Validation and forecasts in models estimated from multiday travel survey

Multiday travel surveys (also known as panel data) have recently assumed high relevance in travel behavior analysis and activity-based modeling. Basically two types of multiday data have been collected: cross-sectional data repeated at separate points in time and data gathered over a continuous period of time. So far, the studies using panel data have focused on the estimation of demand models; little is known about the application and validation of models estimated on repeated measurements. Even the definition of the holdout sample is not obvious in panel data sets. This paper studies issues related to model validation and forecasting of continuous data sets. With both simulated and real data, empirical evidence is provided on the effects that different patterns of correlation have on model forecast and policy analysis. Results show that the way holdout samples are extracted affects the validation results and that the best results are obtained when a percentage of individuals with all of their observations are used. The logit model in the presence of taste heterogeneity could produce biased modal shifts, while failing to account for correlation across observations, did not seem to produce relevant effects on policy analysis. The real case study, estimated by using a 6-week travel diary (MobiDrive), confirms only in part the analysis on simulated data. Results also confirm that in panel data, a model with a better fit might provide a worse validation and forecast.