Mitigating bias in observational vaccine effectiveness studies using simulated comparator populations: Application to...

Measuring vaccine effectiveness (VE) relies on the use of observational study designs. However, achieving robust estimates of direct and indirect VE is frequently compromised by bias, particularly when using syndromic diagnoses of low-specificity.

In order to mitigate confounding between the measured outcome and vaccine uptake, we developed a method to balance comparator populations using individual-level propensity scoring derived from the vaccine-exposed population, and applied it to the unexposed comparator population. Indirect VE was estimated by comparing the unvaccinated vaccine-exposed group with a propensity score-simulated unvaccinated, unexposed group. Direct VE was derived by removing indirect VE from the overall VE. We applied this method to an evaluation of the effectiveness of infant rotavirus vaccination in the UK. Using a general practice cohort of 45,259 live births between May 2010 and December 2015, we calculated indirect and direct VE against consultations for acute gastroenteritis using conventional and vaccination-propensity adjustment comparator populations.

The overall VE during the rotavirus-season (January-May) calculated using mixed-effects Cox regression was 30% [95% confidence intervals (95% CI: 25,35%)]. Use of conventional comparator populations resulted in implausible VE estimates -14% (95% CI: -41,7%) for direct and 29% (95% CI: 14,42%) for indirect effects. Applying our alternative method, direct VE was 26% (95% CI: 1,45%) and indirect VE was 8% (95% CI: -19,29%).

Estimating VE using propensity score simulated comparator populations, particularly for studies using routine health data with syndromic, low-specificity endpoints will aid accurate measurement of the broader public health impact of a vaccine programme.