Aggregation of explanatory factor levels in a binomial logit model: Generalization to the multifactorial unsaturated case

We discuss a situation in which, once a logit model is fitted to the data in a contingency table, some factor levels are grouped. Generally, researchers reapply a logit model on the pooled data, however, this approach leads to the violation of the original distributional assumption, when the probabilities of success of the random variables of aggregation differ. In this paper we suggest an alternative procedure that operates under the unsaturated, multifactorial, binomial, logit model. Based on asymptotic theory and taking advantage of the decrease in the variance when the correct distributional assumption is made, the suggested procedure significantly improves the estimates, reduces the standard error, produces lower residuals and is less likely to reject the goodness of fit test on the model. We present the necessary theory, the results of an extensive simulation designed for this purpose, and the suggested procedure contrasted with the usual approach, through a complete numerical example.