group sequential and adaptive designs in east · group sequential and adaptive designs in east...
Post on 19-Jan-2019
237 Views
Preview:
TRANSCRIPT
Group Sequential and
Adaptive designs in
East
Cyrus MehtaCytel Corporation and Harvard TH ChanApril 25, 2016
Cyrus Mehta Harvard THC April 25, 2016 1 / 101
Objectives of Training
This workshop will teach you how to best use East to design, simulate,and monitor adequately and well controlled trials, while incorporatinggroup sequential and adaptive elements into the design.
East is the industry standard software package for the design of clinicaltrials. It directly addresses many of the requirements given in the FDA“Draft Guidance on Adaptive Design for Clinical Trials for Drugs andBiologics.”
To provide the general background of adaptive designs in clinical trials
To work through some practical examples using East
To inform about regulatory positions regarding adaptive designs
Cyrus Mehta Harvard THC April 25, 2016 2 / 101
Outline
1 Introduction
2 Statistical Methods for Group Sequential DesignsDistribution theoryE�cacy stopping rulesPower and sample size calculationsExample: The CAPTURE TrialFutility stopping rules
3 Survival Designs in EastExample: The JUPITER StudyCalculating the required number of eventsHandling drop-outs, variable accrual and non-constant hazardsNon-proportional hazards
4 Workshops 1 - 2 (Group Sequential)
Cyrus Mehta Harvard THC April 25, 2016 3 / 101
Warm-up Exercise; Recall Fixed SampleSize Session
Cyrus Mehta Harvard THC April 25, 2016 4 / 101
Sample Size Calculation 1
Design a Phase 3 trial for a normal endpoint (di↵erence of means for twoindependent populations). What is the total sample size, n, required?
1 Given two-sided ↵ = 0.05,
2 power 1� � = 0.85, so Z↵2+Z� ⇡ 3,
3 treatment e↵ect � = 0.3,
4 and nuisance parameter � = 1,
5 compute the sample size n = 4
✓Z↵
2+ Z�
�/�
◆2
=?
Cyrus Mehta Harvard THC April 25, 2016 5 / 101
Sample Size Calculation 2
Confirm your answer in East:
Cyrus Mehta Harvard THC April 25, 2016 6 / 101
Sample Size Calculation 3
Select Des1 in Output Preview, and click ’Output Summary’ icon:
Select Des1 in Output Preview, and click ’Save in Workbook’ icon:
Cyrus Mehta Harvard THC April 25, 2016 7 / 101
Sample Size Calculation 4
Select Des1 in the Library, and click ’Details’ icon:
Cyrus Mehta Harvard THC April 25, 2016 8 / 101
Introduction
Introduction to Adaptive Designs
Cyrus Mehta Harvard THC April 25, 2016 9 / 101
What are Adaptive Design Clinical Trials?
“...defined as a study that includes a prospectively planned opportunityfor modification of one or more specified aspects of the study . . . ”
(FDA, 2010)
Cyrus Mehta Harvard THC April 25, 2016 10 / 101
Types of Designs
Traditional Design: fix the sample size in advance, and perform oneanalysis after all subjects have been enrolled and evaluated
Adaptive Design: monitor the accruing data, and make interimdecisions concerning the future course of the study (e.g., stop early,select dose, increase sample size)
Group Sequential Design: a type of adaptive design in which the onlyinterim decision is whether to stop early (for harm, e�cacy, futility)
Cyrus Mehta Harvard THC April 25, 2016 11 / 101
Background Readings
Cyrus Mehta Harvard THC April 25, 2016 12 / 101
Group Sequential Design
Statistical Methods for Group SequentialDesigns
Cyrus Mehta Harvard THC April 25, 2016 13 / 101
Example: The CAPTURE Trial
CAPTURE (1997) was a parallel arm placebo-controlled trial, withbinary primary endpoint
Use East to compute the sample size required to detect a reduction inevent rates (di↵erence of proportions) from 15% to 10% with 80%power using a two-sided level-0.05 test
Cyrus Mehta Harvard THC April 25, 2016 14 / 101
Single-Look Design
Cyrus Mehta Harvard THC April 25, 2016 15 / 101
Ethical Concerns
Very large sample size commitment with no possibility of early terminationfor benefit, harm or futility
Suppose Abciximab is actually beneficial: Do we really have torandomize 683 patients to placebo before we know for sure?
What if Abciximab is no di↵erent than placebo or even harmful? Canwe avoid randomizing 683 patients to a treatment that is no better orworse than placebo?
