The impact of marginal tax rates on taxable income: evidence from state income tax differentials.

Long, James E.

1. Introduction

The relationship of taxable income to the marginal tax rate has important implications for both the revenue consequences of tax policy and the deadweight loss of the income tax.(1) Not surprisingly, then, Feldstein's (1995b) analysis of the 1986 tax reform, in which he concluded that taxable income is highly responsive to changes in the marginal tax rate, has been closely examined and subjected to certain criticisms, which are summarized in Slemrod (1995b), Auten and Carroll (1995), and Goolsbee (1998). Two common criticisms of Feldstein's net-of-tax rate (i.e., one minus the tax rate) elasticity, which averaged either 1.26 or 2.10 depending on how taxable income was measured, were (i) that his panel data contained too few high-income taxpayers (e.g., there were only 57 individuals in the 49 or 50% tax bracket in 1985) and (ii) that his analysis was incomplete because it excluded nontax determinants of taxable income.

Auten and Carroll (1995) addressed these and other criticisms in their reexamination of Feldstein's findings. Their estimate of a "taxes only" and a multivariable model of taxable income using a substantially larger data set (containing 4387 taxpayers in the top two brackets) yielded taxable income elasticities smaller than (in some instances less than half as large as) those reported by Feldstein. The methodological approach of the Feldstein and comparable longitudinal studies, in which changes over time in taxable income of different groups are compared to the changes in the tax rates of the groups, has also been criticized. Goolsbee (1998) argues that it rests on a questionable assumption, namely, that lower-income individuals are a valid control group for higher-income individuals. If high-income individuals have incomes that trend upward at a relatively faster rate or are more likely to be in forms whose timing can be shifted in the short run, then Goolsbee believes that the Feldstein net-of-tax rate elasticities may be substantially overstated, perhaps by 75% or more.

In light of criticisms of existing studies, additional research on the impact of tax rate changes on taxable income is warranted, and a study that uses an alternative approach to measure the tax or net-of-tax rate elasticity may yield independent evidence on this important issue. In this article, the relationship between taxable income and the marginal tax rate is investigated using a cross section of federal individual tax returns filed for tax year 1991. The tax rate effect is identified on the basis of differences in state income tax rates among individuals having the same income and demographic characteristics. Because of reductions in the number of federal income tax brackets during the 1980s, basically all married taxpayers with incomes above about $80,000 faced the same statutory federal marginal tax rate in 1991, which is problematic for measuring tax rate or "price" effects. In contrast, state marginal tax rates in this income range varied considerably (from 0 to 12%), and the state income tax's share of the total income tax rate was substantially higher than in earlier periods. Previous research has exploited the variation in state tax rates to examine the impact of taxation on capital gains (Burman and Randolph 1994; Bogart and Gentry 1995), charitable contributions (Feenberg 1987; Ribar and Wilhelm 1995), individual retirement account (IRA) contributions (Long 1990), and income tax avoidance (Long and Gwartney 1987). However, state tax rate differentials have generally not been used to identify the response of taxable income as a whole to marginal tax rate changes.

2. Methodology, Data, and Preliminary Evidence

As noted above, the basic approach followed in this study was a comparison of the taxable incomes of individual taxpayers who reside in different states and confront unequal state income tax rates. If taxable income is inversely related to the marginal tax rate, as results of recent studies suggest, then taxable income, on average, should be lower in high-tax states than in low-tax states. Data for this comparison were reported on individual tax returns contained in the Internal Revenue Service 1991 Individual Public Use Tax File, a stratified random sample of the Form 1040, Form 1040A, and Form 1040EZ federal tax returns filed for tax year 1991. To identify a taxpayer's state of residence and to measure the state marginal income tax rate, the analysis is restricted to tax returns with an adjusted gross income of less than $200,000.(2) Also excluded are tax returns filed by individuals (i) residing outside the 50 states and District of Columbia and (ii) without positive income.(3) These restrictions resulted in a sample of 66,723 tax returns, 10,839 of which reported incomes in 1991 of $100,000 or more. To make comparisons to the Feldstein and Auten and Carroll studies, analyses were also conducted on a smaller subsample of 30,796 returns filed by married taxpayers under the age of 65 who were not subject to the alternative minimum tax. There were 5292 tax returns in the subsample reporting incomes in the $100,000 to $200,000 range in 1991.

As a preliminary to the more rigorous econometric analyses described below, the following simple test was conducted. Individual tax returns were sorted into eight income groupings and then placed into one of two categories: (i) returns filed by residents of states with maximum state income tax rates of 8% or more and (2) returns of taxpayers from states with maximum rates of less than 3%.(4) A t-test was then performed to determine whether mean taxable income for each income cell differed between high- and low-tax states as defined above. Because state income tax payments are deductible on the federal return, a finding that taxable income is relatively lower in high-tax states might merely reflect differences in state tax liabilities rather than behavioral responses by individuals to higher tax rates.(5) Consequently, the t-tests (and all analyses reported below) were conducted on modified taxable income, defined as follows. If total itemized deductions exclusive of state(6) income taxes paid exceed the standard deduction amount, modified taxable income equals adjusted gross income minus nonincome tax deductions minus personal exemptions. Otherwise, modified taxable income equals adjusted gross income minus the standard deduction minus personal exemptions. In effect, this adjustment results in deductible state income tax payments being added back to federal taxable income.

