DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Adaptive Trial Designs in Pediatric Studies: A Survey of Methods and Practices Alfred Balch, PhD Associate Professor, Pediatrics Division of Clinical Pharmacology & Clinical Trials Office Department of Pediatrics, University of Utah School of Medicine Methods in Progress 18-Oct- 2012 DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Outline • What are adaptive trials and why do we need to do them? • Confirmatory Trials – Group Sequential » Alpha-spending approach » Seamless Phase II-III – Sample Size Re-Estimation • Dose Finding Trials – – – – – Up and Down Designs: 3+3 and others CRM Bayesian Models Adaptive Randomization Using Efficacy and Toxicity: Utility Response DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology What are adaptive Trials? “By adaptive design we refer to a clinical study design that uses accumulating data to decide how to modify aspects of the study as it continues, without undermining the validity and integrity of the trial. The goal of adaptive designs is to learn from the accumulating data and to apply what is learned as quickly as possible. In such trials, changes are made “by design,” and not on an ad hoc basis; therefore, adaptation is a design feature aimed to enhance the trial, not a remedy for inadequate planning. “Adaptive Designs in clinical drug development- An Executive Summary of the PhRMA working Group” (2006) DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology What kind of adaptation? • Number of Patients Enrolled: – Terminate for Futility or for Early Success – Sample Size-Re-Estimation • Variance-Based • Proportion-of-Completers based • Type of Patients Enrolled – Enroll Future Patients with Likely Success • Patient Characteristics: Genotype, Age, Gender – Enroll Patients from Successful Centers – Change Assignment/Sampling to minimize covariate biases (?) • Balance treatment Assignment DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology What kind of adaptation? • Better Medical Practice Based On Partial Information: – “Play the winner” • Preferentially Enroll Patients into “Better” Treatment Groups – Need to preserve patient randomization to treatment, if possible • Dose-Finding- Enroll in “Better Dose” – Better Regimens: Dose-Interval and Route – Optimal (maximal information) Dose • Preferentially Recruit “Better” Patients – Need to preserve randomization to treatment, if possible • Advantageous Co-therapies for all incremental enrolled patients DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology What’s the big deal? • Why were earlier trials inflexible-what are the issues? – Trials for Drug Registration need to maintain nominal level of significance-need to account for multiple looks at the data – This is not achieved for: • Multiple looks at the data without statistical adjustment for multiplicity – Armitage et al.(1969- unadj pval=0.142 for 5 looks nominal 0.05) • Change in Endpoint • Complicated Multiple Hypotheses • Data-Driven changes in primary analysis – The largest “response” may not be reproducible. – In survival-type trials, bad news comes early. Need to be aware, take early results with appropriate skepticism. DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Why do we “Need” more flexibility? Armin Koch: BfArM (Bundesinstituts für Arzneimittel und Medizinprodukte) • Sample size not correct: “Quote marks” and coloring are mine! – For example: Study assumptions were not accurate • Was blinding maintained during sample size recalculation? If not, this is multiple look. • Endpoint not well-chosen: – Wrong doses or treatment – Suboptimal statistical methodology • • (Was blinding maintained when Stats were changed?) Change of objective: – – – – – – Superiority was an over-ambitious aim Change in criteria for inclusion or exclusion CRF’s limited or too extensive composite endpoint does not differentiate between treatments treatment effect in a different variable wrong responder-definition 11 DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Adaptive Method 1:Group-Sequential methods – Preplanned (protocol/SAP) interim looks at the data • k = number of looks (including final) – Usually presented as equal increment group sizes • Z1, Z2, … Zk Calculated Test Statistics at ith look (i=1,k) – Based on all data observed up to look i • • • • • z1,…zk critical values Stop and declare efficacy if Z i > zi for any i (including i=k) Perhaps stop for futility if Z i < -zi (or alternate threshold) For k=1 look this is just standard normal critical value Test needs to adjust z-criterion so overall alpha=0.05 (or 0.025 or…) for this procedure DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Sequential Test Statistics: Normal Case – In normal distribution case, in ith look, adjust Critical value of Zi to allow for multiple looks at the data, taking into account