Impact of Tax Acts on Household Equity Portfolio Choices: A Study on Dividend Clienteles, Study notes of Economics

Download Impact of Tax Acts on Household Equity Portfolio Choices: A Study on Dividend Clienteles and more Study notes Economics in PDF only on Docsity! Office of Tax Analysis Working Paper 102 March 2011 The Dividend Clientele Hypothesis: Evidence from the 2003 Tax Act Laura Kawano A version of this paper was published in February, 2014 in the American Economic Journal: Economic Policy. It is available at: http://pubs.aeaweb.org/doi/pdfplus/10.1257/pol.6.1.114 The OTA Working Papers Series presents original research by the staff of the Office of Tax Analysis. These papers are intended to generate discussion and critical comment while informing and improving the quality of the analysis conducted by the Office. The papers are works in progress and subject to revision. Views and opinions expressed are those of the authors and do not necessarily represent official Treasury positions or policy. Comments are welcome, as are suggestions for improvements, and should be directed to the authors. OTA Working Papers may be quoted without additional permission. The Dividend Clientele Hypothesis: Evidence from the 2003 Tax Act Laura Kawano1 March 2011 Abstract In this paper, I test the dividend clientele hypothesis (DCH) by examining the impact of the Jobs and Growth Tax Relief Reconciliation Act of 2003 (the 2003 tax act) on household portfolio dividend yields. The DCH predicts that the 2003 tax act, which reduced the tax-disadvantage of dividends differentially across the income distribution, would cause high income households to shift their portfolios towards dividend paying stocks relatively more than lower income households. Using the 2001 and 2004 Surveys of Consumer Finances (SCF), I examine how changes in tax rates affect changes in household portfolio dividend yields. I find that the 2003 tax act caused households in the highest (35%) tax bracket to increase their portfolio dividend yields by 1.1 percentage points more than those in the next (33%) tax bracket, and by 2.6 percentage points more than those two tax brackets (28%) below. Compared to a 2.1 percent average dividend yield in 2001, these responses are large and economically significant. Using the 2007 SCF, I find that the reduced variation in dividend tax rates across households caused portfolio dividend yields to become homogeneous within three years of the tax act. Using a battery of sensitivity checks, I verify that these findings are not driven by other explanations for changes in dividend preferences, such as changes in optimism or risk- aversion. 1 This paper was primarily written while I was a graduate student at the University of Michigan. I am grateful to my dissertation committee members, Amy Dittmar, Matthew Shapiro, Joel Slemrod and Jeff Smith for invaluable guidance, to Kevin Moore for assistance with using the Survey of Consumer Finances data, and to Daniel Feenberg for assistance with the NBER TAXSIM model. Charlie Brown, Jim Hines, Sara LaLumia, Sebastien Bradley, Josh Cherry, Osborne Jackson, Ryan Nunn, Todd Pugatch, and participants at the University of Michigan public finance and labor seminars, various colleges and government agencies, and the 2007 Midwest Economics Association Annual Meetings provided useful feedback. Additional comments can be sent to [email protected]. or passively, to high dividend yielding stocks in response to the 2003 tax act. I also find that in the longer term, portfolio dividend yields became quite similar across households. This finding is expected because the distributions of effective dividend and capital gains tax rates were compressed. The differences between the short-term and longer-term responses are interesting and informative regarding the heterogeneity in portfolio adjustments and the importance of adjustment costs. To assess the economic impact of the 2003 tax act, I use the parameter estimates to simu- late the change in portfolio dividend yields caused by the 2003 tax act. I find that households in the top tax bracket more than doubled their portfolio dividend yields (a 115% increase). These top tax bracket households increased their yields by 1.1 percentage points more than those households in the next tax bracket and by 2.6 percentage points more than those two tax brackets below, reflecting the relative intensity of the tax treatment. In addition, the 2003 tax act caused a 0.94 percentage point differential response in portfolio dividend yields across treatment groups, defined by educational attainment measures. Given that av- erage portfolio yields in the 2001 SCF were 2.05%, this represents a large and economically significant response. I run a battery of specification tests to verify that the estimated response to the 2003 tax act is not explained by other factors. I determine that the estimates are robust to different treatment group definitions, to different outlier cut-offs, and to alternative methods of handling imputed values. I find that the main conclusions are unchanged when relaxing the assumptions of the Tobit model. I check that other determinants of household preferences for dividends, such as expectations over the future performance of the economy, did not change differentially across treatment groups over the two periods considered. Understanding the relationship between taxes and investor decisions is important for several reasons. First, such information is useful to corporate financial managers who may consider the tax characteristics of their investors to determine optimal financial policies. 3 Second, because equity holdings and dividend receipts have historically been concentrated in the upper tail of the income distribution, the impact of changing tax rates on household equity portfolios has important implications for the redistributive properties of the tax sys- tem. Indeed, one argument for taxing dividend income at higher rates than capital gains has been that it aids the progressivity of the tax schedule. Lastly, the magnitude of household behavioral responses to changes in the tax structure inform estimates of the efficiency losses of taxation (Galper, Lucke and Toder 1988). For example, the relationship between taxes and portfolio choice is central to tax reform discussions because switching to a comprehensive income tax or a consumption tax would eliminate the differential tax treatment of assets. Because reorganizing investment strategies can be costly, understanding shifts caused by changing tax rates is important to such debates. The remainder of the paper is organized as follows. Section 2 reviews theoretical models of dividend clientele formation. Section 3 summarizes the main components of the 2003 tax act. The data and methodology are described in Section 4. Section 5 provides a presentation and discussion of the results, while section 6 concludes. Appendix A provides a brief overview of a related line of research regarding dividend clienteles, and Appendix B provides detailed descriptions of the sensitivity checks for the main analysis. 2 Theory of dividend clienteles The Modigliani-Miller theorem establishes that in perfect capital markets (i.e., without taxes, transaction or bankruptcy costs, or asymmetric information) a firm’s dividend policy does not affect its value (Modigliani and Miller 1958). In this setting, investors can replicate any stream of dividend payments through the purchase and sale of appropriate equities. Thus, investors view dividend polices as irrelevant and will not pay a premium for any particular policy. However, when investors face different dividend and capital gains tax rates, they 4 have different after-tax valuations for the same asset. Miller and Modigliani hypothesize that such differences lead to the formation of what they termed “dividend clienteles,” in which investors have tax-based preferences over equities that differ only in their dividend policies (Miller and Modigliani 1961). To gain intuition for the mechanism through which investor clienteles emerge, I apply Miller’s (1977) simple clientele model to the case of dividend policies. For simplicity, assume that there are two available stocks: one that does not pay dividends and one that does. Both stocks are assumed to be riskless and there is no available debt security. Also assume that the tax rate on capital gains (τcg) is zero, while the tax rate on dividend income (τdiv) increases with income. The market equilibrium of this model is depicted in Figure 1. Figure 1: Equilibrium in the Miller model This simple model predicts completely specialized portfolios. For a given set of pre-tax returns on the dividend-paying stock (rdiv) and the non-dividend paying stock (rnodiv), the asset demand functions for the dividend stock (Ddiv) and for the non-dividend paying stock 5 That is, given two equities with the same risk exposure, the stock with a higher dividend yield must have a higher expected return to compensate for the tax burden associated with the dividend. Substituting this condition into the investor demand equation yields the following rela- tionship between pre-tax portfolio dividend yields and beta: δip = bi0 + bi1γ0 + bi1γ1βp 1− bi1γ2 (5) This equation implies a linear relationship between efficient portfolio dividend yields and portfolio risk, with the nature of this relationship (i.e., the slope and intercept of this line in dividend-risk space) determined by the relative dividend and capital gains tax rates. For a given level of risk, the compensation required for a higher dividend yield is positively related to the differential in tax rates on dividends and capital gains.4 3 Jobs and Growth Tax Relief Reconciliation Act of 2003 The Jobs and Growth Tax Relief Reconciliation Act of 2003 contained two major components relevant to this study. The first is reductions in long-term capital gains tax rates. The top capital gains marginal tax rate fell from 20% to 15%, while the 10% rate for lower-income individuals fell to 5% (and to zero percent in 2008). The second is that qualified dividends were now taxed at the same statutory rate as capital gains, rather than at the ordinary income marginal tax rate.5 As a result, the top marginal tax rate for dividends fell from 4Without taxes, the “two-fund theorem” states that all investors hold some combination of riskless bonds and the market portfolio, where the proportion in each is determined by risk preference. 