Designing Optimal Social-Health Interventions

EDO AIROLDI – TEMPLE UNIVERSITY

ABSTRACT

Designing experiments that can estimate and disentangle causal and non-causal effects when the units of analysis are connected through a social network is the primary interest, and a challenge, in many modern endeavors at the nexus of science, technology and society. Consider, for instance, (a) HIV testing and awareness campaigns on mobile phones, (b) improving healthcare in rural populations using social interventions, and (c) promoting standard of care practices among US oncologists on a dedicated social media platform. A salient feature of these applications is the multifaceted social network that connects the units of analysis, which translates into the presence of interference (i.e., outcomes are a function of treatment given to multiple units) and homophily (i.e., correlation between outcomes is a function of covariates). In this talk, building on prior work, we introduce and discuss strategies for experimental design centered around a useful role for statistical models. In particular, we wish for certain finite-sample properties of the estimator to hold even if the model completely fails, while we would like to gain efficiency to the extent that certain aspects of the model are correctly specified. We will then overview recent results in this space, including estimation of network effects in the presence of homophily, optimizing randomized saturation designs leveraging graphons, and an application to a large randomized experiment in Honduras.