Mean taxable incomes according to income grouping in the high- and low-tax state samples are reported in Table 1.(7) In all but the lowest income category, average taxable income is relatively lower in states with higher tax rates on individual income. The t-statistics in the taxable income column indicate that the reductions in taxable income associated with high marginal tax rates are statistically different from zero at the 0.01 level in all but one instance. A preview of what phenomena might explain why taxable income is negatively related to the marginal tax rate is provided in the last two columns, which describe how taxpayers in high- and low-tax states differ in terms of (i) deductible expenses, consisting of itemized deductions other than state income taxes paid plus adjustments to income such as IRA contributions; and (ii) the amount of negative incomes (losses) that might reflect tax shelter investments. These activities constitute the avenues for tax avoidance that can be investigated with the type of data used in this study. The pattern of these high-tax/low-tax [TABULAR DATA FOR TABLE 1 OMITTED] differentials strongly suggests that taxable income declines as the marginal tax rate rises because individuals increase tax-deductible expenditures.

Although the results in Table 1 are suggestive, the analysis described here does not constitute a definitive test of the direction or source of the relationship between taxable income and the marginal tax rate. For example, there may be differences between taxpayers living in high-tax and low-tax states, perhaps in age, marital status, or sources of income, that underlie the observed differentials by state in deductible expenditures and taxable income. Furthermore, the state income tax rate may be correlated with other variables that affect taxpayer behavior; hence, the reduction in taxable income should be partially attributed to these variables rather than to high tax rates alone. In the next section, these issues are addressed by a multivariate analysis of taxable income.

3. Econometric Analysis

The econometric model specifies the taxable income reported on individual tax returns to be a function of the state marginal income tax rate and various nontax control variables suggested by previous studies in this area. Following the customary practice in empirical studies of taxpayer behavior, the marginal tax rate is measured on a "first-dollar" basis (Feldstein 1975). This is the rate of taxation on the marginal dollar of positive income and thus measures the tax savings resulting from the first dollar of income tax avoidance. Operationally, each individual taxpayer is assigned the effective state marginal tax rate applicable to preavoidance income, which is defined as total positive income minus the applicable personal exemptions under the state income tax.(8)

Using the formula for the total marginal tax rate found in Feenberg (1987), Ribar and Wilhelm (1995), and elsewhere, the effective state marginal tax rate (SMTR) is computed using the following equation:

SMTR = [(f + [sd.sub.s] - [fst.sub.s] - [fsd.sub.s])/(1 - [t.sub.s]fs)] - f (1)

where f and s are the statutory marginal federal and state rates, respectively; [t.sub.s] is 1 if federal taxes can be deducted on the state return and 0 otherwise; and [d.sub.s] is 1 if itemized deductions are allowed on the state return and 0 otherwise.(9) Feenberg (1987) has persuasively argued that tax price effects be identified by using an instrumental variable estimator to exploit any variation in tax rates independent of personal characteristics that might influence taxpayer behavior. Consequently, in most of the equations reported below, the actual value of SMTR is replaced by its predicted value, which is obtained by regressing SMTR on the maximum statutory marginal tax rate (MAXSR) in the taxpayer's state of residence and dummy variables identifying states that allow itemized deductions and permit federal taxes paid to be deducted.(10) Use of this instrumental variable (IVSMTR) eliminates the intrastate variation in the tax rate by income that occurs under graduated state income taxes.

In a cross-sectional analysis such as this, the taxpayer's reported positive income (INCOME) is probably the most important nontax variable for explaining the level of taxable income. Interacting the tax rate and income variables permits testing for a pattern of tax rate effects. Lindsey (1987) and Long and Gwartney (1987) have found that higher-income taxpayers have higher elasticities of taxable income with respect to the marginal tax rate. The primary demographic characteristics of the taxpayer are measured by dummy variables for married with spouse present (MARRIED) and age 65 or older (AGE65) and the number of exemptions for children or other dependents (DEPS). The percentage of income received as wages and salaries (LABOR) serves as an additional proxy for age, because young individuals tend to receive relatively more income from labor than capital; LABOR can also be regarded as an inverse measure of wealth and the flexibility for altering portfolios to shelter income from taxation (Auten and Carroll 1995).(11) To serve as a proxy for entrepreneurship skills and propensity for risk taking, a dummy variable (BUSINESS) identifying taxpayers who receive income from a proprietorship, partnership, small-business corporation, or rental/royalty properties and pay self-employment taxes is also included in the model.