5Dividends from most foreign corporations, credit unions and banks were excluded from “qualified” dividend income. Non-qualified dividends remained taxed as part of ordinary income. 8 35% to 15%, and from 10% to 5% for lower income individuals.6 This change was applied to dividends from directly held equities and those passed through by a mutual fund or other regulated investment company, partnership, REIT, or common trust fund. Changes to statutory tax rates on capital gains and dividend income are depicted in Figure 2. Prior to the 2003 tax act, high-income individuals had a strong tax incentive to receive equity returns in the form of capital gains rather than dividends. Thus, portfolio dividend yields for high-income households are predicted to be lower than those for low- income households. The 2003 tax act completely closed the gap between dividend and capital gains tax rates, making dividend income more attractive for all households. That the change in the tax treatment was dramatic at high levels of income is also clear in Figure 2. Thus, portfolio dividend yields for higher-income households are predicted to grow by relatively more than those for lower-income households, ceteris paribus. It is this differentially dramatic decrease in the tax treatment of dividend income that is used to identify the effect of dividend and capital gains tax rates on household equity portfolio choices. Figure 2: Statutory tax rates: Married couples filing jointly 6Taxpayers on the Alternative Minimum Tax schedule also benefited from the reduction by facing a reduction from the 28% flat rate to 15%. 9 4 Data and methodology 4.1 Description of data In the main analysis, I use data from the 2001 and 2004 Surveys of Consumer Finances (SCF), a triennial survey conducted by the Federal Reserve Board of Governors that provides repeated cross-sectional data on wealth in the United States.7 In analyzing the longer term household response to the 2003 tax act, I also use 2007 SCF data. The SCF contains detailed household-level information on assets and liabilities, which makes it one of the best data sources for studying household portfolios. The data additionally contain rich information on demographic characteristics and attitudes towards risk and credit. The SCF includes 4,442 households in the 2001 sample, 4,519 in the 2004 sample and 4,418 in the 2007 sample. The sampling methodology of the SCF has two parts to improve coverage of U.S. households. One sample frame is from an area probability weighted sample derived from the Census Bureau’s national sampling frame. The second frame is derived from the IRS Statistics of Income Individual Taxpayer File and is used to oversample high-income households. The oversampling of these households is important for identifying clientele effects since financial asset holdings are concentrated at the top end of the income distribution. Indeed, according to the 2001 SCF, 60.6% of families in the top 10th percentile of the income distribution held stocks, while only 3.5% of families in the bottom 20th percentile held stocks. In 2004, the percentages are 55.0% and 5.1%, respectively (Bucks, Kennickell and Moore 2006). Sampling weights are provided so estimates can be weighted to represent the U.S. household population in each year. The weighted sample represents 106.5, 112.1 and 116.1 million households in the 2001, 2004 and 2007 samples, respectively. All summary statistics and regressions presented in this paper are weighted using the sampling weights. Missing 7Panel data would allow me to observe household-specific changes in portfolios in response to the tax reform. While the SCF contains a panel component for the 1983 - 1989 waves, it does not for the period considered. That the SCF is repeated cross-sectional data rather than panel data does not change the interpretation of the parameter estimates (Heckman and Robb 1985). 10 Figure 3: Empirical tax rate distribution act depends in part on the timing of the tax changes and the surveys. Auerbach and Hassett (2007) document the key events leading to the 2003 tax act. Reductions in dividend tax rates were not seriously discussed prior to December 2002, suggesting that there was no anticipation of such a tax change before that time.14 Notably, capital income tax rate cuts were not part of the 2000 Bush campaign platform. Since dividend income reported in the 2001 SCF sample are derived from equity holdings in 2000, these data are not impacted by the 2003 tax act. By the beginning of 2003, however, households and corporations knew that there was a significant probability that dividends would face a lower tax rate and that when a tax act was passed, the tax cuts would be applied retroactively to the beginning of 2003. The 2004 SCF contains information on dividend receipts from 2003, which are clearly impacted by the 2003 tax act. When the 2003 tax act was first passed, the reduced tax rates were set to expire in 2008. However, the Tax Increase Prevention and Reconciliation Act of 14The first notable mention of the reductions in the press occurred on December 25, 2002, when the Wall Street Journal reported that the Bush administration planned to reduce dividend tax rates by 50 percent. On January 6, 2003, the Wall Street Journal announced the Bush administration’s plans to eliminate dividend taxes. Reductions to capital gains and dividend tax rates were officially proposed on January 7, 2003 by the Bush administration. The Conference Committee version of the 2003 tax act passed the House and Senate on May 23, 2003, and was signed into law on June 20, 2003. 13 2005 extended the reduced tax rates on dividends and capital gains through 2010. A number of demographic characteristics are used to control for non-tax factors in the regression analysis that may influence household choices over portfolio dividend yields. Age categories, an indicator variable for being retired, and educational attainment categories are constructed to correspond to the head of household. Net worth categories and household size are computed for the household unit. Responses to a question about the “amount of financial risk that you or your (spouse/partner) [are] willing to take when you save money or make decisions” are used to construct proxies for risk preference. The risk-aversion indicator variable is set to one if respondents answered that they were “not willing to take financial risks,” and zero otherwise. The “‘moderate risk”, “high risk” and “very high risk” indicator variables equal one if the respondent answered that they were willing to “take average finan- cial risks expecting to earn average returns”, “take above average financial risks expecting to earn above average returns”, and “take substantial financial risks expecting to earn sub- stantial returns” respectively, and zero otherwise. Summary statistics of these variables are presented in Table 1. SCF data are self-reported, so measurement error may be of concern, particularly for sensitive data items such as components of wealth. Measurement error may arise when individuals have to sum up values over several financial accounts or because people are unwilling to accurately report such items. As an overall check of the dividends data, I compare dividend income reported in the SCF with that reported on tax returns provided by the IRS Statistics of Income (SOI) Tax Statistics publications. Unweighted, the dividend income reported in the SCF account for approximately 1% of dividend income reported on tax returns. In the SOI data, 26.3% and 23.3% of tax filers report that they received dividend income in 2000 and 2003, respectively. Of the SCF households, only 16.8% and 15.5% report positive dividend income in the 2001 and 2004 surveys, respectively. This difference could reflect that some households with relatively little dividend income do not remember such 14 Table 1: Summary statistics of demographic and socioeconomic variables Variable 2001 2004 2007 Share of SCF Sample Income (thousands) 0-15 0.14 0.14 0.13 15-25 0.11 0.12 0.13 25-50 0.27 0.26 0.27 50-75 0.16 0.18 0.17 75-100 0.12 0.10 0.11 100-250 0.15 0.17 0.16 250+ 0.03 0.03 0.04 Net worth (thousands) 0-50 0.38 0.38 0.36 50-100 0.12 0.11 0.10 100-250 0.19 0.18 0.19 250-1000 0.23 0.23 0.25 1000+ 0.09 0.09 0.09 Demographic characteristics No degree 0.09 0.09 0.09 High school degree 0.31 0.30 0.32 Some college but no college degree 0.18 0.18 0.18 College degree 0.34 0.37 0.35 Not willing to take financial risks 0.40 0.42 0.42 Female 0.27 0.28 0.28 Married 0.60 0.58 0.59 Household size 2.41 2.39 2.42 Retired 0.19 0.19 0.19 Average Age 48.97 49.56 50.01 Number of households (millions) 106.5 112.1 116.1 Number of observations 4519 4442 4418 Observations are weighted by their sampling weights. Financial data are in 2004 dollars. Demographic characteristics refer to the head of household. Statistics are corrected for multiple imputations. 