P) exist. Once E&P are exhausted, distributions to the extent of contributed capal are returned tax-free. Harris and Kemsley (1999), hereafter referred to as HK, extend the dividend tax capalization view to incorporate taxable and tax-free distributions, and predict that the coefficients from a regression of market value on book value and trailing earnings should vary across firms wh different amounts of contributed capal and also vary across tax regimes. Their 1

4 results, as well as those in follow up research conducted by Collins and Kemsley (2000) and Harris, Hubbard, and Kemsley (2001), find patterns of coefficient variation that are consistent wh full dividend tax capalization, at a rate close to the prevailing top statutory rate on personal ordinary income. These findings, referred to hereafter as the CHHK findings, have met skepticism and cricism. For example, Hanlon, et al. (2002) and Dhaliwal, et al. (2002) question whether other features of the tax code effectively blunt the main thrust of the HK hypothesis. Specifically, most firms cannot distribute nontaxable return of capal because the contributed capal is often buried under a substantial amount of accumulated or current E&P. And if contributed capal is typically returned as a liquidating dividend, the distinction between contributed capal and E&P is lost since all liquidating distributions are treated as share repurchases and taxed at capal gains rates. In addion, they raise concerns about omted correlated variables, and suggest that the CHHK results are not robust. These cricisms have generated a series of rebuttals (e.g., Kemsley, 2002a, and 2002b). At this point, a consensus exists that HK identify a clever way to test dividend tax capalization. However, there is disagreement about what can be learned from the empirical tests, because the conflicting interpretations of the different results are not easily reconciled. 1 This paper asserts that the absence of a formal derivation for the HK predictions regarding the regression coefficients casts doubt on the validy of the analyses conducted so far. Specifically, as indicated in Hanlon et al. (2002), efforts to derive such a formal link suggest that the coefficients on book value and trailing earnings need not vary as predicted by HK. Rather than rely on the HK regression specification, we return to the primive stream of taxable and tax-exempt distributions and derive an alternative regression specification that links the HK hypothesis to predicted patterns for our regression coefficients. This alternative specification, 1 Consider, for example, the different interpretations in the lerature relating to footnote 12 in HK that discusses how the coefficient on trailing earnings should vary wh REBV. That footnote reads as follows: The posive effect of retained earnings on residual income only partially offsets the negative effect of retained earnings on the value of reported book value, except at the lim (i.e. in a perpetuy), where the two effects are perfectly offsetting. 2

5 which uses analyst forecasts of future flows, rather than trailing earnings, has the added benef of being based on a more general valuation model than that implied by the HK model. 2 Our findings contradict the CHHK results. The coefficients we estimate do not follow the patterns predicted by our specification of the HK hypothesis, and we observe no evidence of the valuation multiples reflecting the taxable/nontaxable distinction for corporate distributions. Note, however, that while we reject the HK extension of the dividend tax capalization hypothesis, our methodology is not suable to answer the underlying question that generated this lerature: what is the dividend tax rate capalized in share prices? Some readers may have anticipated our results based on the belief that the stock market does not anticipate nontaxable distributions for most firms. The distinction between taxable and nontaxable distributions may be irrelevant for the vast majory of firms, because of the substantial amount of E&P that would need to be distributed before contributed capal can be returned. To provide evidence on this issue, we examined the taxabily of dividends for all firms included in the Tax Supplements attached to the Standard and Poors Annual Dividend Record for the years 1997 through We were unable to find a single firm wh nontaxable distributions of contributed capal. To provide a setting that increases the likelihood of observing behavior consistent wh the HK hypothesis, we are currently in the process of identifying a subset of firms where nontaxable distributions are not unusual. One promising possibily is the utily industry in the early to mid 1980 s. These firms reported accounting profs and made substantial distributions to shareholders. E&P, however, were depressed by the generous depreciation allowances provided for tax purposes in 1981 under ACRS. In conclusion, while the taxable/nontaxable distribution distinction suggested by HK may provide a viable opportuny to test dividend tax capalization, the results of prior research are unreliable because the tests are based on a regression model that was not derived from the theory 2 The HK model relies on the Ohlson (1995) valuation model that applies only for the special case where residual income (described in the next section) follows an AR(1) time-series process. 3