The estimated relationship between taxable income and the state tax rate (measured either directly or via an instrument) will be biased if the taxable income equation omits variables that affect taxable income and are correlated with the state tax rate. For example, if high-tax states spend more on antipoverty programs than low-tax states, and if public transfers crowd out private charitable donations (Abrams and Schmitz 1984), the tendency for higher tax rates to reduce taxable income (because deductions for charitable contributions would otherwise increase as their tax-price or net cost falls) will be understated. The bias would be in the opposite direction if the state tax rate was positively correlated with an omitted variable that increased itemized deductions. This could occur, for instance, if home prices and annual mortgage interest payments are relatively higher in states with relatively higher personal income tax rates. To account for possible omitted-variables bias, the taxable income model includes two additional control variables, WELFARE, state government expenditures for public welfare in 1991 per $1,000 of personal income in the state, and HOUSEVALUE, median 1990 value of owner-occupied homes in the taxpayer's state, deflated by state personal income per capita to make the variation in HOUSEVALUE reflect interstate differences in housing prices per unit rather than housing consumption levels.

4. Basic Empirical Results

Because nearly 20% of the 66,723 tax returns used to estimate the taxable income model report zero taxable income for 1991, the equation is estimated using the Tobit maximum likelihood technique.(12) The results of estimating the model on the full sample of tax returns (see Appendix) are presented in the first four columns of Table 2. The coefficient of SMTR in Column 1 of Table 2 is highly significant, and the estimate indicates that a one-unit increase in the actual state marginal tax rate is associated with a decline in taxable income of $70 on average.(13) In Column 2, the actual state tax rate is replaced by the instrumental variable estimator; the estimated coefficient of IVSMTR is significant at the 0.01 level and implies that taxable income falls by $80 with each one-unit rise in the tax rate. Column 3 reveals that the tax rate/taxable income relationship is nonlinear in that a tax rate increase has a larger (more negative) impact on taxable income at high income levels than at low incomes.(14) For example, a one-unit increase in the state tax rate reduces taxable income by $249 and $681 at positive incomes of $50,000 and $100,000, respectively. In comparison, the corresponding marginal impacts of a higher tax rate are $-425 and $-876 when the taxable income model omits state characteristics proxied by WELFARE and HOUSEVALUE, which indicates that the tax rate effect is overstated if omitted determinants of taxable income are correlated with the state income tax structure.(15) The key omitted variable is HOUSEVALUE, because omitting WELFARE has little effect on the tax rate coefficients,(16) but as Column 4 shows, the marginal tax rate impact is more than $100 greater (more negative) at every income level when HOUSEVALUE is omitted from the taxable income model.

The estimated coefficients of INCOME and its squared value (INCOMESQ) are highly significant in every equation, and their signs indicate that taxable income rises with positive income but at a diminishing rate, which is not surprising since studies have found that individual components of tax avoidance (such as charitable contributions, interest deductions, or IRA contributions) are directly related to income (Feenberg 1987; Long 1989, 1990), and more tax avoidance results in lower taxable income. Furthermore, under the progressive federal income tax in 1991, an increase in income may have resulted in a higher federal marginal tax rate and hence a lower tax price for deductible expenditures. The demographic and other nontax factors in the taxable income model are generally statistically significant and perform as expected. Specifically, taxable income in 1991 is relatively lower for married couples than for other taxpayers, an additional dependent reduces taxable income by about $2800, and elderly taxpayers report higher taxable incomes than nonelderly persons. The LABOR coefficient is always positive and highly significant, but the implied effect on taxable income is fairly small. Taxpayers who receive income from business sources and pay self-employment taxes report relatively lower taxable incomes, other things equal. The inclusion of LABOR and BUSINESS in the taxable income equation may be criticized on the grounds that the tax rate influences the forms in which personal income is received, but deleting these two explanatory variables has little impact on the estimated coefficients of the state tax rate and other variables.(17)

The control variables WELFARE and HOUSEVALUE are statistically related to taxable income, but their marginal impacts differ in size and direction. For example, according to Column 3 of Table 2, an increase in WELFARE of one standard deviation raises taxable income by about $70, whereas the same increase in HOUSEVALUE is associated with a taxable income decline of around $600. The manner in which these state characteristics affect taxable income would be expected if public antipoverty spending displaced deductible private charitable contributions and if annual mortgage interest expenses were directly related to home prices.

The last two columns of Table 2 show the results of estimating the taxable income equation on the smaller sample of married, nonelderly taxpayers who were not subject to the alternative minimum tax. The coefficients of the IVSMTR terms carry the same signs and high significance levels as they exhibited for the full sample, and their magnitudes are similar, especially when HOUSEVALUE is omitted from the model. According to Column 5, an increase of one percentage point in the state tax rate reduces taxable income by $116 when positive income is $50,000 and by $520 when income is $100,000. The corresponding marginal tax rate effects are much more negative ($-330 and $-751), and the possible estimation bias is therefore larger [TABULAR DATA FOR TABLE 2 OMITTED] than in the full sample case, when the state housing price proxy is omitted from the equation. Note also that the coefficients of WELFARE and HOUSEVALUE are nearly twice as large for the married taxpayer sample than the full sample.

