Correcting for differential item functioning in multi-level regression models in cross-national surveys

• To assess the impact of DIF on background variables in a multi-level regression analysis.
• Results show the impact of DIF cannot be ignored.
• DIF is mitigated using country specific item parameters.
• Different IRT models and person parameter estimation methods are used.
• Impact of DIF is more for PCM than GPCM & more for EAP than WML estimates.

Results of the PISA project have shown that the school average socio-economic status is an important background variable that explains a lot of variance in the student results. However, if the socio-economic variable which is measured at the student level is biased across countries due to a cultural bias, then the aggregated variable (at school level) is also subject to error. In this article, DIF (Differential Item Functioning, i.e., item bias) is mitigated using country specific item parameters. The effect of using this approach is studied on the results from multilevel regression for different measurement models and person parameter estimation procedures. Results showed that for countries affected by DIF the impact on the regression coefficients cannot be ignored. The effect is shown to be more for the PCM than the GPCM and generally more for the EAP estimates than the WML estimates.