The elasticity of taxable income over the 1980s and 1990s.

INTRODUCTION

The degree to which taxes alter U.S. economic activity and tax-reporting behavior is a subject of debate. Estimates of the effect range from extremely large to almost none. For even modest changes to tax rates, the range of estimates implies differences in deadweight loss and income-tax revenue of many tens of billions of dollars. A key variable at the center of recent research is the elasticity of taxable income (ETI), which measures the responsiveness of reported taxable income to changes in marginal tax rates. (1) The ETI, if accurately estimated, can be used to calculate both the change in deadweight loss (2) and the change in income--tax revenue resulting from a change in tax rates. (3) However, in practice, assessing both the efficiency and revenue implications of tax-rate changes is more complex than the formulas suggest. For example, if the ETI differs by income, an accurate assessment of either efficiency or revenue implications requires a breakdown of the responses by income group. (4)

Despite a great deal of variation in ETI estimates, both across studies and within studies that explore different specifications, several recent papers have reported an overall ETI of about 0.40. An often-cited study by Gruber and Saez (2002) examines responses to the tax cuts of 1981 and 1986, and finds an overall estimated ETI of 0.40. However, Kopczuk (2005) finds similarly estimated results to be quite sensitive to sample selection and model specification. Both Giertz (2006) and Helm (2007) also report estimates for the 1990s that are very sensitive to an array of factors. Others (e.g., Saez (2004) and Goolsbee (1999)) report great heterogeneity in estimated responses across time periods.

The estimation portion of this paper first replicates Gruber and Saez's core results by applying their techniques to a dataset that is similar to the one that they used. My results for the 1980s closely parallel Gruber and Saez's results. Applying the same methodology to 1990s data and to data from both the 1980s and 1990s combined, however, yields estimated ETIs that are much smaller than corresponding estimates for the 1980s. In fact, using Gruber and Saez's preferred specification yields an estimated ETI for the 1990s that is a little more than half my corresponding estimate for the 1980s. Weighting regression results by income not only has an enormous impact on the estimates, but yields overall estimates that are driven by a tiny fraction of high-income filers. For example, excluding the 100 most influential observations (just 0.2 percent of the sample), as measured by a dfbeta test, lowers the estimated ETI for the 1980s from 0.37 to 0.08. (5)

While the dfbeta tests suggest that the Continuous Work History Survey (CWHS) estimates may be imprecise because of the small number of very-high-income observations, estimates are generally very similar after adding data from primarily high-income tax fliers to the sample. While the standard errors are much smaller and the estimates more robust with the larger dataset, the fact that estimates are often similar suggests that, despite the small number of very-high-income filers, the CWHS may be a viable dataset for examining behavioral responses to taxation.

The larger dataset also includes additional demographic information. This information (age, gender, and itemization status), using Gruber and Saez's preferred specification, has a positive, albeit modest, affect on the estimated ETI for the 1980s and a negligible affect on the 1990s estimate. When including this information, the larger dataset yields an ETI for the 1980s and 1990s combined of 0.34 with a t-value of over 7.5.

Even with the larger dataset, estimates for the 1980s and 1990s differ greatly. The model with demographics and Gruber and Saez's richest set of controls yields an estimated ETI for the 1980s of 0.43 and for the 1990s, 0.20. (6) Including separate and nonlinear controls for mean reversion and divergence within the income distribution narrows this difference, lowering the 1980s estimate to 0.40 and raising the 1990s estimate to 0.26. Additionally, work by Kopczuk (2005) implies that changes to the tax base since 1986 (IRS, 1979-1998) could account for as much as 14 to 29 percent of this difference. However, this still leaves the vast majority of the difference in estimates between the two time periods unexplained.

When turning to a more encompassing income measure, broad income, substantial variation in estimated elasticities for the 1980s and 1990s is also observed. For the 1980s, the estimated broad income elasticity is 0.21. For the 1980s, the corresponding estimate is 0.13. While the estimated broad income elasticity is much lower for the 1980s than the 1990s, the 1990s estimate represents a larger share of the corresponding taxable income elasticity estimate than does the 1980s estimate. Heterogeneous income elasticity estimates across tax changes is not a new finding. Saez (2004), using aggregated time-series data, finds great variation in income responses to tax changes over years 1960 to 2000. And Goolsbee (1999), using repeated cross-sections of data for selected years between 1920 and 1966, also finds substantial variation in estimated responses across tax changes.

ISSUES IN THE ANALYSIS

While taxes affect income growth, so do many other economic factors. Controlling for non-tax-induced trends in taxable income is a major obstacle to accurately estimating elasticities. The issue of non-tax-related trends in income is given the most attention in this section because the approach used to control for those trends represents the most novel aspect of the model employed in this study--a model developed by Gruber and Saez (2002). The approach also takes into account other factors, such as mean reversion, tax-rate endogeneity, institutional changes (which often coincide with changes in the rate structure), and differences between transitory (or temporary) fluctuations and permanent (or longer-term) responses. (For a discussion of these issues and the related literature, see Giertz (2004) and Slemrod (1998).)

Controlling for Exogenous Trends in Income

The centerpiece of Gruber and Saez's approach is its controls for non-tax-related heterogeneous shifts in income distribution and mean reversion. Over the past 30 years, the distribution of reported income has widened. In fact, that trend accelerated in the 1980s, especially at the top of the distribution. (7) Because people with the highest income pay a disproportionate share of taxes--the top one percent pay approximately one-third of all federal income taxes--their behavior is especially important (see Internal Revenue Service (2004)). Not fully accounting for the portion of that income growth that is unrelated to tax policy can result in large biases. For example, the 1980s cuts in marginal tax rates were greatest at the top of the income distribution and, thus, inversely correlated with the great income growth at the top of the distribution. If the exogenous (non-tax-related) portion of that income growth is not fully accounted for, that trend will bias ETI estimates upward. Because this income trend has been irregular, distributional changes in years without tax changes may not provide useful measures of exogenous shifts that occur during periods with tax changes.

Although changes to the income distribution are widely documented and theories such as heterogeneous (and diverging) returns to education and experience help explain the phenomenon, the underlying driving factors are not well understood, nor are the year-to-year deviations from that trend. (8) The fact that the exogenous-income trend has persisted through periods of both increases and decreases in the level and progressivity of income tax rates suggests that it is, in large part, not a direct response to tax changes. Furthermore, Saez and Veall (2005) find an income trend at the top of the Canadian income distribution that closely parallels that in the U.S.--despite a different pattern of tax changes in Canada with much more modest reductions in marginal tax rates. That notwithstanding, the possibility that the phenomenon results from a longer-run and more-nuanced response to tax changes cannot entirely be ruled out. Note that the progressivity (especially within the top one percent of the income distribution) (9) of both the U.S. and Canadian tax systems has declined substantially since 1970, as the concentration of income held by this group has risen substantially in both countries. By contrast, tax progressivity at the top of the French income distribution has remained stable (or possibly increased), while the income concentration at the top of the French income distribution too has remained relatively stable (Piketty and Saez, 2007).

Controlling for Mean Reversion