1/40
Adaptive and group-sequential methods
for clinical trials – Part II
Group-sequential and combination test methods
Thomas Hamborg
t.hamborg@warwick.ac.uk
Statistics & Epidemiology, Division of Health Sciences, Warwick Medical School
10 January 2012
Table of Contents
1. Introduction
2. Group sequential methods
2.1 General Setup
2.2 Classical Designs
2.3 Alpha-spending functions
2.4 Recursive numerical integration
3. Combination test methods
3.1 Introduction
3.2 Combination functions
3.3 Inverse normal combination function
4. Conditional error method
5. Summary
2/40 Introduction —
Adaptive Designs
I Use interim analyses to assess gradually accumulating data
I Adapt design for remainder of trial
Definition by CHMP(2007)
A study design is called adaptive if statistical
methodology allows the modification of a design
element (e.g. sample-size, randomisation ratio,
number of treatment arms) at an interim analysis with
full control of the type I error.
3/40 Introduction —
Adaptive Designs
Focus on methods for confirmatory trials:
I Sample size re-estimation based on nuisance parameter
estimates
I Sample size re-estimation based on efficacy estimates
(including ’self-designing trials’)
I Selection or modification of hypotheses tested including
subgroup or treatment selection
I Early stopping for futility
I Early stopping for positive results
3/40 Introduction —
Adaptive Designs
Focus on methods for confirmatory trials:
I Sample size re-estimation based on nuisance parameter
estimates
I Sample size re-estimation based on efficacy estimates
(including ’self-designing trials’)
I Selection or modification of hypotheses tested including
subgroup or treatment selection
I Early stopping for futility
I Early stopping for positive results
4/40 Introduction —
Adaptive Designs
Statistical challenge:
Inflation of type I error rate must be avoided!
Sources of potential type I error inflation
I Repeated hypothesis testing with early rejection/acceptance
of null hypotheses
I Adaptation of design/analysis features with combination of
data across stages
I Multiple hypothesis testing
4/40 Introduction —
Adaptive Designs
Statistical challenge:
Inflation of type I error rate must be avoided!
Sources of potential type I error inflation
I Repeated hypothesis testing with early rejection/acceptance
of null hypotheses
I Adaptation of design/analysis features with combination of
data across stages
I Multiple hypothesis testing
5/40 Introduction —
How bad can the naı¨ve approach be?
Two arm randomised (phase III) clinical trial
Compare test statistic for accumulated data at interim to fixed
design critical value
I Significant at 5% level⇒ stop trial
I Not significant⇒ continue with trial
# Looks 2 3 4 10 20 50 ∞
Equally spaced .083 .107 .126 .193 .248 .319 1
Worst case .098 .143 .185 .401 .642 .923 1
Table: Type I error rate by number and timing of interim analyses for
two-sided test at α = 0.05.
6/40 Group sequential methods —
Table of Contents
1. Introduction
2. Group sequential methods
3. Combination test methods
4. Conditional error method
5. Summary
7/40 Group sequential methods — General Setup
General Setup
Repeated hypothesis testing:
Interim analyses to compare treatments E and C
Conduct tests so that overall type I error rate does not exceed
pre-specified type I error rate α
Wish to
I Allow for early stopping
I Do not allow other adaptation
8/40 Group sequential methods — General Setup
Notation
θ Scalar parameter of interest (measure of superiority of E
over C)
Test at jth interim analysis expressed in terms of
I Sj – efficient score statistic
I Ij – (observed) Fisher’s information
Test statistics approximately satisfying S ∼ N(θI, I) exist for
each response type
9/40 Group sequential methods — General Setup
Stopping boundaries
Test H0 : θ = 0
At jth interim analysis
Sj ≥ uj: stop and reject H0
Sj ≤ lj: stop and reject H0 (two-sided test) or
stop and accept H0 (one-sided test)
Stop at the Jth look if not before
{u1, . . . , uJ} and {l1, . . . , lJ} form stopping boundaries
{(l1, u1), . . . , (lJ, uJ)} form continuation region
10/40 Group sequential methods — Classical Designs
