Rounding in Earnings Data

Earnings data are often reported in round numbers. In fact, in the March 1995 Current Population Survey (CPS), 71% of all full-time earnings responses are some multiple of $1,000. Rounding is typically ignored in analyses of earnings data, which effectively treats it as noise in the data. Our GMM estimates of a simple model of rounding indicate that this behavior is highly systematic and correlated with the respondents’ earnings level. We find that the systematic nature of rounding can affect some commonly used statistics based on earnings data. The statistics we investigate in this analysis are inequality summary measures, earnings quantiles, kernel density estimates, and frequency plots of wage adjustments. We find that rounding alters most of these statistics substantially, that is, by more than the typical level of annual changes or reported standard errors.