Introduction and Summary

The distribution of the economic burden of taxes is central to tax policy. In constructing measures of effective tax rates--the share of their income that people pay in taxes--analysts must confront a wide range of issues: what to include in measuring income, the range of taxes to be studied, the shifting and ultimate incidence of taxes (that is, on whom the ultimate burden of a tax falls), whether to study individuals or households, and the period that the analysis is to cover (see Box 1). This Congressional Budget Office (CBO) paper focuses on how the choice of a particular time frame affects distributional assessments of the burden of federal taxes.
 

Box 1. Distributional Analyses of Tax Systems Distributional analyses of tax systems compare the different tax burdens that people bear, typically focusing on how those burdens vary among income groups. For example, in several reports, the Congressional Budget Office (CBO) has estimated how federal taxes are distributed annually among groups of households, which have been defined on the basis of their income. CBO's assessments rely on cross-sectional data covering single years; they do not consider how tax burdens may change over time with taxpayers' circumstances. Such longer-term measures of tax burdens require panel data--information collected annually, for many years, from the same group of people. CBO's previous reports have presented summary statistics that include each group's average effective tax rate--tax liability as a percentage of total income--and the share of federal taxes that it paid. In 2001, for example, households in the lowest fifth (quintile) of the population paid 5.4 percent of their income in federal taxes and about 1 percent of the total federal tax bill. In contrast, households in the top quintile faced an average effective federal tax rate of 26.8 percent and paid about 65 percent of all federal taxes. Several factors strongly influence any estimate of the distribution of taxes. They include the unit of analysis, the components of income measured, the period covered, and assumptions about the incidence of each tax--who actually bears the burden of it. The Unit of Analysis Analysts measure tax burdens over one of four kinds of units: individuals, taxpaying units (people included on a single tax return), families, or households. The key issue from an economic perspective is that taxes reduce consumption, so the appropriate unit of analysis should be some form of consumer unit--a group of people who combine their income to purchase goods and services. Yet even that simplification is problematic because the makeup of the groups often shifts for different purchasing choices. For example, students who share an apartment compose a single group to purchase housing but not necessarily a single group to buy food. Furthermore, consumer units are hard to identify. Available data typically report household and family relationships; to measure tax burdens on the basis of consumer units, analysts must infer the groups that actually make decisions about consumption. The Income Measure People typically think of their income as comprising their pay for working, returns on their investments (interest, dividends, rents, and capital gains), pension payments, and government transfer payments (such as Social Security benefits). However, researchers often use a broader concept, one that also includes the value of in-kind benefits, such as health insurance, food stamps, and housing assistance; income and payroll tax payments that are made by others (such as an employer) but that actually fall on individuals in the form of lower wages and salaries or smaller returns on investments; and the value of "services" provided by durable goods (for example, owner-occupied housing). In its most broadly construed but most difficult-to-measure form, income over a particular period equals the value of consumption of all kinds, including leisure, plus any change in wealth. Some analysts argue that consumption rather than income more accurately represents a person's well-being, even though that measure ignores savings. Larger households generally consume more goods and services than smaller households do, and in distributional studies, many analysts adjust households' income to account for such variation.(1) For example, a couple needs more income than a single person does to consume at the same level. Measuring income per capita (total income divided by number of people) considers increased needs but ignores economies of scale: two people may not be able to live as cheaply as one, but they can live more cheaply together than they can live apart. In general, analysts use various adjusted measures of income that fall somewhere between unadjusted and per capita income. The measure of income that analysts use for a distributional analysis and the adjustments they make for different-sized households can influence their results. Factors affecting those decisions include the purposes of the analysis, availability of data, and methods at hand for valuing noncash income or consumption. The Measurement Period Distributional analyses typically focus on annual measures of income and taxes--because the government generally assesses taxes annually, tax data usually cover a single year, and many taxpayers think about their income and taxes on an annual basis. A person's income may vary, however, and a single year may not accurately indicate a taxpayer's well-being. A spell of unemployment may temporarily lower income, for example, whereas the sale of a business may briefly raise it. People's income also changes relative to that of others. Thus, a single year's assessment of income and taxes may be uncharacteristic of experience over longer periods. Income mobility stems in part from the fact that a person's income over his or her lifetime tends to follow a common pattern: it rises from low levels early in life, peaks during the most productive years, and then falls off during retirement. By measuring income over longer periods, perhaps over an entire lifetime, analysts take income mobility into account, minimize the effects of year-to-year fluctuations, and are better able to assess people's economic well-being. Other factors that change over time affect the taxes that people pay. For example, families grow as they have children, and they shrink as children grow up and leave home. Those fluctuations produce parallel changes in both the number of dependent exemptions taxpayers can claim and their income tax liability. Therefore, even if a person's income is stable over time, other factors may lead to differences between taxes paid in a given year and average taxes paid over a longer period. The Incidence of Taxes Distributional assessments of tax systems require analysts to assign the burden of each tax to individuals. (The reason is that the actual burden of taxes may not fall on the people or businesses that pay them to the government. The burden may instead be shifted to others through lower wages, smaller returns on investments, or higher prices.) The assumptions used in assigning tax burdens can determine the results of an analysis. But researchers disagree about incidence--particularly for the corporate income tax and estate taxes. Furthermore, measuring incidence accurately may mean assessing not just taxes but rather taxes net of the benefits that the collected revenues pay for. For example, considering the distribution of Social Security taxes without accounting for the benefits that the taxpayer will receive in retirement may misrepresent the actual burden of the tax when the program is considered as a whole. Most analysts agree that workers bear both the employee's and employer's share of payroll taxes--employees shoulder the latter tax burden through wages that are lower than they would be in the taxes' absence. There is less agreement about who bears other federal taxes. Some researchers maintain that the personal income tax falls entirely on the people whose income is taxed. Others argue that the portion of the tax that falls on income from capital ultimately reduces workers' income and is thus borne by workers rather than by the owners of capital. Experts similarly disagree about the corporate tax, with some asserting that it falls on income from capital, income from labor, or foreigners. Analysts' views also differ about who actually pays excise taxes and whether the burden of estate taxes falls on the person who dies or on his or her heirs. 1.  Ways to adjust income are discussed in Roberton Williams, David Weiner, and Frank Sammartino, Equivalence Scales, the Income Distribution, and Federal Taxes, Technical Paper 1999-2 (October 1998).

