MICHAEL J. ZENOR and RAJENDRA K. SRIVASTAVA* In this paper, the authors introduce a "latent segment logit" (lSl) model that allows the identification of latent market segments when only macro-level time-series data (e.g., market share or sales, not individual choices) are available. The pro- posed model provides a paramorphic representation of market structure, based on the notion that "structure" implies heterogeneity in preferences and/or respanse to marketing mix elements. It assumes that independence of irrelevant alternatives (IIA) holds within latent segments (i.e., segments are homogeneous) but allows for het- erogeneity across segments. Estimates for segment characteristics (including size, brand preferences, and sensitivity to marketing mix variables) are obtained byap- plying the model to aggregated longitudinal panel data. Validation tests are con- ducted on both the aggregated and disaggregated panel data. Aggregate vali- dation demonstrates that the model is superior to standard market share models in terms of calibration and predictive fit. Disaggregated validation demonstrates that the latent segments recovered by the model account for much of the variation across householdpurchase histories, even though these data were not utilized in the estimation. Inferring Market Structure With Aggregate Data: A latent Segment logit Approach INTRODUCTION Approaches to market structure analysis typically identify relationships between products in terms of their appeal to, or purchase/usage by, segments of consumers (Day, Shocker, and Srivastava 1979). Market segmen- tation approaches, on the other hand, seek to group con- sumers in terms of (1) their preferences for, or purchase/ usage of, products; and/or (2) their response to mar- keting mix strategies of brands. Grover and Srinivasan (1987), in simultaneously in- ferring market structure and segmentation by applying latent class analysis to a brand switching matrix, illus- trate how the concepts of market segmentation and com- petitive market structures are inherently inseparable. But, *Michael J. Zenor is Assistant Professor and Rajendra K. Srivas- tava is Sam Barshop Professor of Marketing at the University of Texas- Austin. This research was supported by a grant from the Dean's Fac- ulty Research Fund, Graduate School of Business, University of Texas-Austin. The authors wish to thank A. C. Nielsen for supply- ing the data used herein, and Patrick Brockett, Vijay Mahajan, Wag- ner Kamakura, and the anonymous JMR reviewers for comments and suggestions on earlier versions of the paper. 369 their approach was subject to stationarity assumptions- i.e., market share stability and a zero-order Markov pro- cess-because they did not explicitly model the poten- tial impact of marketing mix variables. These assump- tions were relaxed by Kamakura and Russell (1989), who developed a disaggregate probabilistic choice model with segmentwise brand preferences and price sensitivities. Their model alleviates potential nonstationarity prob- lems, to the extent that shifts in market shares are the result of price variations (however, the effects of other marketing mix variables, such as displays and advertis- ing, can easily be accommodated). While these approaches represent attractive ways to capture market structure, they require individual-level panel (diary or scanner) data. Model estimation can be cumbersome-especially when it is integrated across several scanner markets. Furthermore, such data can be expensive, prone to bias resulting from unrepresentative panels, and often too "thin" or simply not available for less-frequently purchased products. On the other hand, models based on aggregate (market-level) measures, such as sales or market share, lose information related to seg- ment differences. But decisions and problems faced by managers do not change when analyses based on indi- Journal of Marketing Research Vol. XXX (August 1993), 369-79