WHEN EARLY MODELS MAY BE TOO EARLY - PREDICTING SURVIVAL FROM INTERIM ANALYSES
Author(s)
Ayers D1, Cope S2
1Precision XTract, Vancouver, BC, Canada, 2Precision Xtract, Vancouver, BC, Canada
OBJECTIVES : Health technology assessment (HTA) agencies are increasingly evaluating new interventions based on interim analyses of single-arm clinical trials in tandem with accelerated approvals. However, early models that extrapolate survival before the trial is completed may produce misleading results, particularly where the sample size is small. We sought to highlight the implications of early modeling on basis of limited follow-up and propose alternative options to explore the uncertainty and improve predicted survival. METHODS : Published data was based on a case study of pediatric and young adult acute lymphoblastic leukemia regarding long-term survival for a tisagenlecleucel, chimeric antigen receptor T-cell, where evidence from the phase II ELIANA trial. We examined changes in estimated survival based on the data available at various time points in the trial period and plotted the predicted survival based on each data cut using alternative parametric models. We used the information from early analyses to estimate the event rate, along with assumptions about the censoring rate, to predict the KM curve based on full follow-up. Additionally, we used simulations to identify the influence of limiting data to when a reasonable proportion of the original sample were at risk. RESULTS : Expected survival differed substantially depending the on the data cut. When the tail of the KM curve was restricted results were less sensitive to the amount of follow-up. Upper and lower bounds for the KM curve were obtained, as well as a predicted ‘full follow-up’ curve. Estimates of the variability of censoring and event times were propagated to produce a credible interval for the predicted curve. CONCLUSIONS : It is important to evaluate the stability of KM curves in HTA evaluations of new interventions with limited follow-up, which may benefit from scenarios assessing restricted and/or predicted KM curves that are less likely to be influenced by single events.
Conference/Value in Health Info
2019-11, ISPOR Europe 2019, Copenhagen, Denmark
Code
PCN375
Topic
Clinical Outcomes, Health Technology Assessment, Methodological & Statistical Research
Topic Subcategory
Decision & Deliberative Processes, Modeling and simulation, Relating Intermediate to Long-term Outcomes
Disease
No Specific Disease