Iterative scaling methods for log-linear models expressed using curved exponential families

General log-linear models specified by non-negative integer design matrices have a wide range of applications, for example, small area estimation, image recognition, and text processing. The generalized iterative scaling (GIS) of Darroch and Ratcliff a popular choice for theis MLE computation under such models. The proof of convergence of GIS and its extensions relies heavily on the assumption that a log-linear model is a regular exponential family. We developed an adjusted version of GIS that can be used for curved families as well and gave a proof of its convergence. The application of the new GIS to Monte-Carlo-based power calculations for goodness of fit of log-linear models is proposed and illustrated using data obtained from a real clinical study.

Scientists involved

• Anna Klimova

Publications

Funding

NCT