6 proposed. Whereas we are able to derive a different specification that is directly linked to the HK hypothesis, we are concerned that our tests may not provide a reasonable test of that hypothesis, because nontaxable returns of capal are so rare that a comprehensive sample of firms is unlikely to provide evidence consistent wh that view even if were true. We hope to identify a reasonably large subset of firms that are more likely to be affected by the HK hypothesis and repeat our investigation on those firms. The remainder of this paper is laid out as follows. Section 2 contains a description of the differing views underlying the debate and derives the empirical specification we use to test the HK hypothesis. We describe in Section 3 our sample and the variables used, and our primary results are detailed in Section 4. Section 5 contains sensivy analyses and Section 6 concludes. 2. The competing views and implied valuation relations Before considering the HK extension to the dividend tax capalization hypothesis, we examine the valuation of expected future dividends whout separating the portion of dividends considered return of contributed capal. We begin wh the typical relation between stock prices and the present value of expected pre-tax future dividends and express in terms of after-tax dividends and an after-tax cost of equy capal. (Throughout the paper, after and before tax refer to after and before investor taxes, not corporate taxes.) 1( 1 t) ( 1 + k) d 2 ( 1 t) 2 ( 1 + k) d p = +.. (1) 0 + where p 0 = current price, d n =expected dividends n years out, t=expected investor tax rate on dividends paid, and k=cost of equy, after investor taxes. Figure 1 summarizes alternative views regarding the extent to which dividend taxes are capalized in stock prices; i.e., the extent to which the expected dividend tax rate (t) deviates from zero. For convenience, the discussion ignores capal gains taxes, based on the assumption that they are relatively low. At one extreme, Miller and Scholes (1978) and others argue that 4

7 even though dividends are taxed, the dividend tax rate capalized in stock prices is approximately zero. This low implied tax rate is due to an assumed clientele effect, where highdividend-paying stocks are held by low tax rate investors. At the other extreme, the new view of dividend taxation (e.g., King, 1977, and Bradford, 1981) assumes a fairly high dividend tax capalization rate, close to the top tax rate on dividends (t d ). This full capalization view generates a variety of interesting implications relating primarily to the lower cost of internal equy (retained earnings) relative to external equy (new issues). In between, various features of the tax code that result in lower dividend taxes on reorganizations/liquidations are invoked to argue for a partial capalization view: the dividend tax rate capalized reflects a mix of the higher tax rate on dividends paid prior to termination and a lower tax rate on terminal dividends. Equation (1) can be expressed in terms of book values and residual incomes as described by equation (3) below, provided the clean surplus relation in equation (2) below holds. bv ( 1 t) n = bv 1 + ni d (2) n ni n n ( 1 + k) ( 1 t) n n 1 p = + 0 bv0 (3) n n= 1 (1 t) k * bv where bv n = expected book value of equy at the end of year n, and ni n =expected net income n years out. The HK hypothesis extends the dividend capalization view and argues that contributed capal and retained earnings should be treated differently because returning contributed capal via dividends is tax exempt, whereas payment of earnings is not. Adjusting equation (1) to incorporate this feature results in equation (4) below. ( 1 t) ( 1 + k) ( 1 t) ( 1 + k) dtax1 + cc1 dtax2 + cc2 p = +.. (4) where cc n = tax-exempt portion of dividends in year n, dtax n =taxable portion of dividends in year n. Assuming the addional clean surplus relations detailed in equations (5a) and (5b) below, can be shown that current share prices under the HK hypothesis are described by equation (6). 5

8 cc n = cc 1 cc (5a) n n re n = re 1 + ni dtax (5b) n n n *[ ccn 1 + ren 1( 1 t) ] n ( 1 + k) [ ( )] ni n (1 t) k p = cc0 re0 1 t (6) n= 1 where cc n = expected contributed capal at the end of year n, cc n = tax-exempt portion of dividends in year n, re n = expected retained earnings at the end of year n, dtax n =taxable portion of dividends in year n. Equation (6) is the starting point for the empirical relations tested in CHHK. Since the second term in that relation is not easily estimated at the firm level, HK turn to the pre-tax valuation relation derived in Ohlson (1995) for the special case where pre-tax residual income follows an AR1 time-series process. This assumption allows the sum of pre-tax current book value and pre-tax future expected residual income to be replaced by the following function of current book value (bv 0 ), trailing earnings (ni 0 ), and trailing dividends (d 0 ). p 0 γ 1bv0 + γ 2ni0 + γ 3d 0 = (7) where the coefficients are functions of the cost of equy (k) and the persistence (ω) of pre-tax residual income, representing the parameter that describes the AR1 process. HK then make intuive extensions to suggest that equation (6) can similarly be transformed to generate an after-tax analog of equation (7). Finally, HK extrapolate that logic to suggest that the coefficients α 1 and α 2 from the following valuation relation should increase and decrease wh the ratio of retained earnings to book value (REBV), respectively (see, for example the discussion after equation (4) in HK). p 0 = α 0 + α1bv0 + α 2ni0 + ε (8) While the intuive arguments made in HK to justify their predictions sound reasonable, closer inspection reveals that there is no apparent way to derive those predictions from equation (6). For example, if pre-tax residual income follows an AR1 process and the pre-tax discount rate 6