15 The estimating equation for the treatment effects model of the effect of taxes on portfolio dividend yields that incorporates the Tobit framework is given by: Y ∗ it = Xitβ + ατt(xit) + εit Yit = max{0, Xitβ + ατt(xit) + εit} (6) where Y ∗ is the latent (uncensored) dividend yield, Y is the observed (censored) dividend yield, i corresponds to the household and t denotes the time period. The vector X contains factors other than taxes that may affect household choices over dividend yields. The contin- uous treatment variable is τt(xit), the difference in dividend and capital gains marginal tax rates. It is a function of various household characteristics, such as income, marital status, and family structure. The vector x contains a subset of X. Note that the tax function is indexed only by t because all households face the same tax schedule at a given point in time. That is, two households with the same values of xit face the same tax rates. The parameter of interest is a function of α, the effect of the tax treatment on portfolio dividend yields. Specifically, because this is a corner solution model the marginal effect of interest is that on the observed dividend yield. In principle, α could be identified from a single cross-section of data because it enters the equation linearly and the tax schedule is nonlinear (Scholz 1992). Such identification is weak, however, and thus undesirable in practice. Because all households face the same tax system at a given point in time, two households with the same level of income will face different tax rates only through differences in other characteristics. When variations in economic situations, such as income levels and family structure, are the driving source of variation in marginal tax rates that a household faces, it is difficult to disentangle income effects (and other factors that are correlated with income) from pure tax effects in a single cross-section. Identification of the tax effect is achieved only through the nonlinearities in the tax schedule, which is typically weak in 18 practice. For example, if income impacts dividend yields nonlinearly but we only include the level of income in the regression, then the nonlinearity in the tax schedule used to identify the tax effect is partly due to the nonlinearity of the income effect, and so would confound income effects and tax effects. Instead, the 2003 tax act provides exogenous variation in tax rates that can be used to identify α. Because the SCF is a repeated cross-section rather than a panel, we cannot follow the same individuals over time. Assuming that the two cross-sections are independent, which likely holds given the sampling design of the survey, we can pool the data across the periods and estimate α: Y ∗ i,s = α[τ2003(xi,2004)− τ2000(xi,2001)]I(SCF = 2004) + ατ2000(xi,2001) + ηI(SCF = 2004) +Xi,sβ + εi,s Yi,s = max{0, Y ∗ i,s}, s ∈ (2001, 2004) (7) where I(SCF = 2004) is an indicator variable that equals one if the observation is from the 2004 SCF and zero if the observation is from the 2001 SCF. Note that the year subscripts for the tax function, τ , and its inputs, x, differ by one year to reflect that the survey data contains income information for the previous calendar year. Conditional on the observed variables, α is identified from people with the same vector of X characteristics facing two different sets of tax rates because of the 2003 tax act. The post-treatment indicator variable, I(SCF = 2004), controls for the average difference in portfolio dividend yields across SCF samples. This is important because there is a well- documented increase in the supply of dividends following the 2003 tax act (Chetty and Saez (2005), Brown et al. (2004)). Perhaps most notably, Microsoft initiated a dividend payment for the first time immediately following the 2003 tax act. Such changes in dividend policies affect market prices, so dividend yields are expected to change between the two samples. 19 That firms altered dividend policies and market prices changed in response does not affect the interpretation of the tax effect. This is because the dividend clientele hypothesis regards differences in portfolio dividend yields across investors. It does not matter if the response to the 2003 tax act comes through changes in the numerator or denominator of the dividend yield measure since either reflects the types of equities that a household chooses to hold. Because households can affect their tax rates through their portfolio dividend yield choices, the actual difference in marginal tax rates on dividends and capital gains is en- dogenous. To solve this endogeneity problem, I use instrumental variable techniques to consistently estimate α. Moffitt and Wilhelm (2000) show that when a tax reform changes tax rates by different intensities across groups, a valid grouping variable for a difference-in- differences analysis can instrument for the change in tax rates. The 2003 tax act provides both a natural experiment and a grouping variable. Educational attainment is correlated with permanent income, and thus marginal tax rates (Eissa (1996b), Blundell, Duncan and Meghir (1998), Moffitt and Wilhelm (2000)).18 Because it is unlikely that households ma- nipulated their choice of education in response to the 2003 tax act, particularly in such a short time frame, educational attainment is uncorrelated with transitory income and with behavioral responses to the tax change. I use an indicator for whether the household head has a college degree as the difference-in-differences grouping variable.19 Thus, one of the key identifying assumptions is that non-tax factors that influence dividend yield choices did not change differentially by treatment group across the 2003 tax act. The estimated model is Amemiya’s generalized least squares estimator for a limited de- pendent variable with endogenous regressors (Amemiya (1978), Amemiya (1979)), described by the following system: 18For an example of how difference-in-differences has been used to examine the impact of a tax policy, see Eissa (1996a) and Heckman’s (1996) response to Eissa (1996a). 19If this endogeneity is ignored, the estimated tax effect will be biased upwards (towards zero) because households may reduce their dividend income to reduce their tax liability. Indeed, when I use actual marginal tax rates in the main regressions, the estimated tax effect is closer to zero (and sometimes even positive), though no longer statistically significant. 20 Figure 4: Portfolio dividend yields by educational attainment, 2001 and 2004 p-value on β2 in the following difference-in-differences regression: characteristici,t = β0 + β1collegei,t + β2college ∗ I(SCF = 2004)i,t + uit (9) Importantly, these characteristics are not changing differentially across groups in the two samples. The statistically different change in tax rates does not appear to be due to differ- ential changes in income. Additionally, the proportion of households in each group is stable, so considering the sample of equity-holding households in each education class also appears to be appropriate. Because this analysis focuses on equity-holding households, the assumption that the composition of groups is stable across periods may be violated. Indeed, several studies find that taxes influence stock ownership probabilities (Poterba and Samwick (2002), King and Leape (1998). To test whether the 2003 tax act altered the population of equity-holders, I estimate a difference-in-differences probit for the probability of holding equities. Table 4 presents results from this estimation. The parameters of interest, the coefficients for college ∗ y04 and y04, are not statistically significant and I fail to reject the null hypothesis 23 Table 3: Characteristics of equity-holders, medians by education group No college degree College degree p-value on 2001 2004 2001 2004 diff-in-diff Tax differential 15.8 7.1 22.3 10.0 0.00 Income (thousands, median) 58.9 60.0 103.4 104.3 0.34 Percent with dividend income 34.9 32.3 51.2 49.4 0.85 Not willing to take financial risk 20.1 22.9 8.1 8.6 0.49 Percent married 68.0 63.8 75.3 72.3 0.83 Percent retired 25.3 28.7 16.0 16.9 0.57 Age 51.4 54.6 49.8 51.0 0.23 Household size 2.3 2.3 2.6 2.5 0.47 Number of observations 608 533 1387 1429 Each observations is weighted by its SCF sampling weight. Statistics are corrected for multiple imputations. Demographic characteristics correspond to the head of household. The p-value for test for differences in income corresponds to a test of differences in mean income. that equity-holding households did not change across the two periods. Thus, changes in dividend yields across treatment groups are not likely to be due to the 2003 tax act causing new households to enter equity markets. The validity of a difference-in-differences approach relies on the assumption that the growth rate of the dependent variable would be equal across groups in the absence of treat- ment. Otherwise, the estimated treatment effect may partly reflect other differences across groups. Figure 5 presents household portfolio dividend yields by education groups from the 1992, 1995, 1998 and 2001 SCF samples. The trends in dividend yields look quite similar between the two groups.21 To test this more formally, I run a regression of portfolio dividend yields on a linear trend, a dummy variable for whether or not the head of household has a college degree, and the interaction of the college indicator variable and the linear trend: yield = β0 + β1trend+ β2college+ β3college ∗ trend+ ε. (10) 21The decreasing trend in dividend yields is consistent with the well-documented reduction in firm dividend payments in favor of share repurchases as a means of distributing profits to their shareholders. 