Transcript Impact of Dose Selection Strategies on Probability of

Impact of Dose Selection
Strategies on the Probability of
Success in the Phase III
Zoran Antonijevic
Senior Director
Strategic Development, Biostatistics
Quintiles, Inc.
Other Contributors
Frank Bretz, Novartis
Alex Dmitrienko, Ely Lilly and Company
Vlad Dragalin, Wyeth
Parvin Fardip, Wyeth
Chyi-Hung Hsu, Novartis
Tom Parke, Tessella
Jose Pinheiro, Novartis
1.
2.
3.
4.
Introduction
Design/Methods
Results/Discussion
Conclusions
Introduction
• Selection of dose(s) to advance into the
Phase III is one of the most challenging
decisions during drug development
• It is believed by many that high attrition
rate in the Phase III is largely driven by
inadequate dose selection
Introduction
• Adaptive Dose-Ranging Studies (ADRS)
group formed within PhRMA to develop
new and evaluate existing adaptive doseranging methods and strategies
• This workstreem was formed within the
ADRS with a task to assess the impact of
dose-selection methods/strategies on the
success of the Phase III program
Key Issues for Adaptive Dose-Ranging
Study Design
• Most appropriate method/design; impact
on the PoS, logistics, and cost
• Number of doses to be studied,
• Number of interim looks,
• Optimal size of the Phase II relative to the
size of the Phase III,
• Dose-selection criteria for the Phase III
• Number of doses to take into the Phase III
Approach
• Compare designs/strategies based on the
Phase III outcome
• Measured as the probability of regulatory
approval
• Measured as the Net Present Value (NPV), or
some other financial measurement
Objectives
• Assess the impact of different doseselection methods in the Phase II trials on
the PoS in the Phase III
• PoS defined as probability of drug approval
• Assess the impact of the Phase II sample
size, and number of doses studied on the
PoS in the Phase III
• Compare the performance of Phase III
studies with one vs. two active arms
1.
2.
3.
4.
Introduction
Design/Methods
Results/Discussion
Conclusions
Endpoint
• Change from baseline to Week 6 in a VAS
scale of pain.
• The VAS takes values between 0 (no pain)
and 10 (highest pain) on a continuous
scale.
Phase II Design and Assumptions
• 5 or 9 equally spaced dose levels
• 0=placebo; 2, 4, 6, and 8 active
• 0=placebo; 1, 2,…, 8=active
• Dose-response profiles
•
•
•
•
Linear
Logistic
Quadratic
Emax
• Sample sizes of 150 and 250 total
Efficacy Dose-Response Profiles
0
Umbrella
2
4
6
8
E max
Expected change from baseline in VAS at Week 6
0.0
-0.5
-1.0
-1.5
Linear
Logistic
0.0
-0.5
-1.0
-1.5
0
2
4
6
8
D os e
Safety Penalty Function
2
Umbrella
4
6
8
E max
12
10
Probability of SAE (%)
8
6
Linear
Logistic
12
10
8
6
2
4
6
8
D os e
Dose Selection Methods
• ANOVA with Dunnett’s adjustment
• Flexible design (response-adaptive allocation)
• GADA: Bayesian adaptive dose allocation method
• D-opt: adaptive dose allocation based on the D-opt
criterion
• Flexible analysis
• MCP-Mod; combination of modeling and multiple
comparison procedure
• Multiple Trend Test
• Bayesian model averaging
• Nonparametric linear regression fitting
Phase III Design
• Designs with one or two arms of the test
drug were considered
• Success measured as one positive pivotal
trial at two sided α=0.05
• For design with two active dose arms
Dunnett’s procedure applied to control for
multiplicity
Phase III Sample size
• For one active arm study N=86 per arm
• ∆=1.3; σ=2.6; power 90%
• For two active arms study N=99 per arm
• ∆1=1.3; ∆2=inf. σ=2.6; power 90%
• Dunnett adjusted
Dose Selection Strategy
• For Phase III design with one active dose, select
dose closest to the target efficacy (∆=1.3)
• For Phase III design with two doses of active,
the first dose was selected as described above
• Second dose selected is the one closer to the
target efficacy between doses immediately
above and below the first selected dose
Methods - Efficacy
• For each selected dose and dose
response model we know the “true”
treatment effect.
