Earnings volatility across groups and time.

INTRODUCTION

Earnings volatility is of fundamental interest to many different groups of economists. Researchers studying tax policy are interested in earnings volatility because of the implications for distributional analysis of tax burdens: in a progressive tax system, the more individual incomes vary over time, the more divergent are conclusions about effective tax rates when comparing annual and multi-year measures. (1) Earnings volatility plays a crucial role in macroeconomics because of the impact that earnings uncertainty has on consumption behavior. (2) Recently, earnings volatility has become an interesting policy question in its own right because a growing literature has argued that economic well-being has been adversely affected in recent decades because individual earnings became more volatile. (3) The conclusion that earnings volatility increased is far from universal, however, as other studies have found that variability in earnings growth was flat or even declining since 1980. (4)

Inferences about earnings volatility across groups and time depend on the underlying models of earnings dynamics, data sources, earnings concepts, and sampling strategies. The analysis in this paper suggests that all of these inputs have likely played some role in the divergence of conclusions in previous studies. Our basecase model of earnings dynamics with stable permanent shocks that vary by age and education is consistent with a transitory component that has declined over time. However, altering the earnings concept to include self-employment income and changing the sampling strategy to include observations with minimal labor force attachment has first-order effects on the results. Indeed, those effects may be enough to explain why some studies conclude that earnings volatility is rising.

Three data sets are used in this paper. The first is a one percent longitudinal sample of earners ages 25-55 between 1980-2006 drawn at random from the Social Security Master Earnings File (MEF). The advantage of longitudinal data is that one can measure and analyze changes in log earnings over various time periods, which is the key to our first approach to modeling earnings volatility. The second is another longitudinal sample from the MEF, but this time for the 1940-1960 birth cohorts and linked to Survey of Income and Program Participation (SIPP) data files from the 1990s. For the purposes of this paper, the advantage of the SIPP data is that we know educational attainment, so we can investigate differences in earnings shocks by level of schooling. The third data set is a series of cross sections from the March Current Population Survey (CPS) between 1970-2007. Although one cannot measure changes in earnings using the CPS cross-sections, the data are useful for analyzing the variance of earnings levels over time in a synthetic panel framework and the addition of labor force attachment variables makes it possible to investigate how sampling may be affecting the results.

The starting point for the analysis in this paper is the observation that the variance of earnings growth rates in the MEF longitudinal earnings data declined significantly over the period between the early to mid 1980s and the early to mid 1990s, and has since remained flat. However, a decline in the variance of growth rates does not necessarily imply that volatility has fallen, because labor economists have long recognized the importance of distinguishing permanent and transitory earnings shocks. Permanent shocks reflect differential earnings growth within some reference group that is expected to persist--another way to describe economic mobility. Transitory shocks are temporary (though not necessarily gone after one period) and thus associated with volatility per se.

There are a few different ways to use panel data to separate earnings growth variability into permanent and transitory components. We use an approach suggested by Meghir and Pistaferri (2004) for measuring the variance of permanent shocks that is both intuitive and robust to alternative specifications of the time series properties of transitory shocks. Using that approach, our longitudinal earnings data suggest that (1) the variance of permanent shocks declines with age, (2) the variance of permanent shocks is higher for the college educated, and (3) the variance of permanent shocks has been constant (within age and education groups) over time.

If the variance of permanent shocks within groups has been relatively stable, and overall growth rate variability within groups is falling, that suggests the variance of transitory shocks (volatility) must have fallen. Indeed, that finding would be consistent with evidence from the literature on the "Great Moderation" in macroeconomics (Davis and Kahn, 2008). Also, there was a significant secular decline in U.S. unemployment rates (at least through our sample period), and unemployment is the event one would generally associate with transitory shocks. However, the book is not completely closed on the decrease in transitory earnings shocks; Moffitt and Gottschalk (2008) argue it is possible that transitory shocks actually got bigger but became more serially correlated. Our residual approach--starting with total variation and subtracting permanent shock variation--cannot distinguish between changes in the transitory variance and changes (in the opposite direction) in the covariance of transitory shocks.

In addition to looking at the variability of earnings growth within groups and over time, one can characterize earnings dynamics by looking at the variance of earnings levels within an age group or birth cohort over time. The canonical model that underlies the transitory and permanent decomposition implies that the variance of earnings at any given age and point in time will depend on the variance of transitory shocks at that point in time, the initial earnings dispersion for the age or cohort group in question, and the accumulated permanent shocks since that initial time period. If the stochastic earnings process is stable, one should see stable earnings variances in the synthetic cross section.

The synthetic cross section we develop is based on March CPS data. The advantage of this data relative to our administrative records is that we know self-reported labor force attachment, so we can investigate how restricting the sample to those who worked full-time or eliminating de minimis earnings affects the answers. The CPS data confirms our inferences about a stable permanent shock variance (at least after 1980) while the overall variance of earnings (for our preferred measures) was falling, which is consistent with declining volatility. However, we also show that changing the sampling criteria to include observations with modest labor force attachment and very low self-employment earnings has a dramatic first-order effect on the estimated variances. The sampling criteria are very likely a part of the explanation why researchers have disagreed about trends in earnings volatility.

MEASURING THE VARIABILITY OF EARNINGS GROWTH RATES OVER TIME

Although the concept of earnings volatility may seem straightforward in principle, measuring it in practice requires a number of decisions about data, sampling, and choice of summary statistic. In this section, we use administrative panel data from Social Security earnings records to present some basic findings about the standard deviation of one-period earnings growth rates over time. Although the estimated level of variability at any point in time depends on how the sample is chosen, all of the measures we present suggest a general decline in variability between the early 1980s and mid 1990s, and little change thereafter through the end of our sample in 2006.

There are two main longitudinal data sets used in this section and the two sections that follow. Both data sets are ultimately based on Social Security earnings records in the Master Earnings File (MEF). (5) The first sample is a one percent random draw from the MEF, and the second is a draw from the MEF based on linkages to several panels of the Survey of Income and Program Participation (SIPP). The SIPP-linked data is useful because it introduces more demographic information than is available on the Social Security records. For our purposes, the SIPP provides the level of educational attainment used to further subdivide the sample when looking at various types of earnings shocks in subsequent sections. In both samples, data on wages from W2 reporting are used for all years back to 1980 and data on wages plus self-employment earnings are used for years after 1994. (6)

The decision about which age and/or cohort groups to include in the sample is somewhat dependent on the question being asked. The one percent random sample results presented here are generally based on ages 25-55--what we refer to as prime working years. (7) In the SIPP, the main sample is drawn for birth cohorts 1940-1960. The cohort restriction is set so the sample is mid--career around the points in time when the SIPP linkages are established (1990 through 1996), while generally assuring that the educational attainments are effectively completed for every observation (that is, the links are established for everyone 25 or older at the time of the survey).

The concept of earnings growth variability used in this section is the standard deviation of the one-period change in log earnings. The standard deviation is a convenient statistic, as the extent of variability is summarized in a single number, and the standard deviation is a useful starting point for distinguishing permanent from transitory earnings shocks. Focusing on the standard deviation also has its drawbacks, as it may hide important information about the symmetry of shocks, potentially allowing some very large percentage changes to dominate the results even though their meaning is dubious. In particular, a large percentage change in the standard deviation can mean large dollar changes near the average earnings, but it can also mean relatively small dollar changes at very low earnings. (8)

Print Get the Mag Weekly Updates