Format

Send to

Choose Destination
Pharmacoepidemiol Drug Saf. 2019 Jul;28(7):934-941. doi: 10.1002/pds.4767. Epub 2019 May 27.

A tool for empirical equipoise assessment in multigroup comparative effectiveness research.

Author information

1
Division of Rheumatology, Immunology and Allergy, Department of Medicine, Brigham and Women's Hospital and Harvard Medical School, Boston, Massachusetts, USA.
2
Department of Epidemiology, Harvard T.H. Chan School of Public Health, Boston, Massachusetts, USA.
3
Department of Biostatistics, Harvard T.H. Chan School of Public Health, Boston, Massachusetts, USA.
4
Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women's Hospital and Harvard Medical School, Boston, Massachusetts, USA.

Abstract

PURPOSE:

In observational research, equipoise concerns whether groups being compared are similar enough for valid inference. Empirical equipoise was previously proposed as a tool to assess patient similarity based on propensity scores (PS). We extended this work for multigroup observational studies.

METHODS:

We modified the tool to allow for multinomial exposures such that the proposed definition reduces to the original when there are only two groups. We illustrated how the tool can be used as a method to assess study design within three-group clinical examples. We then conducted three-group simulations to assess how the tool performed in a setting with residual confounding after PS weighting.

RESULTS:

In a clinical example based on rheumatoid arthritis, 44.5% of the sample fell within the region of empirical equipoise when considering first-line biologics, whereas 57.7% did so for second-line biologics, consistent with the expectation that a second-line design results in better equipoise. In a simulation where the unmeasured confounder had the same magnitude of association with the treatment as the measured confounders and a 25% greater association with the outcome, the tool crossed the proposed threshold for empirical equipoise at a residual confounding of 20% on the ratio scale. When the unmeasured variable had a twice larger association with treatment, the tool became less sensitive and crossed the threshold at a residual confounding of 30%.

CONCLUSION:

Our proposed tool may be useful in guiding cohort identification in multigroup observational studies, particularly with similar effects of unmeasured and measured covariates on treatment and outcome.

KEYWORDS:

multigroup comparative effectiveness; multinomial exposure; pharmacoepidemiology; propensity score

Supplemental Content

Full text links

Icon for Wiley Icon for PubMed Central
Loading ...
Support Center