The state income tax rate coefficients reported in Table 2 were subjected to a number of sensitivity tests. As noted earlier, zero taxable income is reported on a substantial fraction of tax returns in the full sample, conceivably because some taxpayers extensively engage in tax avoidance and effectively zero out taxable income but most likely because these returns are filed by individuals who avoid taxation merely by having incomes lower than the threshold above which taxes are imposed. To investigate how the presence of these low-income tax returns in the sample influences the estimated marginal tax rate effect, the taxable income equation was reestimated after deleting tax returns for which positive income was less than or equal to the standard deduction (which varies by marital status) plus the personal exemptions claimed by the primary and secondary taxpayer. The coefficients of the IVSMTR terms remained highly significant, and the estimated negative impacts of the tax rate on taxable income became slightly larger (about 4% so) at incomes above $100,000 and somewhat smaller at lower income levels. The only appreciable consequences of omitting low-income returns from the analyses was an increase in the AGE65 coefficient (to $1405), which may be explained by the fact that elderly taxpayers can claim a relatively larger standard deduction than taxpayers under 65. Furthermore, the estimated tax rate coefficients and significance levels were essentially unchanged when a dummy variable identifying taxpayers subject to the federal alternative minimum tax was added to the taxable income model. Finally, accounting for the fact that property taxes (which are also deductible on the federal return) may be correlated with the state tax rate likewise had only a modest effect on the estimated coefficients of the state tax rate terms.(18)

5. Magnitude and Source of the Tax Rate Effect on Taxable Income

As noted when discussing the tax rate-income interaction terms in Table 2, an increase in the state marginal tax rate reduces taxable income more when positive (potentially taxable) income is high than when income is low. The estimated tax rate impacts, calculated for various income levels using the IVSMTR and IVSMTR x INCOME coefficient estimates in Column 3 of Table 2 are reported in Column 1 of Table 3. These estimates reveal that a one-unit increase in the effective state tax rate reduces taxable income by amounts ranging from $-32 to $-1330. The simple test reported in Table 1 corroborates these conclusions, which were derived from [TABULAR DATA FOR TABLE 3 OMITTED] the multivariate analysis of taxable income. The maximum state income tax rate was about eight percentage points higher for residents of high-tax states than for those of low-tax states, but after accounting for cross-deductibility, the effective state tax rate (as defined in Equation 1 above) differential was as much as 5.8 points or as little as 3.9 points, depending on the income level. Averaging the taxable income differentials for the $25-50 and $50-75 cells in Table 1 and dividing the result by 5.8 yields a taxable income differential per unit of tax rate of $-159; Table 3 reports a tax rate impact of $-249 at $50,000. Similarly, Table 1 implies an average tax rate effect for the $125-150 and $150-175 groups of $-1312, whereas Table 3 suggests that at $150,000, an increase in the marginal tax rate reduces taxable income by $1114.

Given the behavioral responses included in this study, the reduction in taxable income occurs because a higher tax rate prompts taxpayers to either make investments that generate tax losses, spend more on deductible goods and services, and/or claim larger adjustments to income. To investigate which of these types of tax avoidance are most important, the state tax rate and other independent variables included in the taxable income model were posited to be determinants of (i) tax losses and (ii) deductible expenditures as defined in Table 1, and their effects were estimated using the Tobit technique. The coefficients of IVSMTR and IVSMTR x INCOME were not statistically different from zero in the tax losses equation but were highly significant in the deductible expenses equation.(19) Column 2 of Table 3 shows that the predicted increase in deductible expenditures associated with a one-unit increase in the state tax rate constitutes 81-85% of the decrease in taxable income, because of a higher tax rate, at all but the lowest income level. Once more, this finding is highly consistent with the "explanation" of the tax rate-taxable income relationship suggested by the data in Table 1.

These estimates of changes in taxable income and deductions are derived using the Tobit equations that contain HOUSEVALUE and WELFARE. To the extent that the demand for owner-occupied housing is relatively greater because of a high state tax rate and this leads to higher housing prices (and interest deductions), including HOUSEVALUE in the Tobit equations dilutes the state tax rate effect. Hence, the entries in the first two columns of Table 3 can be labeled lower-bound or minimum estimates. Computing the changes in taxable income and deductions using coefficient estimates from Tobit equations that omit HOUSEVALUE, like that in the fourth column of Table 2, yields what can be regarded as upper-bound or maximum tax rate effects. For comparative purposes, these estimates are reported in the last two columns of Table 3. They suggest that a higher tax rate will reduce taxable income by amounts ranging from $-148 to $-1478, with the bulk of tax avoidance taking the form of increased deductible expenditures.

