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.
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 Kawano^1 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 Feenbergfor 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 financeand labor seminars, various colleges and government agencies, and the 2007 Midwest Economics Association Annual Meetings provided useful feedback. Additional comments can be sent to laura.kawano@treasury.gov.
rate calculations and a rich description of portfolio structures, the combination of which is not common to other data sources. In addition, the timing of the 2003 tax act clearly separates tax regimes across the SCF samples. I exploit the resulting exogenous variation in tax rates to identify tax effects rather than relying on variation in a single cross-section. This paper is the first to test for dividend clienteles among the class of individual investors using a natural experiment. The second goal is to quantify the clientele-related economic impact of the 2003 tax act. Because the supply of dividends also changed, this paper is related to earlier studies of firm responses to the 2003 tax act that document the increase in dividend payments (Chetty and Saez (2005), Brown, Liang and Weisbenner (2004)). Note, though, that the overall supply of dividends increased does not inform how these dividends were distributed across households. This question can only be answered by directly considering changes to household portfolios, as is done here. The paper addresses two econometric issues. First, the dependent variable, a household’s portfolio dividend yield, has a mass point at zero. Second, the main regressor of interest, tax rates, is endogenous to investor choices. To account for these issues, I estimate a Tobit-type model with instrumental variables techniques. The natural experiment framework provides an instrumental variable that is preferable to those used in previous research designs. Specif- ically, the different intensities of tax treatment that households face provides the basis for separating households into low- and high-treatment groups used to identify the effects of taxes. I find strong evidence for the dividend clientele hypothesis. I estimate that the rela- tionship between the tax disadvantage of dividend income and household portfolio dividend yields is negative and statistically significant. This suggests both that taxes cause a high degree of investor sorting and that households quickly responded to the tax changes caused by the 2003 tax act. In particular, affluent households shifted their portfolios, either actively
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.
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
(Dnodiv) for an investor with wealth level W are given by:
Ddiv = W, Dnodiv = 0 if (1 − τdiv)rdiv < rnodiv (1) Ddiv = 0, Dnodiv = W if (1 − τdiv)rdiv > rnodiv (2)
Generalizing to the case of multiple equities with varying dividend yields, “high dividend paying stocks will be preferred by tax exempt organizations^1 and low income investors; those stocks yielding more of their return in the form of capital gains will gravitate to the taxpayers in the upper tax brackets” (Miller 1977).^2 This model also shows how clienteles can shift in response to changes in the tax rate structure. The tax rate that defines the cusp for household portfolio specialization in the simple model changes with the progressivity of the tax system. It is important to note that Miller’s (1977) model concerns the equity market equilibrium and not an individual firm’s choice over its payout policy. The model does not predict which firms pay dividends; indeed, from the perspective of any one firm, each clientele is as good as the next. That is, firms do not choose their dividend policy to attract a particular group of investors. Miller’s equilibrium provides intuition for how asset holding clienteles may emerge when investors face differences in tax rates. Yet the model is incomplete because it assumes all assets are riskless. When forming its equity portfolio, a household considers not only the impact of taxes on expected returns but also the riskiness of these holdings. To formally derive the relationship between tax rates and optimal dividend portfolio yields, I combine a model of optimal portfolio dividend yields, which defines the set of after-tax efficient portfolios for an investor with particular tax rates and risk preferences, with the after-tax (^1) There are additional non-tax reasons that tax-exempt institutional investors may form their own clientele. Because institutions are more likely to engage in “due diligence” and equilibrium prices make dividend-paying stocks more attractive to institutional investors, firms may use dividends to signal quality (Allen, Bernardo and Welch 2000). 2 Where foreign investors align in the market for equities will depend on the tax treatment of his income derived from US equities in the US and in their country.
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 = b i 0 + bi 1 γ 0 + bi 1 γ 1 βp 1 − bi 1 γ 2 (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 (^4) Without 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. 5 Dividends from most foreign corporations, credit unions and banks were excluded from “qualified” dividend income. Non-qualified dividends remained taxed as part of ordinary income.