that results are correlated (k interim groups of size m) This is a random walk like process…. DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Pocock’s Method: • Critical Values for K-multiple looks (Pocock (1977)) – Z remains the same for all looks… – Define critical values, Cp(K,α) Pocock Critical Values for 2-Sided Tests K 1 2 3 4 5 10 20 α=0.01 2.576 2.772 2.873 2.939 2.986 3.117 3.225 α=0.05 1.96 2.178 2.289 2.361 2.413 2.555 2.672 α=0.10 1.645 1.875 1.992 2.067 2.122 2.27 2.392 DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology O’Brien and Fleming Boundary O.F. Critical Values for 2-Sided Tests K 1 2 3 4 5 10 20 α=0.01 2.576 2.580 2.595 2.609 2.621 2.660 2.695 α=0.05 1.96 1.977 2.004 2.024 2.040 2.087 2.126 α=0.10 1.645 1.678 1.710 1.733 1.751 1.801 1.842 DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Families of Boundaries: • Wang and Tsiatis (1987)) – Use parameter Δ » Δ=.5 => Pocock » Δ=.0 => O’Brien Fleming – Delta parameter emphasizes/de-emphasizes final Z • Lots of additional work on optimal spending function(s)– Many generalize Wang & Tsiatis further » Kim & Demets(1987) » Hwang, Shih, De Cani (1990) » 2 spending functions based on α and β DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Alpha-Spending Approach K. K. Gordon Lan (PhD in Mathematical Statistics from Columbia University) joined J&J in 2005 after holding positions at (NHLBI/NIH), George Washington University, Pfizer and Sanofi-Aventis. He has published more than 60 research papers in professional journals, and has given He is a Fellow of the ASA. 2012 Deming Conference Website DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Group-Sequential methods (cont) Keaven Anderson has a Ph. D. from Stanford University and is the head of late-stage oncology statistics at Merck Research Laboratories where he has worked since 2003. He is the primary author and maintainer of the open source R package gsDesign for designing group sequential trials. Deming Conf on Applied Statistics DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Group-Sequential methods – Why more looks don’t cost much….Bonferoni is much too conservative: – Random Walks…think coin toss where can win or lose 1 dollar • “Real Case” is continuous analog where Brownian motion replaces Markov Chain – X1…Xn are iid on {-1,1} where • P(Xi = 1) = ½ • P(Xi= -1) = ½ – Sn+1= Sn + Xn+1 – Observe S1, S2, …Sn – How big can Si get before the coin clearly isn’t fair? Test at 0.05 level of significance…. – 2n equally likely outcomes- count paths which exceed a threshold value. This section is optional… DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Group-Sequential methods – Terminate Trial of 5 Tosses for success when a score of 4 is reached. Interim looks at 4 and 5 tosses. (K=2 looks) – For Five Coin Tosses What outcomes give this result? (Assuming a fair trial) – 32 equally likely possible outcomes: • THHHH, HTHHH, ….HHHHT • And for stopping at 4 tosses • HHHHT, HHHHH – But wait, we counted HHHHT twice (really just HHHH) • So 6 total outcomes result in terminate for success • P-value = 6/32 for this procedure under Ho – Note: This doesn’t change with an additional look at n=1 toss! DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Group-Sequential methods – Terminate Trial of 5 Tosses for success when a score of 4 is reached. Interim looks at 4 and 5 tosses. – For Five Coin Tosses What outcomes give this result? (Assuming a fair trial) • THHHH, HTHHH, ….HHHHT • And for stopping at 4 tosses • HHHHT, HHHHH – But wait, we counted HHHHT twice (really just HHHH) • So 6 total outcomes result in terminate for success • P-value = 6/32 for this procedure under Ho – Note: This doesn’t change with an additional look at n=1 toss! DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology -20 -10 0 Sum 10 20 30 Random Walk with 100 coin tosses, Fair Coin 0 20 40 60 number of trials 80 100 DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology What is the power? At .7 10 20 30 Stop for Success at 30 0 What is the level Of significance of This test? Sum 40 50 60 Random Walk with 100 coin tosses, p=.7 0 20 40 60 number of trials 80 100 DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Random Walk with 100 coin tosses, p=.5 5 0 -5 Sum 10 15 interim look at 60 patients 0 20 40 60 80 100 DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Summary: Group-Sequential methods – Under Ho, Adjusted Z scores form a continuous martingale process.