By Vanessa Beddo, PhD (VP, Global Head of Biostatistical Consulting)
Introduction
There are several factors to consider when adding a crossover design feature to an oncology study. These designs typically allow for patients to crossover from placebo or standard of care (SoC) to investigational product after disease progression on their randomized therapy. It is also usually assumed that crossover is not mandated but allowed as per investigator decision. As a result, crossover is not systematic but elective; and, given it takes place post-progression, would be of potentially measurable impact to overall survival (OS) (i.e., no impact to PFS or commonly used response-derived outcomes).
EMA Q&A and Regulatory Considerations for Crossover Designs
Studies have successfully analyzed OS, using standard methodology, regardless of this design feature (i.e., analyzing the endpoint as crossover-naïve). In an EMA Q&A (2018) another situation is discussed: “…the situation where crossover of control patients is not systematic and there is interest in estimating the effect in the (hypothetical) situation that no cross-over would have occurred in the trial, under the assumption that the experimental treatment cannot introduce harm or deterioration of the condition under investigation in the control patients who cross over. In particular, it should be fully justified that this hypothetical effect is a relevant one for regulatory decision making.”
Statistical Considerations and Methods for Analyzing Crossover Designs
Methods specifically mentioned for handling patient crossover include: censoring at the time of cross-over, Inverse Probability of Censoring Weighting (IPCW), Rank Preserving Structural Failure Time models (RPSFT) (Robins and Tsiatis, 1991), and “two-stage” methods. For these analyses methods to be considered valid, the following assumptions should be satisfied:
- Censoring at the time of crossover: The underlying assumption is that the removal of patients from further consideration (censoring) at an individual timepoint will not lead to a bias in treatment effect estimation. The censoring would need to be considered non-informative in this case, which may not be appropriate if the survival of those who cross over differs from those who do not.
- IPCW: As this method censors patients who cross over, and upweights remaining patients in a manner designed to reflect the original population (based on a collection of covariates), the assumption is that the reweighted control arm resembles the original population in terms of performance. Unfortunately, it may be a challenge to evaluate this conclusively.
- RPSFT: This analysis assumes that there exists a “common treatment effect” meaning that the improvement in patient survival would be “line-agnostic” – therapy effectiveness would not be different if administered first (as initially randomized) or second (after progression on placebo) administered first (as initially randomized) or second (after progression on placebo). This may not be wholly appropriate for oncology patients.
Crossover Designs: Case Studies and Analysis Results
As an example of these analysis methods, lenvatinib (2014 FDA submission) showed clinical activity in a phase 2 study involving patients with differentiated thyroid cancer that was refractory to radioiodine (iodine-131). The phase 3 trial, a randomized double-blind design, allowed for patient crossover to lenvatinib at the time of disease progression, patients in the placebo group could then receive open-label lenvatinib. The primary endpoint, which was identified as statistically significant, was progression-free survival. However, the crossover (83% of patients) may have impacted the analysis of OS, a secondary endpoint, which had been based primarily on a log-rank test (p-value 0.1032).
As described in the FDA summary review of lenvatinib, the first method performed for the lenvatinib trial was to correct for the crossover effect, an “as treated” or “on-treatment” analysis (Fox R. et. al 2011) was used to evaluate the endpoint. This method was considered as problematic as it resulted in informative censoring, changing the endpoint outcome. The RPSFT model was also used. However, due to a review of data post-crossover versus the lenvatinib arm (from randomization), the assumption of a “common treatment effect” appeared inappropriate. A simulation method was also implemented. Most methodologies used, with exception of the “as treated” analysis, resulted in survival estimates that reflected a consistent but non-statistically significant (p-value > 0.05) risk reduction for the lenvatinib-randomized patients. Lenvatinib was approved in 2015; labeling reflects the log-rank analysis results.
Another randomized phase 3 trial, ARAMIS, studied darolutamide in men with non-metastatic prostate cancer (2019 FDA submission). The primary endpoint was metastasis-free survival; the study was unblinded after the primary analysis was identified to be successful. Patients were then permitted to crossover to darolutamide treatment. The final OS analysis, using the standard log-rank test, resulted in a reduced risk of death in the darolutamide arm. Regardless, there was interest around the magnitude of effect expected for this treatment. As a result, four sensitivity analyses were planned to adjust for crossover bias in OS: RPSFT; iterative parametric estimation (IPE); censoring at the time of crossover; and IPCW. All four sensitivity analyses resulted in greater risk reductions in favor of darolutamide. The results of the log-rank test, which was successful at the final OS analysis, are included in the approved product label.
