Paper No. 10-18
I Kosmidis and D Firth
Multinomial logit bias reduction via Poisson log-linear model
Date: September 2010
Abstract: It is shown how to obtain the bias-reducing penalized maximum likelihood estimator for the parameters of a multinomial logistic regression by using the equivalent Poisson log-linear model. This allows a simple and computationally efficient implementation of the reduced-bias estimator, using standard software for generalized linear models.
Keywords: Jeffreys prior; leverage; logistic linear regression; Poisson trick.