The effect of omission of variables on choice model transferability

The transferability of disaggregate choice models has received considerable attention in recent years. Past studies observe varying levels of model transferability but do not characterize the conditions which lead to high or low levels of transferability. Differences in transferability appear to arise due to a variety of differences in the characteristics of the estimation and application contexts and the specification of the model used. This paper reports an investigation into the effect of omission of relevant explanatory variables on the level of transfer effectiveness. The relationship between variable omission and its impact on estimation and prediction is analytically formulated. An empirical analysis of transferability among three sectors within an urban area is undertaken to verify and clarify the analytical results. The analytic and empirical investigations indicate that inclusion of relevant variables leads to an absolute improvement in transfer effectiveness which is approximately proportional to the improvement in goodness of fit which is obtained by the corresponding specification improvement in the application environment.