Speaker: Phil Chalmers, York University
Department of Psychology

Title: Profile-likelihood Confidence Intervals in Item Response Theory Models

Abstract: Confidence intervals (CIs) are a fundamental inferential tools that are used to quantify the sampling variability of parameter estimates. In item response theory applications, CIs have been primarily obtained from large-sample Wald-type approaches based on standard error estimates, derived from the observed or expected information matrix, after parameters have been estimated via maximum likelihood. However, an alternative approach to constructing CIs is to quantify sampling variability directly from the likelihood function with a technique known as profile-likelihood confidence intervals (PL CIs; Pawitan, 2001). In this presentation I introduce PL CIs for item response theory models, compare PL CIs to classical large-sample Wald-type CIs, and demonstrate important distinctions among these CIs. CIs are then constructed for estimated parameters and for transformed parameters that are often obtained post-estimation. Monte Carlo simulation results suggest that PL CIs perform considerably better than Wald-type CIs for both non-transformed and transformed parameters.