People pay most federal taxes on the basis of annual accounting; the individual income tax, for example, is generally assessed on taxable income earned during a calendar year.⁽¹⁾ Because of that annual accounting period and because tax laws may change from one year to the next, most analyses of effective tax rates cover individual years. (CBO regularly publishes reports showing the annual distribution of income and taxes for specific years.)⁽²⁾ How much of their income people pay in taxes fluctuates from year to year, in part because a rise or fall in income moves them into higher or lower tax brackets; in part because tax laws change; and in part because their filing status, amount of itemized deductions, or other circumstances change. At the same time, variability in people's income alters their capacity to pay taxes over time. In the face of such changes, focusing exclusively on the impact of taxes on an annual basis may provide a limited view of the relative tax burdens that people bear over time. As a broader view, for example, an analysis could compare the present value of taxes paid over a person's lifetime with the present value of his or her lifetime income.⁽³⁾

But current data do not allow actual lifetime measures of effective tax rates. The longest period for which information is available is that covered by the University of Michigan's Panel Study of Income Dynamics (PSID). The PSID has followed approximately 5,000 families since 1968, interviewing them almost annually, and that 5,000-family sample appears to represent roughly the entire population. But the PSID provides only limited information on taxes and covers relatively few individuals at the upper end of the income distribution--a segment of the population that is central to any analysis of federal tax revenues.

In contrast, the Internal Revenue Service's Statistics of Income (SOI) sample of tax returns provides rich detail on taxes and an extensive random sample of tax filers that includes those at the top of the income distribution. However, because people, particularly those at the bottom of the distribution, move onto and off the tax rolls from year to year, SOI data poorly represent the entire population over time and cannot consistently provide information on low-income taxpayers.

Nevertheless, even with such limitations, the two sources of data, especially when examined in concert, offer the best information with which to examine tax burdens longitudinally--that is, over periods longer than a year.

This analysis compared the distribution of income and taxes measured on an annual basis with the distribution measured longitudinally. Its key findings are as follows:

• Incomes and taxes when measured longitudinally are more evenly distributed than they are when measured annually. In addition, the distribution becomes more equal as the time horizon lengthens.
  
  
• When a one-year measure of income is used, incomes and tax rates for lower quintiles, or fifths, of the income distribution on average are lower than they are when a longer period of measurement is used. Thus, income in the second quintile when measured over 14 years is 26 percent higher than annual income over the same period, and the effective income tax rate is 3.5 percentage points higher.
  
  
• Incomes in the highest quintile are smaller when measured over a longer period than when measured annually. In the top quintile, income measured over 14 years is 4 percent below average annual income. However, for that quintile, the effective tax rate under either measurement period differs by only one-tenth of a percentage point.
  
  
• The effects of illustrative tax policy changes generally appear more evenly distributed when they are measured over multiple years than when they are assessed on an annual basis. For example, as measured on an annual basis, a doubling of the child credit would affect only 3 percent of taxpayers in the second quintile and lower the average tax rate in that quintile by 0.1 percentage points. Under longitudinal measures, almost 9 percent of taxpayers in that second quintile would see an effect, and their effective tax rate would be reduced by 0.2 percentage points. In the average year, 32 percent of taxpayers would be affected by that change, but 52 percent would be affected in at least one of 10 years. A policy that reduced the top two income tax rates would lower the effective rate only in the top quintile--by 1.0 percentage point when measured annually and by 1.1 percentage points when measured longitudinally. In the average year, only 2 percent of taxpayers would be affected by that change, but 6 percent would be affected in at least one of 10 years.

What Is Distributional Analysis?

Equity, or fairness, is a key consideration in evaluating tax policy, although there are many views on what "fairness" means. Distributional analysis, which shows how income and taxes are distributed among individuals, families, or households, is one widely used way to inform judgments about the fairness of changes in the tax system. Policymakers use such analyses to evaluate the existing distribution of tax burdens, study how the distribution has changed over time, and assess proposed changes to tax law.

Several governmental organizations, including CBO, periodically estimate the distribution of income earned and taxes paid in a given year.⁽⁴⁾ The fundamental structure of those analyses is similar, although they differ in their details. In general, the approach is to combine information from a sample of tax returns for a year with survey data for the same year covering people who did not file tax returns. The resulting cross section of data represents the entire population for that year. Analysts assign tax payments to each taxpayer under specific assumptions about who bears the ultimate burden of each tax. (CBO, for instance, assumes that employees bear the burden of the employer's share of the payroll tax and that all owners of capital effectively pay the corporate income tax.) People are arrayed in income groups, usually by using a broad measure of income from both taxable sources (such as wages) and nontaxable sources (such as the health insurance premiums employers pay on behalf of employees). The analyses examine a number of indicators of tax policy, including the average income of each income group and the taxes each group pays--often expressed as a group's effective tax rate and share of total tax payments.

Those analyses of annual data show that the federal tax system is progressive--average effective tax rates increase as income rises. The individual income tax is especially progressive because its graduated rate structure taxes larger incomes at higher rates. Corporate income taxes also appear progressive under CBO's assumptions about incidence because income from capital is concentrated among people in higher-income groups.⁽⁵⁾ Payroll taxes claim a roughly constant share of income for most groups in the income distribution, but that share declines for families with the highest income because their earnings exceed the maximum amount that is subject to tax. Excise taxes--in fact, consumption taxes generally--are regressive because lower-income groups spend a greater proportion of their income than higher-income groups spend on goods that are subject to tax.
 

Longitudinal Versus Annual Measures of Effective Tax Rates

Most distributional analyses consider annual measures of income and taxes. Federal income and payroll taxes are currently levied on income yearly, and annual collection periods for those revenues lead to yearly reporting of tax data. Many major surveys, such as the Bureau of the Census's Current Population Survey, also ask respondents about their income over the past year.

