Attrition bias and inferences regarding earnings properties; evidence from Compustat data

On average, across the years 1980 to 2018, almost 8.5 percent of firms on the Compustat Annual data set, which had earnings observations in year t-1, did not have earnings observations in year t. Because these disappearances were not random, there is attrition bias in estimates of earnings properties that require earnings observations in two consecutive years: namely, earnings persistence (used as an estimate of earnings quality), mean reversion, and accuracy of forecasts of earnings that are based on earnings of the prior year. We suggest three methods for imputing the disappeared earnings observations, which may be useful in future research on earnings properties. We show that conclusions about the properties of earnings change when we use imputed earnings to reduce the effects of attrition bias.