(like a random walk- increments are iid) : E(S n+1 | S1…Sn)= Sn Under Ha there is a drift parameter – Pocock (1977) – O’Brien & Fleming (1979) – Generally O’Brien and Fleming is FDA-preferred; not as wide a boundary in initial looks as Pocock (“spend alpha late”). – Additionally, DSMB or statistician can choose to use alpha-spending (typically spend most of the alpha on final analysis, less in early stage for extreme case in early look). – Regulatory trials total alpha=0.025 DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Interim Looks and Blinding – Logistically don’t want a lot of interim looks at the data, need to clean and lock data for each look and deal with DSMB logisticsMathematically, interim looks with small alpha (big Z) are not a big deal. – Unblinding can be a risk, knowing the trial is not terminated can bias the future conduct of the trial. – Interim looks can save a lot of subjects when the data are a lot better or a lot worse than expected. Also if they are not provided in the protocol, they will happen anyway! – There is excellent free software available to do Group Sequential calculations. DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology gsDesign – Suite of GS software developed by Keaven Anderson at Merck – Example: (Runs under r-statistical programming language) – n.fix is the number reqd. for fixed sample size (dep on delta,sigma) DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology gsDesign Normal test statistics at bounds 3.25 2.99 2.69 2.37 2.03 Normal critical value 2 Bound 0 Lower Upper -2.03 -2 -2.37 -2.69 -2.99 -3.25 -4 N=164 200 N=328 N=492 400 N=656 600 Sample size N=819 800 DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Sample Size Re-Estimation – Variance Based • Suppose initial sample size calculation is based on poorly known estimate of variance? Can we fix this without unblinding the trial? • Variance and Mean estimates for normal distribution are independent • Independent statistician can calculate sample variance for observed data. • Trial statistician can use this estimate directly or use a weighted/Bayesian update of variance estimate to recalculate overall sample size. DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Sample SizeRe-Estimation – Dropout Rate Based: • Independent statistician can calculate dropout rate for observed data. • Protocol can specify number of completed for sample size • What are the unblinding risks?? – Are dropouts biased to treatment? – Will dropouts affect actual trial conduct/termination and hence impact α? DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Adaptive Exploratory Trials • In early trials, dosing frequency, route and total amount may need to be modified based on patient outcomes • Focus on learning (not confirming), successful outcome is a good decision about next trial (if!) and its design; should have ideas about dose, population, indication. DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Adaptive Exploratory Trials – Oldest: Mood & Dixon (1948) up and down • Idea: Converge on Tox rate of (say) 50% – Most common in oncology: 3+3 – Improvements proposed (Ed Korn, NIH, Gezmu & Flourney (2006); • Let R(dj) be the number of toxic responses for group j given dose dj , then R(dj) has a binomial distribution with probability dependent on dose. • Use logistic regression to model Dose-Response DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Traditional Adaptive FIH Oncology Exploratory Trials: 3+3 • Dose follows Fibonacci Sequence – 1,1,2,3,5,8 – Or modified so increments decrease as dose increases • 3+3 Design (somewhat adaptive) – Enter 3 at lowest dose – 0/3 DLT escalate to next dose – 1/3 DLT expand dose cohort to 6: » 1/3 + 0/3 treat next 3 at next higher dose » 1/3+1/3 This is MTD dose » If >2 DLT of 3 then reduce dose and repeat » Stop and DLT when <= 2/6 DLT DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Up and Down Designs- Generally – Define Cu.CL so that next cohort of patients is randomized to dose – dj-1 if R(dj) ≥ Cu – dj+1 if R(dj) ≤ CL – dj if CL < R(dj) <Cu Cu , CL are pre-defined to reflect risk tolerance. DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Up and Down Design (continued) • Features: • Next dose-cohort is based on information gathered from trial (modification: dj+1 can depend on Rj ). Dose sequence defines a Markov Chain • More aggressive dose-expansion when no problems at low dose(s) • Generalizes and Improves much used “3+3 design”. DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Bayesian Adaptive FIH Oncology Exploratory Trials: Continual Reassessment Method(CRM) DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Bayesian Adaptive Ph I-II trial:Escalation With Overdose Control(EWOC) DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Bayesian Adaptive Ph I-II trial:Escalation With Overdose Control(EWOC) DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Bayesian Adaptive Ph I-II trial: EWOC • Let Y be an indicator of DLT – FY(y) be the CDF of Y and define x* and x** such that • P(Y = 1|x = x∗) = 0, • P(Y = 1|x = x∗∗) = 1−ϵ (1−ϵ is “known” to be < ϴ) so prior on ϴ is supported on [x*,x**] Initial dose, d1 less than x**, is chosen DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Bayesian Adaptive