Implicit Institutional Incentives and Individual Decisions: Causal Inference with Deep Learning Models

This paper proposes a methodology to identify tacit organizational incentives from institutional reactions to operational decisions. Football data provide a laboratory for this analysis as referee decisions and their consequences are publicly observable. Our approach relies on comparing the observed length of time between referee appointments with the one predicted if the referee had made a different decision. Deep Learning is instrumental in such analysis as it allows treatment and confounder variables to interact at different layers, which can be used to construct meaningful counterfactual scenarios. We identify institutional incentives for the referee to take gradual (instead of drastic) decisions to send off home team players and deliver the game's expected outcome. Finally, we assess the robustness of the results to other modelling approaches.