Background There are scarce data on best practices to control for confounding in observational studies assessing vaccine effectiveness to prevent COVID-19. We compared the performance of three well-established methods [overlap weighting, inverse probability treatment weighting and propensity score (PS) matching] to minimize confounding when comparing vaccinated and unvaccinated people. Subsequently, we conducted a target trial emulation to study the ability of these methods to replicate COVID-19 vaccine trials. Methods We included all individuals aged ≥75 from primary care records from the UK [Clinical Practice Research Datalink (CPRD) AURUM], who were not infected with or vaccinated against SARS-CoV-2 as of 4 January 2021. Vaccination status was then defined based on first COVID-19 vaccine dose exposure between 4 January 2021 and 28 January 2021. Lasso regression was used to calculate PS. Location, age, prior observation time, regional vaccination rates, testing effort and COVID-19 incidence rates at index date were forced into the PS. Following PS weighting and matching, the three methods were compared for remaining covariate imbalance and residual confounding. Last, a target trial emulation comparing COVID-19 at 3 and 12 weeks after first vaccine dose vs unvaccinated was conducted. Results Vaccinated and unvaccinated cohorts comprised 583 813 and 332 315 individuals for weighting, respectively, and 459 000 individuals in the matched cohorts. Overlap weighting performed best in terms of minimizing confounding and systematic error. Overlap weighting successfully replicated estimates from clinical trials for vaccine effectiveness for ChAdOx1 (57%) and BNT162b2 (75%) at 12 weeks. Conclusion Overlap weighting performed best in our setting. Our results based on overlap weighting replicate previous pivotal trials for the two first COVID-19 vaccines approved in Europe.

中文翻译：

---

COVID-19疫苗有效性研究的观察方法：实证评估和目标试验模拟

背景 在评估疫苗预防 COVID-19 有效性的观察性研究中，关于控制混杂因素的最佳实践的数据很少。我们比较了三种成熟方法[重叠加权、逆概率治疗加权和倾向评分（PS）匹配]的性能，以最大限度地减少比较接种疫苗和未接种疫苗的人时的混杂因素。随后，我们进行了目标试验模拟，以研究这些方法复制 COVID-19 疫苗试验的能力。方法 我们纳入了英国初级保健记录中年龄≥75 岁的所有个体 [临床实践研究数据链 (CPRD) AURUM]，截至 2021 年 1 月 4 日，他们尚未感染 SARS-CoV-2 或未接种过 SARS-CoV-2 疫苗。然后定义了疫苗接种状态基于 2021 年 1 月 4 日至 2021 年 1 月 28 日期间首次接触 COVID-19 疫苗剂量。使用 Lasso 回归来计算 PS。地点、年龄、先前观察时间、区域疫苗接种率、测试工作量和索引日期的 COVID-19 发病率均被强制纳入 PS。在 PS 加权和匹配之后，比较了三种方法的剩余协变量不平衡和残留混杂。最后，进行了一项目标试验模拟，比较了首次接种疫苗后 3 周和 12 周与未接种疫苗时的 COVID-19。结果 已接种疫苗和未接种疫苗的人群分别包含 583 813 名和 332 315 名进行加权的个体，以及匹配队列中的 459 000 名个体。重叠加权在最小化混杂因素和系统误差方面表现最佳。重叠加权成功地复制了临床试验对 ChAdOx1 (57%) 和 BNT162b2 (75%) 在 12 周时疫苗有效性的估计。结论 重叠加权在我们的环境中表现最佳。我们基于重叠加权的结果复制了欧洲批准的首批两种 COVID-19 疫苗之前的关键试验。