William L. Moore (1982), “Predictive power of joint
space models constructed with composition techniques," Journal
of Business Research, 10, 217-236."
There has been considerable interest in the use of composition techniques to build 'joint space' models for predicting consumer preferences. In spite of this interest, Moore was unable to locate studies that examined the predictive ability of these models, that is the ability of the models to predict preferences for new brands. However, he did find two studies that compared the fit of different composition models. (In both cases, factor analytic models outperformed discriminant analytic models.) Moore obtained data on four product categories and compared the ability of various joint space models to predict preferences for brands that were not used to build the spaces. He hypothesized that discriminant models (which build spaces that maximize the among brand variance in perception and minimize the within brand variance) should provide better predictions than factor analytic models (that consider only total variance in perception). Although this strikes me as reasonable, it contradicts the assumptions behind a number of studies. Moore's findings were interesting. First, simpler vector representations of preferences gave better predictions than the more complex and theoretically satisfying 'ideal point' models. He also found that the models constructed with discriminant analyses were more accurate than the joint space models based on factor analysis. The results held up when considering correlation coefficients, as well as mean absolute deviations. (Incidentally, a bit of confusion arose over a typographical error in the second paragraph on his page 229: it should say Table 4 rather than Table 3). Although the results were consistent and statistically significant, the differences were small. Replications of this study would be of great value.