A one-year time horizon, however, provides a limited perspective on the distributional impact of federal taxes. Measuring income and taxes over a longer period, even over a lifetime, may assess overall economic well-being more accurately than an annual measure of income does. Annual income differs from lifetime income for many reasons, some relatively predictable and some transitory. Events in a person's life, such as marriage or divorce, may produce some of the differences; other factors relate to changes in earnings and savings over people's lifetimes. When individuals enter the workforce, they generally earn relatively low wages, but their earnings and other income rise as they gain experience and accumulate assets. Once they retire, their wages fall, and they begin to draw on their savings. Still other differences are transitory: someone may become temporarily unemployed or realize a large capital gain. As income fluctuates, tax payments change correspondingly. Thus, a person's income or taxes in any given year may not accurately represent that person's economic well-being over a longer period.

Because of such fluctuations, annual measures of income group together people who have different lifetime economic circumstances. For example, the bottom income quintile contains many people with a history of low earnings and with limited prospects for earnings in the future. But it may also include a medical student who is working part-time and borrowing against expected long-term income to defray living expenses. And it may contain retirees who have accumulated a substantial amount of wealth but have only limited income. All of those people have low annual income for the specific year being measured. Over their lifetime, however, some of them would not be found in the bottom quintile.

Analyses that use a one-year measure may cause the distribution of income and taxes to appear more unequal than it would seem in a comparable assessment that covered a longer period. In any one year, some people have a temporarily high level of income, and others have a temporarily low level. Measuring income over a longer time frame moderates those temporary effects, and the distribution of income appears more even. Also a factor with an annual measure is that if the size of the share of income that is transitory changes over time, the distribution of annual income might shift, even with no alteration in the distribution of longer-term income.

The extent to which the potential drawbacks of using cross-sectional (annual) analysis will matter depends on the actual pattern of changes in income. If taxpayers' incomes fluctuate greatly from year to year and people's lifetime incomes differ substantially from their income in any one year, then using distributional results based on cross-sectional analysis could lead to very different conclusions than those reached by using a longer-term income measure. However, if effects arising from transitory ups and downs in income and a taxpayer's stage in the life cycle are small relative to other differences among taxpayers, the annual distribution of income and taxes may be similar to the longer-term distribution.

Previous studies have examined the effects of measuring income and taxes over different periods. The Department of the Treasury recently developed a data set for a 10-year panel of families to use in analyzing tax proposals.⁽⁶⁾ Its analysis found that income and taxes were somewhat more evenly distributed when a 10-year measure of income was used in place of an annual measure, but overall, income and taxes were distributed similarly under both measures. Joel Slemrod, a researcher at the University of Michigan, also found that income distributions were similar under the two types of measures. (Slemrod compared "time-exposure" income--income measured over the multiyear 1979-1985 period--with annual income.)⁽⁷⁾ He determined, however, that the share of income going to the bottom group was substantially greater under his longitudinal measure than it was under his annual measure.
 

The Framework of CBO's Analysis

This study followed a group of people over multiple years and compared the distribution of their income and taxes measured on an annual basis and measured longitudinally. Other facets of the analysis--the people considered, the definition of income, and the tax law that was applied to that income--were held constant to isolate the effects of measuring income and taxes over different periods (see Table 1).

Sources of Data

The data for CBO's analysis came from two sources: the Internal Revenue Service's (IRS's) Continuous Work History Sample (CWHS), which is a random sample of individual income tax returns for the same taxpayers each year, and the University of Michigan's Panel Study of Income Dynamics (PSID).⁽⁸⁾ The two data sets have different strengths and weaknesses. The CWHS contains detailed information about income and other tax-related characteristics (such as the amount of itemized deductions), which permits analysts to estimate income taxes accurately. However, the CWHS includes only people who file tax returns, and many lower-income families are thus excluded from the CWHS because they are not required to file returns. The CWHS has another drawback: it provides no information about untaxed income, such as means-tested transfer payments like food stamps or housing assistance. Because such income goes disproportionately to low-income households, its omission may significantly understate those households' total income and economic well-being.

In contrast, the PSID contains less-detailed information about income and no information on other tax-related characteristics, which reduces the accuracy of tax calculations from it. Furthermore, because the PSID is a survey, it is probably less reliable than tax-return data because of errors in reporting by the survey's respondents. In addition, incomes in the PSID are top-coded, or capped at certain levels, to preserve confidentiality, a feature that limits the survey's usefulness in examining upper-income families. However, the advantages of the PSID are that it covers the entire population and captures information about income from major cash transfer programs (such as the Temporary Assistance for Needy Families, or TANF, program).

Because of the relative strengths of the two data sets, CBO relied on the PSID in examining lower-income families and on the CWHS in considering middle- and upper-income taxpayers.

Unit of Analysis

CBO's study considered the income and taxes of taxpaying units, which are generally either individuals or married couples filing joint tax returns.⁽⁹⁾ To make results from the PSID comparable with those from the CWHS, CBO created taxpaying units out of the PSID's families, essentially by including only the income of the head of the family and spouse. For families that had several taxpaying units, such as a multigenerational family, CBO considered only the primary unit (because the PSID does not provide the information necessary to track the income of secondary units). For a substantial majority of families, the primary taxpaying unit, the family, and the household were identical, so the choice of the unit of analysis in this study is unlikely to have a significant impact on its results.

Restrictions on the Sample

In this analysis, multiyear measures of income and taxes were compared with annual measures of those variables. Multiyear measures can be calculated only for taxpayers who appear in every relevant year of data; consequently, CBO restricted its analysis to those taxpayers. Yet that approach does not guarantee a representative sample. If taxpayers for whom data are available every year differ systematically from taxpayers for whom data are available only intermittently, that method might yield a misleading estimate of the experience of taxpayers who dropped out of the study (an issue that is discussed in more detail later).

