A Monte Carlo investigation of the effects of spatial heterogeneity of preferences for discrete choice models

There are reasons researchers may be interested in accounting for spatial heterogeneity of preferences, including avoiding model misspecification and the resulting bias, and deriving spatial maps of willingness-to-pay (WTP), which are relevant for policy-making and environmental management. We employ a Monte Carlo simulation of three econometric approaches to parametrically account for spatial auto-correlation in discrete choice models. The first is based on the analysis of individual-specific estimates of the mixed logit model. The second extends this model to explicitly account for spatial correlation, instead of simply conditioning individual-specific estimates on population-level distributions and individuals’ choices. The third is the geographically weighted multinomial logit model, which incorporates spatial dimensions using geographical weights to estimate location-specific choice models. We analyze the performance of these methods in recovering population-, region- and individual-level preference parameter estimates and implied WTP in the case of spatial autocorrelation. We find that, although ignoring spatial autocorrelation did not significantly bias population-level results of the simple mixed logit model, neither individual-specific estimates nor the geographically weighted multinomial logit model was able to reliably recover the true region- and individual-specific parameters. We show that the spatially-autocorrelated mixed logit proposed in this study is promising and outline possibilities for future development.