In the real-world clinical setting, initiation of pharmacotherapy depends on the clinical status of the patient and progression of the underlying condition. There can be a delay in start of therapy until or after a clinical threshold is reached. We used time-dependent propensity scores derived from a Cox proportional hazards model to match pediatric allo-HCT patients having adenovirus (AdV) viremia who received cidofovir (CDV) to those who did not, while accounting for the immortal time bias due to varying treatment start times.

Patients with AdV viremia who received CDV any time within 2 weeks of crossing ≥1000 copies/mL clinically-relevant threshold were matched with patients who had not yet received CDV at that time, using the hazard component from a Cox proportional hazards model as a propensity score. The model included key time-dependent factors such as viremia, lymphocyte count, and creatinine clearance, in addition to other baseline factors. This model was compared to a logistic regression model with baseline factors traditionally used to derive propensity scores. For each matched pair, the date of CDV initiation among the treated patient was used as the new index date to account for immortal time bias. Clinical outcomes starting from this index date were further examined.

Among the 241 patients, 100 (41.5%) received CDV within 2 weeks of crossing ≥1000 copies/mL threshold. The Cox model identified time-varying viremia as a significant covariate (HR=1.42; 95%CI: 1.24-1.63; p<0.0001) and 91 patients were matched. In contrast, in the logistic regression model, viremia at baseline time-point was not a significant factor for CDV use and 65 patients were matched.

Using propensity scores derived from the Cox proportional hazards model with time dependent covariates enabled accounting for changing clinical parameters leading up to the use of CDV, which are stronger predictors of the treatment decision as compared to baseline factors.