CBO placed no other restrictions on the sample of taxpaying units that it considered. Taxpayers whose family structure changed, whether through marriage, divorce, or having children, were kept in the sample. Similarly, CBO retained taxpayers who reached retirement age and left the workforce as well as taxpayers who entered the workforce for the first time. Effectively, life-cycle transitions were considered part of a set of events that caused income to change over time. As a result, taxpayers who experienced such events were treated no differently than taxpayers who received a pay increase or realized a capital gain.⁽¹⁰⁾

Weighting

Weighting is a statistical tool that helps ensure that estimates derived from a sample can be "generalized"--deemed to apply--to an entire population. In the case of the CWHS, which is a simple random sample, CBO gave each return an equal weight because every tax return has an equal chance of being included in the sample. Weighting was more complex for the PSID data because of that sample's design. Families and individuals have different probabilities of being included in the PSID sample, so the survey includes weights that reflect those probabilities. The weights used in the PSID also adjust for differences in rates of response among the groups; thus, the weight for a family is different in every year of the sample. For its analysis, CBO applied to every year of a period the PSID's longitudinal weight for a family in the last year of that period.⁽¹¹⁾

Definition of Income

The same definitions of income were used throughout the period that CBO examined. The income measure was as broad as possible, given the available data, but falls short of being comprehensive. For the tax-return data, income consists of wages and salaries, business income, taxable and tax-exempt interest, dividends, realized capital gains, unemployment compensation, rent and royalty income, taxable pensions, and other income reported on tax returns. The measure excludes Social Security benefits (because not all recipients are required to report that income on their tax returns); other untaxed income, such as TANF payments (because that information is not captured on a tax return); and in-kind transfers, such as food stamps.

CBO applied two definitions of income to the PSID data: first, taxable income, which includes wages, business income, interest, dividends, pension income, unemployment compensation, and rental income; and second, total income, which equals taxable income plus Social Security benefits and income from several cash transfer programs. Taxable income is the measure most analogous to that used for the CWHS sample, so CBO used it for the most part to facilitate comparisons between the two data sets. Total income is a broader measure that more closely approximates overall economic resources.

Measure of Tax Liability

For this analysis, CBO used a narrow measure that focused on federal individual income taxes. (The measure was intended to simplify the analysis, but in future studies, CBO may broaden the scope of its analysis to include other taxes.) Individual income taxes are the best choice for examining multiyear effective tax rates because they are the largest source of federal revenues, vary the most directly with income, and are most readily attributed to individual taxpayers. Some other taxes, such as payroll taxes, would also be relatively straightforward to study. In contrast, excise taxes would be difficult to include in CBO's analytical framework because few data exist on individuals' consumption of specific goods over long periods.

To better assess changes in income separately from changes in tax law, CBO used a constant-law measure of tax liability: the liability that a taxpaying unit would have incurred if 2000 tax law had applied in the years being examined. CBO thus inflated incomes to 2000 levels and incorporated in its simulations the assumption that the 2000 tax code was in place.⁽¹²⁾ Calculations of taxes using data from the PSID are necessarily more approximate because the survey collects no information about itemized deductions and capital gains.⁽¹³⁾

Construction of Multiyear Measures

For each taxpayer in the data, CBO constructed annual and multiyear measures of income and taxes. Annual measures are simply the values for each year that the data covered. Multiyear, or longitudinal, measures apply to longer periods of various lengths and are the average of annual incomes or taxes over the relevant period. Thus, for example, three-year income in CBO's analysis is one-third of total real (inflation-adjusted) income over a three-year period. A taxpayer who had income of $30,000 in year one and $60,000 in years two and three would have a three-year income of $50,000.

To compute an average effective tax rate, CBO took multiyear measures of taxes, averaged people's tax payments over multiple years, and divided by their average income. Suppose that the taxpayer who had a three-year income of $50,000 had tax liability of $3,000 in year one and $12,000 in years two and three. He or she would have a three-year tax liability of $9,000. Dividing that three-year average liability ($9,000) by three-year income ($50,000) yields a three-year effective tax rate of 18 percent. That multiyear rate represents the share of the taxpayer's total income that he or she would have paid in taxes, given the pattern of income, had 2000 law been in effect over the entire three-year period.

Other methods would generate different longitudinal effective tax rates. One alternative would compute the longitudinal rate by averaging the rates from each year. That method would weight the tax rate in each year equally--even though the tax rate in years when the person's income was higher would contribute more to the multiyear tax burden. Because of the progressivity of the tax system, that alternative method would generally make longitudinal tax rates appear lower than the rates generated by the measure used in this analysis.

Another approach would compute the longitudinal rate by estimating what taxes would have been had the person received his or her average income in every year of the period. Clearly, that method would not capture the interactions between the level of annual income and taxes. Because income taxes are assessed annually and income is subject to a schedule of increasing rates, taxpayers whose income is volatile generally face higher effective rates than those whose income is the same on average but more stable.

Definition of Income Categories

CBO's analysis assigned taxpayers to quintiles on the basis of their income, measured over single years or multiyear periods. The process for the PSID was straightforward because the data represent the entire population. Thus, for each year, analysts ranked all taxpaying units in the PSID cross section by income and divided the units into quintiles of equal size.

Calculating quintiles for the CWHS was more complicated.⁽¹⁴⁾ Because the data exclude families who do not file tax returns and the sample has experienced disproportionate "attrition" (departure from the sample as years passed) among lower-income taxpayers, the CWHS does not represent the full population (see Box 2). As a result, if CBO had constructed quintiles by dividing the CWHS sample into fifths of equal size, the income ranges of those quintiles would have differed from the ranges of quintiles created by dividing a representative sample into fifths. (For example, the bottom fifth of people who filed tax returns in consecutive years have substantially higher incomes than the bottom fifth of the entire population.) The misalignment of income ranges between the two sets of quintiles would have made it difficult to generalize results from a given CWHS quintile to the same quintile of the full population.
 

