(084) Real World Effectiveness of Anticoagulation in Hospitalized COVID-19 Patients using High-Resolution EHR Data in the Presence of Time-varying Confounding: Out with the Old, In with the New

Publication Number: 421

Background: In treatment effectiveness studies using real-world data, confounding is a common concern. The effect of time-varying treatments is popularly estimated using traditional (conditional) Cox models which can produce biased estimates. This bias, however, can be addressed using causal inference methodology such as a marginal structural Cox model (MSM). The National COVID Cohort Collaborative (N3C), a large national EHR database for COVID-19 patients, provides detailed clinical data on time-varying confounders allowing us to compare commonly used conditional Cox models to MSM.

Objectives: To investigate potential bias from the conditional Cox model approach compared to MSMs in a study assessing the effectiveness of time-varying anticoagulation medication (ACT) on venous thromboembolism (VTE).

Methods: Using clinical EHR data within the N3C from 49 institutions in the US, we identified adult patients (≥ 18 yrs) hospitalized for COVID-19 on or after 3/1/2020. We excluded all patients with history of VTE, stroke, or atrial fibrillation as well as institutions with poor data quality. We measured ACT (heparin or enoxaparin) as a time-varying exposure, and hemoglobin and platelets as time-varying confounders. The outcome was the first VTE event during the hospital stay.

We followed patients from the admission to discharge or for 30 days. We estimated the hazard ratio (HR) and 95% CIs for the effectiveness of ACT on VTE using two methods: 1) conditional time-varying Cox model adjusted for baseline and all time-varying confounders and 2) MSM using time-discrete survival analysis in which IPTW weights and model coefficients were estimated using a binomial distribution with complementary log-log link.

Results: Among 153,611 eligible patients (mean age 58, 51% male, 50% White), 121,637 (79%) received ACT at least once during hospitalization. The mean hemoglobin was 11.7 g/dL (SD 2.5) and mean platelet was 252 cells/nL (SD 122) with a mean of 2 measurements per person per day. We observed 4,273 VTE events (3.7 per 1000 person days), 2,598 (3.5 per 1000 person days) of which occurred among patients who were actively receiving ACT. Using a conditional time-varying Cox model, the HR of ACT on VTE was 1.16 (95%CI 1.02,1.32). In contrast, the MSM-based HR was 0.84 (95% CI 0.75,0.95), consistent with the established benefit of ACT in VTE prevention.

Conclusions: We demonstrated the bias from this popular approach using a well-documented prophylactic effect of ACT on VTE. The use of conditional Cox models with time-varying covariates and treatment is known to be biased. Estimating real world effectiveness in such settings requires the use of appropriate methods with a foundation in causal inference as well as precise data on time-varying confounders.