The general conclusions suggested by Table 3 are maintained when the changes in taxable income and deductible expenditures are derived from the Tobit equations estimated using the subsample of married, nonelderly, ordinary tax returns (results not shown). In fact, increases in deductible expenses constitute an even larger component of taxable income reductions than is true for all taxpayers. Additional, and arguably more direct, evidence that most of the response in taxable income to a higher tax rate stems from increased deductions was obtained by exploiting the fact that states differ in the ability of taxpayers to itemize deductions or claim adjustments to income on the state tax rerum. Deductibility indirectly determines the effective state marginal tax rate, as shown by Equation 1, but its role was investigated directly in the following manner. The taxable income and deductible expenditures models were reestimated after two specification changes were made: (i) the effective state income tax rate was replaced by the statutory state tax rate, the s in Equation 1, and (ii) the tax rate and tax rate-income interaction variables were each interacted with a dummy variable identifying states permitting itemized deductions and adjustments to income (PERMIT). The Tobit estimates confirmed, as expected, that a tax rate increase reduces taxable income relatively more (and raises deductible expenses more) in states allowing itemized deductions and income adjustments than in other states.(20)

The conclusion that increases in deductible expenditures are actually a more important form of tax avoidance than tax shelter investments contrasts sharply with the conclusions of Long and Gwartney's (1987) analysis of individual income tax avoidance in 1979. At that time, many upper-income taxpayers faced maximum marginal tax rates in the 50-70% range, accounting losses could be used to offset other income, and capital gains were preferentially taxed. They found that for taxpayers with positive incomes between $80,00 and $120,000 (which corresponds roughly to the $150,000 to $200,000 range in 1991), a one-unit increase in the marginal tax rate raised tax losses by $1820 but increased itemized deductions (exclusive of state income taxes paid) by only $554. Apparently, tax shelters such as real estate partnerships made more sense in 1979 than 1991, the latter a year in which the maximum federal marginal tax rate was only 31%, realized capital gains were almost fully taxed, and certain passive losses were disallowed.(21)

[TABULAR DATA FOR TABLE 4 OMITTED]

6. Elasticity Estimates

To compare the findings of this study with those of related studies, the estimated minimum and maximum tax rate impacts were transformed into elasticities of taxable income with respect to tax and net-of-tax rates.(22) The minimum taxable income elasticities with respect to the income tax rate range from -0.028 in the under $50,000 income group to -0.355 in the top income category (see Table 4). The average elasticity across all income levels is -0.069 when the number of tax returns is the weighting factor and -0.161 when weighting by taxable income in each category. The corresponding maximum estimates are larger in absolute value, ranging with income from -0.212 to -0.395 and averaging -0.236 or -0.290. For the married taxpayer sample, the lower-bound tax rate elasticities are considerably smaller (less negative), especially at lower incomes, but the maximum tax rate elasticities are very close to the corresponding estimates for taxpayers in aggregate. The pattern of these taxable income elasticities - the higher the income, the greater the elasticity - is consistent with tax rate elasticities Long and Gwartney (1987) estimated using a comparable sample of individual income tax returns for tax year 1979. However, taxpayers were apparently much more responsive to tax rate changes in 1979 than 1991, because Long and Gwartney concluded that in 1979, the taxable income elasticity with respect to the marginal tax rate was probably in the -0.6 to -1.5 range for incomes above $60,000.

In terms of the net-of-tax rate, the minimum taxable income elasticities for the full sample rise with income from +0.112 to +0.694; the corresponding estimates for the subsample of married taxpayers are somewhat lower. The maximum net-of-tax elasticities hover in the 0.71 to 0.83 range; they fall rather than rise with gross income because the marginal tax rate impact is especially overstated for lower- and middle-income taxpayers when state housing prices are not held constant in the taxable income model. This explanation is consistent with Internal Revenue Service data showing the prominence of home mortgage interest payments as a federal tax deduction for middle-income taxpayers (Long 1989). Regardless of the sample and model specification considered, all these estimated 1991 elasticities are considerably lower than ones reported by Lindsey (1987), whose analysis of 1982-1984 tax returns suggested a net-of-tax rate elasticity of probably 1.6 to 1.8. Lindsey's higher elasticities are not surprising, because the tax regime he studied offered more opportunities for tax avoidance than existed after the 1986 tax reform. Furthermore, the 1991 net-of-tax rate elasticities are lower than those reported by Feldstein (1995b), which range from 1.04 to 3.05, as well as the elasticity values of 1.5 to 2.3, which Auten and Carroll (1995) obtained when they replicated Feldstein's analysis using the same small number of high-income individuals but a different time interval. As noted in the Introduction, the Feldstein and related elasticities may overstate the responsiveness of taxable income to tax rate changes.