Box 2. Sampling Issues Associated with Longitudinal Data Cross-sectional surveys draw a sample of people that represents the entire population, and they collect information about those people at a single point in time. By comparison, longitudinal, or panel, studies begin with a cross-sectional sample and then follow those people over an extended period, collecting information about them at regular intervals. On the one hand, the multiyear aspect of such research makes longitudinal data valuable for examining dynamic processes or discerning cause-and-effect relationships. On the other hand, the extended nature of panel studies makes it more difficult to maintain a sample that represents the entire population. One way in which a panel may become unrepresentative is through attrition--the disappearance of some people from the panel during some years. If participants who drop out differ fundamentally from those who do not, their omission may make the panel unrepresentative of the full population and bias results based on the panel. Statistical procedures can lessen the effects of attrition, but the problem always afflicts analyses that use panel data. The University of Michigan's Panel Study of Income Dynamics (PSID), on which the Congressional Budget Office (CBO) based a substantial portion of this analysis of effective tax rates, reflects substantial attrition, even though study staff try to maintain contact with all of the families in the sample. About half of the original families in the PSID left the survey during its first 20 years--some moved and could not be located, others died, and still others declined to participate any longer. Several studies have evaluated attrition in the PSID and found that it is concentrated among lower-income respondents.(1) Yet despite the loss of people from the panel, the distribution of income among the PSID's respondents is similar to the income distribution in the Bureau of the Census's March Current Population Survey for various years--which suggests that the PSID has remained roughly representative. Two factors account for that finding. First, the PSID contains attrition-adjusted weights that are intended to partially correct for different rates of attrition among the various demographic groups. Those adjustments are small but generally improve the survey's representativeness. Second, differences in the initial characteristics of the families who leave the study and the families who stay have faded over time, allowing the survivors on some dimensions to adequately represent those who have left. Yet even if cross-sectional data gathered through the PSID remain roughly representative of the larger population, the pattern of changes in income--the focus of CBO's analysis--may differ systematically between those who remained in the sample and those who left it. No external benchmark exists against which that pattern might be compared. One way to measure the possible bias arising from attrition is to compare longitudinal and annual income of PSID respondents who remain in the sample for different lengths of time. Such a comparison shows that patterns are similar whether a respondent is required to be in the sample for nine years or only for three and suggests that attrition does not unduly influence results from the PSID (see Tables B-5 and B-6). Attrition is a bigger problem in the Internal Revenue Service's (IRS's) tax data than in the PSID. Although most Americans file tax returns every year, many people move into and out of the tax system as their income rises and falls. For example, the proportion of taxpayers who appear in data for 1987 from the IRS's Continuous Work History Sample (CWHS) and who also appear in the file for each of the subsequent 13 years is just 56 percent. People who are missing for one or more years may, among other reasons, have died, married and filed joint returns under their spouse's Social Security number, or had income below the level requiring a tax return. (Appendix A discusses attrition in the CWHS in more detail.) Regardless of the reason for their absence, those taxpayers are likely to differ in important ways from people who file a return every year, and their omission from a longitudinal analysis may bias its findings. People who have low income are underrepresented in the CWHS. Their small numbers imply that conclusions about them that are based on the CWHS data may be unreliable--not only because of the small samples but also because low-income people who appear in multiple years may show up only because they are unique in some ways. (Specifically, their low-income status is more likely to be temporary than that of people who file intermittently.) The absence of nonfilers in the CWHS further means that some people who have the most volatile income may not be part of the sample. For example, in a study covering a multiyear period, someone who entered the workforce during that time would not appear in the study's initial years, when they were not employed. And someone who retired might not appear in later years when relying primarily on nontaxable benefits from Social Security. Attrition does appear to affect comparisons of longitudinal and cross-sectional measures based on the CWHS, especially comparisons for the second quintile, or fifth, of the income distribution (see Tables B-1 and B-3). CBO's analysis covered 1987 to 2000. Restricting the study to returns that appear in each year's data over the period results in three-year longitudinal income, measured using the CWHS data, that is approximately 7 percent to 9 percent above average annual income; restricting the analysis to returns that appear in any three-year period leads to three-year longitudinal income that is 3 percent to 4 percent higher. Effective tax rates follow a similar pattern: the difference between longitudinal and cross-sectional rates appears greater when the CWHS is restricted to returns that appear in each of the 14 years. Thus, if an analysis requires that sample members' returns appear in every year of the data, the study's results for the second quintile may overstate the differences between longitudinal and annual income and tax rates. However, comparisons of annual and longitudinal income that use the tax data are qualitatively similar to comparisons that use the PSID. That similarity suggests that conclusions drawn from the tax data may be generally valid, despite the attrition it sustains, if the PSID has in fact remained representative. To further test how attrition might be biasing results, CBO simulated in the PSID the attrition that occurs in the CWHS (see Appendix A for details). For likely filers, differences between cross-sectional and longitudinal income and taxes were comparable for the full population, which suggests that results outside of the lowest quintile might not be driven by attrition and that conclusions drawn from the tax data may be generally valid. Another reason that a sample panel may become unrepresentative is that the underlying population changes over time while the panel remains constant. The sampling design of the CWHS--a simple random sample based on Social Security numbers--ensures that as new people enter the filing population, they will automatically be included in the panel so that the CWHS will reflect that population. The PSID's design requires that it follow all descendants of its initial sample, thus automatically capturing many aspects of shifts in population. The PSID's original design, however, did not include a method to account for new immigrants, and people who came to the United States after the initial sample was drawn in 1968 are not included. As a result, the PSID over time has become less and less representative of the entire population, although changes in the late 1990s added a sample of recent immigrants. 1.  John Fitzgerald, Peter Gottschalk, and Robert Moffitt, "An Analysis of Sample Attrition in Panel Data: The Michigan Panel Study of Income Dynamics," Journal of Human Resources, vol. 33, no. 2 (Spring 1998); Lee A. Lillard and Constantijn W. A. Panis, "Panel Attrition from the Panel Study of Income Dynamics: Household Income, Marital Status, and Mortality," Journal of Human Resources, vol. 33, no. 2 (Spring 1998); and Elena Gouskova and Robert F. Schoeni, "Comparing Estimates of Family Income in the Panel Study of Income Dynamics and March Current Population Survey, 1968-1999" (2002), available at http://psidonline.isr.umich.edu/Guide/Quality/psid_vs_cps.pdf.

To address the CWHS's exclusion of families who do not file returns, CBO used data from a representative cross-section of the entire population to divide the tax-return sample into quintiles. Information from the entire tax-filing population in a given year was first combined with data for the same year from the Current Population Survey, forming a cross section that represented the entire population. The taxpaying units in the cross section were then ranked by income and divided into five equal-sized groups. (The ranking incorporated the assumption that all nonfilers fell into the lowest income quintile.) CBO then used the income range associated with each population quintile to assign members of the CWHS sample to income groups (see Appendix C). For example, if the lowest population-based quintile had income ranging from zero to $15,000, any CWHS taxpaying unit with income in that range would be assigned to the lowest quintile. The "quintiles" that resulted from that procedure did not each contain one-fifth of the CWHS sample; rather, they contained the members of the sample whose level of income put them in that particular population-based fifth. At the same time, the procedure resulted in too few taxpaying units to allow meaningful estimates for the lowest quintile. Consequently, CBO reports no findings for that income category based on the CWHS data.