9 is obtained by setting share price equal to the present value of pre-tax dividends (see footnote 3 below), then α 1 and α 2 should both be 1, and should not vary wh REBV. Rather than assume a certain time-series process for eher pre-tax or after-tax residual income, we return to the more general setup in equation (6) and consider ways to separate the differential value effects of CC and RE. One simple approach is to ignore the (1-t) terms in equation (6), estimate the following relation that combines pre-tax book values and net incomes wh the after-tax cost of capal, and examine how the coefficients δ 1 and δ 2 vary wh REBV. = δ + δ + δ nin k * bvn 1 p = δ + δ + δ + ε ( ) 0 0 1bv bv n 0 2PVFRI (9) n= k Note the importance of the inconsistent combination of pre-tax book values and earnings wh after-tax cost of capal. Had we maintained consistency and used only pre-tax values or only after-tax values, the coefficients δ 1 and δ 2 would equal one, and would not vary wh REBV. 3 By using after-tax discount rates on pre-tax book values and earnings, rather than their after-tax counterparts, we generate regressors that are higher than their true values (in equation (6)). That is, the sum of bv 0 and PVFRI will be higher than current price (p 0 ), and the values of δ 1 and δ 2 adjust for this upward bias by taking on values less than one so that the adjusted sum of bv 0 and PVFRI equals current price. The important innovation we provide is that the upward bias in bv 0 and PVFRI is a function of REBV and as a result the coefficients δ 1 and δ 2 are also functions of REBV. In essence, even though PVFRI as defined above is not an economically meaningful concept, because combines pre-tax book values and earnings wh after-tax discount rates, serves as a convenient vehicle to test the HK hypothesis, by creating a relation between the coefficients and REBV. To generate predictions under the HK hypothesis for how δ 1 and δ 2 vary wh REBV, we compare the terms in equation (9) wh their counterparts in equation (6). The ratio of bv 0 in 3 The pre-tax cost of capal is defined as the discount rate that makes current price equal to the present value of expected pre-tax dividends. This cost of capal is a function of expected dividend payout policy, and is in general not equal to k/(1-t), where k is the after-tax cost of capal and t is the dividend tax rate. 7

10 equation (9) to the corresponding after-tax book value term in equation (6) increases from a minimum of 1 for REBV=0 to a maximum of 1/(1-t) for REBV=1. Similarly, the ratio of the future residual income term in equation (9) to s after-tax counterpart in equation (6) decreases from a value that is more than 1/(1-t) for REBV=0 to a minimum of 1/(1-t) for REBV=1. The specific value of the latter ratio for REBV=0 is a function of the growth rates for net income, RE and CC. To understand better the predicted pattern of variation for δ 1 and δ 2, under the HK hypothesis, we conducted a spreadsheet simulation. We begin wh $1 of contributed capal and consider values of retained earnings varying between $0 and $10, which result in REBV values that range between 0 and For all cases, we assume a dividend payout of 40 percent, a dividend tax capalization rate of 40 percent, and a cost of equy capal of 9% (based on an after-tax risk-free rate of 6 percent). We then consider different ROE values and project future book values (contributed capal and retained earnings), earnings, and dividends for the next 5 years. After that we assume a constant growth in residual income of 3 percent (equal to the expected rate of inflation for the assumed risk-free rate). These assumptions allow us to calculate the variables in equations (6) and (9), which in turn allows us to calculate the values for δ 1 and δ 2 that would be appropriate in each case. The results of our simulation are presented in Figure 2. As predicted, the patterns for δ 1 reported in Panel A indicate a linear decline from a maximum of 1 for REBV=0 to a minimum of 0.6 (equal to 1 minus the tax rate of 40 percent). The same pattern is observed for all assumed values of ROE. The patterns for δ 2, reported in Panel B, are also as predicted: they increase from a low value when REBV=0 to a maximum of 0.6 for REBV=1. As expected, the increasing pattern varies across ROE values, and our simulation suggests that δ 2 increases wh ROE. As 8

11 will be shown later, there is a posive relation between REBV and ROE, and this relation is likely to create a bias in favor of the HK prediction that δ 2 should increase wh REBV. 4 The opposing view is represented by the null hypothesis that δ 1 and δ 2 do not vary wh REBV. Although this view is inconsistent wh the HK extension of the dividend tax capalization view (described by equation (4)), is not inconsistent wh capalization of taxes on future dividends (described by equation (1)); i.e., stock prices could capalize taxes on future dividends but ignore differences between contributed capal and retained earnings for a variety of reasons. The main argument underlying the opposing view is that even though contributed capal is tax-exempt when paid out as dividends, is typically buried under retained earnings and therefore not easily accessed. That is, firms facing investment opportunies that are below the required rate of return for contributed capal, cannot return contributed capal whout first depleting all available retained earnings. Given the potential dislocation caused by such a drastic event, firms are unlikely to undertake such distributions. If so, investors expect most of the contributed capal to be returned at the terminal date, except possibly for firms wh extremely low values of REBV that have relatively easy access to contributed capal. If contributed capal is returned at the terminal date, the taxation of those dividends is likely to ignore differences between contributed capal and retained earnings. In essence, investors ignore the amount of contributed capal because does not impact the taxes expected to be paid on future dividends. 3. Sample formation and variable definions Our sample is generated from the intersection of IBES files (2002 edion), containing monthly consensus forecasts of future earnings up to the end of 2001, and Compustat s annual industrial, full coverage, and research files (2001 edions), containing year-end values of relevant reported financial statement data up to We require that firm-years included in our 4 We attempted to construct portfolios that have similar ROE (to remove the impact of the underlying relation between REBV and ROE) but discovered that our portfolios exhibed a high degree of collineaery which effectively made our regressions meaningless. 9