24 Table 4: Probit model for holding equities Dependent variable: whether the household has equities Estimated Marginal Std. Variable Effect Error p-value College * y04 -0.01 0.02 0.58 College 0.11 0.02 0.00 SCF = 2004 -0.01 0.01 0.41 Retired 0.04 0.02 0.05 Married 0.05 0.02 0.00 Household size -0.06 0.02 0.00 Household size (squared) 0.01 0.00 0.00 Net worth 50,000-100,000 0.10 0.02 0.00 Net worth 100,000-250,000 0.17 0.01 0.00 Net worth 250,000-1,000,000 0.31 0.01 0.00 Net worth >1,000,000 0.56 0.02 0.00 Not willing to take financial risk -0.22 0.01 0.00 Presented estimates are average marginal effects. Standard errors are heteroskedasticity robust. Observations are weighted by their SCF sampling weights. Estimates are corrected for multiple imputations. Age categories are included but estimates are not reported. None are statistically significant. The full table of results is available upon request. A test for the difference in slope coefficients for the two groups over time is equivalent to a test that the coefficient on the interaction term (β3) is zero. In this regression, the p-value for the test that β3 is zero is 0.98 and I fail to reject the null.22 While nonlinear instrumental variables models are not literally estimated in two stages, I run what would be the first stage regression in the linear case to ascertain the instruments’ strength. Table 5 shows select results from this estimation. Because of the different intensi- ties of the 2003 tax changes, we should expect that college-educated households experienced a larger decrease in the tax differential than those without a college education. Indeed, the parameter estimate on the treatment effects variable is negative and statistically significant. The F-statistic for the exclusion restriction is 28.45. Because the critical value of a 5% Wald 22When allowing for a quadratic trend differences in trends across the two groups remains statistically insignificant. The p-values on the linear and quadratic trend-interaction terms are 0.35 and 0.34, respectively. 25 Table 6: Instrumental variable Tobit results Dependent variable: Portfolio dividend yield Instrumental variable: College ∗ y04 (High treatment indicator) Estimated Marginal Effect Std. Variable Effect Error p-value Tax differential -0.31 0.14 0.03 Age 25-35 3.94 1.81 0.03 Age 35-45 4.51 2.2 0.04 Age 45-55 5.42 2.69 0.04 Age 55-65 5.15 2.5 0.04 Over 65 4.92 2.29 0.03 Retired -0.79 0.92 0.39 College 1.87 0.88 0.03 Net worth 50,000-100,000 -0.35 1.32 0.79 Net worth 100,000-250,000 0.05 1.23 0.97 Net worth 250,000-1,000,000 2.58 1.4 0.07 Net worth >1,000,000 5.06 2.18 0.02 Not willing to take financial risk -1.79 1.12 0.11 Willing to take average financial risk -0.34 0.6 0.57 Willing to take high financial risk 0.05 0.64 0.94 SCF = 2004 -4.05 1.74 0.02 Constant -3.07 1.54 0.05 Number of observations 3956 Number of uncensored observations 2379 Marginal effects are effects on observed dividend yields. Standard errors are computed using the Delta Method and are heteroskedasticity-robust. Observations are weighted by their SCF sampling weights. Estimates are corrected for multiple imputations. Included in the regressions but not reported are an indicator for the household head being married and household size (level and square). None are statistically significant at the 5% level. of the 2003 tax act on dividend yields of portfolios of households at different tax brackets, summarized in Table 7. A household in the highest tax bracket would have faced a decrease in the tax rate differential from 34.6 percentage points to 11.25 percentage points, leading portfolio dividend yields to increase by 7.24 (=[11.25-34.6]*-0.31) percentage points. On average, macroeconomic factors are estimated to decrease yields by 4.05 percentage points (the estimate of η) for all households between 2001 and 2004. Thus, the predicted change in observed portfolio dividend yields for households in the highest tax bracket is a 3.19 28 percentage point increase. Relative to an average portfolio yield for households in the top bracket in 2001 of 2.7 percentage points, this is a 115% increase in dividend yields. This constitutes a large and economically substantive response. Similar calculations are done for households in the next two tax brackets, which shows that the tax effect is large and varies substantially with the intensity of the tax treatment.27 Table 7: Effect of the 2003 tax act for select tax brackets Highest Bracket Next Bracket Two below 39.6% 36 % 31% Tax rate differential, 2003 11.25 11.25 11.25 Tax rate differential, 2000 34.60 31.00 26.00 Change in tax rate differential -23.35 -19.75 -14.75 Predicted change in yields 3.1 2.1 0.5 (τ2003 − τ2000) ∗ β̂τ + β̂y04 Average yield in 2001 sample 2.7 6.5 2.4 Percent change 115 32 21 Author’s calculations based on the regression results in Table 6 and SCF data. The previous exercise provides estimates of the impact of the 2003 tax act at particular points in the tax schedule. However, the realized economic impact of the 2003 tax act is better understood as the average portfolio response weighted by the proportion of households at various points of the income distribution.To obtain this estimate, I take households from the 2001 SCF sample and use TAXSIM to compute the tax rates that they would have faced under the 2003 tax rules. This change between a household’s actual tax rates in 2000 and its simulated tax rates for 2003 is exogenous to household decisions in response to the 2003 tax act. I use these simulated tax rate changes and the estimated effect of the dividend and capital gains tax rate differential on portfolio dividend yields to compute the household-specific predicted change in dividend yields caused by the 2003 tax act. Based on these simulations, college-educated households increased their portfolio div- 27The parameter estimates are interpreted as the effect of small changes in tax rates. With large changes to tax rates, these simulated responses are only approximations and unmodeled nonlinearities in the response function could make this estimate inaccurate. However, given the nature of the data, this is still the best way to understand the magnitude of the tax effect. 29 idend yields by 4.26 percentage points with an average yield in 2001 of 1.22% (standard deviation of 2.5%), whereas non-college educated households increased their portfolio div- idend yields by 3.32 percentage points with an average yield of 2.23% in 2001 (standard deviation of 8.53%).28 Thus, the treatment effect of the 2003 tax act is a 0.94 percent- age point differential response in portfolio dividend yields between educational attainment groups. This estimated effect of the 2003 tax act is both economically significant and of plausible magnitude. Figure 6 depicts the actual portfolio dividend yields in 2001 and 2004, along with the predicted dividend yields in 2004 based on these simulations. As before, the predicted dividend yields are the predicted yields scaled by the year fixed effect. The predicted yields broadly match the patterns that are observed in 2004. Figure 6: Comparing simulated change in portfolio dividend yields with actual yields The estimated tax effect is a general equilibrium response that captures both changes to investor demands and changes to the supply of dividends. Because the SCF data is a repeated cross-section and does not contain information on the stocks in a household’s 28This calculation is 1 N ∑N i=1(τ̂2003i − τ2000i )β̂τ , where τ̂2003i is the tax rate differential that household i would have faced under the 2003 tax rules, τ2000i is the tax rate differential for household i in 2000, and β̂τ is the estimated marginal effect of the tax rate differential on portfolio dividend yields. This is computed for all equity-holding households in the 2001 SCF. 30 influence portfolio dividend yields. Risk preferences might matter more for a household’s allocation of wealth between debt and equity, rather than the types of equity that it chooses to hold. Also, self-reported measures of