Unplanned crossover in randomised controlled trials: Consequences for efficacy and safety outcomes

Article type
Authors
D’Amico R1, Bonafede E1, Balduzzi S1, Longo G1, Guarneri V1, Piacentini F1, Moja L2, Liberati A1
1University of Modena and Reggio-Emilia, Italy
2University of Milan, Italy
Abstract
Background: If a randomized controlled trial (RCT) shows results in favour of an intervention, investigators might surmise that the equipoise principle is no longer present. As a consequence, control patients may be offered to cross over to the experimental arm, even if it was unplanned.

Objectives: We assessed: a)the prevalence of the unplanned crossover in oncology trials; b)how it is dealt when analysing data; and c)the potential magnitude and direction of bias.

Methods: RCTs published between 2000 and 2011 in the NEJM, JNCI, Lancet, Lancet Oncology, JCO, Annals of Oncology, assessing the efficacy and safety of treatments in oncologic patients were searched. We explored the relationship between the effect and: a)the time when the crossover occurred; b)the fraction of patients who decided to cross over; and c)their clinical characteristics. The magnitude and direction of bias were investigated through simulations.

Results: Out of 164 papers analyzed so far, in 11RCTs investigators gave patients the opportunity to crossover to the experimental arm. The phenomenon is more frequent in the second half of the decade. All studies compared outcomes by using ITT and 20% used an additional censored analyses. Simulations suggest that ITT analyses dilute the treatment effect size for efficacy and safety, whereas the magnitude and direction of the bias caused by the censored analysis seem to be less predictable. The inverse probability of censoring weighted model was also used, but it is highly speculative since it is dependant on the characteristics of those patients who decided to cross over.

Conclusions: While investigators justify the unplanned crossover of patients on ethical grounds, it can leave the scientific community with increased uncertainty of study results, and may be a condition to inflate the net benefit and risk of an intervention of clinical relevance.