Abstract:
|
Full matching (Hansen 2004) in observational studies benefits from its ability to use more subjects than pair matching and to achieve closer matching than conventional approaches to achieving covariate balance. Nevertheless, where subjects are naturally clustered (Gurm 2013), whether to match within or across clusters reflects the inherent tradeoff between finding close matches based on subject-level factors and minimizing confounding by cluster. In addition, when the exposure of interest varies little within cluster but largely across clusters, matching to achieve covariate balance within cluster becomes impossible. Inability to match within cluster opens the potential for different within-cluster and across-cluster estimates of exposure-outcome association. We present three applications that push the limits of full matching because of (1) variation in exposure prevalence across clusters, (2) modest cluster sizes, and (3) rare outcomes. Full matching, relies on key assumptions about confounding by cluster, and its use in clustered data problems requires careful specification of the causal question and close attention to the assumptions of matching across clusters.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.