Detecting differential item functioning (DIF) is an essential step in establishing measurement invariance in assessments. While traditional methods such as the likelihood ratio test (LRT) are widely used, machine learning approaches have shown promising results but are often computationally expensive and difficult to interpret. This simulation study compares the LRT with two proposed machine learning-based approaches for DIF detection, the Lasso Logistic Regression (Lasso) and Lasso Logistic Regression with interaction effects (Lasso-int), and a random forest (RF) approach. 

Performance was evaluated under conditions of uniform and non-uniform DIF, varying latent trait distributions, and was further assessed using a real dataset. Model performance was assessed using the true positive rate (TPR) and false positive rate (FPR). The LRT demonstrated consistently low FPRs and moderate TPRs, serving as a competitive baseline. The Lasso-int approach showed high TPRs in detecting non-uniform DIF. However, both Lasso and Lasso-int were sensitive to differences in latent trait distributions, yielding high FPRs. The RF method consistently produced low FPRs, but also lower TPRs relative to the other methods. Although the proposed Lasso approaches represent a step toward more interpretable machine learning methods for DIF detection, their high FPRs currently limit their practical applicability.