Alternative Computational Methods for Estimation in Multinomial Logit Response Models. Revision

摘要

Several algorithms have been proposed for the computation of maximum likelihood estimates for contingency tables. Since multinomial logit response models can be treated as special versions of log-linear models, many of these techniques can be used for logit models as well. In this paper we compare, in a qualitative fashion, the relative merits of (1) two variants of Newton's method developed by Fienberg and Stewart; (2) GLIM, as developed by Nelder and Wedderburn; (3) the BMDP program for stepwise logistic regression; and (4) the widely used method of iterative proportional fitting. (Author)