Article type
Year
Abstract
Background: Evidence from randomised controlled trials (RCTs) and systematic reviews (SRs) is usually taken into account when making decisions on which interventions are better to use in clinical practice. RCTs are exposed to bias when investigators offer patients enrolled in a RCT the possibility to cross over from one arm to the other one, without the switch being planned. This phenomenon is referred to as selective cross-over (SCO).
Objectives: Our main objectives were to assess:
1. the prevalence of SCO considering the context of RCTs assessing the efficacy of therapies for breast cancer (BC);
2. whether different statistical methods provide different results, in particular when the outcome of interest is a time-to-event outcome.
Methods: RCTs assessing the efficacy of therapies for BC patients published between January 2000 and December 2015 were searched. Different analysis methods exist, such as the intention-to-treat analysis, the censored analysis and the analysis considering the treatment as a time-varying covariate, or more complex methods, such as the inverse probability of censoring weighting analysis, the Loeys and Goetghebeur estimator, and the rank-preserving structural-failure time models. All the methods were evaluated through simulations, considering scenarios that differed in the proportion of patients crossing-over, their underlying prognosis, and the magnitude of true treatment effect.
Results: Cross-over occurred in the 24% of RCTs identified. Simulations highlighted that complex methods have better performances, especially when the probability of cross-over is assumed to depend on prognosis (i.e. patients with a poor prognosis cross-over more frequently than patients with a good prognosis), but each of them makes assumptions that are not always verifiable or likely to occur in the considered context.
Conclusions: It is important to understand better the bias associated with SCO in RCTs, which can be propagated when the results are meta-analysed in SRs, with important repercussions on patients' health.
Objectives: Our main objectives were to assess:
1. the prevalence of SCO considering the context of RCTs assessing the efficacy of therapies for breast cancer (BC);
2. whether different statistical methods provide different results, in particular when the outcome of interest is a time-to-event outcome.
Methods: RCTs assessing the efficacy of therapies for BC patients published between January 2000 and December 2015 were searched. Different analysis methods exist, such as the intention-to-treat analysis, the censored analysis and the analysis considering the treatment as a time-varying covariate, or more complex methods, such as the inverse probability of censoring weighting analysis, the Loeys and Goetghebeur estimator, and the rank-preserving structural-failure time models. All the methods were evaluated through simulations, considering scenarios that differed in the proportion of patients crossing-over, their underlying prognosis, and the magnitude of true treatment effect.
Results: Cross-over occurred in the 24% of RCTs identified. Simulations highlighted that complex methods have better performances, especially when the probability of cross-over is assumed to depend on prognosis (i.e. patients with a poor prognosis cross-over more frequently than patients with a good prognosis), but each of them makes assumptions that are not always verifiable or likely to occur in the considered context.
Conclusions: It is important to understand better the bias associated with SCO in RCTs, which can be propagated when the results are meta-analysed in SRs, with important repercussions on patients' health.