Modelos multidimensionais de item na avaliação de ingresso da UFLA

Abstract

Item Response Theory (IRT) has been shown as a worhty tool to analyse educational data. However, choosing one out of many models is not a simple task. This is specially true when you can vary the dimension of models for each item. In this thesis we try to enlighten some statistical properties and the analytical potential to interpret multidimensional models of IRT (MIRT). In the first part a review of the literature is done foccusing on compensatory multidimensional models for items. This part bring a didactic effort to present a state of art description of the fully Bayesian MIRT analysis. This analysis is illustrated with data from entrance examination of candidates for Agronomy undergraduate program in the second semestre of 2006 at the Universidade Federal de Lavras (UFLA). The best model in this section is the one that has two dimensions for all items. Second part has two papers. The first discusses a forward type of model selection procedure to identify different dimensions for each model. The second bring a pedagogical analysis based on the same data using two dimensions model. As a final chapter we draw the last remarks on what was learned in our investigation.