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 (^6) Taxpayers on the Alternative Minimum Tax schedule also benefited from the reduction by facing a reduction from the 28% flat rate to 15%.
values are replaced using a multiple imputation technique. These multiple imputations improve the efficiency of the point estimates by increasing the sample size, but as with any imputed values, require that the missing observations be conditionally random. All summary statistics, regressions and their standard errors are corrected for multiple imputations.^8 The dependent variable is a household’s portfolio dividend yield, defined as the ratio of the dollar value of dividend income to the dollar value of taxable equity. This measure represents a household’s weighted-average dividend yield on its taxable equity. Dividend income is the dollar amount of ordinary dividend income received from stocks in taxable accounts in the previous calendar year.^9 Taxable equity is the sum of stocks held directly, stocks held through mutual funds, and stocks held in trusts, annuities, or other managed investment accounts. Equity held in mutual funds is the sum of the full value of stock mutual funds and half the value of combination mutual funds. The full value of other managed assets is included if it is mostly invested in stock, half the value if it is split between stocks and bonds, or stocks and money market accounts, and a third of the value if it is split between stocks, bonds, and money market accounts. The dollar value of equity is the market value at the time of interview, conducted in the second half of the survey year.^10 Stocks held in 401Ks, IRAs or other qualifying retirement accounts, as well as dividend income received from such securities, are not included in this measure. This exclusion is important because the tax rate reductions for dividends do not apply to equities in tax-deferred accounts. However, I am unable to identify if 2004 dividend yields contain stocks shifted between taxable and tax-deferred accounts. All components are adjusted to 2004 dollars. (^8) See Kennickell (1998) for an overview of the multiple imputation methodology. (^9) This value should correspond to item 9 on IRS form 1040 in 2000 and item 9a on IRS form 1040 in 2003/2006, and reported on a 1099-DIV. 10 The 2001 SCF was conducted between May and December 2001, while the 2004 SCF was conducted between June 2004 and February 2005. The difference in timing may bias the yield measure if the equity holdings at the time of the survey are not representative of the equity holdings from which the dividend income was drawn. Unfortunately, there is no information in the survey that informs on the direction of this bias. Small denominator values may create outliers, so sensitivity checks to the influence of outliers are provided in the analysis.
To compute marginal tax rates on dividends and capital gains, I construct household adjusted gross income and deductions information from variables provided in the SCF. Then, I pass a flat file of these variables through the National Bureau of Economic Research’s TAXSIM web program to compute statutory federal marginal tax rates.^11 The effective tax rate on long-term capital gains is lower than the statutory rate because taxes on capital gains are deferred until they are realized and because capital gains that are accrued until death qualify for a “basis step-up,” which excuses the tax liability on such gains. I compute effective long term capital gains tax rates following (King and Fullerton 1984), who argue that the statutory tax rate on capital gains should be halved to account for the option value of tax-deferral, and halved again to account for the step-up basis at death and the selected realization of losses.^12 Figure 3 is a plot of the average effective dividend and capital gains tax rate by income percentile computed from the two samples. This figure shows that the treatment effect is larger for high income households than for lower income households. Because the dividend clientele hypothesis regards the relative tax treatment of dividend income and capital gains, I use the difference in effective dividend and capital gains marginal tax rates as the main re- gressor of interest.^13 The gap between the two lines represents the absolute tax disadvantage of dividends. The validity of using estimates from the SCF surveys to infer the effect of the 2003 tax (^11) Stata programs that convert SCF data into variables required for TAXSIM are available at the NBER website. The TAXSIM programs are found at http://www.nber.org/∼taxsim/to-taxsim/. See Feenberg and Coutts (1993) for a description. State tax rates are a potentially useful source of tax rate variation. How- ever, to maintain anonymity, state identifiers are omitted from the public SCF datasets so this informationcannot be used. (^12) Ivkovic, Poterba and Weisbenner (2005) use individual stock holding data to estimate the effective capital gains tax rates for various stock holding patterns, prospective appreciation rates, and whether stocks were held in taxable or non-taxable accounts. Various assumptions provide a wide range of simulated effective tax rates. They do not have demographic information that might predict effective tax rates, so I use the long-established convention of using 25% of the statutory rate to measure the effective capital gains rate. 13 This is the numerator of the tax rate variable described in equation 3. I use this measure because it nicely captures the relative tax disadvantage of dividends. This is the same tax variable used in Scholz (1992).
