(176) Quantitative Bias Analysis in Comparative Effectiveness Research: A Cautionary Word

Sunday, August 28, 2022

8:00 AM – 1:45 PM CEST

Location: Hall C

Publication Number: 1221

Presenting Author(s)

James Weaver, MS
– Janssen Pharmaceuticals, Inc., University of Oxford, Titusville, New Jersey, United States

Co-Author(s)

Patrick Ryan, BSCE
– Observational Health Data Sciences and Informatics, New York, NY, USA, Titusville, NJ, United States
Victoria Y. Strauss
– Dr, Boehringer Ingelheim Pharma GmbH & Co. KG, Ingelheim am Rhein, Germany
Marc Suchard
– Observational Health Data Sciences and Informatics, New York, NY, USA, United States
Joel N. Swerdel
– Janssen Pharmaceuticals, Inc., Titusville, New Jersey, United States
Daniel Prieto-Alhambra, MD, MSc, PhD, FISPE
– Professor of Pharmaco- and Device Epidemiology, CSM, NDORMS, University of Oxford, University of Oxford, Oxford, United Kingdom

Background: Information bias is acknowledged but rarely corrected in comparative effectiveness research (CER). Quantitative bias analysis (QBA) corrects bias from outcome misclassification using sensitivity, specificity, and incidence proportion, but its performance has been insufficiently assessed with empirical data.

Objectives: We evaluated simple and multidimensional QBA to correct effect estimates subject to outcome misclassification bias in 3 pairwise comparisons of hypertension treatments and their relative risk for ischemic stroke in 4 observational databases.

Methods:

Design: Active comparator new user cohort
Setting: Optum® de-identified Clinformatics® Datamart (Optum), Optum® Electronic Health Record (Optum EHR), IBM MarketScan® Commercial Database (CCAE) and Multi-State Medicaid (MDCD) transformed to a common data model.
Exposures: Angiotensin receptor blockers (ARB), thiazide/thiazide-like diuretics (THZ), and dihydropyridine calcium channel blockers (dCCB) compared to angiotensin-converting enzyme inhibitors (ACE) among patients with hypertension.
Outcome: Inpatient ischemic stroke during 1 and 2 years post-exposure.
Analysis: We estimated outcome sensitivity and specificity with a probabilistic validation study[10.1016/j.jbi.2019.103258] in each database and estimated treatment effects with logistic regression using propensity score (PS) matching without vs with QBA. In QBA, we used outcome sensitivity and specificity estimates non-differential to exposure status to correct observed exposure-outcome 2x2 table counts before fitting outcome models. We then applied multidimensional QBA across a range of sensitivity and specificity values.

Results: Sensitivity ranged from 38.0-58.9% and specificity from 99.0-99.9% across databases. In 1 year follow-up, QBA odds ratio (OR) correction ranged from 0% (1.01 in ACE vs THZ, CCAE) to 31% (1.12 to 1.47 in ACE vs THZ, Optum DOD). Correction magnitude decreased with higher power (2 year follow-up). Multidimensional QBA consistently showed ORs were very sensitive to specificity estimates. For example, in CCAE ACE vs ARB at 49.7% sensitivity and 99.9% specificity, a 10^-3 decrease in specificity invalidated ORs by producing negative contingency table counts.

Conclusions: We applied QBA to a set of rare-event, high-specificity scenarios common in pharmacoepidemiology and found that minimal specificity changes have implausibly large impact on effect estimates. This suggests need for more precise measurement error estimation or alternative methods for measurement error correction in CER.