Quantifying relationships between ordinal categorical variables: Application to metrics tracked by satisfaction barometers

Abstract

Perceptions, attitudes, intentions, etc., are routinely studied for market segmentation, tracking, and predicting consumer reactions. These cognitive, affective, or behavioral components are most often measured by responses to multiple-choice questions: modalities that take the form of a limited number of ordered classes with anchor points, but not always all semantically differentiated. Statistical analysis of such qualitative, ordinal categorical data usually leads us to code them numerically and consider them quantitative. In doing so, we endow them with properties that they do not possess, which entails a risk of bias. To handle them correctly, we propose an econometric parametrization of their relationships, strictly respecting their information content. Combining cumulative and monotonic multinomial regression with random resampling techniques, our methodology allows us to rigorously test whether ordinal categorical feedback metrics can be assimilated to interval measures. We implement this approach conclusively on two large samples derived from the marketing barometers of two competing telecom operators. Thereby, we show to what extent their customers’ willingness to recommend them is explained by their satisfaction with the products and services these two players offer. As a byproduct, this double case study invalidates the very popular Net Promoter Score (NPS), which is supposed to synthesize distributions of degrees to which consumers are likely to participate in the promotion of their favorite brands by word-of-mouth.