Abstract
Background/Aims Observational data often has treatment exposure confounded with baseline covariates. Doubly robust estimation utilizes both a regression model and an additional model for the exposure, often the propensity score, to estimate the causal effect of an exposure on an outcome.
Methods In a study designed to evaluate change in hemoglobin (Hb) with use of angiotensin-converting enzyme inhibitors (ACEI) or angiotensin receptor blockers (ARB), in a primary care patient population, we found that receiving ARB vs. ACEI was associated with several of the baseline covariates. These baseline covariates, Hb, treatment start date, chronic kidney disease, congestive heart failure, diabetes mellitus, hypertension, sex, and age, initially thought to be informative in estimating follow-up (F/U) Hb, were chosen a priori. No medical explanation for the differential ordering of ACEI and ARB based on the level of these covariates was identified, so the treatment effect of ARB relative to ACEI is considered to be truly confounded. We therefore adopted the doubly robust semiparametric efficient estimator of Robins et al. (1994) in a causal analysis of the treatment effect of ARB vs. ACEI on F/U Hgb. The method produces an estimate of the treatment effect by simultaneously incorporating the propensity of a subject to receive ACEI or ARB, given their levels of covariates, and the effects of the covariates upon the response of interest, F/U Hb. It is doubly robust in the sense that it produces an unbiased estimate of the treatment effect if either the outcome or propensity model is correct. A complete-case ANCOVA was conducted to estimate the treatment effect.
Results We found the estimated F/U Hb and bootstrap bias-corrected accelerated (BCa) 95% confidence interval (CI) of ACEI and ARB to be 14.31 (14.21, 14.42) gm/dL and 14.48 (14.33, 14.62) gm/dL, respectively. The causal effect of ARB relative to ACEI and associated BCa CI is estimated to be 0.17 (0.00, 0.31) gm/dL (p = 0.0310).
Discussion The use of doubly robust estimation documented a significant difference between the effects of ACEI and ARB on F/U Hb, despite the association of several of the baseline covariates with the differential ordering of these drugs.
