This paper examines the implications of using the absolute value of discretionary accruals when testing for earnings management. First, we analytically develop the mean and variance of the distribution of absolute discretionary accruals, and show that the expected value is an increasing function of the variance in the underlying error term from the first-stage discretionary accrual estimation model. Second, we highlight several firm characteristics that are related to the error variance in discretionary accrual estimation models. Using simulations, we show that correlation between the earnings management partitioning variable and these firm characteristics leads to an overrejection of the null hypothesis of no earnings management. Third, we provide research design suggestions to help researchers mitigate the potential bias arising from the use of unsigned measures of earnings management. Using these suggestions, we replicate a recent study, and demonstrate that the inferences change after controlling for operating volatility.