Methods for Avoidance of Selection Bias

A Word of Caution: All-Cause Mortality and Residual Confounding in Medicare Claims Data 

Garcia-Albeniz X, Hsu J, Hernan MA. 


This is a cautionary tale about the use of administrative databases for comparative effectiveness and safety. 

Claims databases are a phenomenal resource for healthcare research, and methods have evolved accordingly. For avoidance of selection bias, we have methods that allow target trial emulation. We also have methods that allow us to emulate conditional randomization by adjusting for measured confounding. 

In this study, we used a screening colonoscopy and all-cause mortality scenario where we know that the causal effect of the exposure, if any, was small. 

First, we emulated a target trial of 1-time screening colonoscopy in Medicare beneficiaries by synchronizing eligibility, exposure assignment, and the start of follow-up. Then, we adjusted for confounding using 3 structural approaches (multivariate outcome regression, inverse probability weighting, and propensity scores) and 1 non-structural approach (high-dimensional propensity scores).

The 4 methods yielded the same result: a 30% reduction in the relative risk of death due to a single screening colonoscopy, which is not plausible. Therefore, residual confounding can be intractable for the effect of preventive services on all-cause mortality in administrative databases.