The effect of sampling of alternatives on MNL models: An empirical analysis in the context of shopping-destination choice models

Destination choice, especially shopping-destination choice always has a very large number of alternatives. Estimating such model from the entire population can be very expensive or impossible. By using the real survey data from Dallas-Forth Worth (DFW) area, this project estimates the same empirical shopping-destination choice using different sampling levels and examines the differences in parameter estimates, model fit, etc. A simple (only four explanatory variables) empirical specification with total travel time (TotTime = IVTT+OVTT) to location, retail employment at destination (RetEmp), origin destination adjacent indicator (Adjacent), destination zone high retail employment indicator (HighRet) is used to examine the effect of the sample size on model performance for an MNL model. The final results show a sample size of 50 always generate pretty good estimations for the continuous variables (TotTime and RetEmp), however for the binary variables (Adjacent and HighRet), there is no a fixed correctly sampling size existed. It is encouraging that trail and error process need to be applied to get satisfying results.