Current approaches for studying differential item functioning (DIF) using manifest groups are problematic since these groups are treated as homogeneous in nature. Additionally, manifest variables such as sex and ethnicity are proxies for more fundamental differences - educational advantage/disadvantage attributes. A simulation study was conducted to highlight issues arising from the use of standard DIF detection procedures. Results of this study showed that as the amount of overlap between manifest groups and latent classes decreased, so did the power to correctly identify items with DIF. Furthermore, the true magnitude of the DIF was obscured making it increasingly more difficult to eliminate items on that basis.

After some problems with manifest group approaches for DIF had been identified, a recovery study was conducted using the WINBUGS software in the analysis of the mixed Rasch model for detecting DIF. In this study the mixed Rasch model also showed a lack of power to detect items with DIF when the sample size was small. However, this approach was able to identify the proportion of and ability distribution for each manifest group within latent classes, thereby providing a mechanism for judging the appropriateness of using manifest variables as proxies for latent ones. Finally, a series of protocols was developed for examining DIF using a latent class approach, and these were used to examine differential item functioning on a test of language proficiency for English language learners. Results showed that 74% of Hispanic and 83% of Asian examinees were in one latent class, meaning any DIF found by comparing manifest groups would be an artifact of a relatively small number of examinees. Examination of the output from the latent class analysis provided potentially important insights into the causes of DIF, however covariates were not predictive of latent class membership.