Combining Design and Analysis for Causal Inference under Network Interference

In settings from public health to advertising, practitioners use randomized experiments to estimate the causal effect of a population-wide rollout of a new treatment. Traditional estimators that account for interference (where the treatment of one individual can affect the outcome of another) often have prohibitively high variance. Recent literature provides two, largely disjoint, techniques to address this: leveraging parametric assumptions on the potential outcomes to design better estimators and leveraging structural assumptions on the interference to choose a smarter experimental design. Combining these approaches, we present a pseudoinverse estimator for the total treatment effect in low-order outcome models when the data are collected under general experimental designs. For cluster randomized designs, we show this estimator is unbiased with variance scaling like the smaller of the variance obtained from a low-order assumption and the variance obtained from cluster randomization.

May 26, 2025 2:00 PM

Statistical Society of Canada Annual Meeting (2025)

University of Saskatchewan

107 Wiggins Rd, Saskatoon, SK S7N 2Z4