12 sample satisfy the following condions: 1) IBES consensus eps forecasts for year +1 and +2 and long-term growth forecasts are posive and long-term growth forecasts lie between 0 and 0.5; 2) stock price and number of shares outstanding (both adjusted for stock spls) from IBES are not missing; 3) book value (DATA60), retained earnings (DATA36), net income (DATA172), and number of shares outstanding (DATA25) from Compustat are posive. The IBES estimates are taken from April each year, and the financial statement ems from Compustat correspond to the fiscal year ending just before that April. 5 Following HK, we also deleted observations in the top 1 percent of the distributions of per share price, book value, and retained earnings and observations for which retained earnings to book value ratio (REBV) exceeds 1 or the market-tobook ratio is above 10. The resulting sample consists of 29,003 firm-years. The regressions we estimate are derived from equation (9). Given differences between our sample and those examined in the CHHK papers (our sample period is more recent and we exclude firms whout earnings forecasts on IBES), we also estimate the primary regressions in HK and HHK (corresponding to equation (8)) to confirm that those results are observed in our sample. The variables necessary to estimate these equations are described briefly below as they are introduced (details of all variables are provided in the Appendix). To estimate PVFRI we need forecasts of earnings and book values for all future years, as well as an estimate of the firm-specific cost of equy. We use IBES eps forecasts to generate numbers for years +1 and +2, and then use the IBES eps growth forecasts to generate forecasts for years +3, +4, and +5. To estimate future book values, we assume that the current dividend payout is maintained in the future to generate dividend forecasts from earnings forecasts, and then add forecasted earnings and subtract forecasted dividends to prior year book values to generate year-end book values for years +1 to +5. (see equation (2)). Finally, we assume that the after-tax cost of equy capal (k) equals the yield on long-term AAA municipal (tax-exempt) 5 This procedure allows us to use stock prices collected at the same point in time each year, thus reducing the impact of whin-year fluctuation in stock prices. However, causes variation in the period between the fiscal year-end and the date of the share price. To reduce the impact of this problem, we collect share prices as of April, immediately after December year-end firms (representing the most common year-end) have disclosed their annual reports. 10

13 bonds plus a risk premium of three percent. 6 Having obtained estimates for residual income in years +1 to +5, we then assume that residual income grows thereafter at a constant rate (g), equal to the expected inflation rate (which is proxied by the 10-year risk-free rate less 3 percent). 7 Our estimates for δ 1 and δ 2 are potentially biased because of errors in our estimates for k and g. To avoid generating firm-specific estimates, which could be associated wh large measurement error, we estimate our valuation regressions on portfolios and use market-level estimates for k and g. We believe that the market-level estimate for k is reasonably appropriate for portfolios, and we construct our portfolios to have similar forecast growth (equal to the median forecast growth for our sample in that year) to ensure that market-level estimates for g are also reasonably appropriate. The steps followed to form these portfolios are described next. 1. We first form quintiles based on REBV. 2. To form portfolios wh approximately the same g, we use the eps growth forecast by analysts (for years +1 to +5). The median value of this growth rate each year, g med, is then used as the target value of growth for each portfolio. 3. Each REBV quintile group is spl into low and high growth subgroups based on the median value of the growth proxy that year. 4. Each of the low and high growth subgroups in each of the five REBV quintiles are spl into 20 random equal-sized portfolios. 5. For each of the 200 portfolios, we calculate (aggregate) market value (MV), book value (BV), and earnings for years +1 to +5 (E1-E5). 6. We then calculate implied growth for each of the 200 portfolios as E 5 / E4. 7. We randomly group one low growth and one high growth portfolio into pairs, and then find the weight (w) such that the portfolio s growth equals g med. 8. Use w to get MV, BV, NI and PVFRI for each of the 100 portfolios each year. 6 7 Comparison of the tax-exempt AAA municipal bond yields wh the corresponding taxable AAA corporate bond yields indicates an implied tax rate on interest of about 25 percent. If so, a 3 percent after-tax equy premium corresponds to a before-tax equy premium of 4 percent. This level of equy premium is consistent wh recent estimates mentioned in the lerature. This 3 percent estimate of the real risk-free rate is obtained from yields associated wh inflation-protected Treasury bonds (TIPS). 11

