(279) Time-varying propensity score matching as a method to balance baseline covariates and calendar time effects in a real-world effectiveness analysis
VINCI Director Department of Veterans Affairs, United States
Background: In December 2021, FDA issued an EUA for AZD7442, a combination of monoclonal antibodies (tixagevimab/cilgavimab), as pre-exposure prophylaxis against COVID-19 in moderately/severely immunocompromised (IC) patients. The US Department of Veterans Affairs Health System is conducting an observational study that is providing insight on the use and effectiveness of AZD7442 in evolving pandemic conditions. In addition to balancing systematic differences between the exposed and unexposed cohorts in baseline covariates, balancing calendar time effects between the cohorts is required for robust analyses.
Objectives: To assess the utility of time-varying propensity score (PS) matching to address significant differences between exposed and unexposed patients for an observational effectiveness analysis.
Methods: Patients were grouped by 2-6 calendar week periods (depending on number of AZD7442 administrations) and visit settings (inpatient, outpatient specialty care, and outpatient non-specialty care). Within each time period and visit setting, a PS model was fit using the following pre-specified variables calculated at index date: age, sex, census region, Area Deprivation Index, CCI, VA Frailty Index, smoking status, number of outpatient visits, inpatient visits and emergency department encounters, antihyperglycemic agents, AIDS/HIV, hematologic malignancy, organ transplant, solid tumor malignancy, stem cell transplant, primary immunodeficiency, therapeutically induced immunosuppression, and history of cardiovascular risk. Standardised mean differences (SMD) were used to assess covariate balance overall and within each time interval of matching.
Results: Despite subsequent sample size reduction, application of this type of PS matching resulted in a narrowing of the 95% confidence intervals (CI) of the hazard ratio (HR) as characteristics were added. An analysis without accounting for time period, ADI, frailty or visit setting of AZD7442 administration resulted in a CI ranging 1.05 from lower to upper bound. After accounting for time period, then ADI & frailty, then visit setting, CI range changed to 0.71, then 0.54; and finally 0.41, also directionally moving towards a lower HR. Baseline covariates were generally balanced overall (SMD < 0.1 for all covariates) and within most time intervals.
Conclusions: Time-varying PS matching resulted in a reduction in bias and variance of the HR. Significant pre-matching differences between cohorts were appropriately balanced. Application of this matching approach ensures contemporaneous matching over calendar time and is critical in dynamic cohort studies, particularly with heterogenous patient population and an evolving disease landscape.
