Household Parameter Estimation by dividing observations in segments

Companies collect a lot of panel data to get to know their customers better in order to improve marketing strategies. Companies want to know what kind of customers they need to target to optimize the response. Binary response models, like the Binary Logit model, in a nonlinear setting are normally used to target these questions. One major problem that is not captured by the Binary Logit model is the assumption that consumers differ in their reaction (unobserved heterogeneity). This becomes a problem when the data per household is limited. In literature, a lot of solutions are proposed to deal with this problem. These solutions are based on creating segments. In this paper, a new technique is proposed by estimating the parameters of a Binary Logit model with two segments; the Segment Binary Logit model. Households are optimally divided between these two segment based on their log-likelihood. This is done by actively searching for the optimal division of households. In this paper we analyzed two models with different sets of explanatory variables. Application Study show that the Segment Binary Logit model improves the model fit and the forecasting performance in both cases.