Interim monitoring can help in both these cases
Cyrus Mehta Harvard THC April 25, 2016 16 / 101
Key concepts
1 Understand maximum information, and information fractions
2 Contrast Pocock and O’Brien-Fleming boundaries
3 Choose appropriate error spending functions
4 Compare designs (e.g,. fixed vs group sequential) in East
5 Display boundaries on Z-scale and other scales
6 Calculate expected sample sizes for group sequential designs
Cyrus Mehta Harvard THC April 25, 2016 17 / 101
Topics
1 Introduction
2 Statistical Methods for Group Sequential DesignsDistribution theoryE�cacy stopping rulesPower and sample size calculationsExample: The CAPTURE TrialFutility stopping rules
3 Survival Designs in EastExample: The JUPITER StudyCalculating the required number of eventsHandling drop-outs, variable accrual and non-constant hazardsNon-proportional hazards
4 Workshops 1 - 2 (Group Sequential)
Cyrus Mehta Harvard THC April 25, 2016 18 / 101
Notation
Let � denote the true, unknown treatment e↵ect. We always define � to bea di↵erence between the treatment and control groups. For instance:
a di↵erence of two means
a di↵erence of two binomial probabilities
a log hazard ratio
a log odds ratio
any general coe�cient in a regression model
Cyrus Mehta Harvard THC April 25, 2016 19 / 101
Measures of Information
Monitor the data K times at calendar times ⌧1, ⌧2, . . . , ⌧K
For normal and binomial endpoints let
nj
= sample size at calendar time ⌧j
For time-to-event endpoints let
dj
= number of events at calendar time ⌧j
More generally, in terms of Fisher information let
Ij
= information at calendar time ⌧j
⇡hse
⇣�̂j
⌘i�2
where �̂j
is an e�cient estimate of � at calendar time ⌧j
.
Cyrus Mehta Harvard THC April 25, 2016 20 / 101
The Maximum Information
The maximum information is the precision of �̂ needed to achieve thedesired power
For normal and binomial endpoints let
nmax = the maximum sample size required for the trial
For time-to-event endpoints let
dmax = maximum number of events required for the trial
More generally, in terms of Fisher information, let
Imax ⇡hse
⇣�̂max
⌘i�2= maximum information to be collected
In a group sequential design we will keep the trial open until either the maximuminformation is obtained of a stopping boundary is crossed.
Cyrus Mehta Harvard THC April 25, 2016 21 / 101
The Information Fraction
Define the information fraction tj
at calendar time ⌧j
tj
=
8>>>>>><
>>>>>>:
n
j
nmaxfor normal and binomial
d
j
dmaxfor time-to-event
I
j
Imaxin general
If K is intended to be the last look, we will often denote Imax by IK
,nmax by n
K
, and dmax by dK
We may regard the information fraction t, 0 t 1, as the internaltime axis of the clinical trial
Cyrus Mehta Harvard THC April 25, 2016 22 / 101
Topics
1 Introduction
2 Statistical Methods for Group Sequential DesignsDistribution theoryE�cacy stopping rulesPower and sample size calculationsExample: The CAPTURE TrialFutility stopping rules
3 Survival Designs in EastExample: The JUPITER StudyCalculating the required number of eventsHandling drop-outs, variable accrual and non-constant hazardsNon-proportional hazards
4 Workshops 1 - 2 (Group Sequential)
Cyrus Mehta Harvard THC April 25, 2016 23 / 101
E�cacy Stopping Boundaries
Let {c1, c2, . . . , cK} be the corresponding two-sided stoppingboundaries
Stop the trial and reject H0 at the first tj
such that
| Z (tj
) |� cj
We must select the cj
’s so as to preserve the type-1 error
P0{K\
j=1
| Z (tj
) |< cj
} = 1� ↵
Many cj
’s satisfy this condition.
Cyrus Mehta Harvard THC April 25, 2016 24 / 101
Pocock and O’Brien-Fleming Boundaries
Pocock (1977) was the first to propose a group sequential design: aconstant boundary c
j
= C on the Wald scale.
I An equally spaced 5-look level-0.05 two-sided Pocock test utilizesconstant boundaries c
j
= 2.413 for j = 1, 2, . . . , 5
The next proposal came from O’Brien and Fleming (1979) whointroduced boundaries of the form c
j
= C/ptj
.
I An equally spaced 5-look level-0.05 two-sided O’Brien-Fleming designutilizes the boundaries c
j
= 2.04/ptj
for j = 1, 2, . . . , 5
Cyrus Mehta Harvard THC April 25, 2016 25 / 101
Wang-Tsiatis Family
Wang and Tsiatis (1987) proposed a family of boundary shapes:
cj
= Ct��1/2j
where � is known as a shape parameter
Which values of � yield the Pocock (1977) and O’Brien-Fleming(1979) boundaries, respectively?
Cyrus Mehta Harvard THC April 25, 2016 26 / 101
Exercise: Pocock vs O’Brien-Fleming
In East, plot the Pocock (1977) and O’Brien-Fleming (1979)boundaries on the Wald scale.
What key di↵erences do you notice between these boundaries? Arethey advantages or disadvantages?
