CFE 2019: Start Registration
View Submission - CMStatistics
Title: Simulation-based sensitivity analysis for interference in observational studies with unmeasured links Authors:  Fabrizia Mealli - University of Florence (Italy) [presenting]
Alessandra Mattei - University of Florence (Italy)
Laura Forastiere - Yale University (United States)
Abstract: In causal studies where the commonly invoked "no-interference'' assumption is arguable, ignoring interference may lead to very misleading inferences. In observational studies, information on links between units is usually unavailable and interference cannot be taken into account. In a way, the neighborhood treatment can be seen as an unmeasured confounder. We propose to face this issue by developing a Bayesian simulation-based sensitivity analysis to the violation of the no-interference assumption, where we repeatedly i) draw a set of sensitivity parameters from a prior distribution, ii) simulate potential confounders, and iii) reestimate the posterior distribution of the effect of interest after adjusting for the simulated confounders. We propose a model to generate the unmeasured links, which carries our belief on the level of interference and on the level of association between the individual and the neighborhood treatments. If we assume interference to operate only through a function of the vector of neighbors treatments, after a network is drawn we can compute such function and estimate the direct effect of the treatment taking interference into account. Different functions can be used. This approach has the additional advantage of adjusting for neighborhood and network covariates