Why Bayesian How approaches for CER? Donald A. Berry dberry@mdanderson.org 1 Outline • Bayesian Metaanalysis & CER (ICDs) (ICD) • Adaptive Clinical Trials (I-SPY2) • Modeling in CER using Multifarious Data Sources (CISNET) • Comparing Outcomes—Trials and Tribulations 2 Bayesian Meta-Analysis for Comparative Effectiveness and Informing Coverage Decisions: Application to Implantable Cardioverter Defibrillators* *Berry SM, Ishak J, Luce B, Berry DA. Medical Care (2010). Disclosure: Berry Consultants contract with Boston Scientific via UBC 3 What Bayes Adds Model sources of variation Mortality rates over time: changing hazards Address possible time- dependent effect of ICD Cumulative meta-analysis, illustrate effect of each new study: When was evidence conclusive? Predictive probabilities for future trials 4 Studies Included 5 Bayesian hierarchical modeling of time to death • Model 1: Proportional hazards • Model 2: Time-dependent hazard ratios (modeled separately by year) • Model 3: Hierarchical treatment effects; allow for different treatment effects in different trials 6 Hazard Rates & Survival: Models 1 & 2 Hazard rates Survival probabilities Control ICD Control ICD Model 1 Model 2 7 Results Summary Model 1 Model 2 (Time-dependent) (Proportional) RR ICD+ RR1 RR2 RR3 RR4 RR5 0.777 1.00 0.807 0.713 0.723 0.990 0.877 (0.036) (0.054) (0.063) (0.079) (0.161) (0.215) 0 ICD+ 0.999 8 Relative Risks over Time in Model 1 9 Predictive Probabilities over Time Predicted #3 Predicted #1 Observed RR 10 Some Conclusions • • • • ICD Effective: 23% hazard reduction Effect persistent, consistent Effect clear early on Possible to account for changing patient populations 11 Outline • Bayesian Metaanalysis & CER (ICD) • Adaptive Clinical Trials (I-SPY2) • Modeling in CER using Multifarious Data Sources (CISNET) • Comparing Outcomes—Trials and Tribulations 12 Current use of Bayesian adaptive designs • • • • MDACC (> 300 trials) Device companies (> 25 PMAs)* Drug companies (Most of top 40)** CER? Not yet. *http://www.fda.gov/MedicalDevicesDeviceRegulationandGuidance/ GuidanceDocuments/ucm071072.htm **http://www.fda.gov/downloads/DrugsGuidanceCompliance RegulatoryInformation/Guidances/UCM201790.pdf 13 Two Recent Pubs 14 A Bayesian statistical design was used with a range in sample size from 600 to 1800 patients. 15 16 Bayesian adaptive trials • Stopping early (or late) –Efficacy –Futility • • • • • Dose finding (& dose dropping) Seamless phases Population finding Treatment finding Ramping up accrual 17 Why? • Smaller trials (usually!) • More accurate conclusions and hence better treatment for patients, at lower cost (?) 18 I-SPY 2 Slides from press conference … (Change “Phase 2” to CER; “experimental” to “approved”) 19 Standard Phase 2 Cancer Drug Trials Population of patients Population of patients Experimental arm R A N D O M I Z E Outcome: Tumor shrinkage? Outcome: Longer time disease free 20 Standard Phase 2 Cancer Drug Trials Population of patients Population of patients Experimental drug Consequence: R 60-70% Failure A N of Phase 3 Trials D O M I Z E Outcome: Tumor shrinkage? Outcome: Longer time disease free 21 I-SPY2 TRIAL Population of patients A R D A A N P D T O I M V I E Z L E Y Outcome: Complete response at surgery 22 I-SPY2 TRIAL Population of patients A R D A A N P D T O I M V I E Z L E Y Outcome: Complete response at surgery Arm 2 graduates to small focused Phase 3 trial 23 