| description abstract | Finally, we observe that some of the models with the best predictive accuracy have coefficients with unexpected signs—likely biased due to correlation with unobserved attributes. Despite good prediction accuracy in existing markets, where attribute correlations are similar from year to year, these models may misguide design efforts if the designer makes changes that do not follow correlations in the marketplace. For example, the sign of the coefficient for the gallons per mile (gpm) attribute of the best predictive logit model is negative, suggesting that consumers prefer lower fuel economy, all other attributes being equal. In fact, consumers may purchase vehicles with lower fuel economy because of other features of those vehicles unobserved by the modeler (e.g., size, performance, or styling attributes not captured in the model). The model predicts well if the new market retains such correlations, but a designer who lowers fuel economy alone is not likely to obtain the outcome predicted by the model. Thus, accuracy of predictions in existing markets is not a sufficient condition for use in design. | |
