Missing Data and Censoring in the Analysis of Progression-Free Survival in Oncology Clinical Trials

Abstract

Progression-free survival (PFS) is increasingly used as a primary endpoint in oncology clinical trials. However, trial conduct is often such that PFS data on some patients may be partially missing either due to incomplete follow-up for progression, or due to data that may be collected but confounded by patients stopping randomized therapy or starting alternative therapy prior to progression. Regulatory guidance on how to handle these patients in the analysis and whether to censor these patients differs between agencies. We present results of a reanalysis of 28 Phase III trials from 12 companies or institutions performed by the Pharmaceutical Research and Manufacturers Association-sponsored PFS Expert Team. We show that analyses not adhering to the intention-to-treat principle tend to give hazard ratio estimates further from unity and describe several factors associated with this shift. We present illustrative simulations to support these findings and provide recommendations for the analysis of PFS.

Key Words:

• Censor
• Missing data
• Oncology
• Progression free survival

ACKNOWLEDGMENTS

The authors acknowledge the contribution of the following: Peter Barker, AstraZeneca; Hans Burger, Roche; Will Bushnell, GSK; Paul Bycott, Pfizer; Bill Capra, Genentech; Michelle Casey, GSK; Steve Dahlberg, Amgen; Meredith Goldwasser, Genentech; Shenyang Hong, MedImmune; Dmitri Pavlov, Pfizer; Jane Qian, Abbott; Martin Roessner, Sanofi-Aventis; D. J. Sargent, Mayo Clinic; Raji Swamy, Genentech; Hongliang Shi, Takeda; Zhenming Shun, Sanofi-Aventis; Tao Wang, Pfizer. The authors also acknowledge the referees, whose comments improved the final article.

Notes

¹Some studies (n = 2) had more than 2 arms and are counted as separate studies with each comparison presented separately.

²Multiple myeloma and non-Hodgkins lymphoma.

³Data not provided for three trials.

⁴Data not provided for one trial.

Note. Administrative censoring represents the censoring caused by time of analysis. Under administrative censoring the percentage of patients censored was 15% in the control arm (C), and 15%, 22%, and 28% in the experimental arm (E) when the median event time for this group was 8, 10 and 12 months, respectively. Abbreviations: OHR, observed hazard ratio (geometric mean) and associated (average) 95% CI; Cont., continuous monitoring; and 1 M, 3 M, and 6 M, assessments every 1, 3, or 6 months, respectively.

Note. The minimum of competing event time and censoring time is the observed time. The observation is censored if min(event time, censor time) = censor time; otherwise the observation is an event. Both event time and censoring time follow exponential distributions. Censoring distribution rate parameter (Rate parm.) and observed percent censored presented. Maximum follow-up time was fixed at 40 months. Abbreviations: C (applies later too), Control; E, Experimental; OHR, observed hazard ratio (geometric mean) and associated (average) 95% CI; Cont., continuous monitoring; and 1 M, 3 M, and 6 M, assessments every 1, 3, or 6 months, respectively.

Note. If a subject is considered to be missing, the subject is censored at the last scheduled assessment prior to the simulated event time. Event times follow an exponential distribution. Censoring mechanism: Each subject is censored at the last scheduled assessment prior to the simulated event time with probability p _C and p _E , respectively, depending on their treatment arm. Maximum follow-up time was fixed at 40 months. Abbreviations: C, control; E, experimental; OHR, observed hazard ratio and 95% CI; and 1 M, 3 M, and 6 M, assessments every 1, 3, or 6 months, respectively.