BAYESIAN: ON OVERCOMING NON-CONVERGENCE AND UNREALISTIC PARAMETER ESTIMATES IN ITEM RESPONSE MODELLING

Abstract

Item response modelling is a theoretical frame work organised around the concept of latent traits. Estimation of more than two parameters in binary and multinomial item response modelling using maximum likelihood method may be difficult and produced unrealistic estimates due to non-convergence and high standard error of parameter estimates. Responses from 403 test takers were coded dichotomously and polytomously to illustrate our models with the aids of Stata 17SE on window platform. Bayesian binary item response modelling with prior and hyper-prior were used to overcome non-convergence while Bayesian multinomial modelling was developed using d cat function in Bayesian Inference Using Gibbs sampling to produce a more parsimonious estimates