The Effect of Overdispersion on Regression Based Decision with Application to Churn Analysis on Indonesian Mobile Phone Industry

Abstract

Extra binomial variation or commonly known as overdispersion in logistic regression will provide incorrect conclusions. Overdispersion may be caused by the occurrence of variation in the response probabilities or correlation within the response variable. On the other hand, independent assumption of response variable is required in the logistic regression. In the case of correlated outcomes, although maximum likelihood gives unbiased estimates, their standard errors are underestimate. This study was aimed at showing the effect of overdispersion on the hypothesis test of logistic regression. The example was taken from telecommunication industry to analyze churn of the subscribers. A simple method proposed by William was used to correct the effect of overdispersion by taking inflation factor into consideration. The result showed that the William method adjusted the standard error of estimates and provided more precise conclusion which was important in marketing strategies.