Power assessment in group sequential design with multiple biomarker subgroups for multiplicity problem

Transcription

1 Power assessment in group sequential design with multiple biomarker subgroups for multiplicity problem Lei Yang, Ph.D. Statistical Scientist, Roche (China) Holding Ltd. Aug 30 th 2018, Shanghai Jiao Tong University Page 1

2 GROUP SEQUENTIAL DESIGN & EFFICACY SUPERIORITY INTERIM ANALYSIS Group sequential design Potentially stop trial after interim analysis for futility or efficacy Look into the data multiple times while maintaining integrity Early stopping for efficacy superiority at interim analysis All patients should have timely access to new effective treatment Launch new product early Early analysis may reveal problems (e.g. compliance, accrual rate) Page 2

3 HIERARCHICAL TESTING Group C H 0,BM H Group B 1 1 H 0,BM H & M Group A H 0, All-comer Test primary population at full alpha If significant, test next population with full alpha recycled and so forth Fixed testing sequence ranking by clinical relevance or likelihood of success Strong control of familywise error rate (FWER) due to (1) prospective specification of the testing sequence and (2) no further testing once the sequence breaks Carefully selecting the ordering of the tests is essential FDA Guidance for Industry (2017); Hung et al (2007); Glimm et al (2010); Tamhane et al (2010) Page 3

4 STRAIGHTFORWARD TESTING WITH ONE PRIMARY ANALYSIS ONLY Primary Analysis OS: BM H OS: BM H & M OS: All-comer FPE +32 months FPE: First patient enrolled Page 4

5 NESTED SUBGROUP ANALYSES WITHOUT IA Possible scenarios with 2 BM subgroups: Panel I: non-nested subgroup analyses where subgroups can be disjoint (Panel Ia) or can intersect (Panel Ib); Panel II: nested subgroup analyses Spiessens and Debois (2010) Page 5

6 INTERIM ANALYSIS INTRODUCES MORE COMPLEXITY Interim Analysis Primary Analysis OS: BM H alpha=0.004* OS: BM H OS: BM H & M alpha=0.004 OS: BM H & M OS: All-comer alpha=0.004 OS: All-comer FPE +21 months +32 months *O Brien-Fleming boundary used at 60% information fraction FPE: First patient enrolled Page 6

7 NESTED AND INTERSECTANT SUBGROUP ANALYSES WITH IA Possible scenarios with 2 BM subgroups: Panel I: non-nested subgroup analyses where subgroups can be disjoint (Panel Ia) or can intersect (Panel Ib); Panel II: nested subgroup analyses Spiessens and Debois (2010) Page 7

8 A MIXTURE OF INTERSECTANT AND NESTED SUBGROUPS Interim Analysis Primary Analysis OS: BM H alpha=0.004* OS: BM H OS: BM H & M alpha=0.004 OS: BM H & M OS: All-comer alpha=0.004 OS: All-comer FPE +21 months +32 months *O Brien-Fleming boundary used at 60% information fraction FPE: First patient enrolled Page 8

9 AN ARTIFICIAL PHASE III TIME-TO-EVENT CASE STUDY Drug A vs SOC by randomization ratio 1:1 Population: All-comer & two subgroups defined by predefined cut-offs of CDx assays All-comer (Group A) BM high and medium (Group B) BM high (Group C) Can be extended to more biomarker subgroups Prevalence rates are 75% for Group B and 50% for Group C Primary efficacy endpoint: OS HR assumptions (median OS of drug A vs SOC [months]) Biomarker high: 0.70 (14.3 vs 10) Biomarker high and medium: 0.75 (13.3 vs 10) All-comer: 0.80 (12.5 vs 10) Constant recruitment rate (40 pts/month for all-comer) and no drop-out assumed One-sided test with alpha = Page 9

10 Withou t IA With IA WHEN CORRELATIONS AMONG BIOMARKER SUBGROUPS NOT CONSIDERED Key message HR (median OS [months]) Sample size Marginal power (IA) Events (IA) BM H (Group C) 0.70 (10 vs 14.3) % % BM H&M (Group B) 0.75 (10 vs 13.3) % % All-comer (Group A) 0.80 (10 vs 12.5) % % BM H (Group C) 0.70 (10 vs 14.3) % (40.8%) BM H&M (Group B) 0.75 (10 vs 13.3) % (40.4%) All-comer (Group A) 0.80 (10 vs 12.5) % (31.8%) Sample size determined to give 80% power for all-comer with EPR=70% without considering biomarker subgroups If so, both biomarker subgroups are over-powered when assessed respectively Target number of events slightly larger with IA Event patient ratio, EPR (IA) 307 (184) 68% (41%) 469 (281) 69% (42%) 636 (382) 71% (42%) Question How about if correlations are considered too? Page 10

