Accounting for time-dependent treatment use when developing a prognostic model from observational data: A review of methods

Abstract

Failure to account for time-dependent treatment use when developing a prognostic model can result in biased future predictions. We reviewed currently available methods to account for treatment use when developing a prognostic model. First, we defined the estimands targeted by each method and examined their mechanisms of action with directed acyclic graphs (DAGs). Next, methods were implemented in data from 1,906 patients; 325 received selective β-blockers (SBBs) during follow-up. We demonstrated seven Cox regression modeling strategies: (a) ignoring SBB treatment; (b) excluding SBB users or (c) censoring them when treated; (d) inverse probability of treatment weighting after censoring (IPCW), including SBB treatment as (e) a binary or (f) a time-dependent covariate; and (g) marginal structural modeling (MSM). Using DAGs, we demonstrated IPCW and MSM have the best properties and target a similar estimand. In the case study, compared to (a), approaches (b) and (e) provided predictions that were 1% and 2% higher on average. Performance (c-statistic, Brier score, calibration slope) varied minimally between approaches. Our review of methods confirmed that ignoring treatment is theoretically inferior, but differences between the prediction models obtained using different methods can be modest in practice. Future simulation studies and applications are needed to assess the value of applying IPCW or MSM to adjust for treatments in different treatment and disease settings.