Adaptive Designs for U-shaped ( Umbrella) Dose-Response Yevgen Tymofyeyev Merck & Co. Inc September 12, 2008

Outline Utility function Gradient design Normal Dynamic Linear Models Two stage design Comparison and conclusions

Application Scope Proof-of-Concept and (or) Dose-Ranging Studies when  Dose-response can not be assumed monotonic but rather uni-modal  Efficacy and safety are considered combined by means of some utility function

Examples: Utility Function Utility = efficacy + Coefficient * AE_rate + 2 nd order_term  Utility is evaluated at each dose level Another way to define is by means of a table  Example below Drug Efficacy vs. PLB AE % vs. PBO

Example of Utility function (cont.) Try to modify utility function in order to reduce its variance while preserving the “bulk” structure.

Example (cont.)

Adaptive Design Applicable for ∩ DR maximization Frequent adaptation: Adaptations are made after each cohort of subjects’ responses, data driven  Gradient design Assume ∩ shape dose-response  NDLM No assumption for DR shape, but rather on “smoothness” of DR (dose levels are in order) Two stage design  No assumptions on treatment ordering  Inference is done at the adaptation point

Gradient Design for Umbrella Shaped Dose- Response

Case Study Therapeutic area: Neuroscience Outcome: composite score derived from several tests Objective: to maximize number of subjects assigned to the dose with the highest mean response, the peak dose  improve power for placebo versus the peak dose comparison Assumption: monotonic or (uni-modal) dose-response Proposed Method: Adaptive design that uses Kiefer-Wolfowitz (1951) procedure for finding maximum in the presence of random variability in the function evaluation as proposed by Ivanova et. al.(2008)

Doses Current cohortNext cohort At given point of the study, subjects are randomized to the levels of the current dose pair and placebo only. The next pair is obtained by shifting the current pair according to the estimated slope. Active pair of levels Illustration of the Design

Update rule Let dose j and j+1 constitute the current dose pair. 1.Use isotonic (unimodal) regression or quadratic regression fitted locally to estimate responses at all dose levels using all available data 2.Compute T i.If T > 0.3 then next dose pair (j+1,j+2), i.e. "move up“ ii.If T < -0.3 then next dose pair ( j-1, j), i.e. "move down“ iii.Otherwise, next dose pair ( j, j+1), i.e. “stay” If not possible to “move” dose pair, ( j=1 or j=K-1), change pair’s randomization probabilities from 1:1 to 2:1 (the extreme dose of the pair get twice more subjects) Modification of this rule (including different cutoffs for T) are possible but logic is similar

Inference after Adaptive Allocation Compare PLB to each dose using Dunnett’s adjustment for multiplicity (?) Compare PLB to the dose with the maximum number of subjects assigned (expect inflated type I error) ‘Trend’ Tests are not straightforward since an umbrella shape response is possible.

True Dose Response Scenarios Power Dose Allocation in percents of total sample size (average over simulations) Plb vs all doses (Dunnett) Plb vs max alloc dose PlbD1D2D3PlbD1D2D3 S S S S S S S S Simulation Results

Discussion The standard parallel 4 arms design would require 4*64 subjects to provide 80% to detect difference from placebo 0.5 with no adjustment for multiplicity.  The adaptive design uses 3*64 subjects and results in % power depending on scenario. (25% sample size reduction) Limitations:  Somewhat excessive variability for S2 and S3 for some DR scenarios  Logistical complexity due to the large number of adaptations

Normal Dynamic Linear Models

Introduction How to pool information across dose levels in dose- response analysis ? Solution: Normal Dynamic Linear Model (NDLM)  Bayesian forecaster  parametric model with dynamic unobserved parameters;  forecast derived as probability distributions;  provides facility for incorporation expert information Refer to West and Harrison (1999) NDLM idea: filter or smooth data to estimate unobserved true state parameters

θ0θ0 …etc…θtθt θ t-1 θ2θ2 θ1θ1 θ t+1 Graphical Structure of DLM … Task is to estimate the state vector θ=(θ 1,…, θ K )

Dynamic Linear Models

Dose Idea: At each dose a straight line is fitted. The slope of the line changes by adding an evolution noise, Berry et. al. (2002)

Local Linear Trend Model for Dose Response (ref. to Berry (2002) et al.)

Implementation Details Miller et. al.(2006)  MCMC method Smith et. al. (2006) provide WinBugs code West and Harrison (1999)  An algebraic close form solution that computes posterior distribution of the state parameters

Model Specification (details) ‘Measurements’ errors ε can be easily estimated as the residual from ANOVA type model with dose factor Specification of the variance for the state equation (evolution of slope) is difficult;  MLE estimate form state equation (works well for large sample sizes, not for small)  Other options: (i) use prior for Wt and update it later on; (ii) discount factor for posterior variance of θ as suggested in West and Harrison (1999)

NDLM fit, 200 subjects

Example of Biased NDLM Estimates

Dose allocation Response MSE Ratio (mean / NDLM) at doses plb NaN NDLM versus Simple Mean

