Extrapolation of time-to-event data
- 1. © 2016 PAREXEL International Corporation. All rights reserved. PAREXEL International 195 West Street Waltham, MA 02451 USA + 1 781 487 9900 www.PAREXEL.com EXTRAPOLATION OF TIME-TO-EVENT DATA IN VARIOUS DISEASE AREAS: CURRENT RECOMMENDATIONS AND GAPS INTRODUCTION • Extrapolation of time-to-event outcomes is required to evaluate long-term effect of the new interventions in various disease areas particularly in oncology, diabetes, rheumatoid arthritis, and cardiovascular diseases • Application of incorrect extrapolation methods can lead to rejection of inter ventions by HTA bodies. Therefore, we have conducted a review of methodological papers and guidelines detailing the extrapolation of time-to-event data RESULTS • Literature search using the search terms as “extrapolation" AND "survival“ AND "time-to- event"/“time to event“ resulted in 625 citations • Of the 625 citations, 11 methodological papers and five guidelines were included in this review (Figure 1) • The methodological papers and guidelines discussing extrapolation of time-to-event data were only considered relevant for the review • All the included methodological papers were published in journals Findings from methodological papers • Review of the methodological papers emphasize that generally, single and incorrect extrapolation method is employed by different specialists for time-to-event outcomes • It highlights that a series of extrapolation methods should be employed, which should be then validated using statistical methods like AIC, BIC, and goodness of fit • In general, analysts fits one of the commonly used distributions like exponential and Weibull. However, it is also recommended to test other complex distributions like log-logistic, log-normal, and generalized gamma to get the true distribution in order to avoid economic rejections REFERENCES 1. Andersson TML. Stat Med. 2013;32(30):5286-300 2. Bagust A. Med Decis Making 2013:1-9 3. Baker SG. Biometrics. 2012; 68: 248–257 4. Baker SG. Natl Cancer Inst. 2013;105: 316–320 5. Day SM. Dev Med Child Neurol. 2015;57(12):1105-18 6. Demiris N. Stat Methods Med Res. 2015 Apr;24(2):287-301 7. Drummond M. Med Care 2005;43: II-5–II-14 8. Economics section of the MSAC Guidelines: 1-58 9. Hu B. Stat Med. 2012;31(21):2303-17 10. Hudson HM. Stat Med. 2014;33(10):1621-45 11. Latimer N. Medical Decision Making (2013) 33 (6): 743-754 12. Latimer N. NICE guidelines (DSU TSD 14). June 2011 (last updated March 2013) 13. Lin B. Environ Toxicol Chem. 2009;28(7):1557-66 14. Mittmann et al. CADTH guidelines. Value in Health (2012): 580–585 15. NoMA guidelines. Valid from 13-04-2015, Norwegian Medicines Agency 16. PBAC Guidelines for preparing submissions to the Pharmaceutical Benefits Advisory Committee, version 4.4, June 2013 • Further, assumptions underlying these distributions should also be tested. For example: Assumption of proportional hazards • However, the detailed procedure for extrapolation explicitly has been discussed in the guidelines Findings from the guidelines • A total of five guidelines, NICE from England and Wales, NoMA from Norway, CADTH from Canada, PBAC and MSAC from Australia were included in this review • Across the included guidelines, only NICE guidelines provided a detailed description of the steps to be employed for extrapolation of survival analysis (Figure 2) • NICE guidelines suggest testing a range of survival models including parametric, semi-parametric, and non-parametric models • Further, fit of the models should be checked using visual inspection or various statistical tests like AIC, BIC, and goodness of fit • Same ‘type’ of model should be used in case when parametric models are fitted separately to individual treatment arms. This approach allows using different parameters; however, treatment arm will not follow drastically different distributions • Following which, the methods should be validated using internal and external methods on the basis of clinical plausibility • Source of external data is usually considered from comparable population and longer-follow-up, i.e. non-randomized clinical trials (as suggested in PBAC) and long-term registry data (as recommended in NICE) • NoMA guidelines from Norway are brief of NICE guidelines. In addition to the recommendations from NICE guidelines, NoMA guidelines also provide recommendations for treatment switching • These guidelines suggest that switching data may be contaminated and can commonly overestimate the effect of standard arm compared to the intervention arm • Parameterization and extrapolation in the base case analysis in presence of treatment switching must be based on the KM data • Different correction methods should be applied for sensitivity analyses with correction methods for treatment switching, preferably applying with accompanying discussion on strengths, limitations and assumptions for these methods CONCLUSIONS • Until now, NICE guidelines serve as a gold standard for survival extrapolation. However, there is an unmet need to update the guidelines for extrapolation of different time- to-event outcomes across different therapeutic areas beyond oncology METHODS • Embase® and MEDLINE® databases for methodological paper and country specific websites for guidelines were searched to identify relevant articles • Inclusion criteria: English language studies published post 2005 and assessing extrapolation of time-to-event data were included • Each citation was reviewed by two independent reviewers; any disagreements were resolved by a third reviewer Kamra S1 , Kaur J1 , Kartheek J1 , Sehgal M1 , Siddiqui MK1 , Shukla, P1 1 PAREXEL International, Chandigarh, India Figure 2: Summary of NICE Recommendations Patient level/digitised IPD data available Compare log cumulative hazard plots, quantile-quantile plots or suitable residual plots to allow initial selection of appropriate models Plots are not straight line Plots are not parallel Plots are parallel Consider piecewise or other more flexible models Fit individual model Consider PH/AFT model Compare model fits to select the most appropriate model taking into account the completeness of survival data Complete survival data: • AIC/BIC/Log-cumulative hazard plots/other suitable test of internal validity Incomplete survival data: • Visual inspection/External data/Clinical validity/AIC/BIC/Log-cumulative hazard plots/Other suitable tests for internal and external validity/Consider duration of treatment effect Choose most suitable model based on above analysis. Complete sensitivity analysis using alternate plausible survival models, and taking into account uncertainty in model parameter estimates SOURCE: NICE DSU Technical Support Document 14 AFT: Accelerated Failure Time; AIC: Akaike's Information Criteria; BIC: Bayesian information criteria; IPD: Individual Patient Data; PH: Proportional Hazards Figure 1: Flow diagram 625 citations First-pass (excluded = 552) Second-pass (excluded = 62) 73 citations 11 methodological papers + Five guidelines Country-specific websites