Cyrus Mehta Harvard THC April 25, 2016 27 / 101
Additional Flexibility Required
Wang-Tsiatis boundaries depend on pre-specified values of theinformation fraction t1, t2, . . . , tK
In practice, it might be necessary to alter the spacing or number oflooks after the trial has started, which requires changing remainingboundary values
Lan and DeMets (1983) proposed the error spending functionapproach, where ↵ is treated as a budgeted quantity to be spent
Cyrus Mehta Harvard THC April 25, 2016 28 / 101
The ↵-Spending Function Approach
Specify a monotone increasing function of t for t 2 [0, 1], with↵(0) = 0 and ↵(1) = ↵
Solve recursively for c1, c2, . . . , cK
P0{| Z (t1) |� c1} = ↵(t1)
and for j = 2, . . . ,K
↵(tj�1)+P0{| Z (t1) |< c1, . . . , | Z (tj�1) |< c
j�1, | Z (tj) |� cj
} = ↵(tj
)
Cyrus Mehta Harvard THC April 25, 2016 29 / 101
Exercise: Cumulative ↵ spent and Type Ierror
Consider a linear spending function (↵ = 0.05) with 4 equally-spaced looks:
What are the values of the information fractions: t1, t2, t3, t4 ?
What are the values of cumulative alpha spent: ↵(t1), ↵(t2), ↵(t3),↵(t4)?
What is the probability of committing a Type I error by Look 2?
What is the probability of committing a Type I error at Look 3?
Cyrus Mehta Harvard THC April 25, 2016 30 / 101
Spending Functions and StoppingBoundaries
Lan and DeMets proposed two spending functions
I The LD(OF ) spending function
↵(t) =
(4� 4�(z↵/4)/
pt for two-sided tests
2� 2�(z↵/2)/pt for one-sided tests
yields boundaries that approximate the O’Brien-Fleming boundariesI The LD(PK ) spending function
↵(t) = ↵ log{1 + (e � 1)t}
yields boundaries that approximate the Pocock boundaries
Cyrus Mehta Harvard THC April 25, 2016 31 / 101
LD(OF) vs LD(PK)
Which of these spending functions corresponds to LD(OF) and LD(PK)?
Cyrus Mehta Harvard THC April 25, 2016 32 / 101
Other Families of Spending Functions
Gamma Family
Hwang, Shih, & DeCani (1990) proposed
↵(t) = ↵(1� e��t)
(1� e��), where � 6= 0
� = �4 or �5 (similar to O’Brien-Fleming);� = 1 (similar to Pocock)
Cyrus Mehta Harvard THC April 25, 2016 33 / 101
Topics
1 Introduction
2 Statistical Methods for Group Sequential DesignsDistribution theoryE�cacy stopping rulesPower and sample size calculationsExample: The CAPTURE TrialFutility stopping rules
3 Survival Designs in EastExample: The JUPITER StudyCalculating the required number of eventsHandling drop-outs, variable accrual and non-constant hazardsNon-proportional hazards
4 Workshops 1 - 2 (Group Sequential)
Cyrus Mehta Harvard THC April 25, 2016 34 / 101
Three-Step Procedure for Sample Size /Events
1 Compute boundaries, (c1, c2, . . . cK ), given ↵, spacing, number oflooks and spending function
2 Compute maximum information Imax
needed to achieve the desiredpower, given boundaries, treatment e↵ect � and desired power 1� �
3 Convert Imax
into sample size nmax
or events Dmax
, for givenendpoint, and nuisance parameters
Cyrus Mehta Harvard THC April 25, 2016 35 / 101
Maximum Information Imax required toreject H1 : � = �1 with 1� � power
Jennison and Turnbull (1997) showed that Z (tj
) ⇠ N(⌘ptj
, 1), where
⌘ = �1pImax, and, for any t
j1 < tj2 , cov{Z (tj1),Z (tj2)} =
qt
j1t
j2
Find the value of the drift parameter ⌘ that satisfies the equation
P⌘{K[
j=1
|Z (tj
)| � cj
} = 1� �
Solve for
Imax =
⌘
�1
�2
Cyrus Mehta Harvard THC April 25, 2016 36 / 101
Imax to nmax for Normal Endpoints
The e↵ect size � is estimated by the di↵erence of the two groupsample means
I�1max ⌘ var [�̂(t
K
)] = var [X̄t
(tK
)� X̄c
(tK
)]
= �2
(nmax/2)+ �2
(nmax/2)
= 4�2
nmax
where � is the common standard deviation and assuming balancedallocation
Therefore the relationship between nmax and Imax is
nmax = 4�2Imax
To compute nmax, we need �, a “nuisance parameter”
Cyrus Mehta Harvard THC April 25, 2016 37 / 101
Imax to nmax for Binomial Endpoints
The e↵ect size � is estimated by the di↵erence of the two groupbinomial response rates
I�1max ⌘ var [�̂(t
K
)] = var [⇡̂t
(tK
)� ⇡̂c
(tK
)]
= ⇡t
(1�⇡t
)(nmax/2)
+ ⇡c
(1�⇡c
)(nmax/2)
where we have assumed balanced allocation
Solve for nmax under H1 : ⇡t � ⇡c
= �1 to obtain
nmax = 2[⇡c
(1� ⇡c
) + (⇡c
+ �1)(1� ⇡c
� �1)]Imax
To compute nmax, we need ⇡c
, a “nuisance parameter”
Cyrus Mehta Harvard THC April 25, 2016 38 / 101
Imax to Dmax for Time to Event Endpoints
For time-to-event endpoints, we usually assume the proportionalhazards alternative. In this case, the e↵ect size � is the logarithm ofthe hazard ratio and the variance of �̂
j
at any interim look j isproportional to the number of events D
j
.