Repeated significance tests
Idea: “Raise the bar”
Modify critical values and define stopping boundaries in terms
of critical values
E.g. for upper boundary of two-sided test determine uj,
j = 1, . . . , J so that: Pr
(∪Jj=1Sj ≥ uj) = α/2
11/40 Group sequential methods — Classical Designs
Two classical designs are by Pocock (1977) and O’Brien and
Fleming (1979):
I Pocock: reject H0 if | Sj |≥ uj = uP
√
k
I OBF: reject H0 if | Sj |≥ uj = uOBF
√
K
Pocock boundary corresponds to constant nominal p-value
Nominal OBF p-value is increasing with increasing j
Both designs require finding a single constant (uP, uOBF)⇒
Recursive numerical integration by Armitage, McPherson, Rowe
(1969)
12/40 Group sequential methods — Classical Designs
Stopping boundaries two-sided test I
13/40 Group sequential methods — Classical Designs
Stopping boundaries two-sided test II
14/40 Group sequential methods — Classical Designs
Stopping boundaries one-sided test
15/40 Group sequential methods — Classical Designs
I Many more group sequential tests exist with different
sequences of critical values, e.g. Wang-Tsiatis family
uj = uWT(j/J)∆ , 0 ≤ ∆ ≤ 0.5
I All ensure that overall type I error is equal to pre-specified
value if test statistics have the canonical form
I But all these designs require fixed number and spacing of
interim looks specified at the design stage
I type I errorX power ?
16/40 Group sequential methods — Alpha-spending functions
Spending functions - Idea
Instead of specifying number and timing of looks, specify
spending function α∗ telling how much α to use by information
time Ij/IJ ∈ (0, 1)
I Permits changing the number and spacing of the interim
looks without affecting type I error.
I Specify error spending function at design stage
I Re-calculate boundaries if actual interim monitoring
schedule is altered
17/40 Group sequential methods — Alpha-spending functions
Spending function approach
Specify increasing functions
0 = α∗U(0) ≤ α∗U(1) ≤ . . . ≤ α∗U(J) = α/2
0 = α∗L(0) ≤ α∗L(1) ≤ . . . ≤ α∗L(J) = α/2 (two-sided)
or 0 = α∗L(0) ≤ α∗L(1) ≤ . . . ≤ α∗L(J) = 1− α/2 (one-sided)
Find (l1, u1), . . . , (lJ, uJ) such that
Pr(stop on upper boundary by jthlook | H0) = α∗U(j)
Pr(stop on lower boundary by jthlook | H0) = α∗L(j)
for j = 1, . . . , J.
18/40 Group sequential methods — Alpha-spending functions
Spending function approach
Consider upper boundary only. From previous slide follows
Pr(Sj > uj, Sl < ul, l = 1, . . . , j− 1) = α∗U(j)− α∗U(j− 1)
⇒ Solve recursively for u1, . . . , uJ under H0.
Type I error is preserved because
α∗U(1)+[α
∗
U(2)−α∗U(1)]+. . .+[α∗U(J)−α∗U(J−1)] = α∗U(J) = α
19/40 Group sequential methods — Alpha-spending functions
Examples
Flexible spending function families exist which can also
approximate group sequential boundaries:
I Pocock type: α∗(j) = α log{1 + (e− 1)tj}
I OBF type: α∗(j) = 2{1− Φ(Z1−α/2/√tj)}
I linear: α∗(j) = αtj
I Hwang, Shih, DeCani: α∗(j) = α
{
1−exp(−γtj)
1−exp(−γ)
}
for tj = Ij/IJ
20/40 Group sequential methods — Alpha-spending functions
Alpha-spending illustration
20/40 Group sequential methods — Alpha-spending functions
Alpha-spending illustration
21/40 Group sequential methods — Alpha-spending functions
Final comments
I Timing of looks flexible but assumed to be non-data driven
(not based on observed S1 . . . Sj)
I In practise impact on power and type I error is modest
(Lan, DeMets (1989)) because power depends little on
number of looks
I Asymmetric upper and lower boundaries can be chosen
22/40 Group sequential methods — Recursive numerical integration
Stopping boundary calculation
Obtain boundary by consideration of density of S at each look
First interim analysis: S1 ∼ N(0, I1) (under H0)
f1(z1) =
1√
I1
φ
(
z1√
I1
)
Score statistics for variety of data types can be shown to have
canonical form
Canonical form implies independent increment structure:
Sj − Sj−1 ∼ N(θ(Ij − Ij−1), Ij − Ij−1) ind. of S1, . . . , Sj−1
Stopping boundary calculation
Second interim analysis:
I Can stop at first look (S1 /∈ (l1, u1))
I Or second look (S2 is sum of S1 ∈ (l1, u1) and increment
S2 − S1 )
f2(z2) =
u1∫
l1
1√
I2 − I1
φ
(
z2 − z1√
I2 − I1
)
f1(z1)dz1.