14 Since the market-level parameters, k and g, are not publicly observable, we recognize that our assumed values could be systematically over or understated. To investigate the extent of any bias caused by errors in our assumed values for k and g, we repeat the analysis for alternative assumed values in the viciny of our primary estimates (k equals the after-tax risk-free rate plus 3 percent and g equals the after-tax risk-free rate minus three percent). For k, we consider different values of the after-tax risk premium between 0 and 5 percent, and for g, we consider different growth rates that subtract between 1 and 5 percent from the after-tax risk-free rate. There are other factors that could potentially bias our estimates, especially those for δ 2. First, using too high or too low an assumed value for k and g would bias our PVFRI measure and thereby bias our estimates for δ 2. Second, in addion to systematic over or underestimates of PVFRI, our measure of PVFRI contains considerable random measurement error, which in turn could bias our estimates for δ 2. Finally, given the role of book value in computing PVFRI, we suspect that errors in the assumed values of k and g affect not only our estimates for δ 2, but also those for δ 1. As a result, we do not examine the levels of our estimates of δ 1 and δ 2, and are concerned instead wh variation in those levels across the different REBV quintiles. And given the greater potential for bias in estimates of δ 2, we assign more weight to observed patterns of variation for δ 1, relative to that for δ 2. Finally, based on our earlier discussion of the difficulty of distributing contributed capal when is buried under large amounts of retained earnings, we expect the changes to be greatest for the lowest REBV quintiles, and focus more on the change in δ 1 and δ 2 among the lower REBV quintiles. 4. Primary results Table 1 provides the following univariate statistics relating to the distributions of the different variables we consider: mean, standard deviation, minimum and maximum, and the 1 st, 25 th, 50 th, 75 th, and 99 th percentiles. Panel A contains firm-level data, presented on a per share basis, and Panel B contains portfolio-level data that is aggregated across all firms in each 12

15 portfolio after applying the fractional weights (w and 1-w). Note that the distributions here capture variation both across firms and over time. Examination of the median values of each distribution in Panel A indicate a price per share (P) of $22.79 and a book value per share of $10.34, which is spl approximately evenly between retained earnings ($5.05) and contributed capal. The median REBV of 0.53 also indicates that retained earnings is about one half of total book value. The median ratio of contributed capal to market value (CP) is only 0.21, much lower than the 50 percent noted for book value, because the ratio of book value to market value (BP) is generally less than one for this sample, evidenced by a median of 0.5. When replicating the HK analyses, where all variables including the intercept are scaled by the number of shares outstanding, we use the inverse of the number of shares outstanding (INT) as a regressor and suppress the intercept. The median INT value of 0.05 corresponds to a median number of shares outstanding of 20 million. The median trailing net income per share as reported (NI) is $1.32. The median ratio of eps forecast for next year to book value per share (ROE1) is 15 percent. Since PVFRI is based on pre-tax book values and net incomes, is larger than the true present value of after-tax residual income, and the sum of book value plus PVFRI should exceed price per share. As expected, the median price per share of $22.79 is less than the sum of the median book value of $10.34 and the median PVFRI of $ The median ratio of dividend paid to eps (PAYOUT) is 0.14, and the median beta (estimated over the prior 60 monthly returns) is 1.0. The median ratio of two-yearout forecasted eps to current share price less the pre-tax risk-free rate (EP2), a variable that is inversely related to anticipated growth in residual income, is 0.08, and the median five-year out forecasted growth in eps (LTG) is 15 percent. The distributions for the same variables at the portfolio level, presented in Panel B, indicate much larger magnudes for the medians of the underlying variables (such as P, BV, and so on) since they represent the aggregate values for each of the 100 portfolios formed each year. Also, since the number of firms in each portfolio increases over the sample period, the dispersion of the variables reported in Panel B is exaggerated further. The variable INT (inverse of the 13

16 number of shares outstanding) is dropped because is undefined at the portfolio level. The median values of the ratios in Panel B are generally similar to those reported in Panel A except for PAYOUT, which is slightly higher because the larger firms that generally pay more dividends per dollar of earnings exert greater influence in portfolios. Note that while the median value for LTG in Panel B is similar to that in Panel A, the dispersion in Panel B is much lower because the portfolios are formed such that there is no variation in this growth measure in each year. We present in Table 2 the Pearson and Spearman correlations for different pairs of the variables. To allow comparisons wh the annual regressions we report next, we calculate the correlations each year and then report the mean values, rather than report correlations based on pooling the data across all 17 years. Panel A contains firm-level correlations and Panel B contains portfolio-level correlations. The correlations among price, book value, net income, and REBV reported in Panel A are comparable to those reported for the different CHHK samples. As might be expected, the dependent variable in the valuation regressions are highly related to the regressors (posively related to BV, NI, and PVFRI and negatively related to INT). Turning to the potential of REBV being correlated wh value-relevant variables that are omted from the Ohlson valuation regression used by HK, the relation between REBV and different measures of growth appears mixed. While the correlations wh BM and EP2 are close to zero, the correlation wh LTG is strongly negative (Pearson and Spearman values of 0.29 and 0.25, respectively), suggesting that firms wh lower retained earnings and more contributed capal are associated wh higher forecasted growth. The results in Panel B indicate even higher correlations between P, the dependent variable, and BV and PVFRI, the regressors in equation (9), relative to those reported in Panel A. This increase in correlation is to be expected given that larger firms are weighted more heavily in the portfolio variables. 14

