Analysing ordinal categorical data using the proportional odds model and latent class models

Abstract

Market research is essential for a company to take correct decisions and guarantee future progress. The most important aim of any company is to identify and target different consumers' choices and preferences on a particular product or service. Conjoint analysis has been used extensively for several decades and proved to be an efficient and effective tool, particularly in market research. Through conjoint analysis, it is possible to determine those attributes that truly influence the respondents' choices in the marketplace. One of the application fields of conjoint analysis is market segmentation which addresses consumer preference heterogeneity. In market segmentation a heterogeneous population of customers is represented as a collection of homogeneous subgroups where the customers in each cluster have similar needs and similar views of how to worth a product. To illustrate this methodology, a conjoint analysis is carried out to investigate consumer preferences for different mobile phones having different attributes. The four attributes selected include the brand and price of the mobile phone and whether it has internet access and touchscreen facilities. Using a full factorial design, a number of profiles were generated and the respondents were asked to rate these profiles. Two models were fitted to relate the rating responses to the attributes. The first is the Proportional Odds model, which is appropriate for analyzing ordinal categorical responses and the second model is the Latent Class model which is appropriate for market segmentation. The Proportional Odds models can be thought of as an extension of Logistic regression for dichotomous responses allowing for more than two ordered response categories. These models are mostly used for analyzing rating scores derived from a Likert scale. Estimation is carried through maximum likelihood since ordinary least square estimation does not provide consistent estimator. This model could be fitted either using a logit or a probit link function and the predicted responses are cumulative probabilities. The Latent Class model was fitted to address the heterogeneity in the data by clustering the respondents. The segmentation procedure is carried out using the EM algorithm and allows for a probabilistic classification of respondents into segments and simultaneous estimation of a regression model within each segment. These regression models include both item attributes (brand, price of mobile phones) and demographic variables (gender, occupation, age, locality) in the linear predictor. Latent Class models are appropriate in predicting consumer choices and simultaneously identifying groups of respondents with similar needs and preference which can then be easily targeted.