🤖 AI Summary
This study addresses the limitations of traditional control-based causal inference methods—such as matching and difference-in-differences—in settings characterized by pervasive or structurally ambiguous spillover effects, where reliance on uncontaminated control units impedes accurate identification of both average direct and spillover effects. Within the potential outcomes framework, this work provides the first systematic comparison between control-based and prediction-based counterfactual approaches—including interrupted time series and machine learning control—in terms of their identification capabilities. Through simulation and empirical analyses, the authors demonstrate that in environments with widespread interference, prediction-based methods can more reliably estimate certain causal parameters over short horizons, circumventing the stringent assumption of unperturbed units and thereby offering a promising alternative for causal inference under complex interference.
📝 Abstract
Spillovers and interference pose fundamental challenges for causal inference, as treatment assigned to one unit may affect the outcome of others, violating the no-interference assumption underlying most empirical strategies. Existing approaches, based on partial interference, exposure mapping, spatial, network, or structural frameworks, typically rely on strong assumptions about interaction structures or require the existence of uncontaminated control units to estimate relevant causal parameters. We revisit this identification challenge within the potential outcomes framework and compare the conditions under which causal effects can be identified using two broad classes of counterfactual methods: control-based counterfactual methods (CBCMs), such as matching and difference-in-differences designs, and forecast-based counterfactual methods (FBCMs), including interrupted time-series and machine learning control methods. We show under which circumstances CBCMs and FBCMs identify average direct and spillover effects. Through simulations and an empirical application, we illustrate the main advantages and limitations of each approach. We show that, in the presence of pervasive or ill-defined spillover effects, CBCMs either cannot be used or entail severe identification concerns, whereas FBCMs can more credibly identify some of the causal parameters of interest, at least in the short term.