17 Before estimating our valuation regression, we examine whether the HK and HHK results are observed for our sample. Table 3, Panel A contains the results of replicating the analysis reported in Table 2 of HK. While the magnudes of our estimates for α 1 and α 2 differ from those reported in HK, variation in those coefficients across REBV quintiles is as predicted by the HK hypothesis: α 1 decreases wh REBV and α 2 increases wh REBV. We expect the magnudes of these coefficients to be higher for our sample because of two sample differences that result in higher valuations relative to book value and earnings: our sample includes more recent years (corresponding to the bull markets of the 1990s), and fewer low capalization firms than the HK sample. Despe these sample differences, the values of adjusted R 2 and the mean REBV for different quintiles are comparable to those reported in HK. Examination of the pattern of variation of the values of α 1 and α 2 across the REBV quintiles suggests that the changes occur primarily over the first three REBV quintiles. This pattern is consistent wh the non-linear relation expected if firms wh high REBV are less representative of the HK hypothesis because their contributed capal is buried under retained earnings. Table 3, Panel B, contains the results of replicating the analysis reported in Table 3 of HK. This analysis includes two addional terms that represent the interaction of REBV wh book value and net income to allow for continuous variation in REBV rather than the five REBV partions considered in Panel A. The first subpanel provides the results of estimating the regression across the pooled sample whereas the second subpanel partions the sample period into two: before and after the Tax Reform Act of 1986, when tax rates on dividends decreased substantially. The intuive arguments made by HK predict that the coefficient α 2 on RE (the interaction of BV and REBV) should be negative and the coefficient α 4 on REBVNI (the interaction of NI and REBV) should be posive. Moreover, the absolute value of the coefficient α 2 should equal t, the rate at which dividends are capalized. Our results in Panel B are generally consistent wh the results reported by HK: the coefficients α 2 and α 4 are negative and posive, respectively, and the magnude of α 2, representing the tax rate, declines after However, as wh the replication effort in Panel A, 15

18 the magnudes of the coefficients differ slightly from those reported in HK. Specifically, the magnudes of α 2 are higher in our results, and appear too high relative to the prevailing top tax rates. Table 3, Panel C, describes our efforts to replicate the results reported in Table 2 of HHK. That analysis scales the Ohlson valuation relation by book value, rather than number of shares as in HK, and considers four tax regimes, rather than the two considered in HK. HHK studied five tax rate regimes (between 1975 and 1997): 1) the pre-erta period ( ) when the top personal tax rate was 70 percent; 2) the ERTA period ( ) when the top rate was 50 percent; 3) a transion year (1987) when the top rate was 38.5 percent; 4) the TRA 86 period ( ), when the top rate ranged between 28 and 33 percent, and 5) the OBRA 93 period (1993 onwards), when the top rate increased to 39.6 percent. As our sample covers data from 1985 to 2001, we are unable to examine the first tax rate regime and we have only two years from the second tax rate regime. Similar to the experience of DEMB, we find that our efforts to replicate the HHK results are not as successful as our efforts to replicate the HK results. In fact, our results tend to be more similar to those reported by DEMB, even though their sample is more representative of the HHK sample, relative to ours. Notwhstanding these differences between our results and the HHK results regarding the levels of and changes in magnudes of the different tax regimes, we find that our results are generally supportive of the HK hypothesis: the coefficient α 1 on the interaction between REBV and BV is always negative and the coefficient α 3 on the interaction between NI and REBV is generally posive (all years except 1987). We turn now from the HK specification to tests based on our specification. The results of replacing net income wh PVFRI and estimating portfolio-level regressions rather than firmlevel regressions are reported in the first two panels of Table 4. Panel A contains the average of the coefficient estimates obtained from 17 annual regressions for each of five REBV quintiles, and Panel B contains the results obtained when we pool observations across all 17 years. In 16