Defining income quintiles based on longitudinal income is difficult because multiyear income can only be determined for a subset of the population--taxpayers who appear in every year of the data. That limitation makes it impossible to calculate the longitudinal income ranges for each fifth of the entire population. Instead, CBO created longitudinal quintiles that had the same average number of taxpayers as the average of the cross-sectional quintiles with which they were being compared (the average annual measure). That procedure ensured that observed differences between annual and longitudinal income and taxes for a particular quintile derived solely from changing the measurement period and not from changes in the number of taxpayers in each quintile.

Comparison of Annual and Longitudinal Distributions

CBO averaged values from all the cross sections in a period into a single measure, which it then compared with a longitudinal measure of income and taxes over the same span. That average cross-sectional measure provides a suitable basis for comparison because it covers the same period that the longitudinal measure covers, dividing the same total income and total taxes among the same group of taxpayers. An alternative to the average of annual measures would be to compare the longitudinal results with a single cross section. How to choose that comparison year, however, is not obvious, and whatever choice was made could influence the study's results. (Differences between the longitudinal measure and any one cross section would reflect the shifting of the cross-sectional distribution over time in addition to the length of time over which income and taxes were measured.)

For this analysis, the average annual measure was compared with two different longitudinal measures. The first, longitudinal with longitudinal ranking, combined income and taxes over multiple years into one longitudinal metric for each taxpaying unit and arrayed those units into quintiles on the basis of that metric. Essentially, the measure substitutes a multiyear time horizon for a one-year span, showing distributional results under the longer period.

The second longitudinal measure, longitudinal with first-year ranking, also calculates income and taxes on a multiyear basis, but it arrays taxpaying units into quintiles on the basis of their income in the first year of the measurement period. This second metric shows the income and taxes that accrue to a fixed cross-sectional quintile over multiple years, as taxpayers' income and associated taxes change from one year to the next. By keeping taxpayers in their initial quintile, the measure provides a rough indicator of how likely taxpayers are to be "misclassified" in a cross section.

Limitations of the Analysis

Several shortcomings in the available data limit the conclusions that can be drawn from CBO's analysis. Most important is the issue of attrition. Both the CWHS and the PSID have suffered substantial attrition over time, which may bias results based on them in unknown ways. However, prior research has shown that despite attrition, the PSID remains representative at the cross-sectional level, and CBO's results from the tax data are similar to those from the PSID. Together, that evidence suggests that the attrition bias might not be severe.

A second major limitation involves the length of time that the analysis covered. The longest period over which income was measured was 14 years, a span that CBO considered long enough to mitigate the effects of transitory changes in people's income (such as realized capital gains or spells of unemployment). Even a 14-year period, however, does not fully capture income and tax liabilities over a person's lifetime. The youngest taxpayers will appear in the data only for their middle years, when their income is relatively low. Middle-aged taxpayers will generally show high earnings because they are in their peak income years. Other taxpayers may only appear during their retirement years, and their pensions and income may greatly understate their lifetime income. Unfortunately, no better measures currently exist.

A final cautionary note applies. The distributional methodology used in this analysis differs in some important respects from the methodology that CBO has used in other analyses of effective tax rates (see Table 1). That earlier methodology, which remains CBO's primary approach to distributional analysis, relies on a combination of detailed income and tax data from tax returns and, from the Current Population Survey, information about the family relationships of different taxpaying units and the extent of their nontaxable income. No source of longitudinal data contains the same range of information, so this analysis cannot replicate the kind of results CBO's usual methodology provides. The findings reported below are thus not fully comparable with those from CBO's other distributional analyses. Within the limitations discussed above, however, the general conclusions should be valid.
 

Distribution of Annual and Longitudinal Income and Taxes

In this analysis, CBO examined income distributions and tax rates over several periods of varying lengths. A major finding is that longitudinal measures over such periods show that income and taxes are more evenly distributed among taxpaying units than annual measures show them to be. The largest difference between the two measures appears in the lower-income quintiles. Moreover, the longer the time frame used in longitudinal measurement, the more even the income distribution becomes.

During the three-year period from 1993 to 1995, the distribution of income and taxes as measured annually changed little (see Table 2). For the middle three quintiles, incomes fluctuated around their respective averages over the three years. Income in the highest quintile was flat from 1993 to 1994; then, in 1995, it jumped by more than 5 percent--which pulled up the overall average. Also fluctuating within a fairly narrow range were effective tax rates for the middle three quintiles during the period. And in 1995, when income in the top quintile rose, the effective tax rate increased as well. As a result, the average annual distribution of income and taxes over the period looked much like the distribution in any one year. (Box 3 discusses the income measures used in Table 2.)

Table 2.

At the bottom of the income distribution, income measured longitudinally was higher than average annual income; by comparison, in the highest quintile, longitudinal income was lower. For example, longitudinal income for the second quintile (the lowest quintile was excluded from the analysis) over the three years was $13,650--7 percent higher than the annual average of $12,750 over the same three years. In contrast, longitudinal income for the top quintile--$115,550--was 1 percent below the annual average of $116,750. In each yearly cross section, the lower income quintiles contained people with temporarily low income, and the top quintile contained some individuals with temporarily high income. Measuring income over three years moderates those temporary effects, raising incomes at the bottom and paring them back at the top.

Changes in the share of pretax income for each quintile when measured longitudinally mirrored changes in income levels, rising in the lower quintiles and falling in the top. Even though longitudinal income is noticeably higher at the bottom of the distribution, that quintile's small share of it means that a large percentage increase in income has little effect on overall income shares. For example, the second quintile received 2.3 percent of average annual income and 2.5 percent of longitudinal income. By contrast, the share of longitudinal income received by the highest quintile was close to 1 percentage point smaller than its share of average annual income.

The pattern of effective individual income tax rates was related to that of income. The biggest difference between average annual and longitudinal rates was in the second quintile, for which the changes in income were largest. Because longitudinal income for that group was higher and progressive income tax rates rise as income increases, the associated taxes were also higher.

Extending the analysis over a longer period revealed similar but larger effects (see Table 3 and Table 4). Over the seven years from 1991 to 1997, longitudinal income for the second quintile was 16 percent higher than the average of annual incomes over the same period. Measuring income over the 14 years (1987 to 2000) for which data were available showed longitudinal income exceeding the average annual measure by 26 percent in the second quintile. The differences were smaller for the middle and fourth quintiles; for the highest quintile, longitudinal income was less than average annual income. Again, differences in the shares of income under the two measures mirrored the differences in levels of income. The major effect of lengthening the measurement period was to further narrow the differences between quintiles in levels of income.

