Adjusting overall survival for treatment switch Recommendations of a crossinstitutional statistical working group Claire Watkins Director and Consultant Statistician, Clarostat Consulting Ltd Cytel East User Group Meeting London, 16 th March 2016
Outline The working group Background Methods simple/naïve, complex Example Method selection process Best practice design and analysis 2
Treatment switch subteam of the PSI HTA SIG Remit and membership (past and present) Amongst pharmaceu0cal industry sta0s0cians working with trials including treatment switches: Raise awareness Current sta*s*cal methods Poten*al applica*ons Strengths and limita*ons Promote and share best prac0ce, in both analysis and design Encourage research into new and improved methods [Chair] Claire Watkins (Clarostat Consulting) Pierre Ducournau (Roche) Rachel Hodge (GSK) Xin Huang (Pfizer) Cedric Revil (Roche) 3 Alexandros Sagkriotis (Novartis) Yiyun Tang (Pfizer) Jason Wang (Celgene) Elaine Wright (Roche) [Non-member advisory role] Nicholas Latimer (University of Sheffield)
Key publications 4
Background What do we mean by treatment switch/crossover? Patients in a parallel group RCT may switch or crossover to the alternative treatment at some point before an endpoint of interest occurs. We might build this into the trial protocol e.g. Sunitinib GIST trial (Demetri, 2012) Or it might happen spontaneously due to clinical practice in the region, if the treatment is already on the market e.g. Gefitinib IPASS trial (Fukuoka, 2011) Randomise Randomise Sunitinib Placebo Gefitinib Doublet chemo Disease prog Disease prog Sunitinib Standard clinical practice (inc gef) Survival Survival 5
Potential impact of treatment switch Correlation between PFS and OS in NSCLC Switch prohibited (n=20) R 2 =0.5341 Switch allowed (n=15) R 2 =0.0027 Interaction p=0.019 Hotta et al: Progression-free survival and overall survival in phase III trials of moleculartargeted agents in advanced non-small-cell lung cancer. Lung Cancer 79 (2013) 20 26 6
Switching Regulatory vs HTA viewpoint Regulatory agency Evaluate efficacy in clinical trial Switch happened in clinical trial Do not adjust OS to remove switch - ITT primary May consider switch adjusted OS as a supportive analysis Primary endpoint may not be OS anyway Health Technology Assessment (HTA) Agency Evaluate effectiveness in real world setting Switch real world (mostly) Use plausible methods* to adjust OS to remove switch If no plausible methods, use ITT Key endpoint for lifetime cost effectiveness calculations is OS * Views of plausible methods differ by agency! 7
Why does switching matter for HTA? It all depends on the decision problem In HTA, the decision problem is often to compare: Current clinical practice without new therapy vs Potential future clinical practice including new therapy If there is switching and the new therapy is effective, ITT underestimates this difference How to estimate long term efficacy without switch? 8
Commonly used methods to estimate control arm survival in absence of switch Naive methods 1. Exclude switchers 2. Censor at switch 3. Time varying covariate Simple to apply High levels of bias Complex methods 1. Inverse Probability of Censoring Weighting (IPCW; observational) 2. Rank Preserving Structural Failure Time (RPSFT; randomisation based) 3. Iterative Parameter Estimation (IPE; randomisation based) 4. Two-stage Accelerated Failure Time (AFT) Harder to apply Try to reduce bias External data 9
Naive (1): Exclude switchers Control arm survival Compare to observed experimental arm survival Observed (ITT) control arm Exclude switchers Non switchers Non switchers Switchers S S Switchers S S Key Death time Censor time S Switch time 10 Assumption: Switchers and nonswitchers have the same prognosis (i.e. no confounders)
Naive (2): Censor switchers Control arm survival Compare to observed experimental arm survival Observed (ITT) control arm Censor switchers Non switchers Non switchers Switchers S S Switchers Key Death time Censor time S Switch time 11 Assumption: Switchers and nonswitchers have the same prognosis (i.e. no confounders)
Naive (3): Time varying covariate Non-switchers Switchers Confounding Covariate Covariate Switch Death Death 12 Assumption: Switchers and nonswitchers (with the same covariates) have the same prognosis (i.e. no confounders)