Both the Feldstein and Auten and Carroll studies incorporate a full range of adjustments to tax rate changes, including ones like changes in labor supply and the form of employee compensation and the timing of capital gains realizations, which alter the level of total income reported by individuals. Furthermore, the samples analyzed by these researchers include the truly high-income taxpayers, who have the most incentive as well as ability to manipulate taxable income in response to higher tax rates. In contrast, the taxable income elasticities in the present study capture the impact of a much more limited set of behavioral responses to tax rate changes, and the estimated elasticities do not incorporate the expected greater responsiveness on the part of taxpayers with incomes above $200,000. Consequently, the relatively smaller net-of-tax rate elasticities reported in the present study are not unanticipated. Interestingly, most of the net-of-tax rate elasticities shown in Table 4 are comparable in size to Auten and Carroll's best estimate of the elasticity, a value of around 2/3 (Goolsbee 1998). Therefore, one important implication that can be drawn from the present study is that changes in deductible expenditures are more important than the combined effects of other avoidance responses and real responses, such as factor supply changes, that alter taxable income.(23) In fact, adjustments to deductible expenditures may be the only long-run behavioral response that alters taxable income when tax rates are changed. This viewpoint is consistent with Auten and Carroll's (1995) finding that in some regression specifications, tax rate changes are statistically related to taxable income but not adjusted gross income. A tax rate increase would not affect adjusted gross income if individuals did not alter labor supply, compensation arrangements, capital gains, or tax shelter investments, but a tax-induced increase in deductible expenditures would reduce taxable income. That itemized deductions are the major transmission route for taxpayers to respond to tax rate changes is perhaps the most significant finding of the current study and one that warrants confirmation through additional research.

7. Summary and Conclusions

In this study, the relationship between the marginal tax rate and taxable income was estimated using a cross section of 1991 individual income tax returns in which the tax rate varies across taxpayers with the same gross income and demographic characteristics because of differences in state income taxation. The results confirm Feldstein's recent finding of a negative relationship between taxable income and tax rates, although the responsiveness of taxpayer behavior to tax rate changes found in the present study was considerably less than what Feldstein discovered. Instead, the overall taxable income elasticity estimates derived from cross-sectional data were closer to some of the values reported by Auten and Carroll, who, like Feldstein, also compared tax returns of the same individuals before and after the 1986 tax reform. However, unlike Feldstein, Auten and Carroll's longitudinal (or panel) sample contained a sufficiently large number of high-income individuals to allow their results to be generalized to the actual taxpaying population.

The empirical results presented here are consistent with those of previous studies that have found high-income taxpayers to be more responsive to tax rate changes than low- and middle-income individuals. For example, the elasticity of taxable income with respect to the marginal tax rate was estimated to be no greater than about -0.20 for taxpayers with incomes under $50,000 versus around -0.40 for taxpayers with incomes above $150,000 in 1991. The latter elasticity implies that revenue losses due to a reduction in the tax base would offset a substantial portion of the tax revenue gains due to higher marginal tax rates. Therefore, it is not surprising that the upper-income tax rate increases legislated in 1993 raised far less revenue than would have been generated had there been no behavioral responses by taxpayers to the higher rates.(24) Nor should one be surprised if an across-the-broad cut in marginal tax rates produces a much smaller loss in tax revenue than static projections indicated, especially if taxpayers respond on a full set of margins rather than the limited ones included in this study.

[TABULAR DATA FOR APPENDIX OMITTED]

I thank two anonymous referees and Editor Jonathan Hamilton for helpful comments on earlier drafts.

1 Both Feldstein (1993) and Barro (1993) publicly questioned, in the The Wall Street Journal, the revenue-raising capability of the upper-income tax rate increases contained in the Omnibus Budget Reconciliation Act of 1993. Additionally, Feldstein (1996) refuted criticism that the 15% tax rate cut proposed by 1996 presidential candidate Bob Dole would lead to extreme revenue losses. Feldstein (1995a) has argued that deadweight loss of the income tax depends on the compensated tax rate on taxable income rather than simply labor supply or capital accumulation.

2 Because of confidentiality restrictions, the taxpayer's state of residence is not reported on tax returns with an adjusted gross income above $200,000. The tax returns included in this study reflect a population that is responsible for 99% of all returns filed, 85% of taxable income reported by individuals, 89% of adjusted gross income reported, 88% of itemized deductions, 90% of adjustments to income such as IRA contributions, and 76% of negative incomes or tax losses reported on personal tax returns.

3 This income measure is the sum of all positive and potentially taxable incomes reported on form 1040 before the netting out of any negative income items, such as partnership losses. Because tax return losses often reflect paper (accounting) losses rather than economic losses, total positive income has been used in many studies to proxy actual income (Long and Gwartney 1987; Greenwood 1993; Burman and Randolph 1994; Samwick 1995).

4 There were 19,354 tax returns in the low-tax group (from 12 states) and 22,829 tax returns in the high-tax group (from 17 states).

5 Even in a static framework of no behavioral responses, an increase in the combined marginal tax rate resulting from a rise in the state tax component would likely be associated with a decline in federal taxable income because of larger deductions for state income tax payments. It is not clear whether Feldstein (1995b) or Auten and Carroll (1995) recognized this potential bias or whether steps were taken to address this problem (e.g., restricting the panel sample to taxpayers residing in the same state before and after 1986).

