Two-stage segmentation and finite mixture models for dichotomous responses

Abstract

Market research is essential to assist a company in taking the right decisions and guarantee future progress. The crucial goal of a company is to identify and access the appropriate customers for a particular product or service and recognize the best approach to target these products/services to satisfy the needs of customers. Conjoint analysis measures consumer preferences in market research and has been widely used for decades, proving to be an efficient tool especially in market research. A heterogeneous population of customers in market segmentation is defined as a collection of homogeneous subgroups, where every cluster is made up of customers with similar needs and views of how to worth a product. Several segmentation procedures have been proposed in research to identify these clusters and determine the attributes that influence the respondents’ choices in the marketplace.

This study analyzes customer preferences of hand watches described by three attributes, which include four brands, four prices and two types of watches. A full profile method and full factorial design were used, resulting in thirty-two profiles (incentives) of hand watches. These profiles had distinct attribute level combinations and respondents were asked if they would purchase these items, yielding dichotomous responses. Two segmentation methods were used to relate the respondents’ choices to the attributes, which include the two-stage segmentation and latent class analysis.

The Two-Stage Segmentation procedure involves the estimation of the individual-level parameters, in the first stage, by using binomial logistic regression models, while in the second stage the respondents are clustered into segments based on the similarity of their estimated parameters. This can be carried out using the Two-Step clustering procedure or some other hierarchical and non-hierarchical non-overlapping clustering algorithms.

In latent class analysis, the EM algorithm is used to maximize the expected complete log-likelihood function. The EM algorithm augments the incomplete log-likelihood function by incorporating missing values that represents the unknown cluster membership of each respondent. The posterior probabilities, which are the estimated missing values, are used to allocate respondents in their respective segments. The advantage that Latent class analysis over Two-stage segmentation procedures is that the estimation of regression parameters and segment membership are carried out simultaneously.