CTOS 2012 - Prague BAYESAN STATISTICS FOR CLINICAL INVESTIGATORS Paolo Bruzzi National Institute for Cancer Research Genoa - Italy “Bayesian”? • Predictive value of diagnostic tests • Studies of expression profiles (micorarrays) • Trials with a Bayesian component – Interim analyses – Bayesian design Differences between Conventional and Bayesian Statistics • Meaning of probability • Use of prior evidence Conventional P Probability of an observation Bayesian Probability Probability of a hypothesis Conventional P Probability of observing what was actually observed if H0 true Bayesian Probability Probability of H0/H1/H2/H3… given observed data (and prior distribution) Conventional P Probability of the observed difference (if the experimental therapy does not work) Bayesian Probability Probability that the experimental therapy works/doesn’t work (given observed difference and prior knowledge) Examples - Diagnosis Mr. XY shows a lung nodule at TC • Frequentist Probability that Mr. XY shows a lung nodule (if he doesn’t have lung cancer) • Bayesian Probability that Mr. XY has lung cancer (given his nodule and his prior risk) Examples - Diagnosis Mr. XY shows a lung nodule at TC • Frequentist Probability that Mr. XY shows a lung nodule (if he doesn’t have lung cancer) • Bayesian Probability that Mr. XY has lung cancer (given his nodule and his prior risk) Examples – Clinical trial Control therapy (CT): 5/40 responses Experimental therapy (ET): 10/40 responses • Frequentist Probability to observe 10/40 vs 5/40 if CT and ET are identical • Bayesian Probability that CT and ET are identical (given 5/40, 10/40 and prior knowledge) Examples – Clinical trial Control therapy (CT): 5/40 responses Experimental therapy (ET): 10/40 responses • Frequentist Probability to observe 10/40 vs 5/40 if CT and ET are identical P= 0.15 Not significant Ho (CT & ET identical) NOT REJECTED Examples – Clinical trial Control therapy (CT): 5/40 responses Old drugs, new schedule 10/40 responses • Frequentist Probability to observe 10/40 vs 5/40 if CT and ET are identical P= 0.15 Not significant Ho (CT & ET identical) NOT REJECTED Examples – Clinical trial Control therapy (CT): 5/40 responses Very Promising ther. (ET): 10/40 responses • Frequentist Probability to observe 10/40 vs 5/40 if CT and ET are identical P= 0.15 Not significant Ho (CT & ET identical) NOT REJECTED Examples – Clinical trial Control therapy (CT): 5/40 responses Experimental therapy (ET): 10/40 responses • Bayesian Probability that CT and ET are identical ? (given 5/40, 10/40 and prior knowledge) It depends on prior knowledge! Examples – Clinical trial Control therapy (CT): 5/40 responses Herbal + standard (ET): 10/40 responses • Bayesian Probability that CT and ET are identical (=herbal therapy not effective)? Still high! Examples – Clinical trial Control therapy (CT): 5/40 responses Very Promising ther. (ET): 10/40 responses • Bayesian Probability that CT and ET are identical (that is, new therapy not effective)? Quite low! NOTE The different meaning of Bayesian probability in theory, has • Implications for (clinical) decision analysis • No peculiar implications for rare tumors Differences between Conventional and Bayesian Approaches • Meaning of probability • Use of prior evidence Conventional P Probability of the observed difference (if the experimental therapy does not work) Bayesian Probability Probability that the experimental therapy works/doesn’t work (given observed difference and prior knowledge) Conventional Statistical Reasoning 1. Starting hypothesis (H0): new treatment = standard one Conventional Statistical Reasoning 1. Starting hypothesis (H0): new treatment = standard one 2. To demonstrate: new treatment >> standard, reject null hypothesis Conventional Statistical Reasoning 1. Starting hypothesis (H0): new treatment = standard one 2. To demonstrate: new treatment >> standard, reject null hypothesis 3. To this purpose, only evidence collected within one or more trials aimed at falsifying it can be used Conventional Statistical Reasoning 1. Starting hypothesis (H0): new treatment = standard one 2. To demonstrate: new treatment >> standard, reject null hypothesis 3. To this purpose, only evidence collected within one or more trials aimed at falsifying it can be used -> LARGE SAMPLE SIZE Conventional Statistical Reasoning To this purpose, only evidence