For balanced randomization, it can be shown that, approximately
I�1max ⌘ var [�̂(t
K
)] ⇡ [Dmax/4]�1
Thus
Dmax = 4Imax
Cyrus Mehta Harvard THC April 25, 2016 39 / 101
Topics
1 Introduction
2 Statistical Methods for Group Sequential DesignsDistribution theoryE�cacy stopping rulesPower and sample size calculationsExample: The CAPTURE TrialFutility stopping rules
3 Survival Designs in EastExample: The JUPITER StudyCalculating the required number of eventsHandling drop-outs, variable accrual and non-constant hazardsNon-proportional hazards
4 Workshops 1 - 2 (Group Sequential)
Cyrus Mehta Harvard THC April 25, 2016 40 / 101
Example: The CAPTURE Trial Revisited
CAPTURE (1997) was a parallel arm placebo-controlled trial, withbinary primary endpoint
Use East to compute the sample size required to detect a reduction inevent rates (di↵erence of proportions) from 15% to 10% with 80%power using a two-sided level-0.05 test
Cyrus Mehta Harvard THC April 25, 2016 41 / 101
Single-Look Design
Cyrus Mehta Harvard THC April 25, 2016 42 / 101
Ethical Concerns
Very large sample size commitment with no possibility of early terminationfor benefit, harm or futility
Suppose Abciximab is actually beneficial: Do we really have torandomize 683 patients to placebo before we know for sure?
What if Abciximab is no di↵erent than placebo or even harmful? Canwe avoid randomizing 683 patients to a treatment that is no better orworse than placebo?
Interim monitoring can help in both these cases
Cyrus Mehta Harvard THC April 25, 2016 43 / 101
Sample Size for Three-Look Design
The investigators planned to take up to two interim looks using theLD(OF ) spending function. The three looks will be equally spaced
Use East to compute the maximum sample size for this groupsequential design
Cyrus Mehta Harvard THC April 25, 2016 44 / 101
Summary of Three-Look Design
Cyrus Mehta Harvard THC April 25, 2016 45 / 101
Detailed Properties of Three-Look Design
The sum of incremental boundary crossing probabilities under[H0 : � = 0] is 0.05, the type-1 error
The sum of incremental boundary crossing probabilities under[H1 : � = �0.05] is 0.8, the power
Cyrus Mehta Harvard THC April 25, 2016 46 / 101
Exercise: Expected Sample Size
1 Let n1, n2, and n3 be the sample sizes at Looks 1, 2, and 3,respectively.
2 Let p1, p2, and p3 be the incremental boundary crossing probabilitiesunder H1 at Looks 1, 2, and 3, respectively.
3 Write down an expression to calculate the expected sample size, fromthe variables: n1, n2, n3, p1, p2, p3. Confirm by calculation.
Cyrus Mehta Harvard THC April 25, 2016 47 / 101
Topics
1 Introduction
2 Statistical Methods for Group Sequential DesignsDistribution theoryE�cacy stopping rulesPower and sample size calculationsExample: The CAPTURE TrialFutility stopping rules
3 Survival Designs in EastExample: The JUPITER StudyCalculating the required number of eventsHandling drop-outs, variable accrual and non-constant hazardsNon-proportional hazards
4 Workshops 1 - 2 (Group Sequential)
Cyrus Mehta Harvard THC April 25, 2016 48 / 101
Futility Boundaries in East
An e�cacy boundary o↵ers possibility of early stopping and savings insample size if H1 is true, but what if H0 is true?
Need to add a futility stopping rule.
Many options for futility boundaries in East (for one-sided test type):
Cyrus Mehta Harvard THC April 25, 2016 49 / 101
The �-Spending Function Approach
(Pampallona, Tsiatis, and Kim, 2001)
Just as we use an ↵-spending function to generate e�cacy boundariesthat preserve type-1 error, we can also use a �-spending function togenerate futility boundaries that control type-2 error
The probability of crossing the e�cacy boundary under H0 is ↵
The probability of crossing the futility boundary under H1 is �
We force the two boundaries to meet at the last look to ensure thateither the null or the alternative hypothesis is rejected
Cyrus Mehta Harvard THC April 25, 2016 50 / 101
↵-Spending and �-Spending
For the CAPTURE trial, we continue to use the LD(OF ) spendingfunction for ↵, but spend � using the �(�2) spending function
Cyrus Mehta Harvard THC April 25, 2016 51 / 101
Both E�cacy and Futility Boundaries
The two boundaries meet at l3 = u3 = �1.993, exactly as in thee�cacy boundary only designThus the futility boundary can be safely ignored (Non-Binding) ifcrossed at an early look without inflating the study’s type-1 error
Cyrus Mehta Harvard THC April 25, 2016 52 / 101
Savings under both H0 and H1
This design protects the sample size both if � = 0 and if � = �0.05
Why is the attained ↵ lower than the specified ↵?
