@EdWorkingPaper{ai25-1303,
title = "Beg to DIFfer: Resolving Statistical Complications of Intersectional DIF Analyses",
author = "Lily An, Edward J Kim",
institution = "Annenberg Institute at Brown University",
number = "1303",
year = "2025",
month = "October",
URL = "http://www.edworkingpapers.com/ai25-1303",
abstract = {Modern test developers conduct differential item functioning (DIF) analyses to ensure fairness in educational and psychological testing. To address previously unrecognized biases, researchers have recently demonstrated the importance of conducting intersectional DIF analyses that attend to the intersectional nature of test-takers’ multiple identities. However, these intersectional DIF approaches overlook how overlapping identity categories affect the statistical validity of DIF analyses. As the related tests violate independence, typical p-value corrections used in intersectional DIF analyses such as Bonferroni yield overly conservative family wise error rates (FWER) which limit statistical power to identify true DIF. Additionally, DIF on one dimension can spuriously cause DIF to appear while testing another demographic dimension with high overlap, a phenomenon we call signal interference. These concerns are particularly aggravated in intersectional DIF. We offer an approach utilizing parametric bootstrapping that adjusts significance levels of DIF detection processes to yield the intended Type I error rates. Using simulations studies, we illustrate the statistical complications of intersectional DIF analyses and the ability of our proposed method to resolve them.},
}