Behavioral inference from non-stationary policies: Theory and application to ridehailing drivers during COVID-19 lockdowns

In the aftermath of a disruptive event like the onset of the COVID-19 pandemic, it is important for policymakers to quickly understand how people are changing their behavior and their goals in response to the event. Choice modeling is often applied to infer the relationship between preference and behavior, but it assumes that the underlying relationship is stationary: that decisions are drawn from the same model over time. However, when observed decisions outcomes are non-stationary in time because, for example, the agent is changing their behavioral policy over time, existing methods fail to recognize the intent behind these changes. To this end, we introduce a non-parametric sequentially-valid online statistical hypothesis test to identify entities in the urban environment that ride-sourcing drivers increasingly sought out or avoided over the initial months of the COVID-19 pandemic. We recover concrete and intuitive behavioral patterns across drivers to demonstrate that this procedure can be used to detect behavioral trends as they are emerging.