11 CANONICAL JOINT DISTRIBUTION WITHOUT IA Distribution of OS effect estimates θ A = log HR A ~N θ A, var(θ A ), I A = var(θ A ) 1 = N A 4 θ B = log HR B ~N θ B, var(θ B ), I B = var(θ B ) 1 = N B 4 θ C = log HR C ~N θ C, var(θ C ), I C = var(θ C ) 1 = N C 4 with I A, I B, I C are information levels in Group A, B and C and N A, N B, N C are numbers of events respectively Under null hypotheses H 0,C : θ C = 0; H 0,B : θ B = 0; H 0,A : θ A = 0, logrank test statistics (Z C = θ C I C, Z B = θ B I B, Z A = θ A I A ) has approximately canonical joint distribution Z C Z B Z A ~N θ C I C θ B I B, θ A I A 1 τ CB τ CA τ CB 1 τ BA τ CA τ BA 1 where τ CB = I C I B, τ CA = I C I A, τ BA = I B I A are information fractions Define score statistics (S C, S B, S A ) = (Z C I 1C, Z B I 1B, Z A I 1A ), then (S C, S B S C, S A S B ) are independent For the case study without IA, τ CB = 0.64, τ CA = 0.46, τ BA = 0.72 Spiessens and Debois (2010); Jennison and Turnbull (1997, 2000) Page 11

12 OUTCOMES WHEN WITHOUT IA BM H PA BM H & M Allcomer Outcome of label claiming All-comer at PA BM H&M at PA Probability of outcome Accumulated probability 74.46% 74.5% (Marginal power for All-comer: 80.0%) 7.95% 82.4% (Marginal power for BM H&M: 87.3%) BM H at PA 5.18% 87.6% (Marginal power for BM H: 87.6%) Hypothesis testing Reject H 0,C, H 0,B, H 0,A under alternative hypotheses at PA Reject H 0,C, H 0,B under alternative hypotheses at PA Reject H 0,C under alternative hypotheses at PA 4 - No 12.41% 100.0% - Key message Powers are lower when taking correlations into consideration (-5.5% for all-comer and -4.9% for BM H&M), when there is no IA Difference might be larger if assumptions changed (HR, prevalence rates, etc) Page 12

13 CANONICAL JOINT DISTRIBUTION WITH IA Distribution of OS effect estimates θ ka = log HR ka ~N θ A, var(θ ka ), I ka = var(θ ka ) 1 = N ka 4 θ kb = log HR kb ~N θ B, var(θ kb ), I kb = var(θ kb ) 1 = N kb 4 θ kc = log HR kc ~N θ C, var(θ kc ), I kc = var(θ kc ) 1 = N kc 4 with I ka, I kb, I kc are information levels in Group A, B and C and N ka, N kb, N kc are numbers of events respectively at kth stage, where k 1, 2 with 1 for IA, 2 for PA Under null hypotheses H 0,kC : θ C = 0; H 0,kB : θ B = 0; H 0,kA : θ A = 0, logrank test statistics (Z C = θ 1C I 1C, Z B = θ 1B I 1B, Z A = θ 1A I 1A, Z C = θ 2C I 2C, Z B = θ 2B I 2B, Z A = θ 2A I 2A ) has approximately canonical joint distribution where are information fractions Z C Z B Z A Z C Z B Z A ~N θ C θ B θ A θ C θ B θ A I 1C I 1B I 1A I 2C I 2B I 2A, 1 τ CB τ CA τ CB 1 τ BA τ CA τ BA 1 τ CC τ BC τ AC τ CB τ BB τ AB τ CA τ BA τ AA τ CC τ CB τ CA τ BC τ BB τ BA τ AC τ AB τ AA 1 τ C B τ C A τ C B 1 τ B A τ C A τ B A 1 τ CB = I 1C I 1B, τ CA = I 1C I 1A, τ CC = I 1C I 2C, τ CB = I 1C I 2B, τ CA = I 1C I 2A, τ BA = I 1B I 1A, τ BC = (I 1C I 1B ) (I 1C I 2C ), τ BB = I 1B I 2B, τ BA = I 1B I 2A, τ AC = (I 1C I 1A ) I 1C I 2C, τ AB = (I 1B I 1A ) (I 1B I 2B ), τ AA = I 1A I 2A, τ C B = I 2C I 2B, τ C A = I 2C I 2A, τ B A = I 2B I 2A, Define score statistics (S C, S B, S A, S C, S B, S A ) = (Z C I 1C, Z B I 1B, Z A I 1A, Z C I 2C, Z B I 2B, Z A I 2A ), then (S C, S B S C, S A S B, S C S A, S B S C, S A S B ) are not independent anymore For the case study with IA, τ CB = 0.64, τ CA = 0.46, τ CC = 0.60, τ CB = 0.38, τ CA = 0.28, τ BA = 0.72, τ BC = 0.38, τ BB = 0.60, τ BA = 0.43, τ AC = 0.28, τ AB = 0.43, τ AA = 0.60, τ C B = 0.64, τ C A = 0.46, τ B A = 0.72 Spiessens and Debois (2010); Jennison and Turnbull (1997, 2000) Page 13