Table 3.

Table 4.

Tax rates for the middle three quintiles were higher under the longitudinal measures than under the annual measures, reflecting the quintiles' higher levels of income. In the top income group, longitudinal tax rates were either unchanged from or slightly higher than annually measured rates, despite the lower longitudinal income for that group. Under the annual measures, capital gains for the top quintile made up a larger share of income than they accounted for under the longitudinal measure. Since gains are taxed at lower rates than those that apply to ordinary income, they reduce the overall effective tax rate to a greater extent under the annual measures than under the longitudinal ones.

Analysis of income and taxes using data from the Panel Study of Income Dynamics also found that distributions were more even under longitudinal measures than under annual measures (see Box 4). Examining different periods or varying the restrictions applied to the sample yielded similar conclusions about the relative distribution of annual and longitudinal income and taxes.⁽¹⁵⁾
 

Box 4. Measuring Income Longitudinally Using the Panel Study of Income Dynamics The Internal Revenue Service's Continuous Work History Sample (CWHS) is not well suited to examining the income of lower-income families, many of whom are not required to file tax returns in some years. By comparison, the University of Michigan's Panel Study of Income Dynamics (PSID) avoids some of the pitfalls inherent in the tax data, making it a superior source of information for examining the financial well-being of lower-income families. As part of its analysis, the Congressional Budget Office (CBO) compared annual and longitudinal income from 1987 to 1995, using a measure of taxable income based on the CWHS and measures of taxable income and total income based on the PSID. The same general pattern appeared in both the CWHS and PSID measures: income measured longitudinally was more evenly distributed among quintiles than was income measured annually. Longitudinal income was substantially higher than annual income among families in the lower quintiles and slightly lower than annual income for families in the highest quintile (see the figure below). Varying the definition of income produced substantially different results in the lowest quintile. Taxable income is very volatile for that income group; consequently, when measured longitudinally, it is almost double the amount of average annual taxable income. In contrast, total income measured longitudinally is only 25 percent higher than total income measured on an annual basis. Average annual taxable income among families in the bottom quintile is very low, averaging $1,650 over the nine-year period, so small changes in the dollar value of taxable income for that group produce large percentage changes. Because the bottom quintile relies more heavily on income from nontaxable sources, total income for that group varies less than does taxable income. For the upper four quintiles, the three measures show similar differences between annual and longitudinal income regardless of whether data from the CWHS or PSID are used and whether total or taxable income is examined. Upper-income families are fully represented in both data sets, and they receive relatively little income from nontaxable sources. As a result, the choice of data set and income measure has little effect on the relationship between annual and longitudinal income and taxes for such families.   Difference Between Annual and Longitudinal Income Under Different Definitions of Income, 1987 to 1995 (Percent)

 

Distributional Effects of Illustrative Policy Changes Under Longitudinal Measures

Changes in tax law affect taxpayers at different points in the income distribution in varying ways. Because people's financial circumstances fluctuate from year to year, changes in policy will affect a particular taxpaying unit differently over time. The distributional impact of a policy measured longitudinally will look somewhat different from the policy's effect as measured on an annual basis.

CBO's analysis examined the distributional effects of three illustrative policy changes: reducing the top two individual income tax rates, doubling the value of the child credit, and making the child credit fully refundable. The effects of the first two policies were simulated using data from the CWHS because those policies would primarily affect middle- to high-income families.⁽¹⁶⁾ The third policy would affect mainly low-income families; therefore, CBO used data from the PSID for that simulation.

To assess the impact of the changes in policy over 10 years, CBO used three different measures of the distribution of income: average annual, longitudinal, and longitudinal with annual, or first-year, ranking (see Box 3). For each policy, CBO compared the tax liabilities that each measure generated with a fixed base that reflected tax law in 2000. Comparing the results of the three measures demonstrated the effect of moving from an annual to a longitudinal assessment.

• The average annual measure combined the distributional effects for 10 different years measured separately. For each year, CBO ranked taxpayers into quintiles on the basis of that year's income; the effects of the illustrative policy change were then simulated and allocated to those cross-sectional quintiles. (That method is similar to approaches that are currently used to estimate the distributional effects of changes in tax policy.)
  
  
• The longitudinal measure showed the 10-year change in taxes, assigning taxpayers to their multiyear income quintile. That approach substituted a 10-year time frame for the annual period that is currently used and showed how tax changes affect the tax burdens of each income class.
  
  
• The third method, a longitudinal measure with first-year ranking, arrayed taxpayers into quintiles by their income in the first year but calculated their income and taxes for the full 10-year period. That approach measures how the policy changes affect a fixed cross-sectional income quintile over 10 years.

Reduce the Top Two Individual Income Tax Rates

The first illustrative scenario would reduce the top two tax rates under 2000 law--36 percent and 39.6 percent--to 31 percent. Those top rates, which were created by the Omnibus Budget Reconciliation Act of 1993 (OBRA-93), apply to the highest-income taxpayers. In 2000, the 36 percent rate applied to taxable income above $132,600 for single taxpayers and above $161,450 for married taxpayers filing a joint return. The 39.6 percent rate applied to both single and joint filers who had income above $288,350. This scenario would undo the changes made by OBRA-93, taxing income previously taxed at 36 percent or 39.6 percent at the new maximum rate of 31 percent.

The cross-sectional analysis of that change shows that all of its effects are concentrated in the top of the income distribution (see Table 5). According to the average annual measure, the effective tax rate would change only for the highest quintile, dropping by 1 percentage point over 10 years. Taxpayers in the top 1 percent of the income distribution would see a reduction of 3.2 percentage points in their effective tax rate.

Table 5.

If taxpayers are ranked by their income in the first year of the measurement period and then income and tax changes are measured longitudinally, a somewhat different picture emerges (see the second column in the table). The scenario's effects would still be highly concentrated but at the same time slightly more dispersed. Again, the effective tax rate would change only for taxpayers in the highest quintile in the first year, but some people in the lower quintiles also would be affected by the lower rates--for example, 0.4 percent of taxpayers in the fourth quintile would face a lower tax rate. In addition, some people who had lower income in the base year would have higher income in subsequent years and benefit from the tax change at that point.

