Causal inference with randomised clinical trials of chemotherapy: The importance of well-documented treatment side-effects

Abstract

Introduction Randomised Controlled Trials (RCTs) are universally considered as the most reliable way to demonstrate and assess causal relationships between treatments and outcome; science-based medicine is rooted in them. Spurious relationships between the outcome and a timefixed treatment-variable are eliminated by randomising patients over two or more arms of the trial. Hence, the randomisation procedure initiates the process by which treatment and outcomes of interest should be interpreted in a causal way. However, treatment is not always administered as intended, not least because of the occurrence of side effects and adverse events. In RCTs of chemotherapy, for example, the treatment administered may differ from the intended one because of the application of either cycle delays or dose reductions. Background Opposite to the∗intention-to-treat approach∗, a statistical analysis based on actual treatment data might be problematic due to the presence of the so-called∗treatment-adjustment bias∗. Exposure to chemotherapy agents may in fact be reduced and/or delayed as a consequence of previous-treatment side-effects. In particular, both reductions and delays contribute to lowering the value of the so-called Received Dose Intensity. Methods Inverse Probability-of-Treatment Weighting (IPTW) is a general methodology for removing treatment-adjustment bias. Working under the hypothesis of∗No Unmeasured Confounding∗, it creates a pseudopopulation by weighting each patient with the inverse probability of observing a certain treatment administration given the past treatment and toxicity history. However, a review of data collected from RCTs on osteosarcoma suggests that treatment side-effects may not be sufficiently well-documented. The pseudo-population created by IPTW has the following two properties: 1. Pseudo-patients' past toxicity-history no longer predicts exposure to chemotherapy in the next cycle; 2. The causal effect of treatment modifications on outcome is the same in both original and pseudopopulations. Results Using data from Medical Research Council trial BO06 (European Organisation for Research and Treatment of Cancer trial 80931) we will illustrate the use of IPTW and Marginal Structural Models (MSMs) for estimating the causal effect of dose reductions on Event-free Survival (EFS). The use of IPTW and MSMs allows to move beyond intention-to-treat and unbiasedly estimate the effect of treatment modifications on EFS. Conclusions We demonstrate that, even with complex and entangled data such as those collected in a RCT of chemotherapy, constructing and estimating a causal model is possible, provided that side-effects are well documented. When this is not the case, the removal of treatmentadjustment bias via IPTW might be problematic, if not prevented at all by the presence of unmeasured confounder. As such, good-quality toxicity data should be regarded as important enablers of causal modelling in RCTs of chemotherapy.