(A) new approach for large scale loglinear modeling

Abstract

A log-linear modelling takes a long time if the data involves many variables and if we try to deal with all the variables at once. Fienberg and Kim (1999) investigated the relationship between log-linear model and its conditional, and we will show how this relationship is employed for finding the actual model structure based on a collection of conditional log-linear structures. We show that parts of model structure can be read from conditional model structures and so that we can find the actual model structure by putting the conditional log-linear structures together as long as the conditionals are collected under some reasonable condition. The method proposed in the thesis is applied to simulated data with a strong indication that the method may be very useful for a large scale model searching. The result of this thesis can be applied for modelling all the hierarchicallog-linear models.