Until May 2023, the COVID-19 pandemic had led to 160 million cases and 1.14 million deaths in the United States of America [1]. The spread of the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) has been slowed down after the wide injection of vaccines. The effect of vaccination on preventing disease is a treatment effect estimation problem. Many studies are conducted to study the efficacy of COVID-19 vaccinations [2–7], and most of them are strictly designed clinical trials. However, valid and rigorous observational vaccine effectiveness studies are needed. In this project, the observational EHR data from Optum® de-identified COVID-19 Electronic Health Record dataset (2007-2021) will be used to estimate COVID-19 vaccination efficacy for preventing infection. The causal inference by counterfactual reasoning method would be applied as the analysis framework for observational vaccine effectiveness studies. Vaccine effectiveness (VE) estimation can be considered a treatment effect estimation problem in which the vaccine is a treatment against a particular virus or pathogen infection. Vaccine effectiveness (VE) estimation has been a key study area for infectious diseases. Accurate estimation of the extent of the waning of vaccine-induced protection over time is an important public health need [8]. Vaccine efficacy (VE) is generally estimated by: VE = 1-RR where is the measurement of relative risk in the vaccinated compared with the unvaccinated group. The time-varying VE will be estimated by VE (t) = 1 – RR (t), where RR (t) is the rate ratio at a specific time point [9]. When estimating vaccine effectiveness in observational studies, scientists often encounter biases, including confounding/treatment selection bias, censoring bias, event-induced covariate shifts, competing events, prior immunity, emergence of new virus variants, healthy vaccinee bias, healthcare seeking behavior, and waning immunity. Confounding bias means that exposure and outcome share a common cause. Selection bias generally exists for observational studies, i.e., the subjects are not randomly assigned to the treatment group or control group. In addition to confounding or treatment selection bias, sources of covariate shifts in learning treatment-specific hazard functions and vaccine effectiveness estimation include censoring bias, which indicates the dependence of the censoring hazard on the covariates; event-induced shifts if the event-hazard depends on covariates; and competing events, when the effect of a treatment on an event of interest can only be jointly assessed with the presence of risk due to a competing event [10,21&22].

To address these biases, we adapt recently developed generalization bounds from domain adaptation and treatment effect estimation to time-to-event data, and apply the counterfactual causal inference method with deep neural network-based balancing representations and the framework of ERM (SurvITE) [10] to vaccine effectiveness estimation from the time-to-event data which are extracted from Optum® COVID-19 data. To evaluate its performance, the adapted SurvITE method is compared with the results from a Cox regression model [11] and a Random survival forest (RSF) model [12]. The preliminary results showed that the vaccine effectiveness curves estimated by the SurvITE method and RSF model were closed, but the vaccine effectiveness curve estimated by the Cox regression model was much higher than that estimated by the SurvITE method and RSF model.

Problem formulation

For full article, please go through the link: https://www.pubtexto.com/pdf/?causal-inference-for-estimation-of-vaccine-effects-from-timetoevent-data