Understanding causal effects of ride-sourcing subsidy: A novel generative adversarial networks approach

Ride-sourcing platforms offer subsidies for drivers to ensure stable supply capacity for on-demand ride services. These subsidies guarantee minimum surges in advance to help drivers plan to drive, thus more efficiently coping with the spatial-temporal imbalance between supply and demand than surge wage. Understanding the causal effects of ride-sourcing subsidy is a prerequisite for precision targeting when deploying driver promotion activities. We propose a novel generative adversarial networks based model GANIRC to estimate the individual treatment effect (ITE) of ride-sourcing subsidy on drivers' online time. It captures counterfactual distributions via adversarial learning and generates individual response curves (IRCs) in the form of neural networks. The generator acts as a metafunction of responses, and the analytical constraints of IRCs are relaxed. Wasserstein distance is adopted to measure the closeness between counterfactual and factual distributions, and improve the stability of the minimax game between the generator and discriminator. In addition to the theoretical proof that our model approximately minimizes the upper bound on the ITE estimation error, the model performance is verified by leveraging ride-sourcing subsidy activity data in four cities with varying scales. Results show that our proposed model can generate accurate counterfactual distributions based on limited observational data. In the case study of the four cities, the drivers with neoclassical behavior, who increase working time with income, dominate in the ride-sourcing market, and inactive drivers respond more positively to subsidies than active drivers. Very few drivers in large cities are identified to have income-targeting behavior, who have income targets and stop working when reaching these targets. Our study reconciles the controversy of whether neoclassical theory or income-targeting theory is appropriate for ride-sourcing drivers, and has practical implications for understanding behavior and optimizing ride-sourcing operations.