Focused Model Selection for Markov Chain Models: With an Application to Armed Conflict Data

This thesis is devoted to the development of a focused information criterion for dynamic multinomial logit models. The achievements of the thesis are fourfold. First, a dynamic multinomial logit model is defined which admits the possibility of model misspecification. Then, approximate large sample distributions of maximum likelihood estimates of this model are deduced. The deduction is done both for correctly specified models and for misspecified models. On the basis of these approximate distributions, the Focused Information Criterion is constructed. The performance of the developed Focused Information Criterion is investigated through simulation experiments. It is shown that the developed information criterion indeed aims at selecting the models giving the most precise estimate of the focus parameter. As an application of the developed methodology, armed conflict data are analyzed. The focus parameter in this analysis is the probability of conflict escalation. The findings show that the level of democracy has no significant e ect on conflict escalation probabilities.