I-SPY2 TRIAL Population of patients A R D A A N P D T O I M V I E Z L E Y Outcome: Complete response at surgery Arm 3 drops for futility 24 I-SPY2 TRIAL Population of patients A R D A A N P D T O I M V I E Z L E Y Outcome: Complete response at surgery Arm 5 graduates to small focused Phase 3 trial 25 I-SPY2 TRIAL Population of patients A R D A A N P D T O I M V I E Z L E Y Outcome: Complete response at surgery Arm 6 is added to the mix 26 Outline • Bayesian Metaanalysis & CER (ICD) • Adaptive Clinical Trials (I-SPY2) • Modeling in CER using Multifarious Data Sources (CISNET) • Comparing Outcomes—Trials and Tribulations 27 28 29 CNN: Statistical Blitz Helps Pin Down Mammography Benefits 30 Fig. 1, Berry JNCI 1998 Updates K S C O E H G M U 31 Fig. 2, Berry JNCI 1998 U 32 33 CISNET from NEJM Women 40-79 Node-positive BC 34 CISNET from NEJM 35 Percent reductions in BC mortality due to adjuvant Rx and screening 30 Due to Adjuvant Treatment 25 E 20 R W M 15 S G D 10 5 0 0 5 10 15 20 Due to Screening 25 30 36 Model(s) M 37 Accepted simulations E R M G W S D 38 Model M: Prior to Posterior (2 of several parameters) “the posterior mean effect of tamoxifen is 0.37, corresponding to a 37% decrease in the hazard of breast cancer mortality due to the use of 5 years of tamoxifen for ER-positive tumors in actual clinical practice.” Prior Posterior Posterior Prior 39 Breast Cancer Mortality Future BC mortality HP2010 BC Mortality / 100,000 Population 45 Background 40 T 14 - AI 10 T 14 - AI 40 35 T 14 - AI 10 - M Age 40+ T 14 - AI 40 - M Age 40+ 30 T 14 - AI 10 - M Age 50+ T 14 - AI 40 - M Age 50+ 25 20 T 40 - AI 10 HP 2010 Target T 40 - AI 40 T 40 - AI 10 - M Age 40+ T 40 - AI 40 - M Age 40+ 15 T 40 - AI 10 - M Age 50+ T 40 - AI 40 - M Age 50+ 10 2000 Rate Target 5 0 1975 Truth 1980 1985 1990 1995 2000 Year Year 2005 2010 2015 2020 40 Keeping track of costs (and their uncertainties) is straightforward with Bayesian simulations 41 Outline • Bayesian Metaanalysis & CER (ICD) • Adaptive Clinical Trials (I-SPY2) • Modeling in CER using Multifarious Data Sources (CISNET) • Comparing Outcomes—Trials and Tribulations 42 Newsweek: “What You Don’t Know Might Kill You” “The right doctors can make all the difference when it comes to treating cancer. So why don't we know who they are?” 43 Survival Outcomes, by Disease Stage Us: Them: 44 Local Artifact Truth is no difference 60% longer Comm Central “Will Rogers Effect” 100% longer Regional Advanced Comm Central 33% longer Community Central (years) survival(years) Mediansurvival Median Comparing Outcomes Overall Stage 45 Local Comm Central Comm Central Community Central Median survival (years) Using Central Staging Regional Advanced Overall Stage 46 Median survival (years) 10 5 0 Local Comm Central Comm Comm Central Central Community Community Central Central Using Community Staging 25 20 15 Regional Advanced Overall Stage 47 Back to Newsweek “A spokesperson for M.D. Anderson Cancer Center in Houston said, ‘We do not have outcomes data at this time,’ while a physician there explained that doctors don't want to release data ‘that's difficult for people to interpret.’” 48 What would Bayes do? Model disease stage, build experiments to bolster weak parts of the model. 49 Outline • Bayesian Metaanalysis & CER (ICD) • Adaptive Clinical Trials (I-SPY2) • Modeling in CER using Multifarious Data Sources (CISNET) • Comparing Outcomes—Trials and Tribulations 50