Application of the MIMIC model to detect and

Research Questions v RQ 1 – Is there differential item functioning between gender on

Differential Item Functioning v Differences in probability of endorsing an item between two groups

MIMIC Model v Generalized SEM which incorporates covariates on items and latent traits v.

Scale Purification v Items used to measure latent trait might exhibit DIF and should

Predictive Model v Interested in items exhibiting largest DIF based on effect size v

Judgment Time v Within items, interested in largest absolute drop in effect size from

Judgment Time v Depends on your interest as a researcher v. Substantive, focused on

Slides: 13

Download presentation

Application of the MIMIC model to detect and predict differential item functioning Kevin Krost, M. A. Josh Cohen, Ph. D. Virginia Tech, Educational Research and Evaluation

Research Questions v RQ 1 – Is there differential item functioning between gender on this assessment? v RQ 2 – What attitudinal factors predict or mediate differential item functioning and/or gender differences among mathematics?

Differential Item Functioning v Differences in probability of endorsing an item between two groups after controlling for ability v. Can indicate item bias v. Difficult to explain and determine cause

MIMIC Model v Generalized SEM which incorporates covariates on items and latent traits v. Family and link function for items v Effect of grouping variable on latent trait indicates group differences v. Items indicate DIF v. Covariates DIF can predict item and mediate

MIMIC Model

Scale Purification v Items used to measure latent trait might exhibit DIF and should be removed from scale (Wang, Shih, & Yang, 2009) v. Iterative procedure v. Items detected as exhibiting DIF removed from scale and modeled again v Time-consuming!

Purified Model

Predictive Model

Predictive Model v Interested in items exhibiting largest DIF based on effect size v Enter all covariates to predict item and mediate DIF v. Based on results, drop nonsignificant covariates to create more parsimonious model v. Drop all non-significant variables or one at a time?

Judgment Time v Within items, interested in largest absolute drop in effect size from model to model v. Most interested when grouping variable becomes non-significant v After most parsimonious models obtained, interested in covariates significantly predicting item v. Again, interested in full mediation

Judgment Time v Depends on your interest as a researcher v. Substantive, focused on relationship of covariates and mediation of DIF v. Professional/test developer, interested in determining if DIF indicates item bias or simply spurious v. Both, depending on research interests

Any questions? THANK YOU!