Bayesian Decision Analysis for Dose Allocation and Optimal Stopping (Ref. to Miller et. al.(2006) Sampling from the posterior of the state vector allows to program dose allocation as a decision problem using a utility function approach (not to confuse with efficacy + safety utility)  Use the utility function that reflects the key parameter of the dose/response curve, e.g. the posterior variance of the mean response at the ED95 variance at the most effective safe dose  Utility function is evaluated by MC method by sampling from θ posterior  Select next dose that maximizes the utility function Bayesian decision approach for early stopping

Bayesian Decision Analysis (Cont.): Utility function for dose allocation (details)

Example of NDLM Dose allocation: utility = response variance at dose of interest (most effective safe dose)

Discussion as for NDLM use Potentially very broad scope applications  Flexible dose response learner including arbitrary DR relationship possible incorporation of longitudinal modeling  Framework for Bayesian decision analysis Adaptive dose allocation Early stopping for futility or efficacy  Probabilistic statements regarding features derived from the modeled dose-response

Two Stage Design

Design Description N – fixed total sample size; K – number of treatment arms including placebo 1 st stage (pilot):  Equal allocation of r*N subjects to all arms  Analysis to select the best (compared to placebo) arm 2 nd stage (confirmation):  Equal allocation of (1-r)*N subjects to the selected arm and placebo  Final inference by combing responses from both stages (one- sided testing) Posch –Bauer method is used to control type 1 error in the strong sense  BAUER & KIESER 1999, HOMMEL, 2001, POSCH ET AL. 2005

Adaptive Closed Test Scheme for 3 treatments Arm and Placebo Hypoth. p-value 1st Stage p-value 2nd Stage C(p,q) < α (1)p1q1x (2)p2 (3)p3 (1,2)p12q12x (1,3)p13q13x (1,2,3)p123q123x Combination Intersection Multiplicity

Combination Test Let p and q be respective p-values for testing any null hypothesis H from stage 1 and 2 respectively Note: Type I error is controlled if, under H0, p and q are independent and Unif.(0,1)  E.g. Two different sets of subjects at stages 1 and 2 Combination function: (inverse normal)

Closed Testing Procedure To ensure strong control of type I error  Probability of selecting a treatment arm that is no better than placebo and concluding its superiority is less than given α under all possible response configuration of other arms H 1, H 2 null hypothesis to test Use local level α test for H 1, H 2 and H 12 =H 1 ∩H 2 Closure principle : The test at multiple level α Reject Hj, j = 1, 2 If H 12 and Hj are rejected at local level α.

Test of the Intersection Hypothesis 1 st state:  Bonferroni test p12 = min{ 2 min { p1, p2}, 1}, p13- similar p123 = min{ 3 min {p1, p2, p3}, 1}  Simes test (1986); Let p1 < p2 < p3 p12 = min { 2p1, p2}; p13 = min {2p1,p3} p123 = min { 3 p1, 3/2 p2, p3} 2 nd stage  B/c just one dose is selected q(1,2) = q(1,3) = q(1,2,3) = q(1)

Optimal Split of the Total Sample Size into Two Stages True response Prop. of total sample size used in 1 st Stage, r Probability of rejecting H0 Prob. Of selecting Max Dose D1D2D3D

Comparison of 2-stage and Gradient (KW) Designs for a 4 Treatment Arms Study Scenario True Response Probability of Rejecting H0 Prob. of Selecting Max Dose PLBD1*D2*D3two-stageKWtwo-stageKW NA * Marks the starting dose pair for KW design

Comments on Posch –Bauer Method Very flexible method  several combination functions and methods for multiplicity adjustments are available  Permits data dependent changes sample size re-estimation arm dropping several arms can be selected into stage 2 furthermore, it is not necessary to pre-specify adaptation rule from stat. methodology point of view, but is necessary from regulatory prospective.

Conclusions as for Two Stage Design Posch-Bauer method is  Robust Strong control of alpha No assumptions on dose-response relation  Powerful  Simple implementation; Just single interim analysis

References Berry D, Muller P, Grieve A, Smith M, Park T, Blazek R, Mitchard N, Krams M (2002) Adaptive Bayesian designs for dose-ranging trials. In Case studies in Bayesian statistics V. Springer: Berlin pp Bauer P, Kieser M. (1999). Combining different phases in the development of medical treatments within a single trial. Statistics in Medicine, 18: Ivanova A, Lie K, Snyder E, Snavely D.(2008) An Adaptive Crossover Design for Identifying the Dose with the Best Efficacy/Tolerability Profile (in preparation) Hommel G. Adaptive modifications of hypotheses after an interim analysis. Biometrical Journal, 43(5):581–589, 2001 Muller P, Berry D, Grieve A, Krams M. A Bayesian Decision-Theoretic Dose-Finding Trial (2006) Decision Analysis 3(4): Posch M, Koenig F, Brannath W, Dunger-Baldauf C, Bauer P (2005). Testing and estimation in flexible group sequential designs with adaptive treatment selection. Stat. Medicine, 24: Smith M, Jones I, Morris M, Grieve A, Ten K (2006) Implementation of a Bayesian adaptive design in a proof of concept study. Pharmaceutical Statistics 5: West M and Harrison P (1997) Bayesian Forecasting and Dynamic Models (2 nd edn). Springer: New York.