Complex (1): IPCW (weight non-switched times) Control arm survival Compare to observed experimental arm survival Observed (ITT) control arm IPC weighted Non switchers Non switchers WEIGHT WEIGHT S WEIGHT Switchers S Switchers WEIGHT Key Death time Censor time S Switch time 13 Assumption: The variables in the weight calculation fully capture all reasons for switching that are also linked to survival (i.e. no unmeasured confounders) Weights represent how switch-like a patient is that has not yet switched
Statistical detail: IPCW An observational data-based two stage modelling approach Modelling stage 1: Population Control arm patients Model Output Modelling stage 2: Population All patients Model Output Probability of switch over time conditional on baseline and time varying factors (propensity score) Time varying subject specific IPC weights Survival analysis with IPC weights applied to control arm patients (censor at switch) IPC weighted Hazard Ratio and KM 14
Complex (2): RPSFT (adjust post switch times) Control arm survival Compare to observed experimental arm survival Observed (ITT) control arm RPSFT adjusted Non switchers Non switchers Switchers S S Switchers Key 15 Time off experimental Death time Censor time S Switch time Time on experimental Time off experimental Treatment Time on multiplier x experimental Assumptions: Each cycle of treatment extends survival by a constant amount. Balanced arms due to randomisation
Statistical detail: RPSFTM (Robins 1991) Accelerated Failure Time model structure Works by estimating the treatment effect that balances counterfactual survival time in the absence of treatment [T(0)] between randomised arms T(0) = T off + e ψ T on Survival time without experimental Time off experimental Treatment effect Time on experimental Treatment effect <1 is beneficial (time to death is slowed down on trt) Core assumption: Trt effect is the same regardless of when treated 16
Statistical detail: RPSFTM Need to define two states: On and Off treatment Treatment effect estimated using G-estimation (non-parametric) P-value is similar to ITT analysis! What does the constant treatment effect assumption mean? An example Take two identical patients, Bob and Steve On best supportive care (control) only, their survival would be 6m and they would progress at 4m Give them both 1 cycle of experimental treatment Bob has his treatment at the start Steve has his treatment at 4m after progression Constant treatment effect => Bob and Steve still have identical survival (e.g. 7m each) even though treated at different times
Complex (3): IPE (adjust post switch times) Essentially the same as RPSFTM Key difference: - Treatment effect is estimated using a parametric model (e.g. Weibull) instead of non-parametric G-estimation - Additional distributional assumption required - May reduce model fitting difficulties of RPSFTM In practice, RPSFTM and IPE results usually very similar
Complex (4): 2-Stage AFT (observational study) Control arm survival Compare to observed experimental arm survival Observed (ITT) control arm 2-stage AFT adjusted Non switchers Switchers 19 P P P P S Key Death time Censor time SSwitch time PProgression time S Not valid if switch can occur before progression Treat control arm as observational study post progression Re-baseline at progression Collect covariate data at progression Calculate effect of switch treatment adjusting for covariates Adjust switcher data and compare randomised arms Assumptions: The covariates fully capture all reasons for switching that are also linked to survival No time-dependent confounding between P and S (i.e. no unmeasured confounders)
Case study: Sunitinib vs Placebo, GIST Overall Survival (Final, 2008) Overall Survival Probability (%) 100 90 80 70 60 50 40 30 20 10 0 Total deaths 176 90 Sunitinib (N=243) Median 72.7 weeks 95% CI (61.3, 83.0) Placebo (N=118) Median 64.9 weeks 95% CI (45.7, 96.0) Hazard Ratio=0.876 95% CI (0.679, 1.129) p=0.306 103 (87.3%) patients switched from placebo to sunitinib treatment 0 26 52 78 104 130 156 182 208 234 Time (Week)
Naïve Analyses ITT (naïve) Cox Model Hazard Ratio (SU/PB), 95% CI and p-value 0.876 (0.679, 1.129), p=0.306 Dropping switchers 0.315 (0.178, 0.555), p<0.0001 Censoring at switch 0.825 (0.454, 1.499), p=0.527 Time-dependent treatment covariate 0.934 (0.520, 1.679), p=0.820
Overall Survival (Final, 2008) Crossover Adjusted by RPSFT Overall Survival Probability (%) 100 90 80 70 60 50 40 30 20 10 0 Sunitinib (N=207) Placebo (N=105) Sunitinib (N=243) Median 72.7 weeks 95% CI (61.3, 83.0) Placebo (N=118) Median* 39.0weeks 95% CI (28.0, 54.1) Hazard Ratio=0.505 95% CI** (0.262, 1.134) p=0.306 0 26 52 78 104 130 156 182 208 234 Time (Week) *Estimated by RPSFT model ** Empirical 95% CI obtained using bootstrap samples.