risk preferences may not accurately reflect cross- sectional differences across households. Because older households may be more financially sophisticated due to prolonged expe- rience with financial markets, older college-educated households may respond more quickly to tax policy changes than others. To account for this possibility, I run the same instru- mental variables Tobit specification including interaction terms between the age categories and retired indicator variable with the treatment group indicator. In this specification, pa- rameter estimates on the interaction terms are not statistically significantly different from zero and the estimated tax rate differential effect remains roughly the same. The effects of these controls on portfolio dividend yields do not appear to change over the two periods considered. In addition, optimism over the future state of the economy has been shown to influence portfolio choices, particularly the decision of whether to hold stocks (Kezdi and Willis 2003). If investors believe that dividends signal safety, then optimistic households may choose lower dividend yields, ceteris paribus. Responses to the question, “Over the next five years, do you expect the U.S. economy as a whole to perform better, worse, or about the same as it has over the past five years?” are used to construct indicator variables for households who believe the economy will perform better, worse and about the same. When including this measure of optimism in the main regression, the parameter estimate on the tax rate differential is similar at -0.34 (std. error = 0.15). “Optimistic” households have lower dividend yields relative to households who believe the economy will perform about the same or worse. The parameter estimate on this indicator variable is -0.88 (std. error = 0.37), which is statistically significant at the 5% level.32 32There are other factors that may influence household portfolio dividend yields but are not included because they are endogenous to portfolio choices. The 2003 tax act may have changed where households 33 Several demographic characteristics may have had differential effects on portfolio yields over time. For example, older households may respond differently to a tax change because portfolio choices are influenced by a desire to finance current consumption. To account for this possibility, I run the instrumental variables Tobit model including interaction terms between the age categories and retired indicator variable and the treatment group indicator. In this specification, parameter estimates on the interaction terms are not statistically sig- nificantly different from zero, and the estimated tax rate differential effect remains roughly the same. 5.2 Predicted effect of the 2003 Tax Act sunset provisions The Bush tax cuts of 2001 (the Economic Growth Tax Relief Reconciliation Act of 2001, which reduced ordinary income tax rates for most taxpayers and created a new tax bracket for lowest levels of income) and 2003 are set to expire at the end of 2010. If Congress does not act, dividend income will again be taxed as ordinary income at pre-2001 tax rates and long term capital gains tax rates will increase.33 I consider the effects of these tax increases implied by the estimates of this study. I simulate marginal and average tax rates that households in the 2007 SCF would face in 2011 by adjusting income variables to 2001 dollars using Consumer Price Index from the Bureau of Labor Statistics and computing tax locate their dividend-yielding equities, i.e., between taxable or tax-deferred accounts. See Shoven and Sialm (2003) for a discussion of the optimal location of equity securites. Also, concentrated equity holdings in mutual funds may restrict a household’s ability to adjust portfolio dividend yields. That these variables are not included may cause bias if the omitted variables are correlated with the included regressors. To check for this possibility, I re-estimate the regression including these additional regressors. Though not presented here, results from these alternative specifications are available upon request from the author. In each, the magnitude of the estimate of the tax rate differential effect remains roughly the same and the parameter estimate on the additional variable is statistically insignificant. These results indicate that excluding these variables is not problematic for interpreting the main estimation results as consistent for the causal effect. There may, of course, remain other factors not considered that make such an interpretation invalid. 33Marginal tax rates on dividend income would increase from 15% to 39.6% for those in the highest tax bracket and from 0% to 15% for those in the lowest tax bracket. The top statutory capital gains tax rate will increase from 15% to 20%, and the lowest statutory capital gains tax rates of 0% will increase to 10%. 34 rates under the 2001 tax rules.34 For comparison, I first consider the implications of the tax reversals if dividend clientele effects are ignored, i.e., assuming that households do not adjust their equity portfolios (actively or passively) in response to the tax increases. Households in the 2007 SCF received $148 billion in dividend income in 2006 and paid $22.2 billion in taxes on that income.35 The 2011 average tax rates and dividend receipt patterns in 2007 imply that dividend tax revenue would increase to $38.3 billion in 2011.36 This paper shows, however, that households will shift their portfolios away from dividend paying stocks in response to the tax rate increases. Moreover, higher income households will shift away from these stocks by more than lower income households because of their relatively large tax increases. For each household, I compute the change in the dividend and capital gains tax rate differential that they would face in 2011 and the predicted change in portfolio dividend yields.37 Given the simulated change in dividend and capital gains tax rate differentials and holding the level of equity holdings constant, predicted dividend tax revenues from individuals will only increase to $23.6 billion, less than 62% of the anticipated dividend tax revenues when clientele effects are ignored. If portfolio adjustments are hindered by transaction costs or other adjustment costs, then the increase in dividend tax revenues 34I compute average dividend tax rates as the ratio of federal income tax liability to federal taxable income, both of which outputs from the TAXSIM model. For households that have negative average tax rates, I treat them as though their average tax rate is zero. 35Recall that all summary statistics are weighted by SCF sampling weights and income variables correspond to the calendar year prior to the survey. This level of dividend income, again, is less than the amount reported in the SOI, which reports that $199 billion in ordinary dividends was reported by individuals in 2006. 36This exercise holds dividend payout rates constant between 2007 and 2011. There are, however, several reasons to expect that firms will decrease dividend payments. First, Chetty and Saez (2005) find that firms increased dividend payments in response to the 2003 dividend tax cuts. Thus it is likely that firms will decrease dividend payments as dividends become more costly to their investors. This effect is somewhat hindered by evidence of negative investor responses to dividend payment decreases. Secondly, even if total dividend payments do not change, firms will likely accelerate dividend payments to 2010 so that there are lower dividend payments in 2011. Lastly, dividend payouts in 2011 may decrease for nontax reasons. In particular, the financial crisis and recession in the intervening years make profit distributions even less likely. 37A household’s predicted portfolio dividend yield in 2011 is given by Ŷ ield(i,2011) = Y ieldi,2007 + α̂ · ∆τ(i,2011−2007) + η̂2011, where α̂ is the estimated effect of a 1-percentage point change in the dividend and capital gains tax rate differential, ∆τ(i,2011−2007) is the simulated change in the tax rate differential because of the tax rate reversal, and η̂2011 is a year fixed effect, which would include changes in market prices that result from changes in asset demand. For this simulation, I assume that η̂2011 = −η̂2004. That is, average yields are assumed to return to their pre-treatment levels. 35 Table 8: Instrumental variables Tobit model, 2001 and 2007 Estimated Marginal Std. Variable Effect Error p-value Tax differential -0.07 0.05 0.12 Age 25-35 -0.31 0.57 0.58 Age 35-45 0.14 0.58 0.82 Age 45-55 0.28 0.59 0.63 Age 55-65 0.41 0.6 0.49 Over 65 0.75 0.65 0.25 Retired -0.09 0.36 0.80 Married -0.12 0.20 0.54 Household size -0.25 0.19 0.17 Household size (squared) 0.04 0.04 0.23 College 0.72 0.25 0.00 Net worth 50,000-100,000 0.89 0.46 0.06 Net worth 100,000-250,000 0.73 0.42 0.08 Net worth 250,000-1,000,000 1.56 0.67 0.02 Net worth >1,000,000 2.44 0.93 0.01 Not willing to take financial risk -1.14 0.36 0.00 Willing to take average financial risk -0.60 0.27 0.03 Willing to take high financial risk -0.36 0.29 0.21 SCF = 2007 -0.69 0.66 0.29 Constant -0.48 0.95 0.61 Number of observations 5810 Standard errors are computed using the Delta Method and are heteroskedasticity-robust. Observations are weighted by their SCF sampling weights. Estimates are corrected for multiple imputations. implies that changes in firm dividend policies are immediately capitalized into stock prices. Six years may be too long a period for examining longer-term responses when the equity market adjusts quite quickly. Many other factors may have changed in that period that make it difficult to interpret conditional changes in dividend yields as a tax effect. 