Spatial Transferability of Person-Level Daily Activity Generation and Time Use Models

An empirical assessment is made of the spatial transferability of person-level daily activity generation and time use models in regions in Florida and between Florida and California. The empirical models are for unemployed adults and are based on the multiple discrete-continuous extreme value structure. The prediction properties of the model are examined first. The results shed new light on the prediction properties of the multiple discrete-continuous extreme value model that have implications for transferability and that provide insight into how the model structure could be improved. Two approaches to transferring models are evaluated—naïve transfer and updating model constants—with measures such as log likelihood–based metrics, aggregate predictive ability, and model sensitivity to changes in demographic characteristics. Results suggest that accurate prediction of aggregate observed patterns is not an adequate yardstick with which to assess transferability; emphasis should be placed on model sensitivity to changes in explanatory variables. Updating of constants improves a transferred model's aggregate prediction ability, but not necessarily its policy sensitivity. The extent of transferability between regions within a state is greater than that across states. Within Florida, there is greater transferability between urban regions (especially between Southeast Florida and Central Florida) than between urban and rural regions.