Créer une présentation
Télécharger la présentation

Télécharger la présentation
## WHY BAYES? INNOVATIONS IN CLINICAL TRIAL DESIGN & ANALYSIS

- - - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - - -

**WHY BAYES?INNOVATIONS IN CLINICAL TRIAL DESIGN & ANALYSIS**Donald A. Berry dberry@mdanderson.org**Conclusion These data add to the growing evidence that**supports the regular use of aspirin and other NSAIDs … as effective chemopreventive agents for breast cancer.**Results Ever use of aspirin or other NSAIDs … was**reported in 301 cases (20.9%) and 345 controls (24.3%) (odds ratio 0.80, 95% CI 0.66-0.97).**Bayesian analysis?**• Naïve Bayesian analysis of “Results” is wrong • Gives Bayesians a bad name • Any naïve frequentist analysis is also wrong**What is Bayesian analysis?**Bayes' theorem: '(q|X) (q)*f(X|q) • Assess prior (subjective, include available evidence) • Construct model f for data**Implication: The Likelihood Principle**Where X is observed data, the likelihood function LX() = f(X|) contains all the information in an experiment relevant for inferences about **Short version of LP: Take data at face value**• Data: • Among cases: 301/1442 • Among controls: 345/1420 • But “Data” is deceptive • These are not the full data**The data**• Methods: • “Population-based case-control study of breast cancer” • “Study design published previously” • Aspirin/NSAIDs? (2.25-hr ?naire) • Includes superficial data: • Among cases: 301/1442 • Among controls: 345/1420 • Other studies (& fact published!!)**Silent multiplicities**• Are the most difficult problems in statistical inference • Can render what we do irrelevant —and wrong! **Which city is furthest north?**• Portland, OR • Portland, ME • Milan, Italy • Vladivostok, Russia**Beating a dead horse . . .**• Piattelli-Palmarini (inevitable illusions) asks: “I have just tossed a coin 7 times.” Which did I get? 1: THHTHTT 2: TTTTTTT • Most people say 1. But “the probabilities are totally even” • Most people are right; he’s totally wrong! • Data: He presented us with 1 & 2! • Piattelli-Palmarini (inevitable illusions) asks: “I have just tossed a coin 7 times.” Which did I get? 1: THHTHTT 2: TTTTTTT • Most people say 1. But “the probabilities are totally even” • Most people are right; he’s totally wrong! • Data: He presented us with 1 & 2!**THHTHTT or TTTTTTT?**• LR = Bayes factor of 1 over 2 = P(Wrote 1&2 | Got 1) P(Wrote 1&2 | Got 2) • LR > 1 P(Got 1|Wrote 1&2) > 1/2 • Eg: LR = (1/2)/(1/42) = 21 P(Got 1|Wrote 1&2) = 21/22 = 95% • [Probs “totally even” if a coin was used to generate the alternative sequence]**Marker/dose interaction**Marker negative Marker positive**Proportional hazards model**Variable Comp RelRisk P #PosNodes 10/1 2.7 <0.001 MenoStatus pre/post 1.5 0.05 TumorSize T2/T1 2.6 <0.001 Dose –– –– NS Marker 50/0 4.0 <0.001 MarkerxDose –– –– <0.001 This analysis is wrong!**Data at face value?**• How identified? • Why am I showing you these results? • What am I not showing you? • What related studies show?**Solutions?**• Short answer: I don’t know! • A solution: • Supervise experiment yourself • Become an expert on substance • Partial solution: • Supervise supervisors • Learn as much substance as you can • Danger: You risk projecting yourself as uniquely scientific**A consequence**• Statisticians come to believe NOTHING!!**OUTLINE**• Silent multiplicities • Bayes and predictive probabilities • Bayes as a frequentist tool • Adaptive designs: • Adaptive randomization • Investigating many phase II drugs • Seamless Phase II/III trial • Adaptive dose-response • Extraim analysis • Trial design as decision analysis**http://www.cfsan.fda.gov/~frf/bayesdl.html**http://www.prous.com/bayesian2004/**BAYES AND PREDICTIVE PROBABILITY**• Critical component of experimental design • In monitoring trials**Example calculation**• Data: 13 A's and 4 B's • Likelihood p13 (1–p)4**Laplace’s rule of succession**P(A wins next pair|data)= EP(A wins next pair|data, p)= E(p|data)= mean of Beta(14, 5)= 14/19**Suppose 17 more observations**P(A wins x of 17 | data) = EP(A wins x | data, p) = E[px(1–p)17–x| data, p] ( ) 17 x **Best fitting binomial vs. predictive probabilities**Binomial, p=14/19 Predictive, p ~ beta(14,5)**Example: Baxter’s DCLHb & predictive probabilities**• Diaspirin Cross-Linked Hemoglobin • Blood substitute; emergency trauma • Randomized controlled trial (1996+) • Treatment: DCLHb • Control: saline • N = 850 (= 2x425) • Endpoint: death**Waiver of informed consent**• Data Monitoring Committee • First DMC meeting: DCLHb Saline Dead 21 (43%) 8 (20%) Alive 28 33 Total 49 41 • P-value? No formal interim analysis**Predictive probability of future results (after n = 850)**• Probability of significant survival benefit for DCLHb after 850 patients: 0.00045 • DMC paused trial: Covariates? • No imbalance • DMC stopped trial**OUTLINE**• Silent multiplicities • Bayes and predictive probabilities • Bayes as a frequentist tool • Adaptive designs: • Adaptive randomization • Investigating many phase II drugs • Seamless Phase II/III trial • Adaptive dose-response • Extraim analysis • Trial design as decision analysis**BAYES AS A FREQUENTIST TOOL**• Design a Bayesian trial • Check operating characteristics • Adjust design to get = 0.05 • frequentist design • That’s fine! • We have 50+ such trials at MDACC**OUTLINE**• Silent multiplicities • Bayes and predictive probabilities • Bayes as a frequentist tool • Adaptive designs: • Adaptive randomization • Investigating many phase II drugs • Seamless Phase II/III trial • Adaptive dose-response • Extraim analysis • Trial design as decision analysis**ADAPTIVE DESIGN**• Look at accumulating data … without blushing • Update probabilities • Find predictive probabilities • Modify future course of trial • Give details in protocol • Simulate to find operating characteristics**OUTLINE**• Silent multiplicities • Bayes and predictive probabilities • Bayes as a frequentist tool • Adaptive designs: • Adaptive randomization • Investigating many phase II drugs • Seamless Phase II/III trial • Adaptive dose-response • Extraim analysis • Trial design as decision analysis**Giles, et al JCO (2003)**• Troxacitabine (T) in acute myeloid leukemia (AML) when combined with cytarabine (A) or idarubicin (I) • Adaptive randomization to: IA vs TA vs TI • Max n = 75 • End point: CR (time to CR < 50 days)**Randomization**• Adaptive • Assign 1/3 to IA (standard) throughout (unless only 2 arms) • Adaptive to TA and TI based on current results • Final results **Drop**TI Compare n = 75**Summary of results**CR rates: • IA: 10/18 = 56% • TA: 3/11 = 27% • TI: 0/5 = 0% Criticisms . . .**OUTLINE**• Silent multiplicities • Bayes and predictive probabilities • Bayes as a frequentist tool • Adaptive designs: • Adaptive randomization • Investigating many phase II drugs • Seamless Phase II/III trial • Adaptive dose-response • Extraim analysis • Trial design as decision analysis**Example: Adaptive allocation of therapies**• Design for phase II: Many drugs • Advanced breast cancer (MDA); endpoint is tumor response • Goals: • Treat effectively • Learn quickly**Comparison: Standard designs**• One drug (or dose) at a time; no drug/dose comparisons • Typical comparison by null hypothesis: response rate = 20% • Progress is slow!**Standard designs**• One stage, 14 patients: • If 0 responses then stop • If ≥ 1 response then phase III • Two stages, first stage 20 patients: • If ≤ 4 or ≥ 9 responses then stop • Else second set of 20 patients**An adaptive allocation**• When assigning next patient, find r = P(rate ≥ 20%|data) for each drug [Or, r = P(drug is best|data)] • Assign drugs in proportion to r • Add drugs as become available • Drop drugs that have small r • Drugs with large r phase III**Suppose 10 drugs, 200 patients**• 9 drugs have mix of response rates 20% & 40%, 1 (“nugget”) has 60% • Standard 2-stage design finds nugget with probability < 70% (After 110 patients on average) • Adaptive design finds nugget with probability > 99% (After about 50 patients on average) • Adaptive also better at finding 40%**Suppose 100 drugs, 2000 patients**• 99 drugs have mix of response rates 20% & 40%, 1 (“nugget”) has 60% • Standard 2-stage design finds nugget with probability < 70% (After 1100 patients on average) • Adaptive design finds nugget with probability > 99% (After about 500 patients on average) • Adaptive also better at finding 40%