38 5.4 Sensitivity analysis 5.4.1 Model specification and sample selection I perform a number of sensitivity checks of the main results, which are described in detail in Appendix C. I verify that the magnitude of the dividend and capital gains tax rate differ- ential effect remains unchanged when using more flexible education attainment measures to instrument for marginal tax rates, using alternative cut-points to determine outliers (both to the right and left of the cut-point used in the main analysis), dropping imputed values, and excluding households whose heads are particularly young. Specification tests for the Tobit model are also provided in Appendix C. As a general diagnostic check, I find that coefficients from a probit model of the household being at the mass point and standardized coefficients from the Tobit model are roughly the same. The Tobit model assumes that the marginal effect of an explanatory variable is the same at both the extensive and intensive margins. To relax this assumption, I estimate a hurdle model which separately estimates the probability of being at the mass point and the relationship between the dependent and explanatory variables for observations away from the mass point. Simulations of the response to the 2003 tax act reveal that the magnitude of the estimated treatment effect is unchanged in this more flexible model. 5.4.2 Alternative explanations for changing dividend demand A key identifying assumption is that non-tax factors that influence investor preferences for dividends did not change differentially across treatment groups. However, there are several events between 2001 and 2004 that may have influenced preferences. For example, accounting scandals at Enron and PriceWaterhouseCoopers may have led to higher demand for dividends as agency problems were of increasing concern.40 The effects of such concerns 40Baker and Wurgler (2004) propose a “catering theory” of dividends, where the salient preferences of investors affect firm dividend payout policies. Interestingly, they reject that taxes influence demands for dividends in favor of other preferences. Relatedly, Becker, Ivkovich and Weisbenner (2009) find that firm 39 should be capitalized into market prices, and likely do not affect investors differentially. However, if higher income households were relatively more responsive to changes in such non-tax factors, then these changes are included in the estimated tax effect and biases the estimate away from zero (i.e., in favor of finding a dividend clientele effect). To test whether non-tax preferences for dividends changed differentially across treatment groups, I identify several questions in the SCF about household attitudes that may proxy for non-tax preferences. First, because investors may associate dividends with safety, then risk- averse investors may choose equity portfolios with a higher dividend yield, ceteris paribus. To account for changes in risk preferences, I use the risk-averse indicator variable from the main regressions as a dependent variable. To further assess risk preferences, I use a question that asks respondents to choose on a scale from 1 to 5 how strongly they agree with the following statement, “Compared with other people of [my] generation and background, [I] have been lucky in [my] financial affairs.” Those who “disagree somewhat” or “disagree strongly” are coded to consider themselves financially unlucky. I posit that those who are not willing to take financial risks and those who believe themselves to be financially unlucky prefer high dividend yield stocks. Changes in respondents’ subjective expectations over the future state of the economy may lead to changes in portfolio choices. Two SCF questions aim to ascertain such beliefs. The first asks, “Over the next five years, do you expect the U.S. economy as a whole to perform better, worse, or about the same as it has over the past five years?” The second asks, “Five years from now, do you think interest rates will be higher, lower, or about the same as today?” From these questions, I construct an indicator variables for whether the household believes the economy will get worse and an indicator variable for whether the household believes that interest rates will increase. For both of these variables, an affirmative response dividend payout policies are related to the age of residents in the headquarters’ location. These studies suggest that there is a causal link between the non-tax based dividend preferences of a firm’s investors and that firm’s payout policy. 40 rates (Feldstein, Slemrod and Yitzhaki 1980). Thus, when trying to isolate the impact of taxes on portfolio dividend yields, the effect of taxes on the timing of capital gains realizations leads to confounding variation in the dependent variable of interest. These results may also be biased if excluded factors not available from tax returns, such as wealth, demographic characteristics and risk preferences, are correlated with tax rates and portfolio choices. Two studies use tax return data and find that, consistent with the dividend clientele hypothesis, dividend yields fall as the marginal tax rate on dividend income rises (Blume, Crockett and Friend (1974), Chaplinsky and Seyhun (1987)). Brokerage house data contain equity holding information, but marginal tax rate infor- mation is limited because individuals report their income only within a small set of ranges. In addition, data from a single firm may not be representative of a household’s investments if they hold accounts outside that brokerage house. Two studies use 1960s data on individ- ual portfolio positions from a large national retail brokerage house (Pettit (1977), Lewellen, Stanley, Lease and Schlarbaum (1978)). The limited variation in marginal tax rates along with the differences in empirical methodologies are the likely reasons for their conflicting conclusions drawn from the same data.42 Graham and Kumar (2006) use 1990s brokerage house data and find that the relationship between income and portfolios is consistent with the dividend clientele hypothesis. Examining stock holding patterns around the Revenue Reconciliation Act of 1993, they document that changes to dividend yields across income groups are consistent with tax-based dividend clienteles. While they provide the only other study to use a natural experiment, they cannot distinguish tax effects from income effects. Scholz (1992) uses the 1983 SCF so, like my study, is able to accurately compute marginal tax rates and portfolio dividend yields. He finds that the relationship between tax rates and portfolio dividend yields supports dividend clientele effects.43 There are limitations to using 42Pettit uses a linear regression model and finds evidence for a clientele effect, whereas Lewellen, et al. use linear discriminant analysis and conclude there is not sufficient evidence to support the dividend clientele hypothesis. 43Scholz (1994) provides descriptive evidence for dividend clienteles by examining portfolio dividend yields 43 a single cross-section to study tax effects, however, as explained in section 5. In addition, the tax rate instrument used, the rate assuming that all households have the same portfolio dividend yield, is endogenous if households simultaneously make choices over labor and investment income.44 He estimates a large effect of taxes on portfolio dividend yields that is three times larger than that found in this study, a magnitude that may be implausibly large (Poterba 2002).45 To compare my estimates to Scholz (1992) and better understand the gains from using a natural experiments framework, I estimate my model using each SCF cross-section separately. Because the high-treatment indicator variable is no longer available as an instrument, I use an instrument based off of the tax rates that apply to the “first dollar” of investment income.46 These results are presented in Table 15 in Appendix C, where they are also described in greater detail. The estimated magnitude of the tax effect is much smaller when using a single cross-section, and is no longer statistically significantly different from zero when using the 2004 SCF. Together, these findings are consistent with the weaker identification of the tax effect using a single cross-section, and suggest that the instrument used when estimating on a single cross-section is endogenous. The difference in magnitude found in Scholz’s (1992) study also reflects the relative prevalence of dividends as a means distributing profits to shareholders in the 1980’s. Two studies test for dividend clienteles using the 2003 tax acts. Both of these studies focus by income decile and by marginal tax rate ranges in two SCF samples around the Tax Reform Act of 1986. He provides tabulations that show that households in the highest ranges of the income distribution have below-average dividend yields. 