6 Technically, both state and local income taxes paid are excluded because there is no exact way to separate state from local income tax payments. For comparative purposes, all the empirical work described below was also conducted on data adjusted in the following ad hoc manner. In states where local governments levy individual income taxes, the ratio of statewide local income tax revenues to total (state plus local) income tax revenues was used to deflate the deduction for income taxes paid on itemized returns. Local income tax revenues comprised more than 10% of total revenues in only five states (Kentucky, Maryland, New York, Ohio, and Pennsylvania). Estimates derived from this adjusted data were not materially different from the results reported in Tables 1-4.

7 The Public Use Tax File is a stratified random sample of individual tax returns; consequently, the proportions of returns within various strata, which are defined on the basis of factors such as level of income and size of business or farm receipts, do not reflect population proportions. The sampling weight provided for each tax return was used to make the mean incomes, and other variables included in Table I reflect population values.

8 When necessary, positive income is adjusted to account for the fact that some states exempt some incomes reported on the federal tax return from state taxation (e.g., social security and pension benefits and unemployment compensation).

9 In measuring the federal tax rate, preavoidance taxable income is also defined as total positive income minus the personal exemptions allowance. The details of state income taxes necessary to calculate the effective state marginal rate (i.e., income brackets and marginal rates, personal exemptions, exclusions, and deductibility provisions) are reported by the Advisory Commission on Intergovernmental Relations (1992). For taxpayers residing in New Hampshire and Tennessee, states that tax interest and dividends only, the state income tax rate is assumed to be zero. Alternatively, tax returns from residents of these two states can be excluded without altering the basic findings and conclusions of this article.

10 Burman and Randolph (1994) suggest that the top tax rate captures most of the important differences in state income tax structures. The regression results are: (with standard errors in parentheses)

[Mathematical Expression Omitted]

11 When constructing LABOR, those forms of income that are statutorily excluded from federal taxation (exempt interest, a portion of pensions and annuities, and part of social security benefits) were included in the denominator.

12 Tobit estimation (Tobin 1958) is the generally preferred technique for estimating a model in which the range of the dependent variable is constrained or truncated. Individual observations were weighted using the sampling factors provided in the data set. Hausman and Wise (1981) argue that estimation of a model in which the endogenous variable (in this case, taxable income) is a component or derivative of exogenous variables that define strata will lead to biased and inconsistent estimates of population parameters unless the observations are weighted on the basis of sampling proportions.

13 Technically, the partial derivative of taxable income with respect to the state tax rate is equal to the product of the Tobit coefficient of SMTR and the predicted probability that taxable income is positive. Because the probability always exceeds 0.99, for practical purposes the impact of each variable on taxable income can be described by the variables's estimated coefficient alone.

14 Following a reader's suggestion, the nonlinearity of the tax rate effect was also tested by including the squared value of the state tax rate rather than a tax rate-income interaction term. The coefficient of the squared tax rate was negative and highly significant, revealing that an increase in the tax rate reduces taxable income more at high tax rates than at low tax rates. Because a high state tax rate does not necessarily imply a high income level, the tax rate-income interaction term is better suited for investigating tax avoidance differences by income level.

15 The Pearson correlation coefficient (probability) between HOUSEVALUE and the maximum state marginal rate is +0.57(0.0001); that between WELFARE and the state maximum rate is -0.27(0.0001).

16 This can be seen by comparing the coefficients of the tax rate terms in Equation 3 to the following estimates when WELFARE is omitted:

[175** IVSMTR - 8.68**IVSMTR x INCOME (22.18) (0.47)

17 The tax rate coefficients changed to

190**IVSMTR - 9.16**IVSMTR x INCOME (22.83) (0.48)

Other definitions of BUSINESS were also used (e.g., receiving income from schedules C, F, or E even though no self-employed taxes were paid) without altering the estimated tax rate-taxable income relationship.

18 This fear was unfounded because the simple correlation between the maximum state marginal tax rate and the average effective property tax rate on single-family homes was virtually zero. Consequently, the IVSMTR coefficients shown in column 3 of Table 2 were literally unchanged when the taxable income equation also included the state's effective property tax rate, as shown below:

[Mathematical Expression Omitted]

An alternative approach used was to add back property tax deductions in the manner used for state income taxes paid. This resulted in the following relevant coefficients:

[Mathematical Expression Omitted]

These coefficients suggest a relatively larger (more negative) impact of a tax rate increase on taxable income than reported in Table 2.

19 The pertinent estimates from the tax losses model are

[Mathematical Expression Omitted]

The corresponding estimates in the deductions equation are

[Mathematical Expression Omitted]

20 The pertinent coefficient estimates from the taxable income model are

[Mathematical Expression Omitted]

The corresponding coefficients in the deductible expenditures equation are

[Mathematical Expression Omitted]

Complete Tobit results are available from the author.

21 Feldstein (1995b, p. 559) also makes this point. See Samwick (1995) for research on the role of passive loss rules in contributing to the decline in tax-sheltered investments after the 1986 tax reform.

