Abstract:
|
Causal inference using propensity score matching (PSM) relies on several assumptions including positivity, the stable unit treatment value assumption, and conditional exchangeability. When data are collected over geographic space, the presence of spatial patterning in unobserved confounders may be possible, violating the conditional exchangeability assumption. Ignoring this spatial patterning and this source of confounding can easily lead to biased estimates of causal effects. We proposed and investigated a Bayesian framework that includes accounting for space at two levels of the PSM analysis using the neighborhood adjacency matrix of areal data to potentially recover the unconfoundedness assumption. We compared the model to models that 1) did not account for space and 2) accounted for space at only one level of the PSM. We found that including space in at least one stage of the analysis leads to smaller bias and improved coverage probabilities than ignoring space. We applied the proposed framework to study the impact of PM2.5 on premature mortality from the County Health Rankings dataset.
|