Cyrus Mehta Harvard THC April 25, 2016 53 / 101
The Role of Simulations
1 Sensitivity analysis for changes in assumptions (eg., treatment e↵ect �or nuisance parameters)
2 Breakdown of asymptotics when samples are small
3 Operating characteristics in complex survival designs, adaptive samplesize re-estimation, or dose selection trials, are possible only viasimulation
4 Visual communication of properties and variability of study design
Cyrus Mehta Harvard THC April 25, 2016 54 / 101
Demo: Simulate the CAPTURE Trial
Cyrus Mehta Harvard THC April 25, 2016 55 / 101
Interim Monitoring of CAPTURE
The Capture trial was designed for three equally spaced looks withnmax = 1383 and stopping boundaries derived from the LD(OF) spendingfunction. But the the trial was actually monitored with unequal spacing asshown below:
Look (Nj
/1383) Resp. Resp. Wald Old Newj N
j
tj
Plcbo Abcix Zj
cj
cj
1 350 0.253 30/175 14/175 -2.6046 -3.71 -4.30482 700 0.506 55/353 37/347 -1.9332 -2.5114 -2.94293 1050 0.759 84/532 55/518 -2.485 -1.993 -2.3426
Look 3 was unplanned and occured before the planned end of the trial
Cyrus Mehta Harvard THC April 25, 2016 56 / 101
Demo: Monitor the CAPTURE Trial
Cyrus Mehta Harvard THC April 25, 2016 57 / 101
Group Sequential Designs forTime-to-Event Trials
Designing and Simulating for SurvivalEndpoints
Cyrus Mehta Harvard THC April 25, 2016 58 / 101
Survival Studies
For studies with survival or time-to-event endpoints, the asymptoticdistribution theory and the derivation of stopping boundaries remainsthe same
There are however some special considerations
I The information is directly proportional to the number of eventsI Thus the number of events, not the number of patients, determines the
power of the studyI If study duration is not fixed, there is a trade-o↵ between sample size
and study duration.
Cyrus Mehta Harvard THC April 25, 2016 59 / 101
Survival Studies
If we recruit more patients to the study, we obtain the requirednumber of events sooner and the total study duration is reduced
Cyrus Mehta Harvard THC April 25, 2016 60 / 101
Topics
1 Introduction
2 Statistical Methods for Group Sequential DesignsDistribution theoryE�cacy stopping rulesPower and sample size calculationsExample: The CAPTURE TrialFutility stopping rules
3 Survival Designs in EastExample: The JUPITER StudyCalculating the required number of eventsHandling drop-outs, variable accrual and non-constant hazardsNon-proportional hazards
4 Workshops 1 - 2 (Group Sequential)
Cyrus Mehta Harvard THC April 25, 2016 61 / 101
Example: The JUPITER study
Justification for the Use of Statins in Prevention: an InterventionTrial Evaluating Rosuvastatin (JUPITER) examined the question ofwhether treatment with 20 mg of rosuvastatin daily, as compared withplacebo, would reduce the rate of first major cardiovascular events(Ridker et al., 2008)
Cyrus Mehta Harvard THC April 25, 2016 62 / 101
Example: The JUPITER study (cont.)
JUPITER was a randomized, double-blind, placebo-controlled,multicenter trial conducted between 2003 and 2008 by AstraZenecaat 1315 sites in 26 countries
Composite primary endpoint: occurrence of a first majorcardiovascular event, defined as nonfatal myocardial infarction,nonfatal stroke, hospitalization for unstable angina, an arterialrevascularization procedure, or confirmed death from cardiovascularcauses
Cyrus Mehta Harvard THC April 25, 2016 63 / 101
Example: The JUPITER study (cont.)
Designed for statistical power of 90% to detect a 25% reduction inthe rate of the primary end point, with a two-sided significance levelof 0.05
We are interested in a 25% reduction in the rate of the primaryendpoint, i.e. a hazard ratio �
t
/�c
= 0.75; the e↵ect size is thus� = � ln(�
t
/�c
) = 0.2877
Baseline hazard rate in placebo arm of 0.0077
Study to complete in 7.5 years, with 4 years accrual and 3.5 years offollow-up
Two interim analyses are planned with LD(OF ) spending functiondefined boundaries at 37.5% and 75% of the information
How do we design such a trial?