collected within one or more trials aimed at falsifying it can be used -> LARGE SAMPLE SIZE No use of – External evidence – Evidence in favor of… Example Question: Efficacy of radiochemotherapy in a tumor type very rare in a site (e.g. squamous histology in stomach c.) External evidence: RX+CTX is effective in squamous cancers in more common sites Evidence in favor of..: The observed response rate is very high (e.g. 6/10) Does this information affect …. - the sample size of the phase III trial aimed to assess RT+CTX in squamous gastric c. ? - the analysis of its results (p value)? Squamous gastric cancer Planning a trial of RT+CTX Analysing its results (p value) Squamous gastric cancer Planning a trial of RT+CTX Herbal therapy Analysing its results (p value) Squamous gastric cancer Planning a trial of RT+CTX Herbal therapy Analysing its results (p value) Conventional (frequentist) statistical reasoning Exclusive reliance on experimental evidence Large Trials Large Trials Implication in rare tumors: Generic Selection criteria (All STS’s + Stage + treatment line) - Appropriate for chemotherapy trials - Possibly inappropriate for trials of Targeted Therapies Conventional (frequentist) statistical reasoning Experimental evidence Bayesian statistical reasoning Experimental evidence + Previous Knowledge Bayesian Approach Prior Evidence + Experimental Evidence Posterior Probability Distribution Conventional Probability Probability of a positive test given disease/no disease (Sensitivity, specificity) Bayesian Probability Probability of disease given test result and disease prevalence (Predictive value) Previous Knowledge? • Biological rationale • Evidence of activity • Efficacy in other diseases with similarities • Efficacy in other stages of the same disease Prior evidence in Bayesian statistics • Needed in order to compute posterior probability Prior evidence in Bayesian statistics • Needed in order to compute posterior probability • It must be transformed into a probability distribution (shape, mean, median, standard deviation, percentiles, etc) Prior evidence in Bayesian statistics • Needed in order to compute posterior probability • It must be transformed into a probability distribution • Based on – Objective information – Subjective explicit beliefs – Both Prior evidence in Bayesian statistics • Note: The difference between Bayesian and conventional statistics decreases with increasing strength of the empirical evidence Prior evidence in Bayesian statistics Difference between Bayesian and conventional statistics Young man, never smoked, no family history Prior probability of lung c.=1/100.000 - Nodule at routine x-ray - Suspicious lesion at TC - Histological confirmation after biopsy Posterior Probability? Prior evidence in Bayesian statistics Difference between Bayesian and conventional statistics Epidermoid lung cancer: Palliative care prolongs survival: A priori: 1-30% Small monocentric trial: HR= 0.75 (0.55-0.95) Large multicentric trial: HR= 0.8 (0.7 -0.9) Meta-analysis of 8 trials:HR= 0.82 (0.76-0.88) Prior evidence in Bayesian statistics Frequent tumors Prior direct evidence Estimates Indirect evidence? Empirical evid. Prior evidence in Bayesian statistics Rare tumors Prior direct evidence Estimates ? Indirect evidence! Empirical evid. Need to use all the available information -Direct evidence -Experimental -Prior -Indirect evidence Example Mortality Tumor X Nil vs A 15% vs 12.5% N=12000 P = 0.0007 H0 Rejected: A is effective in X Example Mortality Tumor X Nil vs A 15% vs 12.5% N=12000 P = 0.0007 Tumor Y N= 240 Nil vs A 15% vs 7.5% P=0.066 H0 not rejected: A not shown effective in y Prior Information: Mortality Tumor X Nil vs A 15% vs 12.5% N=12000 P = 0.0007 Tumor Y N= 240 Nil vs A 15% vs 7.5% P=0.066 Prior Information: X and Y are BRAF+ Mortality Tumor X Nil vs A 15% vs 12.5% N=12000 P = 0.0007 Tumor Y N= 240 Nil vs A 15% vs 7.5% P=0.066 Prior Information: X and Y are BRAF+ A = Anti BRAF Mortality Tumor X Nil vs A 15% vs 12.5% N=12000 P = 0.0007 Tumor Y N= 240 Nil vs A 15% vs 7.5% P=0.066 Prior Information: X and Y are BRAF+ A = Anti BRAF Mortality Tumor X Nil vs A 15% vs 12.5% N=12000 P = 0.0007 Tumor Y N= 240 Nil vs A 15% vs 7.5% P=0.066 INTERPRETATION? A different example 1. Study Backgound Uterine Sarcomas Stage I-II - High rate of distant relapses - Adjuvant pelvic RT in US: No impact on OS - Adjuvant CTX: Scanty and inconsistent data: 3 cycles of + RT compared with an historical control group of RT alone: 3-year DFS: RT alone= 43% CT+RT = 76% 2. Study Aims …a randomized trial to confirm or not a benefit in terms of PFS and OS. …a multicentric phase III study, comparing API chemotherapy regimen followed by RT versus RT alone for patients with localized US after complete surgery. 