2C-03 - Use of Proximal Causal Inference to Address Unmeasured Confounding in Target Trial Emulation: Empirical Example on Influenza Vaccination and Stroke

Background: High-quality observational studies suggested protective effect of influenza (FLU) vaccine against stroke, although not conclusive enough due to vulnerability to unmeasured confounding (UC). The recently proposed proximal causal inference (PCI) method may mitigate UC by leveraging a pair of negative control exposure (NCE) and negative control outcome (NCO), based on stringent assumptions.

Objectives: To assess the performance of PCI on mitigating UC, hypothesizing the protection of FLU vaccine on stroke among older adults as an empirical example.

Methods: This study applied sequentially designed target trial emulation with the Yinzhou Regional Health Care Database, China. Participants were categorized as FLU vaccinees or non-vaccinees based on their vaccination regimen during the 1-week enrollment period of each sequential trial. Propensity score (PS) matching was performed by 6 different sets (scenarios, S) of variables, i.e., incrementally adding variables from key demographic (S1), lifestyle & lab test (S2), comorbidity (S3), co-medication (S4), healthcare utilization (S5) to frailty indicator (full covariate, S6), representing varying levels of UC. On the PS-matched cohorts of each scenario: ① Conventional univariate Poisson regression was performed; ② PCI with injury as NCO and otolaryngology visit as NCE was applied via 2-stage Poisson regression. The incidence rate ratio (IRR) of FLU vaccine on stroke was compared between conventional regression and PCI method. Multiple sensitivity analyses were conducted with regard to the assumptions of PCI, including replacing the NCO with vitamin deficiency, altering the NCE to gastroesophageal reflux visit, and adding extra NCO with vitamin deficiency.

Results: A total of 170,194 FLU vaccinees and 8,017,735 non-vaccinees from 50 emulated trials were included in the analysis. Conventional Poisson regression showed protective effect (IRR: 0.95 [95% CI 0.93-0.98]) of FLU vaccine under full covariate (S6), but the UC moved the effect to null in scenarios excluding various covariates (S1-S4), with IRR ranging from 1.00 [0.97-1.03] to 1.03 [1.00-1.07]. In contrast, PCI analyses consistently showed FLU vaccination was associated with a reduced risk of stroke across all of the covariate scenarios (S1-S6), with IRR ranging from 0.85 [0.83-0.88] to 0.91 [0.88-0.94]. When negative control was substituted or added in sensitivity analyses, the PCI results were in line with the main analyses.

Conclusions: These findings corroborated prior evidence on the potential protective effect of influenza vaccine against stroke. Compared with PS matching, PCI can further mitigate UC in vaccine effectiveness study.