Use of external data Control arm without exposure to experimental RCT RWE Retrospective historical Prospective contemporaneous Ideally Use as external validation of analyses, or quasi-control Comparability to pivotal RCT control arm is key - Patient characteristics, eligibility criteria, treatment, time, location, clinical practice, outcome 23
Method selection process Specific process proposed in NICE DSU Technical Support Document 16 (Latimer and Abrams) Key steps (in general): 1. Can the model be fitted with available data? 2. If yes, is the model appropriate given the switching mechanism? 3. If yes, are the assumptions reasonable? 4. If yes, are the results plausible?
Summary of data collection requirements for complex methods Data RPSFTM /IPE 2-stage IPCW Date of starting switch treatment!!! Date of death/censoring!!! Date of stopping switch treatment!* All baseline covariates that may!! influence switch decision or OS All time varying covariates that may! influence switch decision or OS, collected until switch or death/censoring All time varying covariates that may! influence switch decision or OS, collected until secondary baseline Date of secondary baseline (usually disease progression)!! * for on-treatment approach only 25 Increasing data collection burden
Summary of key assumptions for switch adjustment methods Method ITT Exclude switchers Censor switchers Time varying covariate IPCW RPSFTM IPE 2-stage Key assumptions Switch treatment ineffective No confounders (unlikely) No confounders (unlikely) No confounders (unlikely) No unmeasured confounders Constant treatment effect Balance due to randomisation Constant treatment effect Balance due to randomisation Parametric distribution No unmeasured confounders (stronger assumption than IPCW as fewer covariates in model) 26 If you spend long enough on an island of one-eyed trolls, you will eventually be able to decide which is the most beautiful
Best practice recommendations - Trial design Up front planning at design stage Be clear about the question, rationale, customer Describe intent to do analyses in protocol Define switch treatment (drug(s), dose(s) etc) Consider external data sources Collect the data to enable robust analysis 27
Best practice recommendations - Analysis Pre-specify preferred adjustment method in protocol/ analysis plan with justification Follow a transparent model selection process Don t use naive as sole or primary method Don t use IPCW/2-stage AFT if most patients switch or near perfect predictor of switch Don t use RPSFTM if near equal exposure or survival Always present ITT Provide results from a range of sensitivity analyses 28 External data assess and take steps to maximise comparability
Summary and looking to the future No method is universally best Situation specific assessment required Best practice recommendations have been outlined Be clear on the decision problem Collect the right data Up front planning Further work and research is needed: 29 Multiple switches Multiple switch treatments Different or less strong assumptions Binary or continuous data Software
30 Q & A
References (1) Watkins C et al. Adjusting overall survival for treatment switches: Commonly used methods and practical application. Pharm Stats 2013 Nov-Dec;12(6):348-57 Latimer NR and Abrams KR. NICE DSU Technical Support Document 16: Adjusting survival time estimates in the presence of treatment switching. (2014). Available from http://www.nicedsu.org.uk Demetri GD et al. Complete longitudinal analyses of the randomized, placebo-controlled, Phase III trial of sunitinib in patients with gastrointestinal stromal tumor following imatinib failure. Clin Cancer Res 2012; 18:3170-3179 Fukuoka M et al. Biomarker analyses and final overall survival results from a Phase III, randomized, open label, first-line study of gefitinib versus carboplatin/paclitaxel in clinically selected patients with advanced non-small cell lung cancer in Asia (IPASS). J Clin Oncol 2011; 29(21):2866-2874 Hotta et al. Progression-free survival and overall survival in phase III trials of molecular-targeted agents in advanced non-small-cell lung cancer. Lung Cancer 2013; 79:20-26 Robins J and Finkelstein D. Correcting for Noncompliance and Dependent Censoring in an AIDS Clinical Trial with Inverse Probability of Censoring Weighted (IPCW) Log-Rank Tests. Biometrics 2000; 56(3):779-788 D Agostino RB et al. Relation of pooled logistic regression to time dependent Cox regression analysis: The Framingham Heart Study. Stat Med 1990; 9:1501-1515 Fewell Z et al. Controlling for a time-dependent confounding using marginal structural models. The Stata Journal 2004; 4(4):402-420 Hernan MA et al. Marginal structural models to estimate the causal effect of zidovudine on the survival of HIVpositive men. Epidemiology 2000; 11(5):561-570 31
References (2) Robins JM and Tsiatis A. Correcting for non-compliers in randomised trials using rank-preserving structural failure time models. Communications in Statistics - Theory and Methods 1991; 20:2609-2631. White IR et al. strbee: Randomization-based efficacy estimator. The Stata Journal 2002; 2(2):140-150 Huang X and Xu Q. Adjusting the Crossover Effect in Survival Analysis Using a Rank Preserving Structural Failure Time Model: The Case of Sunitinib GIST Trial. MRC HTMR workshop, Feb 2012, available here 32