A Study of Structural Modeling Using Plausible Value Imputation

Data from computerized adaptive tests can be used to evaluate hypotheses about student proficiency, such as hypotheses about differences between groups of students (e.g., gender, racial/ethnic group, previous education, preparatory training) and hypotheses concerning the development of proficiency. Such hypotheses can be evaluated by analysis of variance and regression models for item response theory (IRT) proficiency parameters. Several methods for the estimation of such models are compared with respect to their power for the detection of effects: methods based on plausible value imputation and methods based on marginal maximum likelihood estimation. The power of the methods is evaluated for models for dichotomously scored items, polytomously scored items, and an IRT model for responses to dichotomous items combined with the RTs. Simulation studies show that a relatively simple plausible value imputation method that ignores the covariance between measurement occasions does not perform worse than more advanced methods.