Experimental Evidence on Structural State Dependence in Demand

Abstract

Research in marketing and economics has documented structural state dependence on demand, i.e., a causal effect of past on current brand purchases. Our work re-evaluates the evidence for structural state dependence and assesses the credibility of the extant literature and its identifying assumptions. Using a natural couponing experiment with 223 brands, we estimate the local average treatment effect (LATE) of current brand purchases on future brand choices. A novelty of our work is that we use an encouragement design to identify state dependence, with randomized variation in coupon exposure at the customer level. This design eliminates two threats to identification. First, past and current brand choices are both based on idiosyncratic customer preferences (selection); second, brand switching may result from serially correlated unobserved factors (spurious state dependence). In contrast to past work, our estimator is nonparametric and only requires two assumptions: monotonicity and an exclusion restriction. The exclusion restriction could be violated if, for example, coupons have an advertising effect and directly influence future purchases. To assess the robustness of our analysis to such a potential violation, we exploit the random variation in coupon discount percentages in our data and use the discount variation as an instrument while conditioning on coupon exposure. The corresponding results are consistent with our main analysis. Finally, we estimate structural brand choice models following the recent literature and predict the LATEs implied by the model estimates. We find strong consistency between the results of the two nonparametric methods and the model-based approach: On average, the LATEs have similar magnitudes, and the nonparametric and model-predicted LATEs are significantly correlated. In sum, our findings provide strong evidence for structural state dependence and lend credibility to the structural choice model approach. This credibility is practically important because structural models are designed for policy evaluation, such as predicting the future customer value given different prices and promotions, and for studying switching costs and equilibrium prices.

Keywords: Structural state dependence, brand choice, natural field experiment, coupons, local average treatment effect