19 Panel C, we estimate annual regressions, similar to those in Panel A, but do so at the firm-level rather than at the portfolio level. The results in Panel A indicate relatively ltle variation in the mean estimates for δ 1 across the five REBV quintiles, and no systematic relation wh REBV. Specifically, the mean value of δ 1 is about 1.32 for 3 of the REBV quintiles, declines to 1.15 for the first quintile, and increases to 1.45 for the third REBV quintile. The fairly low standard errors based on the timeseries distribution of the 17 estimated coefficients indicates relatively stable estimates over time. While the mean value of δ 2 also appears to vary whin a narrow band between 0.58 and 0.66 for all but the fifth REBV quintile, we detect an increasing pattern between the third and fifth quintile (0.58 to 0.66 to 0.71). Although this increasing pattern is consistent wh the HK hypothesis, runs counter to our expectation that we should observe greater increases in δ 2 for the lower REBV quintiles (since high REBV firms have greater difficulty in paying out contributed capal). In fact, we observe a decrease in δ 2 between the first and second REBV quintile. In sum, we believe these results are generally inconsistent wh the HK hypothesis: the patterns observed for δ 1, which is expected to provide more reliable tests than δ 2, clearly do not indicate a negative relation wh REBV, and the patterns observed for δ 2 are at best only partially consistent wh that hypothesis. The results in Panel B of Table 4 confirm that our overall conclusions regarding the validy of the HK hypothesis are not altered when we estimate our regressions using data pooled across all 17 years, rather than estimate the annual regressions reported in Panel A. Again, the coefficients δ 1 and δ 2 do not systematically decrease and increase, respectively, wh REBV. While we note some evidence of a decrease in δ 1 over the last three quintiles, the patterns observed for both coefficients over the first three quintiles (which are more likely to be representative of the HK hypothesis) are inconsistent wh the predictions of the HK hypothesis. The results reported in Panel C examine the effect of estimating firm-level regressions rather than portfolio-level regressions. The market-level estimates for k and g that we use to construct our PVFRI measure are likely to contain more error at the firm-level, and we provide 17

20 these results merely to supplement the primary results reported in Panel A. We provide two sets of results, corresponding to estimating the regression in equation (9) based on scaling all variables including the intercept by the number of shares outstanding and also based on unscaled values. In both subpanels, the coefficient estimates exhib remarkably ltle variation wh REBV, and we see no systematic evidence of a decreasing pattern for δ 1 or an increasing pattern for δ Sensivy analyses Given potential concerns about the assumed values of k and g being systematically under or overstated, we consider next whether our results are sensive to alternative assumed values for these two parameters. In essence, we repeat the analysis in Panel A of Table 4 for values of k and g other than those we believe are our best estimates. To recall, in Panel A of Table 4, we assume that k equals the after-tax risk-free rate plus 3 percent, and g equals the after-tax risk-free rate minus 3 percent. The results reported in Panel A of Table 5 are based on varying the risk premium underlying k between 0 and 5 percent in 1 percent increments, while holding g constant at the after-tax risk-free rate minus 3 percent. We report the mean annual estimates for δ 1 and δ 2 along wh the associated time-series standard errors for the other five cases, and report our primary results (from Table 4, Panel A) in the right-most column for ease of comparison. We find a general increase in the level of both estimates as we increase the assumed risk premium. However, examination of the variation of δ 1 across REBV quintiles whin each set of assumptions does not reveal a case where δ 1 systematically decreases wh REBV. Examination of estimates for δ 2 reveals some evidence of a systematic increase wh REBV, however that increase occurs after the third REBV quintile, not for the low REBV quintiles. The results reported in Panel B of Table 5 describe the impact of varying g while holding k constant at the after-tax risk-free rate plus 3 percent. Our primary results from Panel A of Table 4 are again reported in the right-most column, and we vary the adjustment subtracted from the after-tax risk-free rate between 1 and 5 percent in 1 percent increments. As wh the results in 18

21 Panel A, we see no evidence of a systematic decrease in δ 1 as REBV increases, and while we observe some evidence of an increase for δ 2, occurs after the third quintile. Overall, given that we believe that our δ 1 results are more reliable, we find that our conclusion regarding our evidence being generally inconsistent wh the HK hypothesis is robust to alternative measures of g and k. 6. Conclusions Identifying the extent to which dividend taxes are capalized in share prices is an important research topic in finance and economics. Harris and Kemsley (1999) offer an innovative procedure designed to shed light on this question. While the conceptual underpinning of their proposed methodology is straightforward (contributed capal should be worth more than retained earnings since return of contributed capal via dividends is tax exempt) they do not derive the valuation regression they estimate (price on book value and trailing earnings) and as a result there is some confusion about the predictions they make for how the estimated coefficients should vary wh the proportion of book value represented by contributed capal. We offer instead a different specification that allows clear predictions about how the estimated coefficients should vary wh the proportion of contributed capal. Our results suggest that market participants behave as if they do not discriminate between contributed capal and retained earnings. While our methodology offers evidence on the extension of the dividend capalization hypothesis proposed by Harris and Kemsley (1999), we are unable to draw inferences regarding the extent to which dividend taxes are capalized in share prices. Despe the advantages of our methodology over that used in the lerature, has some potential problems: requires estimates of future book values, earnings, dividends, and the aftertax cost of equy. Given the absence of publicly available estimates of the stock market s implied after-tax cost of equy capal as well as publicly available forecasts for periods beyond the first few years, we rely on generic market-level assumptions for growth past available forecasts and the cost of equy capal. Even though we take steps to migate the impact of 19

