Average direct and indirect causal effects under interference

Summary We propose a definition for the average indirect effect of binary treatment in potential outcomes model causal inference under cross-unit interference. Our is analogous to standard direct and can be expressed without needing compare across multiple randomized experiments. show that proposed satisfies decomposition theorem stating Bernoulli trial, sum effects always corresponds policy intervention infinitesimally increases probabilities. also consider number parametric models interference find our nonparametric remains natural estimand when re-expressed context these models.