Much recent research has been directed at earnings management via the manipulation of real activities (i.e., real earnings management or REM). While all tests of REM hinge critically on the underlying model of abnormal real activities there is no systematic evidence on the properties of commonly-used REM measures or on the specification of statistical tests based on REM measures. This study provides such evidence by documenting which measures of REM lead to the best-specified tests of real activity management across a wide variety of settings, and which measures do not. Such evidence facilitates a keener understanding of the evidence prior studies have provided on REM in addition to informing future researchers on the trade-offs involved in measuring REM. In general, our results indicate that the traditional REM measures used (to date) in the literature tend to be severely mis-specified in the sense that they indicate the presence of abnormal real activities too often in samples constructed at random (i.e., their Type I error rates differ dramatically from 5%). To address such specification problems we develop alternative measures of REM based on performance-matching. While the performance-matched REM measures may not be well-specified in each and every possible case they tend to provide better specified tests across a wide variety of different research settings when compared to the traditional REM measures. Our findings suggest that use of performance-matched REM measures is likely to provide a more reliable basis from which to draw inferences about REM. In some settings, however, performance-matched REM measures may provide conservative tests of REM-related hypotheses.