Article type
Year
Abstract
Background: Recently, several randomised controlled trials (RCTs) in which unplanned patient crossover to the experimental arm occurs during the study, have been published in the literature (1, 2). This phenomenon is referred to as selective crossover and is increasingly common.
Objectives: To assess the impact of different data analysis approaches that are widely used to deal with selective crossover: Intention to treat (ITT) and censored analyses.
Methods: The magnitude and direction of bias were investigated through simulations. For each method we explored the relationship between the effect size and (i) time when the crossover occurred, (ii) the fraction of patients who decided to cross over and (iii) their prognostic characteristics (patients selected at random or at high/medium/low risks).
Results: Simulations based on 1000 replications suggest that the ITT analyses tend to dilute the treatment effect size (the bias direction goes towards the null hypothesis) and the magnitude of this dilution is highly dependent on patient characteristics. Best estimates are obtained when patients at low risk switch to the treatment arm. The censored analysis provides estimates for which the direction and magnitude of bias are less predictable.
Conclusions: The analyses most used in RCTs in which selective crossover occurs, tend to be biased. The assessment of the reliability of these results is highly relevant for its consequences on patient health care.
References
1. The BIG 1-98 Collaborative Group. A comparison of letrozole and tamoxifen in postmenopausal women with early breast cancer. N Engl J Med 2005;353:2747–2757.
2. Piccart-Gebhart MJ, Procter M, Leyland-Jones B, Goldhirsch A, Untch M, Smith I. Trastuzumab after ajuvant chemotherapy in HER2-positive cancer. N Engl J Med 2005;353:1659–1672.
Objectives: To assess the impact of different data analysis approaches that are widely used to deal with selective crossover: Intention to treat (ITT) and censored analyses.
Methods: The magnitude and direction of bias were investigated through simulations. For each method we explored the relationship between the effect size and (i) time when the crossover occurred, (ii) the fraction of patients who decided to cross over and (iii) their prognostic characteristics (patients selected at random or at high/medium/low risks).
Results: Simulations based on 1000 replications suggest that the ITT analyses tend to dilute the treatment effect size (the bias direction goes towards the null hypothesis) and the magnitude of this dilution is highly dependent on patient characteristics. Best estimates are obtained when patients at low risk switch to the treatment arm. The censored analysis provides estimates for which the direction and magnitude of bias are less predictable.
Conclusions: The analyses most used in RCTs in which selective crossover occurs, tend to be biased. The assessment of the reliability of these results is highly relevant for its consequences on patient health care.
References
1. The BIG 1-98 Collaborative Group. A comparison of letrozole and tamoxifen in postmenopausal women with early breast cancer. N Engl J Med 2005;353:2747–2757.
2. Piccart-Gebhart MJ, Procter M, Leyland-Jones B, Goldhirsch A, Untch M, Smith I. Trastuzumab after ajuvant chemotherapy in HER2-positive cancer. N Engl J Med 2005;353:1659–1672.