JMIR Publications

ABSTRACT

Background:

Real-world COVID-19 vaccine effectiveness (VE) studies are investigating exposures of increasing complexity accounting for time since vaccination, for example. These studies require methods that adjust for confounding that arises when morbidities and demographics are associated with vaccination and risk of outcome events. Methods based on propensity scores (PS) are well-suited to this when the exposure is dichotomous, but present challenges when the exposure is multinomial.

Objective:

This simulation study investigates alternative methods to adjust for confounding in VE studies that have a test-negative design.

Methods:

Adjustment for a disease risk score (DRS) is compared with multivariable logistic regression. The performance of VE estimators is evaluated across a multinomial vaccination exposure in simulated datasets.

Results:

Bias in VE estimates from multivariable models ranged from -5.3% to 6.1% across 4 levels of vaccination. Standard errors of VE estimates were unbiased and 95% coverage probabilities were attained in most scenarios. Bias in VE estimates from DRS-adjusted models was low, ranging from -2.2% to 4.2%. However, the DRS-adjusted models underestimated the standard errors of VE estimates, with coverage ranging from 87.8% to 94.5%.

Conclusions:

Overall, models using a DRS to adjust for confounding performed adequately but not as well as the multivariable models that adjusted for covariates individually.