3. Statistical considerations • Primary Endpoint: 3-yrs EFS (OS secondary) • Sample Size: 256 patients (128 in each arm) - 80% power - Delta = 20% diff. in EFS (from 35% to 55%) - 2-sided 5% significance level 4. Results • In 8 years (October 2001- July 2009) 81 patients randomized in 19 institutions “Study was stopped because of lack of recruitment” • 3-year DFS = 55% vs 41% (P = 0.048) • 3 years OS = 81% in arm A vs 69% (P= 0.41) Conclusions “We have shown for the first time a statistical impact of adjuvant chemotherapy on DFS in this population of 81 patients without impact on OS yet.” Conventional analysis • 3-year DFS = 55% vs 41% (P = 0.048) Probability of observing this difference if CT not effective (H0) on DFS = 4.8% -> Reject H0 • 3 years OS = 81% vs 69% (p=0.41) Probability of observing this difference if CT not effective (H0) = 41% -> “No difference” in OS Bayesian Analysis of OS Background: a) Expected effect on DFS: 20% improvement b) Any effect on OS mediated by the effect on DFS c) Observed effect on DFS = 14% improvement in 3-year DFS (55% vs 41%) LIKELY EFFECT ON OS? Predicted and Observed Effect of CT on OS Endpoint Delta (95% CI) EFS 14% (1% to 27%) Predicted Effect on OS? -20% -10% 0 RT better +10% +20% +30% +40% +50% CT better Predicted and Observed Effect of CT on OS Endpoint Delta (95% CI) EFS 14% (1% to 27%) Observed Effect on OS OS 12% -20% -10% 0 RT better +10% +20% +30% +40% +50% CT better Predicted and Observed Effect of CT on OS Endpoint Delta (95% CI) EFS 14% (1% to 27%) However…. OS 12% (-18% to 42%) -20% -10% 0 RT better +10% +20% +30% +40% +50% CT better Bayesian Analysis of OS a) Expected effect on DFS: 20% improvement b) Any effect on OS mediated by the effect on DFS c) Observed effect on DFS = 14% improvement in 3-year DFS (55% vs 41%) LIKELY EFFECT ON OS? Likely effect on OS? Needed step: Make assumptions on the type and strength of the association between DFS and OS Assumptions used (based on Adjuvant studies in Breast c. and Colon c.): 1. Effect on OS 20%weaker than effect on DFS 2. Weak correlation (R2 = 0.50) 3. (Sensitivity analyses) Bayesian Analysis of OS a) Expected effect on DFS: 20% improvement b) Any effect on OS mediated by the effect on DFS c) Observed effect on DFS = 14% improvement in 3-year DFS (55% vs 41%) LIKELY EFFECT ON OS = Prior Probability distribution: 5-14% improvement Bayesian Analysis of OS Prior Probability distribution: 5-14% improvement + Observed Effect on OS: 12% improvement = Posterior Probability of effect on OS Predicted and Observed Effect of CT on OS Endpoint Delta EFS 14% (1% to 27%) OS (95% CI) 12% (-18% to 42%) -20% -10% 0 RT better +10% +20% +30% +40% +50% CT better Predicted and Observed Effect of CT on OS Endpoint Delta (95% CI) EFS 14% (1% to 27%) OS Credible Effect on OS 10% (1% to 20%) -20% -10% 0 RT better +10% +20% +30% +40% +50% CT better Predicted and Observed Effect of CT on OS Endpoint Delta (95% CI) EFS 14% (1% to 27%) OS NOTE 10% (1% to 20%) -20% -10% 0 RT better +10% +20% +30% +40% +50% CT better Hazard Ratios for adjuvant therapies Endpoint CT +RT RT HR 0.67 0.56 0.54 French Study, Sarcomas 3-yrs DFS 55% 41% 3-yrs OS 81% 69% 5-yrs OS 72% 55% Breast c OS Colon c. OS 0.6-0.8 0.7-0.9 Conclusions of the study • Conventional analysis: Significant effect on EFS, no effect on OS • Bayesian analysis: a remarkable effect on EFS (10-20% absolute improvement) and on Overall Survival (5-15% absolute improvement) Conclusion of the presentation In Rare tumors - Limited prior direct evidence - Limited experimental evidence IT IS NECESSARY TO USE BAYESIAN STATISTICS BAYESIAN STATISTICS - Not a mysterious statistical religion for adepts who trust in it - Not a confortable container where everyone can accomodate his crazy beliefs -Not a sneaky statistical machinery to fool referees and regulatory agencies BAYESIAN STATISTICS It is simply - a tool to reproduce human reasoning, - by connecting decisions to knowledge, - in a transparent fashion