Measuring the tax scenario's effects longitudinally would result in no change in effective rates outside of the top quintile (see the third column in the table), although a few taxpayers outside of that income group would face reduced rates. Relative to the cross-sectional measure, a smaller share of taxpayers in the top percentile would be affected by this rate reduction policy under the longitudinal measure. But there is little difference between the two measures in the share of filers in the top 5 percent who would feel an effect, implying movement within the distribution's top few percentiles.

Many more people would be affected by the reduction in rates in any one year than would be affected in the average year. Over the 10-year period, the reduction would affect an average of 2.4 percent of taxpayers. But 6 percent would see a drop in their effective tax rate in at least one year, implying significant movement the falling rates. At the same time, the overall distribution of effects, measured by the change in the effective tax rate, would show smaller changes. That lesser impact would largely result because people who moved briefly into the affected income range would generally see a smaller drop in rates than those who were more permanently in that range.

Doubling the Child Credit

A second tax policy scenario would increase the child credit from $500 to $1,000 for each child aged 17 or younger in a family. As under 2000 law, the credit would begin to phase out when a single taxpayer's income exceeded $75,000 and a married taxpayer's income exceeded $110,000. The credit would not be refundable--that is, its value could be no greater than the taxpayer's income tax liability. Thus, taxpayers who owed little or no income tax before applying the credit would be little affected by it.

Under an average annual measure, the distributional effects of doubling the child credit would be concentrated in the top three quintiles (see Table 6). Taxpayers at either end of the income distribution (that is, the bottom two quintiles and the top percentiles) would be less likely to be affected by the change than would taxpayers in the middle of the distribution. Taxpayers with lower income do not owe enough income tax to take full advantage of the credit, whereas taxpayers with the highest income fall above the phaseout range and hence cannot claim the credit.

Table 6.

Measuring the policy's impact longitudinally shows a distribution of effects that is similar to the distribution under annual measures, with the biggest change in rates in the middle and fourth quintiles. Under the longitudinal measures, the policy change would affect more taxpayers at the top and bottom of the income distribution than would be affected under the cross-sectional measure. For example, under the average annual measure, only 3.2 percent of taxpayers in the second quintile would receive the higher credit, compared with 8.6 percent under the longitudinal measure. As their income rose and fell, both low-income and the highest-income taxpayers would move into and out of the income range in which they could receive the credit. The effects of the illustrative policy appear most dispersed when measured longitudinally with first-year ranking.

Many more taxpayers would receive the credit in at least one year of the period than would receive it in the average year. Just under one-third of filers would get the credit in the average year; by contrast, more than half would receive it in at least one year. Income dynamics partially explain that outcome. In any given year, many taxpayers would have income that was too high or too low to receive the credit. But over a multiyear period, fluctuations in income would move them into the range in which the credit would affect their tax liability. Demographic changes would also contribute to the credit's being more widely distributed among the quintiles. Over the 10-year window, some families would have newborn children, making them newly eligible for the credit, whereas other families would see their children turn 18 and thus no longer qualify for it.

Under the average annual measure, only 3 percent of taxpayers in the second quintile would receive the higher credit, compared with 31 percent who would receive it in at least one year under the longitudinal measure. Yet that difference may overstate the degree to which those taxpayers would experience a lower effective tax rate, because they would benefit only in years in which their income was high enough that they would owe income tax. Only 9 percent of taxpayers in the second longitudinal quintile would receive the credit in the average year, which implies that the average recipient in that quintile would get the credit in only three of the 10 years possible. In contrast, recipients in the third longitudinal quintile would receive the credit on average in five of the 10 years, while those in the fourth and fifth quintiles would get the credit in nearly seven years of the period.

Make the Child Credit Fully Refundable

Under the tax laws prevailing in 2000, the child credit is not refundable, so families that have no income tax liability cannot receive it. This third policy scenario would make the $500 child credit fully refundable, allowing such families--those with no earnings or income tax liability--to claim the full credit.⁽¹⁷⁾

Analysis of that policy using an average annual measure showed that the policy's effects would be heavily concentrated in the bottom of the income distribution (see Table 7). The change would lower the effective tax rate in the bottom quintile by 2.8 percentage points; it would cut the effective rate in the second and middle quintiles by 0.6 and 0.1 percentage points, respectively. Some families in the fourth quintile who had very low levels of taxable income compared with their total income would also be affected by the change, although not by enough to change the quintile's effective tax rate.

Table 7.

Under the longitudinal measures, the distribution of the policy change's effects among the quintiles was similar to that under the average annual measure but somewhat more dispersed. The effective rate as measured using any of the three methods would change by more than a percentage point in the bottom quintile, by about half a percentage point in the second quintile, by one-tenth of a percentage point in the middle quintile, and not at all in the fourth and fifth quintiles. Under the longitudinal measures, some families in the fourth and highest quintiles would be affected by the policy because they may experience spells of unemployment, businesses losses, or some other decline in their income. Those circumstances might push their income below the level at which they would owe federal income taxes, allowing them to claim the credit because of its refundability.

The change in the effective rate for the lowest quintile is smaller when measured longitudinally than when measured annually. Lower-income families would receive about the same amount of credit under both longitudinal measures. However, because their longitudinal income would be much higher than their annual income, the tax reduction would represent a smaller percentage of their income under the longitudinal measures.

Many more families (over 16 percent) would receive the credit in at least one year than would receive it in the average year (fewer than 7 percent). As with the previous policy scenario, both fluctuations in income and changes in the families' composition would cause different families to benefit from the credit over time.

Conclusion

For each of the illustrative tax policy scenarios, the effects of the change were distributed more broadly when they were assessed under the longitudinal measures than when evaluated under an annual metric. CBO's results showed that income quintiles that were unaffected by a change in tax policy in a single year might be affected over longer periods, as income rose and fell. However, the choice of either a longitudinal or an annual metric to measure a potential policy's effects would not dramatically shift the overall distribution of any of the changes CBO examined. Policies that when measured annually would primarily affect the upper or lower portions of the income distribution showed roughly similar patterns in the longitudinal distributions.