S
I
u1
l1
Stopping boundary calculation
Second interim analysis:
I Can stop at first look (S1 /∈ (l1, u1))
I Or second look (S2 is sum of S1 ∈ (l1, u1) and increment
S2 − S1 )
f2(z2) =
u1∫
l1
1√
I2 − I1
φ
(
z2 − z1√
I2 − I1
)
f1(z1)dz1.
S
I
u1
l1
u2
l2
Stopping boundary calculation
Whole boundary (and density of final S) can be found
recursively:
fj(zj) =
uj−1∫
lj−1
1√
Ij − Ij−1
φ
(
zj − zj−1√
Ij − Ij−1
)
fj−1(zj−1)dz1.
S
I
u1
l1
u2
l2
u3
l3
Same technique can be used for calculation of maximum sample
size and repeated significance testing
25/40 Combination test methods —
Table of Contents
1. Introduction
2. Group sequential methods
3. Combination test methods
4. Conditional error method
5. Summary
26/40 Combination test methods — Introduction
The basic idea
Combining evidence from different stages by aggregating
stage-wise (not cumulative!) test statistics in pre-defined way
Under null hypothesis distribution of these statistics is known⇒
type I error rate can be controlled
Consider two-stage design for illustration below
27/40 Combination test methods — Introduction
Example: Fisher’s product
Combine evidence from two stages via the product of
independent (one-sided) p-values from two stages p1 × p2
Under H0 without early stopping: ln(p1 × p2)−1/2 ∼ χ24
Or in general reject H0 if
− ln(p1 . . . pJ) > χ22J(1− α/2)
28/40 Combination test methods — Introduction
Fisher’s product with early stopping
Let α1 < α early rejection boundary and β1 > α early
acceptance boundary, 0 ≤ α1 < α < β1 ≤ 1
Bauer and Ko¨hne (1994) idea: overall type I error rate is
retained for second stage critical value Cα2 given by
Cα2 =
α− α1
ln(β1)− ln(α1)
I Can find Cα2 for pre-specified α1, β1
I Or find α1,Cα2 jointly subject to constraint
29/40 Combination test methods — Combination functions
Combination function
Base decision on stage-wise p-values assumed i.i.d. U[0, 1]
(strictly only p-clud)
Proceed to 2nd stage if α1 < p1 ≤ β1
Base 2nd stage decision on a combination function C(p1, p2)
Reject if C(p1, p2) ≤ c so that
α1 +
β1∫
α1
1∫
0
1{C(x, y) ≤ c} dx dy = α
30/40 Combination test methods — Combination functions
Examples
Popular combination functions:
I Fisher’s product
C(p1, p2) = p1 × p2
I Weighted inverse normal (WIN)
C(p1, p2) = 1− Φ
(
ω1Φ
−1 (1− p1) + ω2Φ−1 (1− p2)
)
with ω1, ω2 pre-planned such that ω21 + ω
2
2 = 1
I Proschan & Hunsberger method
C(p1, p2) = 1− Φ
((
Φ−1 (1− p1)
)2
+
(
max[0,Φ−1 (1− p2)]
)2)
Combination function illustration
1
10 p1
p2
c
c
reject H0
fail to reject H0
area = 1− α
Figure: Rejection regions for Fisher’s product (red) and inverse normal
(blue) combination function
32/40 Combination test methods — Inverse normal combination function
Inverse normal method – details
WIN combination function has special relation to group
sequential designs
Assume one-sided p-values, then
Yj = Wj Φ−1 (1− pj) ∼ N(0,W2j )
Now put W2j = (Ij − Ij−1), so that
Yj =
√
(Ij − Ij−1)Φ−1
(
Φ
(
(Sj − Sj−1) /
√
Ij − Ij−1
))
= (Sj − Sj−1)
33/40 Combination test methods — Inverse normal combination function
Inverse normal method – details
It follows that
Sj = Y1 + . . .+ Yj = Sj and Ij = W21 + . . .+ W
2
j = Ij
Inverse normal approach and group sequential approach are
identical in this case!
