Causal survival analysis: A guide to estimating intention-to-treat and per-protocol effects from randomized clinical trials with non-adherence

Abstract

When reporting results from randomized experiments, researchers often choose to present a per-protocol effect in addition to an intention-to-treat effect. However, these per-protocol effects are often described retrospectively, for example, comparing outcomes among individuals who adhered to their assigned treatment strategy throughout the study. This retrospective definition of a per-protocol effect is often confounded and cannot be interpreted causally because it encounters treatment-confounder feedback loops, where past confounders affect future treatment, and current treatment affects future confounders. Per-protocol effects estimated using this method are highly susceptible to the placebo paradox, also called the “healthy adherers” bias, where individuals who adhere to placebo appear to have better survival than those who don’t. This result is generally not due to a benefit of placebo, but rather is most often the result of uncontrolled confounding. Here, we aim to provide an overview to causal inference for survival outcomes with time-varying exposures for static interventions using inverse probability weighting. The basic concepts described here can also apply to other types of exposure strategies, although these may require additional design or analytic considerations. We provide a workshop guide with solutions manual, fully reproducible R, SAS, and Stata code, and a simulated dataset on a GitHub repository for the reader to explore.

Considerations for choosing an imputation method for addressing sparse measurement issues dictated by the study design - An illustration from per-protocol analysis in pragmatic trials

“…For estimation of per-protocol effect using the IPAW model, we organize the data in a long-format, where we record data about pre-randomization factors, as well as observations of each time-varying variables, such as changes in treatment and post-randomization factors by time index (e.g., could be in months). 22 Under MCAR assumption, we have considered two imputation methods, LOCF and multiple imputation, to impute values for the months where measurements of any post-randomization factor were not recorded. In the following comparisons, zip plots have provided visual illustrations of the overall performances of the imputation methods in most scenarios.…”

“…Tutorials, editorials, and guideline documents are being written to raise awareness, and to fill the gap in the renewed interest about this method in recent years. 1,2,22 The strength of this method lies in utilizing high-quality pre and post-randomization factors in the analysis that can predict the adherence pattern, and appropriately adjusting those factors in the analysis via inverse probability weighting. However, infrequent measurement of the post-randomization factors can threaten the validity of this method.…”

“…We are primarily interested in the per-protocol effect of sustained adherence to treatment versus standard of care. The estimation of this contrast Pr(Yt+1a¯=1|Z=1) versus Pr(Yt+1a¯=0|Z=0) required following three steps 4,22 . We first censored the participants at the point of protocol deviation related to treatment (i.e., first time A t ≠ Z ).…”

“…(2) Same is true when there exists confounding, and all of the necessary confounders and measured and adjusted. (3) However, bias levels from both approaches increase with the increment of the gap in measurement frequency, when there exists an unmeasured baseline confounder, that leads to treatment-condfounder feedback 22 . Interestingly, the model SEs and empirical SEs from both LOCF and multiple imputation approaches were comparable, bias from multiple imputation approach was slightly elevated compared to that from the LOCF approach.…”

Ticagrelor Monotherapy or Dual Antiplatelet Therapy After Drug‐Eluting Stent Implantation: Per‐Protocol Analysis of the GLOBAL LEADERS Trial

“… 19 , 20 , 21 We accounted for this by using time‐varying censoring weights. 25 For the calculation of the censoring weights, we restructured the data set such that each subject provided monthly observation periods. At each monthly time period, time‐varying month‐specific information and baseline information from the trial start was included in the data set.…”