BACKGROUND: Prognostic models can help clinicians make decisionsover whether to prescribe certain medications for individual patients.For this purpose, models should estimate what a patient's outcomerisk would be if they were to remain unexposed to a certain treat-ment. However, data used to develop prognostic models typicallyinclude treated individuals and standard methods fail to accountfor this.
OBJECTIVES: To compare methods for utilizing routinely collected infor-mation on medication use in prognostic model development andassess the benefits for risk prediction, using the prescription of selec-tiveβ‐blockers (SBBs) in chronic obstructive pulmonary disease(COPD) patients as a case study.
METHODS: Clinical and prescription information on COPD patients wereidentified from the Utrecht General Practitioners (UGP) database. Wecompared 7 approaches to model time‐varying treatments: 1) ignoringtreatment, 2) excluding treated patients, 3) censoring treated patients,4) inverse probability of treatment weighting after censoring (IPTW),modeling treatment as a 5) binary or 6) time‐varying covariate and 7)marginal structural modeling (MSM). First, directed acyclic graphs(DAGS) were used to assess the theoretical properties of the differentapproaches. Next, models to predict 5‐year mortality risk without SBBuse were developed in the UGP data using the 7 approaches. The abso-lute risk predictions and overall performance (calibration slope, c‐statis-tic, Brier score) of the models were compared.
RESULTS: Based on DAGs, we found IPTW and MSMs have the besttheoretical properties, as they can account for the effects of treat-ment and avoid selection bias. In our case study, 1906 patients wereincluded in the analysis and 325 received SBBs during follow‐up.Compared to ignoring SBBs, approaches (2) and (5) provided predic-tions that were 1% and 2% higher on average. Measures of modelperformance varied minimally between approaches.
CONCLUSIONS: We found IPTW and MSMs are preferred in theory, butfound little difference between methods in our case study. Futurestudies should consider using data on time‐varying medication use tobetter model prognosis.
