Practical and ethical advantages of Bayesian approaches in adaptive clinical trial designs Kristian Thorlund
Background This talk was previously given as an invited talk at a DSEN sponsored meeting on innovative clinical trials designs, Jan 31-2014, Ottawa 2
Background Health care is moving towards an era with a growing number of patient sub-populations that are rare and difficult to study in clinical trials Within Canada, there is a very high demand for clinical trial methodologists that can think beyond the conventional parallel design RCT 3
Objectives Generate a healthy debate on how to move forward and be innovative, yet conscientious, about trial designs in Canada Illustrate the type of thinking that may be required for trial designs in challenging populations 4
Outline Teaser Bayesian stereotypes and efficiency reflections Introduction to Bayesian analysis Introduction to Adaptive designs Illustrative example: hepatitis C 5
ECMO trials Phase I - CMT: 4/10 deaths; ECMO: 0/9 deaths Phase II ECMO 1/10 death Within the scientific literature this was criticized on both sides No patients should have been randomized to CMT Not enough patients were randomized to CMT Was O Rourke s approach a breaking point? 4 babies deaths was all the staff could take
Traditional hierarchy of evidence RCTs & Systematic reviews Cohort Studies Case-series Clinical expert opinions 12
In small populations Observational studies RCTs and SRs expert opinion 13
In small populations Highly Efficient Study Design Not so efficient study design 14
Bayesian Statistics Are Bayesian statistics the answer? when and how? 15
Concepts of Bayesian Statistics (gross oversimplification) Bayesian analysis is unique for allowing data to be mixed with prior (external) evidence or opinions. By contrast, the conventional Frequentist discipline of statistics is only data driven 16
Bayesian Stereotypes Not fully data driven Too complex Hard to trust due to subjectivity Goes against conventional EBM Too philosophical Buzz word.. Used by the cool statisticians Incorporates evidence of lower quality 17
Bayesian Stereotypes Please forget these stereotypes! Bayesian analysis is flexible it can be what you want, need and/or require of it to be Easy or complex Rigorous or exploratory Objective or subjective Broad or narrow 18
In small populations Highly Efficient Study Design Not so efficient study design 19
In small populations Bayesian Adaptive Designs Conventional parallel design RCTs 20
Concepts of Bayesian Statistics (gross oversimplification) Bayesian analysis is unique for allowing data to be mixed with prior (external) evidence or opinions. By contrast, the conventional Frequentist discipline of statistics is only data driven 21
Bayesian Statistical Inferences Priors take the form of probability distributions, and are mixed with the likelihood to shape a posterior probability distribution from which inferences are drawn Frequentist drawn inferences solely from the likelihood (but necessitates parametric distributional assumptions) 22
Bayesian priors Available data from RCTs (e.g. response to Tx) Drug A vs Drug B Difference in response 23
Bayesian priors + external data: clinical expert survey/obs. study Drug A vs Drug B Difference in response 24
Bayesian posterior distributions Adding the two together forms a posterior distribution 25
Bayesian philosophy (again gross oversimplification) Bayesians try to answer given the data, what are the likely effects and degrees of uncertainty Frequentist try to answer how likely are the estimated treatment effects to be observed in a study attempting to replicate the current data? 26
Bayesian thinking and adaptive designs Considering the underlying philosophies, Bayesian statistics and adaptive clinical trials seem to go hand in hand. Frequentist methods in adaptive clinical trials are conceptually counter-intuitive. Note: many researchers apply frequentist methods, but Bayesian thinking (without knowing it) 27
Adaptive designs Adaptive designs typically allow for flexible change in randomization ratio or elimination of treatment groups after interim analyses 28
Adaptive designs For example, a trial examining placebo, low dose, mid dose, and high dose, may start with 1:1:1:1 randomisation, but gradually randomise more and more to higher doses based on better observed responses in the mid and high dose arms This is Bayesian thinking relying only on data, or in other words, non-informed Bayesian statistics (i.e., employed priors are non-informative) 29
Adaptive designs Consider the placebo, low, mid, high dose example. After data is available for 200 patients, you make a decision on whether to change the 1:1:1:1 randomisation ratio 30
Adaptive designs Assume you see the following responses. What would your new randomization ratio be? 23/50 20/49 11/49 5/52 Placebo Low Dose Mid Dose High Dose 31