14 A PRESPECIFIED TESTING RULE WITH IA Start from testing BM H in IA Overall hierarchical rule* If significant, go to next subgroup in IA, and stop when Allcomer in IA is significant If not significant for any test in IA, go to the same subgroup in PA; then stop if not significant; otherwise move forward until All-comer in PA is significant Key message The so-called overall hierarchical rule achieves strong control of FWER in the case of with an IA *Glimm et al (2010) Page 14

15 SIMPLIFIED TESTING STRATEGY Interim Analysis Primary Analysis OS: BM H alpha=0.004 OS: BM H OS: BM H & M alpha=0.004 OS: BM H & M OS: All-comer alpha=0.004 OS: All-comer Key message Dimensions reduced (n+1<2n*); nested subgroups only Score statistics with independent increment structure *n is the number of tests in either IA or PA with always n>=2. In the artificial case study, n=3. Page 15

16 OUTCOMES WHEN WITH IA BM H BM H & M IA PA Outcome of label claiming BM H BM H & M Allcomer Allcomer Probability of outcome Accumulated probability Power of IA All-comer at IA 22.39% 22.4% (Marginal power for Allcomer: 31.8%) BM H&M at IA, All-comer at PA 8.52% 30.9% BM H&M at IA 0.02% 30.9% (Marginal power for BM H&M: 40.4%) BM H at IA, Allcomer at PA BM H at IA, BM H&M at PA 9.75% 40.7% 0.10% 40.8% BM H at IA 0.02% 40.8% (Marginal power for BM H: 40.8%) Key message Powers are lower when taking correlations into consideration with IA (-9.4% for all-comer and -9.5% for BM H&H respectively at IA) Difference might be larger if assumptions changed Page 16

17 NON-PROPORTIONAL HAZARD RATIO MARGINAL POWER With out IA Sampl e size Event patient ratio, EPR (IA) Events (IA) PH: Marginal power (IA) BM H (Group C) % % 62.7% BM H&M (Group B) % % 62.2% All-comer (Group A) % % 56.8% NPH(assume 3M efficacy effect delay): Marginal power (IA) With IA BM H (Group C) % (41%) 307 (184) 87.6% (40.8%) BM H&M (Group B) % (42%) 469 (281) 87.3% (40.4%) All-comer (Group A) % (42%) 636 (382) 80.0% (31.8%) 62.7% (13.9%) 62.2% (12.1%) 56.8% (8.1%) Key message Power decrease are 23%~25% for PA when assuming 3 months efficacy effect delay while larger for IA (27%~49%) Decrease depends on specific assumption of treatment effect delay Page 17

18 NON-PROPORTIONAL HAZARD RATIO A TENTATIVE STRATEGY - How about if correlations are considered too? - Unfortunately NO theory established yet when assuming non-proportional HR! A tentative strategy To approximate the NPH model with a PH model by re-estimating a rough proportional HR* from NPH simulation Pros Very straightforward for audience to understand impacts of NPH Make cross-table comparison possible Cons Lack of solid theory support and good accuracy Need further investigation on the HR reestimation * e.g., with a relatively stable proportional HR around the timing of PA and beyond Page 18

19 NON-PROPORTIONAL HAZARD RATIO POWER FINDINGS BM H PA/IA BM H & M Allcomer Outcome of label claiming All-comer at PA BM H&M at PA PH Power of PA (IA)/Assumed HR NPH* Power of PA (IA)/Reestimated HR 74.5% (22.4%)/ % (7.9%)/ % (30.9%)/ % (12.2%)/ BM H at PA 87.6% (40.8%)/ % (19.0%)/0.85 Key message Either taking correlation into account or NPH will lower down powers for both IA and PA When taking correlations into account, power decrease from PH to NPH is usually larger than the case of marginal power for PA (for IA, it is on the contrary) When assuming the NPH model, power decrease from marginal power to correlated one is usually larger than the case of PH model for PA (for IA, it is on the contrary) IA power for NPH considering correlation >= marginal IA power for NPH is mostlikely due to poor approximation of the tentative strategy (intuitively it should be < ) More investigation needed to understand the why * Assuming there is a 3 months efficacy effect delay Page 19

20 TOPICS FOR FURTHER EXTENSION Not uncommon in CIT New theory needs to be further developed Nonproportional HR assumption Secondary endpoint included in testing sequence Different alpha spending methods for primary & secondary? Correlation Power might be increased for secondary endpoint e.g., O B-F boundary for BM H vs (refined) Pocock boundary for others? Power might be increased for BM H&M and/or all-comer Different alpha spending methods over subgroups New testing method beyond hierarchical testing e.g., weighted test with loop-back or weighted test exploiting correlation Power might be increased Maurer and Bretz (2013); Xi et al (2017); Tamhane et al (2018) Page 20

21 TAKE-AWAY MESSAGES Interim analysis brings more complexity to group sequential design with multiple biomarker subgroups Correlations among subgroups should be considered to calculate power Scenario planning for claiming study success at IA/PA is be pre-discussed prior to study readouts Group sequential testing method, non-proportional HR and alpha spending methods need to be fully discussed prior to study conduct Page 21

22 THANK YOU Page 22