Drug target Mendelian Randomisation vs propensity score weighting for causal inference: estimating vaccine-induced antibody effect on COVID-19 outcomes

Background: Causal inference using observational data is prone to confounding and bias. Mendelian randomisation (MR), using genetic instruments as “randomised” exposures, emerges as a novel method to establish causality. However, its application in pharmacoepidemiology remains limited.

Objectives: In this study, we performed a head-to-head comparison of propensity score weighting vs MR analysis to determine the effects of antibody positivity to protect against COVID-19 outcomes within 6 months after COVID-19 vaccination.

Methods: We included participants from UK Biobank who received 1+ doses of COVID-19 vaccines and then enrolled in the Self-Test Antibody study. Binary SARS-CoV-2 antibodies response (positive or negative) were tested using two validated lateral flow arrays. Individuals’ genetic information was measured, and comorbidity, medications, and COVID-19 outcomes were obtained via linkage to electronic health records. In the cohort study, we followed participants from their antibody test date (index date) until the earliest of: outcome (COVID-19 infection or hospitalization), death, 6 months after index date or the end of study (30th November 2021). We used propensity score weighting and Cox proportional hazards to estimate the protective effect of antibody positivity. In the MR study, we, based on prior clinical knowledge, selected HLA gene variations as instruments, and used MR analytic approach to derive the estimand. For both studies, we excluded those previously infected or vaccinated < 1 week or >12 weeks before antibody testing.

Results: Of 142,502 eligible individuals, 57,081 (40%) were antibody positive and 85,421 (60%) were negative following vaccination. Average ages were 70.04 (SD: 70.04) and 68.08 (7.45) years respectively. The positive group had more two-dose vaccine recipients and shorter time since the latest dose compared to the negative. After propensity score weighting, all covariates were balanced (standardised mean difference < 0.1). Over 6 months, there were 4,159 COVID-19 infections and 261 hospitalisations. The estimated hazard ratio (HR) for antibody positivity was 0.63 (95% CI 0.58-0.68) for COVID-19 infection and 0.48 (95% CI 0.34-0.68) for hospitalisation.
In the MR analysis, 24 independent HLA alleles were identified as associated with antibody positivity. The estimated causal odds ratio (OR) of positivity was 0.69 (95% CI 0.53-0.89) for COVID-19 infection and 0.52 (95% CI 0.20-1.31) for hospitalisation.

Conclusions: Our study is the first to compare propensity score cohort methods with Mendelian randomisation to estimate the causal effect of antibody positivity on COVID-19 outcomes. The accuracy of resulting estimation was similar, but relatively less precise in MR method.