Cyrus Mehta Harvard THC April 25, 2016 64 / 101
Topics
1 Introduction
2 Statistical Methods for Group Sequential DesignsDistribution theoryE�cacy stopping rulesPower and sample size calculationsExample: The CAPTURE TrialFutility stopping rules
3 Survival Designs in EastExample: The JUPITER StudyCalculating the required number of eventsHandling drop-outs, variable accrual and non-constant hazardsNon-proportional hazards
4 Workshops 1 - 2 (Group Sequential)
Cyrus Mehta Harvard THC April 25, 2016 65 / 101
Maximum Number of Events
To convert information into number of events, we use Schoenfeld’sapproximation (Biometrika, 1981)
Dmax =Imax
p(1� p)
where Dmax is the maximum number of events required, p is theproportion of these events that occur on the treatment arm, and(1� p) is the proportion of these events that occur on the control arm
If the null hypothesis holds, we can set p = r , where r is the fractionof patients randomized to the treatment arm. For balancedrandomization where p = 1/2 we have
Dmax = 4Imax
Cyrus Mehta Harvard THC April 25, 2016 66 / 101
Equating Required and Expected Numberof Events to Estimate Sample Size and
Study Duration
For a K look design, if Dmax events are desired, then the study mustremain open until a boundary is crossed or calendar time ⌧
K
where ⌧K
satisfies the relationship
D�1(⌧K ) = Dmax
where D�1(⌧K ) is the expected number of events at time ⌧K
We can derive an expression for D�1(⌧K ) from exponentialassumptions and thereby estimate sample size and study duration
Cyrus Mehta Harvard THC April 25, 2016 67 / 101
JUPITER Study Design in East
We can use the Logrank Test Given Accrual Duration and
Study Duration design in East to obtain a 3-look GSD for theJUPITER trial
East tells us that the required number of events is Dmax = 517 aspreviously determined, and that we will require 14,229 subjectsaccrued over 4 years
Cyrus Mehta Harvard THC April 25, 2016 68 / 101
Early Stopping and Expected StudyDuration
E (Study Duration | H1) = 0.0703⇥ 4.042+0.6186⇥ 6.109+0.2114⇥ 7.5+(1� 0.0703� 0.6185� 0.2114)⇥ 7.5
= 6.397
Cyrus Mehta Harvard THC April 25, 2016 69 / 101
Planning for DMC Meetings
The Events vs. Time Chart is useful in planning for interim analyses& DMC meetings; eg., after the second look, 389 events should occurroughly 6.1 years into the study assuming the treatment e↵ect is �1
Cyrus Mehta Harvard THC April 25, 2016 70 / 101
Topics
1 Introduction
2 Statistical Methods for Group Sequential DesignsDistribution theoryE�cacy stopping rulesPower and sample size calculationsExample: The CAPTURE TrialFutility stopping rules
3 Survival Designs in EastExample: The JUPITER StudyCalculating the required number of eventsHandling drop-outs, variable accrual and non-constant hazardsNon-proportional hazards
4 Workshops 1 - 2 (Group Sequential)
Cyrus Mehta Harvard THC April 25, 2016 71 / 101
JUPITER Design Dropouts
Dropouts: 5% per year in both the treatment and placebo arms
Like events: Dropouts assumed exponential, with correspondinghazard rates
Like Cumulative % survival: Cumulative % dropout calculated forpatient time (from study entry), not calendar time (from study start)
Cyrus Mehta Harvard THC April 25, 2016 72 / 101
JUPITER Design Accruals
Non constant accrual (piece-wise linear): 15% by end of year 1; 35%by end of year 2; 65% by end of year 3
Cyrus Mehta Harvard THC April 25, 2016 73 / 101
Revisited JUPITER Design in East
Note that since we have not changed the e↵ect size �1 that we arepowering the trial for, Dmax = 517 has not changed. However, thedropouts and slow starting accrual means we now need 17,344patients if we want to finish the study within 7.5 years
Cyrus Mehta Harvard THC April 25, 2016 74 / 101
Topics
1 Introduction
2 Statistical Methods for Group Sequential DesignsDistribution theoryE�cacy stopping rulesPower and sample size calculationsExample: The CAPTURE TrialFutility stopping rules
3 Survival Designs in EastExample: The JUPITER StudyCalculating the required number of eventsHandling drop-outs, variable accrual and non-constant hazardsNon-proportional hazards
4 Workshops 1 - 2 (Group Sequential)
Cyrus Mehta Harvard THC April 25, 2016 75 / 101
Non-Proportional Hazards
All survival designs in East assume the proportional hazardsassumption holds
Non-proportional hazards may arise in clinical trials for many reasons(e.g., delayed treatment e↵ects)
In East, simulated data may be generated according tonon-proportional hazards
Thus, it is possible to evaluate the impact of non-proportional hazardson the power of a study
Cyrus Mehta Harvard THC April 25, 2016 76 / 101
Demo: Simulate the JUPITER Trial
Cyrus Mehta Harvard THC April 25, 2016 77 / 101
JUPITER stops early
Cyrus Mehta Harvard THC April 25, 2016 78 / 101
Monitoring the JUPITER Study
At the first interim analysis in September 2007, the e�cacy boundarywas crossed. However, the DMC voted to continue the trial for anadditional 6 months. Thus, the next interim analysis in March 2008was not originally planned.
