Sample-size restrictions limit the contingency table approaches based on asymptotic distributions, such as the Mantel-Haenszel (MH) procedure, for detecting differential item functioning (DIF) in many practical applications. Within this framework, the present study investigated the power and Type I error performance of empirical and inferential criteria for DIF detection in small samples. Sample sizes (50/50, 100/50, 200/50, and 100/100 for the reference and focal groups, respectively), ability distributions (equal and unequal), and amount of DIF (moderate and high) were manipulated. The results show the advantages of employing theMHchi-square statistic using high levels of significance ([alpha] = .20) as opposed to the empirical criteria (cutoffs for categorizing DIF based on the magnitude of the MH common odds ratio estimator and the standardized p-difference statistic). Some considerations concerning Type I and Type II errors are made.