• For a given design of the Phase III
program (sample size, sig. level for test…)
we can determine the associated power
corresponding to the “true” effect
Methods - Safety
• Likewise, for an assumed safety doseresponse model (probability of having
unacceptable safety in the trial) we can
also determine the probability of failing for
safety for each dose.
Methods - PoS
• Assuming that efficacy and safety
successes are independent the probability
of a successful Phase III program for a
given dose/model combination is:
Prob(success|d,m) = [(power(dose,m)
* (1 - safety.prob(dose))]
1.
2.
3.
4.
Introduction
Design/Methods
Results/Discussion
Conclusions
Probability of Success for Efficacy
1 dos e
70
80
2 dos es
90
70
80
N = 250
l o g i s ti c
N = 250
u m b re l l a
N = 250
lin e a r
N = 250
Em a x
N = 150
l o g i s ti c
N = 150
u m b re l l a
N = 150
lin e a r
N = 150
Em a x
LOC FIT
BMA
MTT
MC PMod
GAD A
D opt
AN OVA
LOC FIT
BMA
MTT
MC PMod
GAD A
D opt
AN OVA
70
80
90
70
80
Av er age pow er ( %)
90
90
Probability of Acceptable Safety Profile
1 dos e
75
80
85
2 dos es
90
75
80
85
N = 250
l o g i s ti c
N = 250
u m b re l l a
N = 250
lin e a r
N = 250
Em a x
N = 150
l o g i s ti c
N = 150
u m b re l l a
N = 150
lin e a r
N = 150
Em a x
LOC FIT
BMA
MTT
MC PMod
GAD A
D opt
AN OVA
LOC FIT
BMA
MTT
MC PMod
GAD A
D opt
AN OVA
75
80
85
90
75
80
85
Av er age s afety pr obability ( %)
90
90
Overall Probability of Success
1 dos e
2 dos es
55 60 65 70 75 80
55 60 65 70 75 80
N = 250
l o g i s ti c
N = 250
u m b re l l a
N = 250
lin e a r
N = 250
Em a x
N = 150
l o g i s ti c
N = 150
u m b re l l a
N = 150
lin e a r
N = 150
Em a x
LOC FIT
BMA
MTT
MC PMod
GAD A
D opt
AN OVA
LOC FIT
BMA
MTT
MC PMod
GAD A
D opt
AN OVA
55 60 65 70 75 80
55 60 65 70 75 80
Av er age s uc c es s pr obability ( %)
Discussion
• PoS slightly (but consistently) better for
Phase II design with 250 patients vs. 150
patients
• Methods with response-adaptive
randomization component, particularly
GADA, consistently outperform other
designs on the overall PoS
• These designs are also generally less
affected by the Phase II sample size
Discussion
• Design with two doses improves the
probability of a positive efficacy result
• This is not surprising given the sample size
calculation method
• This design also improves the chance of
selecting at least one safe arm due to the
“distribution of risk”
• Resulting PoS for two active doses
improved over design with one active dose
Probability of Success for Efficacy
nD os e = 5
70
80
nD os e = 9
90
70
80
2 doses
l o g i s ti c
2 doses
u m b re l l a
2 doses
lin e a r
2 doses
Em a x
1 dose
l o g i s ti c
1 dose
u m b re l l a
1 dose
lin e a r
1 dose
Em a x
LOC FIT
BMA
MTT
MC PMod
GAD A
D opt
AN OVA
LOC FIT
BMA
MTT
MC PMod
GAD A
D opt
AN OVA
70
80
90
70
80
Av er age pow er ( %)
90
90
Probability of Acceptable Safety Profile
nD os e = 5
70
75
80
85
nD os e = 9
90
70
75
80
2 doses
l o g i s ti c
2 doses
u m b re l l a
2 doses
lin e a r
2 doses
Em a x
1 dose
l o g i s ti c
1 dose
u m b re l l a
1 dose
lin e a r
1 dose
Em a x
LOC FIT
BMA
MTT
MC PMod
GAD A
D opt
AN OVA
LOC FIT
BMA
MTT
MC PMod
GAD A
D opt
AN OVA
70
75
80
85
90
70
75
80
85
Av er age s afety pr obability ( %)
90
85
90
Overall Probability of Success
nD os e = 5
nD os e = 9
55 60 65 70 75 80
55 60 65 70 75 80
2 doses
l o g i s ti c
2 doses
u m b re l l a
