An Evaluation of Unsigned Discretionary Accrual Measures after Relieving the Impact of Influential observations on Accrual Model Estimation and Hypothesis Testing

This study evaluates the accrual models after mitigating the impact of influential observations using MM-estimation, a robust regression method. I find that, relative to ordinary least squares (OLS) regression using winsorized data, MM-estimation reduces the magnitude of the unsigned abnormal accruals for the majority of the observations and dramatically reduces the type I error rates in hypothesis testing. Also, relative to their robust counterparts, the small OLS unsigned abnormal accruals have low test power due to the noise caused by influential observations. Using a prior study, I show that influential observations that are 1 percent of the sample cause a type I error. I also show that small robust unsigned abnormal accruals have higher power than their OLS counterparts