Measuring Discretionary Accruals: Are ROA-Matched Models Better than the Original Jones Type Models?

Abstract

Kothari et al. (2005) propose ROA-matched models to estimate discretionary accruals for samples skewed toward firms with good or poor performance to reduce the frequency of Type I errors. Many researchers, however, use such models for unskewed samples. In these cases, ROA-matched models have no advantage over the original Jones type models in frequency of Type I errors. This study shows, more importantly, ROA-matched Jones type models systematically underestimate discretionary accruals for firms that manage earnings. Thus, researchers using discretionary accruals estimated from such a model as an independent variable in regression analysis may obtain a coefficient estimate with a different sign from the true coefficient. Moreover, the ROA-matched models are likely to have a much higher frequency of Type II errors than the original Jones type models. The results of empirical tests are consistent with this prediction. Depending on the extent of earnings management, the ROA-matched models often are less than half as likely to detect earnings management as the original Jones type models for random (unskewed) samples. The results hold for both annual and quarterly data, and regardless of whether accruals are estimated using the balance sheet or income statement approach. Researchers therefore should use the ROA-matched models in empirical research with caution, and only for situations where the partitioning variable is highly correlated with performance.