I Critical values α1, β1, c can be computed with standard
group sequential software, e.g. α1 = 1− Φ
(
l1/
√
I1
)
I Combination function weight link: ω2j = (Ij − Ij−1)/IJ
I Extends to more than two stages
34/40 Conditional error method —
Table of Contents
1. Introduction
2. Group sequential methods
3. Combination test methods
4. Conditional error method
5. Summary
35/40 Conditional error method —
The basic idea
Let Pr(reject H0 | p1) = A(p1) conditional error function (given
observed stage 1 p-value)
Early acceptance⇒ A = 0
Early rejection⇒ A = 1
Reject 2nd stage if p2 ≤ A(p1)
Control the conditional type I error:
Mueller and Scha¨fer (2004) showed that overall type I error is
preserved for any data-dependent change in trial if conditional
type I error is preserved, i.e. EH0(A) ≤ α
⇒ Design determined by observed statistic
36/40 Conditional error method —
Combination function link
Conditional error function corresponding to combination test
given by
A(p1) =
1 p1 ≤ α1
maxx∈[0,1]{x | C(p1, x) ≤ c} α1 < p1 ≤ β1
0 p1 > β1
General case
36/40 Conditional error method —
Combination function link
Conditional error function corresponding to combination test
given by
A(p1) =
1 p1 ≤ α1
c/p1 α1 < p1 ≤ β1
0 p1 > β1
Fisher’s product
36/40 Conditional error method —
Combination function link
Conditional error function corresponding to combination test
given by
A(p1) =
1 p1 ≤ α1
1− Φ
(
Φ−1(1−c)−ω1Φ−1(1−p1)
ω2
)
α1 < p1 ≤ β1
0 p1 > β1
Weighted inverse normal
37/40 Summary —
Table of Contents
1. Introduction
2. Group sequential methods
3. Combination test methods
4. Conditional error method
5. Summary
38/40 Summary —
Pros and Cons of the two approaches
Kelly et al. (2005) compared Fisher’s combination, inverse
normal and group sequential method
I For two-stage design with unaltered interim the
continuation regions coincide and the power is very similar
I For more stages power is slightly less for Fisher’s
combination method
I Little reason to prefer one approach to another (if no design
modification allowed)
Using combination test method in unplanned manner typically
10–25 % less efficient than group sequential test (Jennison and
Turnbull (2003))
39/40 Summary —
Summary
I Group sequential methods generally regarded as well
understood
I Some research questions remain (optimal designs,
overrunning, use of secondary endpoint information)
I Combination test methods somewhat less well understood
and potentially less efficient but offer additional flexibility
I Approaches not discussed: Continuous boundaries (e.g.
triangular test), repeated confidence intervals
40/40 Summary —
Books
I Jennison and Turnbull (2000) Group sequential methods
with applications to clinical trials. Chapman and Hall/CRC.
I Proschan, Lan and Wittes (2006) Statistical monitoring of
clinical trials. Springer.
I Chang (2007) Adaptive design theory and implementation
using SAS and R. Chapman and Hall/CRC.
I Pong and Chow (2010) Handbook of Adaptive Designs in
Pharmaceutical and Clinical Development. Chapman and
Hall/CRC.
Introduction
Group sequential methods
General Setup
Classical Designs
Alpha-spending functions
Recursive numerical integration
Combination test methods
Introduction
Combination functions
Inverse normal combination function
Conditional error method
Summary