Proposal for an Improvement of the Parameter Estimation Method in Multinomial Logit Model Using Weight of Evidence

抄録

Modeling multiple-choice behavior is an important research topic in the field of social sciences, including consumer behavior research. A model that has been studied for a long time is the multinomial logit model; however, it has been shown to have problems, such as the need to set different parameters for each choice used as the objective variable. Furthermore, the number of parameters increases exponentially as the number of choices considered in the model increases. This results in unstable parameter estimates and makes variable selection difficult. In particular, when using discrete variables consisting of many categories as explanatory variables, it is necessary to set dummy variables for each category, and the effect of multicollinearity may further destabilize parameter estimates. To address these issues, we focused on cases where explanatory variables are composed of discrete variables and converted the categories contained in the discrete variables into numerical variables using the Weight of Evidence. Furthermore, we attempted to reduce the number of parameters in the model. In addition, by including only variables having a strong causal relationship with choice behavior based on the Information Value and correlation coefficients to the variable list, we proposed improvements to the model to simplify variable selection in the multinomial logit model. We then report the results of applying conventional methods and the method proposed in this study to real data, demonstrating the effectiveness of the proposed method.