Adaptive designs Placebo Low dose Mid dose High Dose 50 patients/arm 100 patients/arm 32
Adaptive designs Placebo Low dose Mid dose High Dose 50 patients/arm 100 patients/arm 33
Adaptive designs Placebo Low dose Mid dose High Dose 50 patients/arm 100 patients/arm 34
Bayesian thinking in rare diseases No major design challenges in larger populations However, most diseases in pediatrics and rare populations are both rare and trialist face several difficulties enrolling a large number of patients 35
Adaptive designs in pediatrics Typically only a few centers (e.g., SickKids hospitals) can enroll patients and this limits the number of feasible candidate patients compared with adult trials Adapting to small numbers is prone to errors by the play of chance. Thus, without innovative twists these designs typically have little value for informing clinical practice 36
Adaptive designs in pediatrics Consider again a trial that starts with 1:1:1:1 randomisation to placebo, low, mid, and high dose. Say you can feasibly only enroll 50 patients over the trial period of 3 years. That is, no more than 50 patients will be available for any clinical study over the next 3 years. How realible is randomisation adjustments then half-way through? How useful it is? (pros and cons?) 37
Adaptive designs Which adjustments would be comfortable with the following numbers? 2/6 3/5 3/6 1/7 Placebo Low Dose Mid Dose High Dose 38
Adaptive designs You only have 26 patients left and need to make the most our of the evidence. 2/6 3/5 3/6 1/7 Placebo Low Dose Mid Dose High Dose 39
Adaptive designs What is external evidence on the same trial population was available on placebo? 3/5 3/6 2/6 3/19 4/35 1/7 External Evidence on Placebo Placebo Low Dose Mid Dose High Dose 40
Adaptive designs How many more patients do we need to randomise to placebo with this evidence? 3/5 3/6 2/6 3/19 4/35 1/7 External Evidence on Placebo Placebo Low Dose Mid Dose High Dose 41
Conventional vs Bayesian Adaptive Consider two scenarios 1. We ignore the external evidence and keep randomising 1:1:1:1 2. We include the external placebo evidence, stop randomizing to placebo and randomise 1:1:1 with the three doses 42
Scenario #1 Maintain 1:1:1:1 randomisation, no statisticallly significant difference detected 4/12 6/11 7/14 2/13 Placebo Low Dose Mid Dose High Dose P=0.11 43
Scenario #2 Stop randomising to placebo 9/17 9/18 6/18 3/19 4/35 1/7 External Evidence on Placebo Placebo Low Dose Mid Dose High Dose 44
Scenario #1 versus Scenario #2 #1 Placebo Response P=0.11 Mid Dose Response 0% 10% 20% 30% 40% 50% 60% 70% 80% #2 Placebo Response Evidence P>0.05 0% 10% 20% 30% 40% 50% 60% 70% 80% 45
Scenario #1 versus Scenario #2 #1 Placebo Response P=0.11 Mid Dose Response 0% 10% 20% 30% 40% 50% 60% 70% 80% Posterior Placebo Response #2 P<0.05 0% 10% 20% 30% 40% 50% 60% 70% 80% 46
Motivating example - HCV An increasing number of highly potent agents are available for treating hepatitis C in adults Conventional therapy, peginterferon+ribavirin is known to eradicate the virus children at the same rate as in adults and have similar or better safety profile than in adults (Druyts et al CID 2012, systematic review of 8 trials) 47
Motivating example - HCV Some of the most potent newer direct acting agents (DAAs) eradicate the virus in 90% of adults without co-administration of interferon (and are thus more safe) How are these agents likely to work in children? Can they reduce adverse events (e.g., anemia)? Can they avoid reduction in growth? 48
Motivating example - HCV Expected efficacy 90% 80-90% 50% Peg-Riba DAA+ Peg-Riba DAA 49
Motivating example - HCV Choosing the randomisation scheme n? n? n? Peg-Riba DAA+ Peg-Riba DAA 50
Motivating example - HCV Choosing the randomisation scheme n? n? 50% n? Peg-Riba External Evidence Peg-Riba DAA+ Peg-Riba DAA 51
Motivating example - HCV Expected safety (anemia, neutropenia, rash, ) 20% 20% 25% 10% Peg-Riba External Evidence Peg-Riba DAA+ Peg-Riba DAA 52
Motivating example - HCV Expected safety (anemia, neutropenia, rash, ) 20% n? n? n? Peg-Riba External Evidence Peg-Riba DAA+ Peg-Riba DAA 53
Motivating example - HCV Bayesian adaptive design: Borrows strength from systematic review to stop placebo randomization Could also borrow strength from adult population (confirm signal) Stop early for benefit/safety based on Bayesian significance 54
This is where we started Not fully data driven Too complex Hard to trust due to subjectivity Goes against conventional EBM Too philosophical Buzz word.. Used by the cool statisticians Incorporates evidence of lower quality 55
Hopefully this is where we are going Incorporates all relevant evidence Allows for transparent analysis Takes the step into the era beyond EBM Sufficiently flexible Helps us think about efficiency Only used by the cool statisticians (same as before) 56
THANK YOU! 57