44The direction of bias from using this instrument is ambiguous because it depends on the relationship between labor and investment income. Absent substitution effects between dividend and non-dividend income, the tax rate will fall for marginal individuals who reduce their dividend income to reduce their tax liabilities. This would cause an upward bias in the estimated tax effect. 45Scholz concludes that moving from a system with no taxes to a one with a 50-percent marginal tax rate, portfolio dividend yields are predicted to increase by 5.4 percentage points. This simulation is difficult to interpret because we should expect that when tax rates are similar across households, there are no tax-based dividend clienteles. 46This is equivalent to the instrument in Scholz (1992) if portfolio yields are assumed to be zero. The results do not change substantively if the instrument is constructed assuming that all households receive the average yield on their portfolios. 44 Table 10: Instrumental variables Tobit on single cross-sections 2001 2004 Est. Marg. Std. Est. Marg. Std. Variable Effect Error p-value Effect Error p-value Tax differential -0.03 0.02 0.09 0.00 0.03 0.97 Age 25-35 2.52 1.20 0.04 1.54 1.06 0.15 Age 35-45 2.61 1.27 0.04 1.48 1.13 0.19 Age 45-55 3.04 1.41 0.03 1.81 1.26 0.15 Age 55-65 3.00 1.30 0.02 1.57 1.17 0.18 Over 65 3.33 1.43 0.02 2.23 1.37 0.10 Retired 0.19 0.48 0.70 0.10 0.40 0.80 Married 0.07 0.51 0.89 -0.46 0.31 0.13 Household size -0.61 0.33 0.07 -0.21 0.30 0.47 Household size (squared) 0.07 0.07 0.30 0.02 0.05 0.66 College 0.73 0.36 0.04 0.64 0.31 0.04 Net worth 50,000-100,000 0.23 1.19 0.85 0.44 0.85 0.61 Net worth 100,000-250,000 0.18 0.94 0.84 0.75 0.71 0.29 Net worth 250,000-1,000,000 1.41 1.03 0.17 1.60 0.83 0.06 Net worth >1,000,000 2.19 1.08 0.04 2.48 0.83 0.00 Not willing to take financial risk -0.64 0.54 0.23 -0.68 0.36 0.06 Constant -4.38 0.55 0.00 -4.00 0.63 0.00 F-statistic on instrument 852.77 626.61 Standard errors are heteroskedasticity-robust. Observations are weighted by their SCF sampling weights. Parameter estimates from the probit model reported are average marginal effects. Estimates are corrected for multiple imputations. on changes to individual equity holding patterns in the aggregate rather than differential changes in equity holding patterns across individual investors, which is done in this paper. Desai and Dharmapala (2007) exploit that the 2003 tax act lowered the tax treatment on dividends from US firms and only extended this preferential treatment to a subset of foreign firms. They estimate the impact of the tax policy change on US investor equity holdings in affected and unaffected countries and find a large response to the 2003 tax act. Blouin, Raedy and Shackelford (2010) examine the relationship between changes in dividend payout policies and changes in equity holding patterns among insiders, mutual funds, and individual investors. They find that firm executives, but not other individual investors, rebalanced their equity portfolios in response to the dividend tax cuts. Because they collapse individual 45 responses to the 2003 tax act. Lastly, there may be other clienteles in the market that are important for a complete analysis of the effect of taxes on portfolio choices over dividend yields. For example, many institutional investors, a growing proportion of investors, are tax exempt and so may form another dividend clientele. To better understand the overall impact of the 2003 tax act, future work should be done to assess the impact of the tax act on institutional investors’ portfolios. In addition, this paper focuses on clientele effects within equity portfolios. There may be other tax-based clienteles that form across other financial assets. 48 A Review of market-based approaches When investors have heterogeneous after-tax valuations of dividends and capital gains, they may adjust their trading behavior around ex-dividend days to capture or avoid upcoming dividend payments. Such adjustments imply that a share’s price drop around its ex-dividend day relative to the dividend payment is related to the tax rates of its investors, controlling for other market fluctuations. If tax-based dividend clienteles exist, then the tax rates implied by these price changes will differ across equities according to their dividend yields. Using this intuition, Elton and Gruber (1970) derive a test for dividend clienteles and find strong evidence for the existence of dividend clienteles. Since Elton and Gruber’s (1970) seminal study, over one hundred articles regarding ex-dividend pricing behaviors have been published, with mixed results. An incomplete list of studies includes: Litzenberger and Ramaswamy (1979) Litzenberger and Ramaswamy (1980), Litzenberger and Ramaswamy (1982) and Auerbach (1983), that find evidence in favor of dividend clienteles, and Black and Scholes (1974), and Gordon and Bradford (1980), Miller and Scholes (1982), and Michaely (1991) that find they cannot reject the null hypothesis that dividends and capital gains are valued equally. While the ex-dividend day studies may summarize the impact of taxes on aggregate mar- ket behavior, they do not identify a direct link between investor behavior and taxes, which would require micro-level data on stock holdings and tax rates. In addition, interpreting these ex-dividend day results are complicated by several factors. First, the coincidence of ex-dividend days and dividend announcement days may lead to a spurious correlation be- tween returns and dividend yields (Miller and Scholes (1982), Gordon and Bradford (1980)). Second, the interpretation of the ex-dividend studies depends on whether a stock’s “typical” investors are setting prices around ex-dividend days. If price changes are driven by short- term investors, the price movements contain little information about the characteristics of a firm’s long-term investors. The return on a stock may be a function of the interactions 49 between multiple classes of investors, so it is difficult to obtain information about clienteles from market price movements (Michaely and Vila 1995). Finally, these studies do not ac- count for transaction costs or risk aversion because they are not available from stock market data. 50 Table 12: Results using different cut-offs for outliers and excluding imputed values Tax differential No. of obs. deleted Est. Marg. Std. Effect Error p-value Include all observations -0.97 0.78 0.21 0 Drop if yield > 2000 -1.16 0.63 0.07 6 Drop if yield > 1500 -0.29 0.15 0.06 7 Drop if yield > 1000 -0.31 0.14 0.03 9 Drop if yield > 500 -0.30 0.14 0.03 11 Drop if yield > 300 -0.28 0.12 0.02 14 Drop imputed values -0.32 0.13 0.01 717 This table presents select results from instrumental variable Tobit regressions using different samples based on changing cut-offs for outliers and by dropping imputed values. Table 13. A general test of whether the Tobit model is mis-specified is done by comparing these coefficients. The estimated coefficients are all of the same sign, as expected. They are also generally similar in magnitude, except for the net worth categories. The Tobit model restricts the effect of the explanatory variables to be the same for both the extensive margin of whether to receive dividends and the intensive margin of the portfolio dividend yield. To relax this assumption, I run a hurdle model that separately estimates a probit model for having a positive dividend yield and an instrumental variables regression of dividend yields on the uncensored observations. To help account for heteroskedasticity in portfolio dividend yields, the dependent variable in the instrumental variables regression is the log of a household’s portfolio dividend yield. Results from the hurdle model are presented in Table 14. That most coefficients are of the same sign indicates that the variables have the same directional effect on both the decision to receive dividends and the choice over dividend yields. The exceptions are the indicator variable for being retired (though not statistically different from zero) and the net worth categories. Interestingly, the tax rate differential effect is five times larger in the instrumental variables regression than in the probit model. Moreover, it is statistically significant at the 10% level in the instrumental variables regression, but 53 Table 13: Comparing probit and standardized Tobit estimates Est. Coeff. Std. Coeff. Variable from Probit from Tobit Tax differential -0.03 -0.05 Age 25-35 0.52 0.65 Age 35-45 0.51 0.75 Age 45-55 0.60 0.89 Age 55-65 0.56 0.85 Over 65 0.64 0.82 Retired 0.13 -0.13 Married -0.04 -0.04 Household size -0.12 -0.08 Household size (squared) 0.01 0.01 College 0.35 0.31 Net worth 50,000-100,000 0.27 -0.06 Net worth 100,000-250,000 0.38 0.01 Net worth 250,000-1,000,000 0.89 0.43 Net worth >1,000,000 1.50 0.84 Not willing to take financial risk -0.41 -0.30 Willing to take average financial risk -0.09 -0.06 Willing to take high financial risk 0.09 0.01 SCF = 2004 -0.35 -0.67 Constant -0.92 -0.52 Coefficients from the Tobit model are standardized by the estimated standard deviation of the error term. Observations are weighted by their SCF sampling weights. Parameter estimates are corrected for multiple imputations. not significantly