Suppose that in March 2008, 142 events had been observed on theRosuvastatin arm against 251 events on the placebo arm
The estimated hazard ratio was 0.56 so that �̂1 = ln(0.56). The
standard error can be estimated as se(�̂) = I�1/21 =
p4/D1 where
D1 = 142 + 251 = 393
The JUPITER study was terminated early for overwhelming e�cacyin the primary endpoint: HR 0.56 (95% CI 0.46-0.69) P=<0.00001(Ridker, 2008)
Cyrus Mehta Harvard THC April 25, 2016 79 / 101
Demo: Monitor the JUPITER Trial
Cyrus Mehta Harvard THC April 25, 2016 80 / 101
Demo: Monitor the JUPITER Trial
Cyrus Mehta Harvard THC April 25, 2016 81 / 101
General Recommendations for Design
East provides great flexibility, but in practice, most group sequentialdesigns have a few common features:
1 Conservative ↵-spending: Discourages termination during earlyadjustments (patient treatment, trial management). Also, wait untilenough safety data have been collected
2 Non-binding futility: DMCs may wish to continue a trial even after thefutility boundary has been crossed (eg., to follow secondary endpoint)
Cyrus Mehta Harvard THC April 25, 2016 82 / 101
General Recommendations for Design(cont.)
3 Delay time of first interim: The first interim analysis is usuallyperformed after enough data to provide stable estimates
4 Anticipate and avoid over-runs: The final interim analysis should beperformed while it is still possible to reduce the sample size.
5 Fewer interim analyses: Usually between 2 and 3. (Large trials withmortality endpoints might have as many as 6)
Cyrus Mehta Harvard THC April 25, 2016 83 / 101
Questions?
Cyrus Mehta Harvard THC April 25, 2016 84 / 101
Workshop 1
Design of a Phase 3 Clinical Trial inHypertension
Cyrus Mehta Harvard THC April 25, 2016 85 / 101
Scenario
You work for Lifebeat Inc., a multinational pharmaceutical companywhich has in-licensed a product called ANTIHYPE
Your team has been tasked with developing a phase 3 clinical studydesign for this product in the essential hypertension indication toserve in a New Drug Application submission for the US Food andDrug Administration regulatory agency.
Your team is told that the phase 3 trial must be completed within twoyears
Cyrus Mehta Harvard THC April 25, 2016 86 / 101
Fixed Sample Design
Primary endpoint: Change from baseline to Week 6 (1.5 months) inthe 24-hour mean systolic blood pressure by ambulatory bloodpressure monitorBased on preliminary phase 2 study results and the commercial team’sfeedback, you should power your trial for a treatment e↵ect of� = �1.5 mm Hg when compared to placebo in a 1 : 1 allocationA common standard deviation of � = 6 mm Hg can be assumedEnrollment rate can be assumed to be 40 patients per month (Click“Include Options”, then“Accrual/Dropout Info”)
Question 1: Creating a fixed-sample design (Design 1) with 90%power and two-sided ↵ = 0.05 type-1 error. What is the sample sizein each group?(Note that the sample size provided by East is the total sample size)
NTreatment
: NPlacebo
:
Question 2 : How long will the study last?
Study Duration :Cyrus Mehta Harvard THC April 25, 2016 87 / 101
Three-Look Group Sequential Design
Convert Design 1 to a 3-look group sequential study design with thedefault LD(OF ) spending function (Design 2)
Question 3: What is the maximum total sample size of this 3-lookgroup sequential design? (And what is the penalty in maximumsample size for Design 2, compared to Design 1)
Nmax :
Examine the boundaries chart, the spending function chart, and thepower chart by clicking on their respective icons.
Question 4: From the spending function chart, what amount of alphawould have been spent by information fraction = 0.5?
alpha spent :
Cyrus Mehta Harvard THC April 25, 2016 88 / 101
Exploring the Group Sequential DesignProperties
Observe that the boundaries chart can be displayed on various scales.The delta scale is particularly useful for communicating with clinicians(NB: contrast with design delta)
Question 5: What is final critical value on the Z-scale and delta-scalefor the 3-look design (cf. for the 1-look design)?
critical values :
Cyrus Mehta Harvard THC April 25, 2016 89 / 101
Further Exploration of Design Properties
Question 6: What is the expected savings in sample size under H1 forDesign 2 compared to Design 1?
Savings:
Do you know how to calculate the expected sample size by hand?
Question 7: If the treatment e↵ect is indeed �1 = �1.5, with whatsample size are you most likely to stop the trial and reject H0? Withwhat probability does this occur?
Sample Size : probability :
Cyrus Mehta Harvard THC April 25, 2016 90 / 101
Sensitivity of the Design to Number andSpacing of Looks
Suppose you want to take only 2 looks, not 3. And suppose you wantto space them unequally, after 75% and 100% of the information hasaccrued (Design 3).
Question 8: What is the maximum sample size of this 2-look groupsequential design (Design 3)?