2 doses
lin e a r
2 doses
Em a x
1 dose
l o g i s ti c
1 dose
u m b re l l a
1 dose
lin e a r
1 dose
Em a x
LOC FIT
BMA
MTT
MC PMod
GAD A
D opt
AN OVA
LOC FIT
BMA
MTT
MC PMod
GAD A
D opt
AN OVA
55 60 65 70 75 80
55 60 65 70 75 80
Av er age s uc c es s pr obability ( %)
Discussion
• For a fixed sample size (N=250) design
with 5 doses performed better on the
overall PoS than design with 9 doses,
other than for response-adaptive designs
(GADA & Dopt)
• For GADA & Dopt designs with 5 and 9
doses performed similarly
Discussion
• Power for efficacy, safety, as well as the
overall PoS better for the Phase III design
with two active doses, whether 5 or 9
doses were studied in the Phase II
Distribution of Selected Dose
Logistic, N = 250
2
9 doses
ANOVA
4
6
8
2
9 doses
Dopt
9 doses
GADA
4
6
8
2
9 doses
MCPMod
9 doses
MTT
4
6
8
9 doses
BMA
9 doses
LOCFI T
60
40
% Trials
20
0
5 doses
ANOVA
5 doses
Dopt
5 doses
GADA
5 doses
MCPMod
5 doses
MTT
5 doses
BMA
5 doses
LOCFI T
60
40
20
0
2
4
6
8
2
4
6
8
2
D os e s elec ted
4
6
8
2
4
6
8
Distribution of Selected Dose
Linear, N = 250
2
9 doses
ANOVA
4
6
8
2
9 doses
Dopt
9 doses
GADA
4
6
8
2
9 doses
MCPMod
9 doses
MTT
4
6
8
9 doses
BMA
9 doses
LOCFI T
40
30
20
% Trials
10
0
5 doses
ANOVA
5 doses
Dopt
5 doses
GADA
5 doses
MCPMod
5 doses
MTT
5 doses
BMA
5 doses
LOCFI T
40
30
20
10
0
2
4
6
8
2
4
6
8
2
D os e s elec ted
4
6
8
2
4
6
8
Distribution of Selected Dose
Umbrella, N = 250
2
9 doses
ANOVA
4
6
8
2
9 doses
Dopt
9 doses
GADA
4
6
8
2
9 doses
MCPMod
9 doses
MTT
4
6
8
9 doses
BMA
9 doses
LOCFI T
50
40
30
20
% Trials
10
0
5 doses
ANOVA
5 doses
Dopt
5 doses
GADA
5 doses
MCPMod
5 doses
MTT
5 doses
BMA
5 doses
LOCFI T
50
40
30
20
10
0
2
4
6
8
2
4
6
8
2
D os e s elec ted
4
6
8
2
4
6
8
Distribution of Selected Dose
Emax, N = 250
2
9 doses
ANOVA
4
6
8
2
9 doses
Dopt
9 doses
GADA
4
6
8
2
9 doses
MCPMod
9 doses
MTT
4
6
8
9 doses
BMA
9 doses
LOCFI T
50
40
30
20
% Trials
10
0
5 doses
ANOVA
5 doses
Dopt
5 doses
GADA
5 doses
MCPMod
5 doses
MTT
5 doses
BMA
5 doses
LOCFI T
50
40
30
20
10
0
2
4
6
8
2
4
6
8
2
D os e s elec ted
4
6
8
2
4
6
8
Discussion
• With exception of the logistic response
profile our dose-selection criterion misses
(usually stops short of) the dose with
highest PoS.
• Similar distributions have been observed
regardless of the method used, or the
number of doses studied.
1.
2.
3.
4.
Introduction
Design/Methods
Results/Discussion
Conclusions
Conclusions
• Methods with response-adaptive
randomization component, particularly
GADA, consistently outperform other
designs on the overall PoS
• Only a small gain in the Phase III PoS is
observed if the Phase II sample size is
increased from 150 to 250
Conclusions
• Design with smaller number of doses
performed better
• The overall PoS for a design with two
active doses is consistently much higher
than that of a design with one active dose
Conclusions
• Increasing the sample size generally
results in an improved PoS.
• Increasing the sample size also results in
the increased costs. It is therefore
important to study when and by how much
to increase investments in the program.
Conclusions
• Our current dose-selection criteria
(focused on MED) usually fails to detect
the optimal dose for the Phase III
• Future research to include different dose
selection approaches, including approaches
based on the utility (safety + efficacy)
criterion.