22 The elasticity of taxable income (TI) with respect to the tax rate (t) and the net-of-tax rate (1 - t) equals ([Delta]TI/[Delta]t) (t/TI) and [[Delta]TI/[Delta](1 - t)] (1 - t/TI), where [[Delta]TI/[Delta](1 - t)] = (-[Delta]TI/[Delta]t). The coefficients of the IVSMTR terms in columns 3-6 of Table 2 provide estimates of [Delta]TI/[Delta]t, and the marginal tax rate (t) used in the elasticity calculation is the total (federal plus state) tax rate.

23 See Slemrod (1995a) for a brief discussion of the dimensions of responses to tax rate changes (real, avoidance, and timing).

24 Feldstein and Feenberg (1995) found that the 1993 tax rate increases raised only about one-third the static tax revenue gains.

References

Abrams, Burton, and Mark Schmitz. 1984. The crowding-out effect of government transfers on private charitable contributions: Cross section evidence. National Tax Journal 37:563-8.

Advisory Commission on Intergovernmental Relations (ACIR). 1992. Significant features of fiscal federalism. Washington, DC: Advisory Commission on Intergovernmental Relations.

Auten, Gerald, and Robert Carroll. 1995. Behavior of the affluent and the 1986 Tax Reform Act. In 1994 Proceedings of the eight-sixth annual conference, National Tax Association. Columbus, OH: National Tax Association.

Barro, Robert. 1993. Higher taxes, lower revenues. The Wall Street Journal, 9 July, A8.

Bogart, William T., and William M. Gentry. 1995. Capital gains taxes and realizations: Evidence from interstate comparisons. The Review of Economics and Statistics 77:267-82.

Burman, Leonard E., and William G. Randolph. 1994. Measuring permanent responses to capital-gains tax changes in panel data. American Economic Review 84:794-809.

Feenberg, Daniel. 1987. Are tax price models really identified: The case of charitable giving. National Tax Journal 40:629-33.

Feldstein, Martin S. 1975. The income tax and charitable contributions: Part I - Aggregate and distributional effects. National Tax Journal 28:81-100.

Feldstein, Martin S. 1993. Clinton's revenue mirage. The Wall Street Journal 6 April, A14.

Feldstein, Martin S. 1995a. Tax avoidance and the deadweight loss of the income tax. NBER Working Paper No. 5055.

Feldstein, Martin S. 1995b. The effect of marginal tax rates on taxable income: A panel study of the 1986 Tax Reform Act. Journal of Political Economy 103:551-72.

Feldstein, Martin S. 1996. A case for cutting marginal rates. The Wall Street Journal 9 August, A10.

Feldstein, Martin, and Daniel Feenberg. 1995. The effect of increased tax rates on taxable income and economic efficiency: A preliminary analysis of the 1993 tax rate increases. Paper presented at the NBER Conference on Tax Policy and the Economy, Washington, DC.

Goolsbee, Austan. 1998. It's not about money: Why natural experiments don't work on the rich. Working Paper No. 98-7. Ann Arbor, MI: Office of Tax Policy Research, University of Michigan.

Greenwood, Daphne T 1993. Price and income elasticities of charitable giving: How should income be measured? Public Finance Quarterly 21:196-209.

Hausman, Jerry, and David Wise. 1981. Stratification on endogenous variables and estimation: The Gary Income Maintenance Experiment. In Structural analysis of discrete data with econometric applications, edited by C. Manski and D. McFadden. Cambridge, MA: MIT Press, pp. 365-91.

Internal Revenue Service. 1994. 1991 Individual Public Use Tax File. Washington, DC: Internal Revenue Service, Statistics of Income Division.

Lindsey, Lawrence B. 1987. Individual taxpayer response to tax cuts, 1982-1984: With implications for the revenue maximizing tax rate. Journal of Public Economics 33:173-206.

Long, James E. 1989. The effect of the 1986 Tax Act on personal interest deductions. Public Finance Quarterly 17:243-63.

Long, James E. 1990. Marginal tax rates and IRA contributions. National Tax Journal 43:143-53.

Long, James E., and James D. Gwartney. 1987. Income tax avoidance: Evidence from individual tax returns. National Tax Journal 40:517-32.

Ribar, David C., and Mark O. Wilhelm. 1995. Charitable contributions to international relief and development. National Tax Journal 48:229-44.

Samwick, Andrew A. 1995. Tax shelters and passive losses after the Tax Reform Act of 1986. NBER Working Paper No. 5171.

Slemrod, Joel. 1995a. Income creation or income shifting? Behavioral responses to the Tax Reform Act of 1986. American Economic Review Papers and Proceedings 85:175-80.

Slemrod, Joel. 1995b. High-income families and the tax changes of the 1980s: The anatomy of behavioral response. NBER Working Paper No. 5218.

Tobin, James. 1958. Estimation of relationships for limited dependent variables. Econometrica 26:24-36.