Special QME Seminar: Sources of Error in IRT Trait Estimation: Effects on Trait Score Bias and Confidence Interval Coverage Rates

Item response theory (IRT) models item response probabilities as a function of item characteristics and latent trait scores. Within an IRT framework, trait score misestimation results from (1) random error, (2) the trait score estimation method, (3) errors in item parameter estimation, and (4) model misspecification. Through a simulation study, I explore the relative effects of these error sources on the confidence interval coverage rates for trait scores. Overall, coverage rates are most accurate for central trait scores using non-Bayesian methods. Moreover, certain types of model misspecification lead to severely biased trait estimates and poor confidence interval coverage. Implications are discussed in the context of computerized adaptive testing.

Friday, April 22, 2016 - 10:00am