What Do We Really Know About Wages? New Evidence From A Natural Experiment

Lilliard, Smith and Welch (1986) note the rising proportion of workers whose earnings go unreported in the Current Population Survey and argue that the Census Bureau's imputation procedure fails to account for the positive correlation between nonresponse and income. As a consequence, an imputation or reporting bias is substituted for the nonresponse bias with unclear implications for what we really know about wages. In 1989 the processing system for compiling CPS data was substantially revised. To permit users of the public use tapes to judge the comparability of the data under the two processing systems, the Census Bureau issued two March 1989 CPS tapes containing earnings imputed under the old and the new system. The distinguishing characteristic of the old procedure was that if one piece of information was missing from an interview, then all items were imputed. The new imputation procedure imputes only the true missing values. Combining individual level data from both processing systems provides a natural experiment that sheds new light on Lilliard, Smith, and Welch's question--What Do we Really Know About Wages? We exploit the earnings data for the same individuals in several ways. First, we investigate erasure bias by estimating earnings equations based on both the old and the new CPS processing system data. Second, we analyze reporting bias by estimating a nonresponse probit and using the inverse Mills ratio to correct the earnings equation for reporting bias. Third, we explore the important question of model bias in the imputation of earnings. We find that for white males the effect of correcting for reporting bias dramatically increases the returns to education. In addition to stressing the importance of reporting bias in the pre-1989 data, our results indicate that failure to correct for nonparticipation is a serious and continuing problem in the post-1989 CPS processing system.