Evaluación de criterios de bondad de ajuste para modelos logit marginales

Abstract

This thesis presents a study about the performance of goodness of fit statistics for marginal logit models with correlated binary data. Some of them are descriptive measures, usually based on the concept of information lost when an approximating model is used to describe the reality, while others are global goodness of fit statistics. In general they are natural extensions of those available for conventional logit models, i.e. under the assumption of independent binary data. In order to evaluate these measures, a simulation study is carried out using a simple data generation algorithm with exchangeable correlation structure. The Akaike criterion extension stands out among the descriptive statistics and it is also a measure useful to select the best covariable subset to include in the model. Among the global goodness of fit statistics, the extensions of the Pearson statistic and the unweighted sums of squares statistic have the best behavior in terms of type I error rates but they have low power.