different from zero in the probit model. This suggests that taxes may be important for determining dividend yields at the intensive margin rather than at the extensive margin. Thus, shifts to dividend clienteles caused by the 2003 tax act are likely confined to shifts among clienteles with some dividend income, rather than inducing more households to receive dividends. Simulations of the impact of the 2003 tax act on household portfolio dividend yields produce similar results to those generated by the instrumental variables Tobit model. The high-treatment (college educated) group is predicted to increase its portfolio dividend yield by 4.53 percentage points while the low-treatment (non-college educated) group is predicted 54 Table 14: Hurdle model for household portfolio dividend yields Probit IV Regression Dependent variable: Indicator for yield > 0 Log Dividend Yield Est Std. Est Std. Variable Coeff Error p-value Coeff Error p-value Tax differential -0.02 0.03 0.53 -0.11 0.06 0.09 Age 25-35 0.49 0.25 0.05 0.73 0.53 0.17 Age 35-45 0.48 0.26 0.06 1.03 0.53 0.05 Age 45-55 0.57 0.25 0.03 1.09 0.51 0.03 Age 55-65 0.50 0.25 0.04 1.20 0.51 0.02 Over 65 0.58 0.27 0.03 1.25 0.54 0.02 Retired 0.14 0.19 0.44 -0.36 0.26 0.17 Married -0.03 0.10 0.76 -0.29 0.31 0.34 Household size -0.16 0.10 0.10 -0.11 0.37 0.76 Household size (squared) 0.02 0.01 0.13 0.03 0.05 0.62 College 0.36 0.09 0.00 0.53 0.27 0.05 Net worth 50,000-100,000 0.33 0.16 0.04 -0.73 0.53 0.17 Net worth 100,000-250,000 0.41 0.14 0.00 -1.05 0.45 0.02 Net worth 250,000-1,000,000 0.91 0.18 0.00 -0.76 0.51 0.13 Net worth >1,000,000 1.49 0.28 0.00 -0.39 0.66 0.56 Not willing to take financial risk -0.36 0.11 0.00 -0.27 0.29 0.35 SCF = 2004 -0.33 0.37 0.38 -1.49 0.79 0.06 Number of observations 3956 2379 Standard errors are heteroskedasticity-robust. Observations are weighted by their SCF sampling weights. Estimates from the probit model are average marginal effects. Estimates are corrected for multiple imputations. to increase by 3.25 percentage points. Thus, there is an estimated 1.28 percentage point differential increase across treatment groups. The more flexible model provides very similar results to the Tobit model. B.4 Analysis using single cross-sections Results from regressions based on a single SCF cross-section are provided in Table 15. The components for the instrumental variable for tax rates are computed using TAXSIM. Specif- ically, I compute the marginal tax rate that applies to a household’s last dollar of taxable income less capital gains, dividend income and interest income. The difference in these div- 55 References Allen, F., A. Bernardo, and I. Welch, “A theory of dividends based on tax clienteles,” Journal of Finance, 2000, LV (6). Amemiya, T., “The estimation of a simultaneous-equation generalized probit model ,” Econometrica, 1978, 46. , “The estimation of a simultaneous-equation Tobit model ,” International Economic Review, 1979, 20. Antoniewicz, R., “A comparison of the household sector from the Flow of Funds Accounts and the Survey of Consumer Finances,” 1996. FEDS Working paper No. 96-26. Auerbach, A., “Stockholder tax rates and firm attributes,” Journal of Public Economics, 1983, 21. and K. Hassett, “The 2003 dividend tax cuts and the value of the firm: An event study,” in A. Auerbach, J. Hines, and J. Slemrod, eds., Taxing corporate income in the 21st century, Cambridge University Press, 2007. and M. King, “Taxation, portfolio choice, and debt-equity ratios: A general equilib- rium model,” Quarterly Journal of Economics, Nov 1983, 98 (4). Baker, M. and J. Wurgler, “A catering theory of dividends,” The Journal of Finance, Nov 2004, 59 (3). Becker, B., Z. Ivkovich, and S. Weisbenner, “Local dividend clienteles,” NBER Work- ing Paper 15175, July 2009. Bernheim, B., “Tax policy and the dividend puzzle,” The RAND Journal of Economics, 1991, 22 (4). 58 Bhattacharya, S., “Imperfect information, dividend policy, and ‘the bird in the hand’ fallacy,” The Bell Journal of Economics, 1979, 1. Black, F. and M. Scholes, “The effects of dividend yield and dividend policy on common stock prices and returns,” Journal of FInancial Economics, 1974. Blouin, J., J. Raedy, and D. Shackelford, “Dividends, share repurchases and tax clien- teles: Evidence from the 2003 reductions in shareholder taxes,” 2010. NBER Working Paper 16129. Blume, M., E. Crockett, and I. Friend, “Stockholders in the United States: Character- istics and trends,” Survey of Current Business, 1974. Blundell, R., A. Duncan, and C. Meghir, “Estimating labor supply responses using tax reforms,” Econometrica, 1998, 87. Brav, A., J. Graham, C. Harvey, and R. Michaely, “Payout policy in the 21st cen- tury,” Journal of Financial Economics, 2005, 77. Brennan, M., “Taxes, market valuation and corporate financial policy,” National Tax Journal, 1970, 25. Brown, J., N. Liang, and S. Weisbenner, “Executive financial incentives and payout policy: Firm responses to the 2003 dividend tax cut,” FEDS Working Paper 2006-14, 2004. Bucks, B., A. Kennickell, and K. Moore, “Recent changes in U.S. family finances: Evidence from the 2001 and 2004 Surveys of Consumer Finances,” Federal Reserve Bulletin, 2006. Cameron, A.C. and P. Trivedi, Microeconomics methods and applications, Cambridge University Press, 2006. 59 Chaplinsky, S. and H. Seyhun, “Tax rationality and the demand for dividends,” 1987. Working Paper, University of Michigan. Chetty, R. and E. Saez, “Dividend taxes and corporate behavior: Evidence from the 2003 dividend tax cut,” Quarterly Journal of Economics, August 2005, CXX (3). Desai, M. and D. Dharmapala, “Taxes and portfolio choice: Evidence from JGTRRA’s treatment of international dividends,” NBER Working Paper No. 13281, July 2007. Eissa, N., “Labor supply and the Economic Recovery Tax Reform of 1981,” in M. Feldstein and J. Poterba, eds., Empirical Foundations of Household Taxation, Chicago, 1996. , “Tax reforms and labor supply,” in J. Poterba, ed., Tax Policy and the Economy, Vol. 10, MIT Press, 1996. Elton, E. and M. Gruber, “Marginal stockholder tax rates and the clientele effect,” Review of Economics and Statistics, 1970, 52. Feenberg, D. and E. Coutts, “An introduction to the TAXSIM model,” Journal of Policy Analysis and Management, 1993, 12 (1). Feldstein, M., J. Slemrod, and S. Yitzhaki, “The effects of taxation on the selling of corporate stock and the realization of capital gains,” Quarterly Journal of Economics, 1980, 94 (1). Galper, H., R. Lucke, and E. Toder, “A general equilibrium analysis of tax reform,” in H. Aaron, H. Galper, and J. Pechman, eds., Uneasy compromise: Problems of a hybrid income-consumption tax, Brookings, 1988. Gordon, R. and D. Bradford, “Taxation and the stock market valuation of capital gains and dividends,” Journal of Public Economics, 1980, 14. 60 and J. Vila, “Investors’ heterogeneity, prices, and volume around the ex-dividend day,” Journal of Financial and Quantitative Analysis, 1995, 30. Miller, M., “Debt and taxes,” Journal of Finance, 1977, 32. and F. Modigliani, “Dividend policy, growth, and the valuation of shares,” Journal of Business, 1961. and M. Scholes, “Dividends and taxes: Some empirical evidence,” Journal of Political Economy, 1982, 90. Modigliani, F. and M. Miller, “The cost of capital, corporation finance and the theory of investment,” American Economic Review, 1958. Moffitt, R. and M. Wilhelm, “Taxation and labor supply decisions of the affluent,” in J. Slemrod, ed., Does Atlas Shrug? The Economic Consequences of Taxing the Rich, Russell Sage Foundation and Harvard University Press, 2000. Newey, W., “Efficient estimation of limited dependent variable models with endogenous explanatory variables,” Journal of Econometrics, 1987, 36. Pettit, R., “Taxes, transactions costs and the clientele effect of dividends,” Journal of Financial Economics, 1977. Poterba, J., “Taxation, risk-taking, and household portfolio behavior,” in A. Auerbach and M. Feldstein, eds., Handbook of Public Economics, Vol. 3, Russell Sage Foundation and Harvard University Press, 2002. and A. Samwick, “Taxation and household portfolio composition: US evidence from the 1980s and 1990s,” Journal of Public Economics, 2002. Scholz, J. K., “A direct examination of the dividend clientele hypothesis,” Journal of Public Economics, 1992, 49. 63 , “Tax progressivity and household portfolios,” in J. Slemrod, ed., Tax Progressivity and Income Inequality, Cambridge University Press, 1994. Shefrin, H. and M. Statman, “Explaining investor preference for cash dividends,” Journal of Financial Economics, 1984, 13. and R. Thaler, “The behavioral life-cycle hypothesis,” Economic Inquiry, 1988, 26. Shoven, J. and C. Sialm, “Asset location in tax-deferred and conventional savings ac- counts,” Journal of Public Economics, 2003, 88. Smith, R. and R. Blundell, “An exogeneity test for a simultaneous equation Tobit model with an application to labor supply,” Econometrica, 1986. Stock, J. and M. Yogo, “Testing for weak instruments in linear IV regression,” Nov 2002. NBER Working Paper No. T0284. Strickland, D., “Determinants of institutional ownership: Implications for dividend clien- teles,” 1997. Working paper, Ohio State University. Wooldridge, J., Econometric analysis of cross section and panel data, Massachusetts In- stitute of Technology, 2002. 64