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
Table 1: Summary statistics of demographic and socioeconomic variables
Share of SCF Sample Observations are weighted by their sampling weights. Financial data are in 2004 dollars. Demographic characteristics refer to the head
- Variable
- 0-15 0.14 0.14 0. Income (thousands)
- 15-25 0.11 0.12 0.
- 25-50 0.27 0.26 0.
- 50-75 0.16 0.18 0.
- 75-100 0.12 0.10 0.
- 100-250 0.15 0.17 0.
- 250+ 0.03 0.03 0.
- 0-50 0.38 0.38 0. Net worth (thousands)
- 50-100 0.12 0.11 0.
- 100-250 0.19 0.18 0.
- 250-1000 0.23 0.23 0.
- 1000+ 0.09 0.09 0.
- No degree 0.09 0.09 0. Demographic characteristics
- High school degree 0.31 0.30 0.
- Some college but no college degree 0.18 0.18 0.
- College degree 0.34 0.37 0.
- Not willing to take financial risks 0.40 0.42 0.
- Female 0.27 0.28 0.
- Married 0.60 0.58 0.
- Household size 2.41 2.39 2.
- Retired 0.19 0.19 0.
- Average Age 48.97 49.56 50.
- Number of households (millions) 106.5 112.1 116.
- Number of observations
Table 2: Dividend receipts and equity holdings by income Percentage of Percentage of Dividends as a Income total dividends total equity percent of income Percentile 2001 2004 2007 2001 2004 2007 2001 2004 2007 0-10 1.21 0.98 0.84 0.80 0.97 1.49 1.42 0.90 0. 10-20 0.57 0.35 0.60 0.60 0.41 0.64 0.43 0.28 0. 20-30 1.57 1.20 1.19 0.90 1.01 1.04 0.74 0.46 0. 30-40 2.30 0.88 0.93 2.28 1.84 2.70 0.70 0.28 0. 40-50 3.31 1.15 1.87 1.85 2.23 1.45 0.97 0.28 0. 50-60 5.18 3.30 1.29 3.48 3.93 1.80 1.13 0.70 0. 60-70 6.13 4.93 4.58 5.82 5.98 4.30 0.96 0.73 0. 70-80 4.33 4.23 6.34 6.34 5.99 7.57 0.53 0.50 0. 80-90 10.54 6.45 8.60 9.86 7.96 7.12 1.04 0.59 0. 90-95 13.39 7.36 5.44 10.89 8.61 6.89 1.88 0.88 0. 95-99 23.59 26.64 24.92 28.58 24.69 25.14 2.09 2.15 2. 99-100 27.88 42.53 43.39 28.59 36.39 39.86 2.03 3.32 3. Source: Author’s calculations using SCF data. Observations are weighted by their SCF sampling weights. Statistics are corrected for multiple imputations.
4.2 Estimation strategy
To examine the existence of tax-based dividend clienteles, I consider the relationship be- tween household portfolio dividend yields and tax rates. Because I am interested in the mix of equities that households choose to hold, rather than the choice of whether to hold equities, I focus on equity-holding households in the main analysis. Additionally, I exclude 7 observations with dividend yields of over 1000%.^17 I use several other cut-off values in the sensitivity analysis to ensure that the main estimates are robust to this choice. Since many equity-holding households do not receive dividend income, there is a mass point in the dependent variable at zero. I treat these observations with dividend yields equal to zero as households for whom no dividend income is preferred to receiving some. This suggests a censored regression model (Type II Tobit) that Wooldridge (2002) calls the “corner solution model” because there is a mass point that results from household optimization. (^17) These large outliers likely arise because some households who received dividend income in the year prior to the survey liquidated their equity holdings by the time of the survey. When excluding households with yields over 1000%, the maximum dividend yield is 650%.
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
