In customer value analysis (CVA), a company conducts sample surveys of its customers and of its competitors' customers to determine the relative performance of the company on many attributes ranging from product quality and technology to pricing and sales support. The data discussed in this paper are from a quarterly survey run at Lucent Technologies.We have built a Bayesian model for the data that is partly hierarchical and has a time series component. By ``model'' we mean the full specification of information that allows the computation of posterior distributions of the data, sharp specifications such as independent errors with normal distributions and diffuse specifications such as probability distributions on parameters arising from sharp specifications. The model includes the following:
- random respondent means modeled by a t-distribution;
- random respondent variances modeled by an exponential distribution;
- company-attribute means through time modeled by integrated moving average gaussian processes;
- variation about the company-attribute and respondent means modeled by a compound normal distribution where the normal variance is proportional to the respondent variance and thus is mixed by an exponential.
The model is the first full description of CVA data; it provides both a characterization of the performance of the specific companies in the survey as well as a mechanism for studying some of the basic notions of CVA theory.
To analyze the CVA data we had to develop some new ideas about various methodological issues and about approaches to data analysis:
- a Bayesian theory of data exploration that provides an overall guide for methods used to explore data for the purpose of making decisions about sharp specifications in a model;
- computational methods for integrated moving-average time series models;
- Modeling data with random locations and scales using compound normal distributions;
- a posteriori modeling via regression and factor analysis.
Our data exploration, our model checking, and our study of posterior distributions make heavy use of trellis display, a relatively new framework for multivariable data display.