Evaluating Bias Associated With Inadequate Variable Selection When Using Inverse Probability of Censoring Weights to Adjust for Treatment Switches in Clinical Trials
Author(s)
Xuan J1, Mt-Isa S2, Latimer N3, Yorke-Edwards V4, Vandormael K5, White IR4
1University College London, London, LON, UK, 2MSD, Zurich, Switzerland, 3University of Sheffield & Delta Hat Limited, Sheffield, DBY, Great Britain, 4University College London, London, UK, 5MSD, Brussels, Belgium
Presentation Documents
OBJECTIVES: Treatment switching is common in randomised controlled trials. Participants may switch between randomised treatments, or onto other treatments. When treatment switches do not represent treatment pathways that would be observed in clinical practice, adjustment analyses may be used to estimate counterfactual outcomes to inform healthcare decision making. Since simply censoring switchers is prone to selection bias, inverse probability of censoring weights (IPCW) is a common method used to correct for this by giving extra weight to uncensored individuals who had similar prognostic characteristics to censored individuals. Such weights are computed by modelling selected covariates. IPCW relies on the no unmeasured confounding (NUC) assumption, and selecting variables can be challenging. In this study, we explore the behaviour of IPCW under conditions where too few, too many, and mis-specified covariates are included in the weighting models.
METHODS: We simulated data based on realistic trial settings. Scenarios were designed to vary treatment switches in one or both arms with different prevalence, correlation between two confounders, effect of each confounder, and sample size. IPCW with different combinations of covariates (including confounders and non-confounding factors) were investigated.
RESULTS: IPCW with too many covariates but satisfied NUC performed well; and applications with too few or mis-specified confounders yielded biased estimates. Residual confounding caused by omitting confounders remained the main source of bias. Including unnecessary non-confounding factors for treatment switches or outcome increased bias and standard errors but to an acceptable level.
CONCLUSIONS: IPCW is crucially reliant upon the NUC assumption. Including variables in weighting models that are not confounders is sub-optimal but is preferable to excluding confounding variables. When IPCW may be needed, it is important to ensure complete data collection on all potential confounders to allow more reliable estimates to be obtained.
Conference/Value in Health Info
Value in Health, Volume 26, Issue 11, S2 (December 2023)
Code
MSR156
Topic
Methodological & Statistical Research, Patient-Centered Research, Study Approaches
Topic Subcategory
Adherence, Persistence, & Compliance, Clinical Trials, Confounding, Selection Bias Correction, Causal Inference
Disease
No Additional Disease & Conditions/Specialized Treatment Areas