Estimating Models of Trends in Income Volatility with the PSID: New Results and Comparisons to the Literature
Abstract
This Abstract serves two functions: (1) An Introduction to the entire Session and a description of the motivation of the Session, and (2) An Abstract for this specific paper.(1) Introduction to the Session. There is by now a very large literature on estimating earnings dynamics models in the U.S., most of which uses error components models of various types which decompose the variance of earnings over the life cycle into permanent and transitory components (Lillard and Willis, ECMA, 1978 is usually cited as the original). A particular focus of this literature has been whether earnings instability and volatility has increased in the U.S. over time. Gottschalk and Moffitt (BPEA, 1994) was the first paper in this literature but there has been a stream of subsequent estimates in the literature (reviewed by Moffitt and Zhang, May AER P&P, 2018). However, while most estimates have been performed with the Panel Study of Income Dynamics (PSID), estimates are also available for other data sets, including matched Current Population Survey (CPS), the Survey of Income and Program Participation (SIPP), and administrative data sets such as the Longitudinal Employer-Household Dynamics (LEHD) or Social Security earnings data.
As the review by Moffitt and Zhang showed (this paper is attached to this submission), the estimates of trends are often different across data sets. Whether estimated by what Moffitt and Zhang call “gross volatility,” which is just the variance of some measure of dispersion of the change in earnings from t to t+1, or by the variance of the transitory component in a model which has a permanent-transitory decomposition, instability in annual male earnings rose in the U.S. in the late 1970s and early 1980s, experienced a flat or declining period through the mid-2000s, and then resumed its increase. Matched CPS data show somewhat similar trends but the turning points are not the same. SIPP data show a very different trend, and administrative data on earnings show downward trends.
This is an important issue, for whether earnings instability has been going up, down, or neither is clearly something that needs to be established. The problem is that the different studies have used different models of earnings dynamics and have used different measures of instability, and have used different samples, variable definitions, and other differences. The goal of this session is to make headway on this problem by bringing together 4 papers, each using one of the major data sets, to estimate common earnings dynamics models, using similar sampling frames and exclusions. Further, assuming there are still differences in instability trends, the papers in the session will attempt to explore reasons for those differences, including possible biases of various kinds.
The four data sets are the PSID, matched CPS’s, the SIPP, and an administrative data set, the LEHD, drawn from wage records in the UI system. There will be a separate paper on each. The four authors will work jointly in a two-stage process. First, the authors will estimate common earnings dynamics models on their data sets, using as similar sample compositions and variable definitions as possible, and then exchange their results in Fall, 2018. Second, they will then work jointly, as a team, to explore different hypotheses for any differences in findings. A partial list of differences to explore will include biases due to attrition, imputation for missing data, measurement error, and biases due to selection into the particular samples. If successful, the papers in the session will jointly establish strong evidence on what the trend in earnings instability in the U.S. actually has been.
(2) This specific paper. This specific paper will contribute to this literature by providing new estimates of trends in earnings volatility with the PSID and by (1) attempting to reconcile estimates from the PSID with those of other data sets and (2) using common models used by other papers in the session with other data sets. Differences in trends in different data sets could be the result of either some underlying difference in the data or a difference in the models used. The paper will investigate both using PSID data from 1970 to 2015.
With regard to new estimates, we will provide new estimates of trends in instability and volatility of female earnings and family income, which have rarely been examined with the PSID. This will also help us make new comparisons to trends in other data sets, which often have computed trends in those variables.
With regard to the differences in the underlying data, we will conduct a wide variety of tests of estimates from the PSID which will address some of its potential deficiencies. One leading possibility is that there is attrition bias in the PSID, which by now has reached over 50 percent. Tests for selection on observables and unobservables will be conducted to determine how estimates of earnings instability are affected. Another is that methods of imputation for missing earnings--a common issue in all survey data sets, including the others in this session--may affect estimates of earnings volatility. We will examine the effect of the PSID’s methods of imputation, consider alternative methods, and will attempt to reach conclusions about whether is likely that estimates of earnings instability are affected by this issue. We will also explore other ways to test for measurement error. A third issue is that the PSID data typically provide only information on heads of household and their spouses, which makes comparability difficult with administrative data sets like the LEHD and SSA which cannot identify headship. However, The PSID has recently released a public file of earnings on other members of the family in 2005-2015, and we will use that to explore this issue in our work. Our examination of family incomes instability trends and how they compare to those in other data sets will also help in this regard, because family income includes all family members’ earnings. We will also endeaver to mimic the sampling frames of the CPS, SIPP, and LEHD papers in this session.
With regard to the use of common models, we will estimate econometric models of earnings dynamics that have been used in other data sets to determine if this could be a source of difference. The decomposition of earnings levels or changes into permanent and transitory components requires assumptions which can differ across studies. We will also extend a new model with more flexible forms and less restrictive assumptions which we introduced in Moffitt and Zhang (2018). Finally, we will work with the authors of the other papers in this session to estimate common models of earnings dynamics.