Additional Considerations for Crossover Designs
These are two examples in which randomized, double-blind trials incorporated a crossover feature. Not all studies incorporate this. A recently published commentary (Gyawali, 2023) suggests that crossover should either be mandated or prohibited depending on the context. The author’s position is that to accurately assess the magnitude of OS improvement, a proper, up-to-date standard-of-care reference arm should include (or even mandate) the use of the investigational treatment if it is already approved for subsequent lines of therapy. Thus, crossover should be included as a design feature when a therapy has already been approved for use after progression of randomized treatment.
Further to this, Haslam and Prasad (2018) specifically discussed the KEYNOTE-024 trial, where patients with non-small-cell lung cancer (NSCLC) and high expression of PD-L1 received first-line pembrolizumab or a platinum-based chemotherapy. Prior to KEYNOTE, PD-1 antibody therapy had shown improved OS over docetaxel in second-line NSCLC. As a result, the studied intervention could then potentially be considered standard of care for a line subsequent to that under study. As a result, for this trial, patients could cross over to pembrolizumab after disease progression. In this context, crossover seems particularly valuable when researching whether there is a survival advantage to administering an investigational drug upfront versus giving the same drug as a subsequent line.
Conclusion
When considering crossover for inclusion into a study design, it would be prudent to keep in mind the following:
- Even when OS is successful, it will be desirable to produce an estimate of risk reduction that corrects for crossover bias.
- Depending on the scientific question and stakeholder perspective, crossover may be very highly desirable in certain circumstances.
- The regulatory and analysis impact of crossover to the OS endpoint should be carefully evaluated prior to proceeding.
-
Impacts to estimand definitions should also be considered.
-
Should crossover, or other design complexities, be up for consideration for your study, Allucent’s biostatistical consultants can assist with an evaluation of your protocol to ensure its alignment with your regulatory and corporate objectives.
For more reading on the important role of a statistician when developing your study designs check out this blog: The Role of Statisticians in Study Design and Why to Engage them Early | Allucent
References
- EMA Q&A. Question and answer on adjustment for cross-over in estimating effects in oncology trials. 13 December 2018. Accessed 12MAR2024.
- Robins JM, Tsiatis AA. Correcting for non-compliance in randomized trials using rank structural failure time models. Commun Stat Theory Methods. 1991; 20(8):2609-31.
- Fox R, Lucinda B, Abrams K: Evaluation of methods to adjust for treatment switching in clinical trials. Trials 2011, 12 (Suppl 1): A139
- Branson M, Whitehead J. Estimating a treatment effect in survival studies in which patients switch treatment. Stat Med 2002;21:2449–63. https://doi.org/10.1002/sim.1219.
- Schlumberger M, Tahara M, Wirth LJ, et al. Lenvatinib versus placebo in radioiodine-refractory thyroid cancer. N Engl J Med 2015;372:621-30. DOI: 10.1056/NEJMoa1406470
- FDA, Statistical Review and Evaluation, NDA/BLA# 206947, Drug Name: Lenvatinib, Indication: The treatment of patients with progressive, radioiodine-refractory differentiated thyroid cancer, Submission date: August 14, 2014. Accessed 12MAR2024.
- Fox R, Lucinda B, Abrams K: Evaluation of methods to adjust for treatment switching in clinical trials. Trials 2011, 12 (Suppl 1): A139
- N. Shore, K. Fizazi, T. Tammela, M. Luz, M. P. Salas, P. Ouellette, et. Al. Effect of crossover from placebo to darolutamide on overall survival in men with non-metastatic prostate cancer: sensitivity analyses from the randomised phase 3 ARAMIS study, Eur J Cancer. 2023 (195) 113342.
- Watkins C, Huang X, Latimer N, Tang Y, Wright E. Adjusting overall survival for treatment switches: commonly used methods and practical application. Pharm Stat. 2013; 12(6): 348-357.
- Gyawali B, Problematic crossovers in cancer drug trials. Nature Reviews Clinical Oncology. 2023; 20: 815–816.
- Haslam H, Prasad V, When is crossover desirable in cancer drug trials and when is it problematic? European Society for Medical Oncology. 2018; 29(5): 1079-1080.