22 errors in these assumptions (by constructing portfolios wh similar forecast growth) and conduct extensive sensivy analyses to confirm the robustness of our findings, we recognize that our conclusions are tentative and awa confirmation from alternative methodologies and samples. 20

23 Appendix Variable definions DATAxx refers to the corresponding data em from the annual Compustat file, and the IBES forecast data are taken as of April each year from the summary files. Variable Name Definion MV Market value of equy=stock price*shares outstanding (from IBES actual data file), as of analysts forecast date P Stock price per share=mv / Compustat number of shares outstanding (DATA25) BV Book value of equy per share=data60 / DATA25 RE Retained earnings per share=data36 / DATA25 INT Intercept, scaled by # of shares outstanding=1/data25 NI Net income per share=data172 / DATA25 k After-personal-tax required rate of return; proxied by Moody s AAA municipal bond yield as of April each year (from DRI Database) plus 3 percent REBV Retained earnings to book value ratio= DATA36 / DATA60 REBVNI =REBV*NI EP2 Two-year-out forward E/P ratio adjusted for RF; Analysts forecast earnings for year +2 (EPS2), scaled by share price (P), minus RF, the pre-investor-tax risk-free rate of return, proxied by 10-year Treasury bond yield as of April each year. PVRI1_5 Present value of year +1 to +5 residual income per share; 5 nin k * bv n 1 n n= 1 ( 1+ k ) PVTV ( NI )( ) 5 k * BV4 1+ g Present value of terminal value; measured, where g is 5 ( k g)( 1+ k) growth rate in residual income beyond year 5, the terminal year. We assume g equals Moody s AAA municipal bond yield as of April each year 3%. PVFRI Present value of all future residual income; (PVRI1_5 + PVTV) / DATA25 PAYOUT Dividend payout ratio; indicated annual dividends per share / trailing EPS (IADIVPS / FYA0EPS from IBES actual files). FYA0EPS is 0 for 188 firm-level observations (no portfolio observations). In that case, is replaced by BV*12%. PAYOUT is Winsorized at 1. BM Book-to-market ratio = BV/MV BETA CAPM beta; estimated using prior 60 monthly returns in a market model regression. Set to equal 1 if insufficient data available to estimate BETA. CP Contributed capal / market value of equy ROE1 Analysts forecast of earnings for year +1 / book value of equy For firm-level analyses, these variables refer to per share data, unless indicated otherwise. For portfolio-level analyses, the variables refer to portfolio aggregates. See text for a description of the procedure followed to form portfolios. 21

24

25 Figure 1 Alternative views of the tax rate on dividends (t) that is capalized in share prices (the top statutory dividend tax rate is t d, and capal gains taxes are assumed zero for convenience) Full capalization: t t d (e.g. King 1977) this is the new view, wh a variety of implications (e.g.. the cost of internal equy is lower than that of external equy). Partial capalization: 0<t<t d (e.g. DEMB, 2001) even though dividend taxes are expected to be paid on future dividends, taxes on liquidating dividends will be zero. No capalization: t 0 (e.g. Miller and Scholes, 1978) stocks wh high dividend payout are held by tax exempt investors, and this clientele effect causes the dividend tax rate capalized in share prices to be low. special case of new view CC RE (e.g. HK, 1999) while taxes on dividends are capalized, the fraction of future dividends that is contributed capal is returned tax-free. 23

26 Figure 2 Simulation results to develop predictions for relation between REBV and δ 1 and δ 2. Panel A: variation of δ1 (coeff on BV) wh REBV value of δ similar pattern for all ROE ratio of RE to BV Panel B: variation of δ2 (coeff on PVFRI) wh REBV value of δ ROE=15% ROE=14% ROE=12% ratio of RE to BV 24

27 Table 1 Descriptive statistics for sample Panel A. Firm Level Variables VARIABLE MEAN STD MAX P99 Q3 MEDIAN Q1 P1 MIN N P ,003 BV ,003 RE ,003 REBV ,003 CP ,003 BM ,003 INT ,003 NI ,003 ROE ,003 PVFRI ,003 PAYOUT ,003 BETA ,003 EP ,003 LTG ,003 25

28 Table 1 (continued) Descriptive statistics for sample Panel B. Portfolio Level Variables VARIABLE MEAN STD MAX P99 Q3 MEDIAN Q1 P1 MIN N P 11,348 14, ,921 69,765 13,148 6,827 3,859 1, ,700 BV 4,631 4,656 70,760 22,717 5,699 3,292 2, ,700 RE 2,582 3,223 42,848 15,656 3,172 1, ,700 REBV ,700 CP ,700 BM ,700 NI ,700 ROE ,700 PVFRI 8,154 10, ,077 51,268 9,874 4,980 2, ,707 1,700 PAYOUT ,700 BETA ,700 EP ,700 LTG ,700 26