A comparative study of the traditional two-stage segmentation procedure and latent class models in respondents' preference for bonds

Abstract

In recent years, market research has become an essential tool in a company' s decision making process. One of the goals of market research is to explain and predict customer preference judgments and choice behaviour. An effective way of conducting a market research is through conjoint analysis whose objective is to determine the attributes that influence respondents' choices in the marketplace. In such a study respondents are asked to rate a set of potential products or services having different attribute levels. These evaluations can be used to create market segmentation models. This research study analyzes customer preferences for investment bonds described by four attributes; coupon rate, currency, redemption term and price. Two approaches are employed to analyse the respondents' rating scores. The first approach is the traditional Two-Stage Segmentation Procedure and the second approach is Latent Class Analysis. The Two-Stage Segmentation Procedure involves the estimation of individual level parameter using a least squares regression of respondents' preference ratings. In the second stage, respondents are clustered into segments based on the similarity of their estimated parameters. Several non-overlapping clustering techniques, including both hierarchical and non-hierarchical methods are contrasted to establish the best clustering procedure. In Latent Class Analysis prediction models are derived by maximizing the expected loglikelihood function, given that the number of clusters is known. This EM algorithm yields posterior probabilities that are used to allocate respondents to segments. The key advantage of Latent Class Analysis over the Two-Stage Segmentation Procedure is that segmentation and prediction are conducted simultaneously. A validation of the two models is provided using choice data. Four bond profiles are presented to the respondents, which they have to assess by choosing the item they prefer most. A Latent Class Multinomial Logit model is fitted to the data to predict the number of preferences for each profile. By classifying the respondents by gender, marital status and age, the predicted frequencies are then compared with the actual frequencies. Empirical evidence suggests that Latent Class Models are appropriate in predicting consumer choice behaviour.