Efficient Nonparametric Estimation of Stochastic Policy Effects with Clustered Interference.

Area of Science:

• Causal inference
• Epidemiology
• Biostatistics

Background:

• Interference, where one unit's treatment affects another's outcome, is complex in clustered settings.
• Existing causal estimands for clustered interference often lack real-world applicability or rely on restrictive parametric models.
• Quantifying treatment effects under clustered interference from observational data remains a significant challenge.

Purpose of the Study:

• To propose novel causal estimands for clustered interference based on propensity score distribution modifications.
• To develop nonparametric estimators for these new causal estimands, avoiding parametric assumptions.
• To evaluate the real-world relevance and statistical performance of the proposed methods.

Main Methods:

• Development of new causal estimands by modifying propensity score distributions.
• Construction of nonparametric sample splitting estimators for enhanced flexibility and data-adaptability.
• Consistency, asymptotic normality, and efficiency analysis of the proposed estimators, achieving parametric convergence rates.

Main Results:

• The proposed nonparametric estimators demonstrate good finite sample performance in simulations.
• New causal estimands offer potentially greater real-world relevance compared to existing methods.
• The methods are validated through application to water, sanitation, and hygiene (WASH) interventions and child diarrhea in Senegal.

Conclusions:

• The study presents a novel framework for causal inference under clustered interference using propensity score modifications.
• Nonparametric estimation provides a flexible and statistically efficient approach for analyzing observational data with clustered interference.
• The findings have implications for public health research, particularly in evaluating interventions like WASH facilities.