Sample Size :
Cyrus Mehta Harvard THC April 25, 2016 91 / 101
Adding in a Futility Boundary
Edit the original 3-look equally-spaced design (Design 2). Change to aone-sided with ↵ = 0.025. Add a non-binding futility boundary, withan aggressive �-spending function: �(�2). This will be Design 4
Question 9: What is the maximum total sample size of this design?
Nmax :
Question 10: At the first look, what is the futility boundary on theCP delta1 scale?
CP boundary :
Cyrus Mehta Harvard THC April 25, 2016 92 / 101
Simulation
Let’s simulate Design 4 under di↵erent conditions.
Question 11: Verify the operating characteristics of this design bysimulating the study under H0 : � = 0 and under H1 : � = �1.5. Usethe same fixed seed for each simulation.
↵̂ : 1� �̂ :
Question 12: Investigate each of the following scenarios (1) � = �0.5; (2) � = �1.8 ; (3) � = �1.5;� = 9.5. Provide below the estimatedpower, the probability of stopping early for e�cacy, and theprobability of stopping for futility for each of these settings. You mayneed to complete some calculations by hand.
Scenario Power Pr(early e�cacy) Pr(early futility)
(1) � = �0.5(2) � = �1.8(3) � = 9.5
Cyrus Mehta Harvard THC April 25, 2016 93 / 101
Workshop 2
Design of a Phase 3 Survival Study inSmall Cell Lung Cancer
Cyrus Mehta Harvard THC April 25, 2016 94 / 101
Scenario
You now work for Cancergene Inc., a multinational pharmaceuticalcompany which has developed an in-house product calledONCOBLAST
Your team has been tasked with developing a phase 3 clinical studydesign for this product in the small cell lung cancer indication to servein a Market Authorisation Application submission for the EuropeanMedicines Agency regulatory agency.
Your final safety database must include at least 500 patients.
Your team is told that the phase 3 trial must be completed within fiveyears.
Cyrus Mehta Harvard THC April 25, 2016 95 / 101
Fixed Sample Design
Primary endpoint: overall survival (OS)Based on early phase results, the medical literature, and thecommercial team’s feedback, you should power your trial for a 30%improvement in median survival time for ONCOBLAST whenadministered with Standard of Care (SOC) relative to the SOC aloneMedian survival time on the SOC is assumed to be 11 monthsThe study must enroll within 3 years and be completed within 5 years.Note that you are using Logrank Given Accrual Duration and StudyDuration.
Question 1: Create a fixed-sample design (Design 1) with 90% powerand one-sided ↵ = 0.025 type-1 error. How many events do you needto observe?
D :
Question 2: How many subjects do you need to enroll in the 3 yearperiod to complete the study within 5 years?
N :
Cyrus Mehta Harvard THC April 25, 2016 96 / 101
Three-Look Group Sequential Design
Convert this design to a 3-look group sequential study design. Choosethe Lan-Demets (OF) boundary family with equal spacing of the looks(Design 2)
Question 3: What is the maximum number of events needed to powerthis 3-look group sequential design?
Dmax :
Question 4: How many patients will you need to enroll to achievestudy completion within 5 years?
Nmax :
Cyrus Mehta Harvard THC April 25, 2016 97 / 101
Additional Assumptions
Add in a Lan-Demets(OF) non-binding futility boundary (Design 3)
Plan for 3% dropout per year in each arm [NB: Use“Dropout Rates”]
Also, accural is slower than expected: 20% by end year 1; 50% by endyear 2; and the rest by end year 3
Question 5: With these added assumptions, what is the maximumnumber of events needed to power this trial?
Dmax :
Question 6: How many patients will you need to enroll to achievestudy completion within 5 years?
Nmax :
Cyrus Mehta Harvard THC April 25, 2016 98 / 101
Exploration of the Design Characteristics
Question 7: In Design 3, select Charts:Power vs. Sample size. Enter500 in the Sample Size field. What power is achieved at 500 patients,the minimum needed for your safety database?
Power :
Question 8: In the table from Design Details, identify the timing fromstudy start of your two interim analyses if the treatment e↵ect isindeed �1. Alternatively, find these times from the events/accruals vs.time Chart.
Look #1 : Look #2:
Cyrus Mehta Harvard THC April 25, 2016 99 / 101
Survival simulations
In the ‘Simulation Control Info’ tab, set a fixed random number seed.
Question 9: Simulate Design 3 under the following three hypotheses:(1) HR = 0.9; (2) HR = .5, and (3)HR = 1. What is the power andexpected sample size under each scenario?
Scenario Power E(Sample Size)
(1) HR = .9(2) HR = 0.5(3) HR = 1
Cyrus Mehta Harvard THC April 25, 2016 100 / 101
Non-Proportional Hazards
Now suppose the baseline hazard rate in the control arm remainsunchanged, but that HR = 1.0 in the first 6 months, yet HR = 0.72thereafter
Question 10: Simulate this study under the Design 3. What is thesimulated power of the trial?
Power :
Question 11: How often do we stop early for futility (1st or 2nd look)?
Pr(futility) :
Cyrus Mehta Harvard THC April 25, 2016 101 / 101
top related