Ph I-II trial: EWOC DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Bayesian Adaptive Ph I-II trial: EWOC • EWOC software is available free André Rogatko Ph.D Director, Biostatistics and Bioinformatics at Cedar Sinai. Dr. Rogatko has more than 100 published peer-reviewed articles From 2010 Deming Conference Website DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Not to be confused with EWOK! DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Traditional Non-Oncology FIH • Dose Increase by ½ log10 – – – – – 1,3,10,30,100 Single Ascending Dose (SAD) then Multiple Ascending Dose (MAD) Emphasis on PK, PD, PG Endpoints First Dose Based on Animal Tox • 6+2 Design – 2 subjects receive placebo at each dose » Control for safety and PD/Response » Stop with dose 1-2 DLT (in active trtd) – Not adaptive-may change dose, monitoring, etc., but usually not in a pre-planned way. DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Traditional Non-Oncology FIH – MCP-Mod- Adjust Dose-Reponse Function – Use Scheffe’ to allow multiple looks at Dose-response significance – Jose Pinheiro DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Bayesian non-oncology FIH • Adaptive Randomization – Use CRM strategy, but with utility replacing safety » Utility function can combine safety/efficacy – First Dose Response Based on Animal Tox, update DR curve with human data as cohort data becomes possible – Preferentially randomize into “best dose” – May need to span low doses to get human data. – SAE is still a stopper! – Design can be parallel or nested, dosing can be direct dose, or PKbased (safe concentrations, AUC, Cmax) DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Adaptive design methods • Innovations in clinical trials methodology have the potential to: – Improve the quality of knowledge, protect human subjects, and improve the efficiency of clinical research • Adaptive trials are often ethically required and will happen whether planned or not. – Better to plan them as far as possible DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Review: Adaptive Trials • Adaptive designs can be classified into two categories: – Exploratory/Dose Finding – Confirmatory • An adaptive design allows modifications to be made to trial and/or statistical procedures of on-going clinical trials. • Adaptation may/should be pre-specified by a rule: DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Concerns • Concern that the actual patient population after the adaptations could deviate from the originally target patient population • Could lead to an increased overall type I error rate (to erroneously claim efficacy for an infective drug). • Analysis needs to be adjusted appropriately for adaptive design • Major adaptations of trial and/or statistical procedures of ongoing trials may result in a totally different trial that is unable to address the scientific/medical questions the trial intends to answer. DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Benefits of all adaptive trials – Patient welfare is paramount, need to be able to ethically enroll patients in ongoing doses/arms of trials – In pharmaceutical/clinical research and development has become popular due to its flexibility and efficiency for identifying potential signals of clinical benefit of the test treatment under investigation. – The value of adaptive clinical trial methods in exploratory phase (phase II) clinical research is generally well accepted. DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Resources* Websites: http://cran.r-project.org/web/packages/gsDesign/gsDesign.pdf Really great documentation of Keaven Anderson’s free gsDesign suite for the free r statistical language. https://biostatistics.mdanderson.org/SoftwareDownload/ Good variety of programs, including Adaptive Bayesian designs http://cran.r-project.org/web/packages/CRM/CRM.pdf R Program to do CRM –check other places around CRAN as well! http://arxiv.org/pdf/1011.6479.pdf Andre Rogotku “Escalation with Overdose Control”-software also available from the author (Cedar-Sinai) * Presentation will be made available after class 46 DEPARTMENT OF PEDIATRICS Division of Clinical Pharmacology Resources Knowledgeable about adaptive design and helpful for this talk! Dr. Michael Spigarelli Professor: Pediatrics, Internal Medicine and Pharmacy Dr. Catherine M. T. Sherwin, PhD, FCP Assistant Professor, Pediatrics Books Statistical Methods for Dose-Finding Experiments Ed. Sylvie Chevret (2006) John Wiley and Sons Good collection of Adaptive Design methods mostly for early development. Bayesian Adaptive Methods for Clinical Trials Scott M. Berry, et al. (2011) CRC Press All Bayesian- good approach, esp. for Oncology Group Sequential Methods with Applications to Clinical Trials C.Jennison, B. Turnbull (2000) CRC Press Nice Summary of GS methods and underlying theory-well developed 47