Title: Multiplicity Adjustment For Clinical Trials with Two Doses of an Active Treatment and Multiple Primary and Secondary Endpoints
1Multiplicity Adjustment For Clinical Trials
with Two Doses of an Active Treatment and
Multiple Primary and Secondary Endpoints
- Hui Quan1, Tom Capizzi1, Ji Zhang2
- 1Merck Research Laboratories
- 2Sanofi-Synthelabo Research
- FDA/Industry Statistics Workshop
- September 21-23, 2004
2Outline
- Background and clinical trial setting
- Procedures for a two-dimensional problem
- Procedures for a three-dimensional problem
- Numerical examples for comparing powers
- Real data example
- Procedures without overall strong control
- Discussion
3Background
- Multiple primary and secondary endpoints are
often simultaneously considered in Phase III
trials. - Positive findings on secondary endpoints could be
for - 1 Supporting primary results or Hypotheses
generating - 2 Additional claims or description in drug
label - For category 2 ? recent regulatory guidances and
publications emphasized to have Type I error rate
controlled for secondary endpoints.
4Background (2)
- Different approaches for interpreting trial
results - Conventional standard for positive trial
positive effect on primary endpoints. - Others a trial as positive even when effect is
positive only on a secondary endpoint ? secondary
endpoints must be considered in multiplicity
adjustment. - For example Moyé (2000) proposed a PAAS for
controlling the FWE rate aF with aP for the
primary endpoints and aF-aPaS for the secondary
endpoints.
5Trial Setting
- Trial examples
- Parathyroid hormone fracture trial (Neer et al.
2000) 20 and 40 µg vs. placebo on vertebral
fracture and other fractures - Rofecoxib RA trial (Geunese, et al. 2002) 25 and
50 mg vs. placebo on four primary and other
secondary endpoints - Estrogen Alzheimer trial (Mulnard, et al. 2000)
0.625 and 1.25 mg vs. placebo on CGIC and other
scores - Metformin and rosiglitazone combination trial
(Fonseca, et al. 2000) Rosiglitazone 4 and 8 mg
vs. placebo on glucose and lipid endpoints
6Trial Setting (2)
- Three-dimensional multiplicity adjustment
problem - Two doses
- Two priority categories of endpoints primary
secondary - Multiple endpoints in each priority category
- Higher dose-the primary dose, lower dose-the
secondary dose (no monotonic order on parameter
space for dose response is assumed for any
endpoint) - Trial positive only if positive on one or more
primary endpoints except when PAAS is used for
multiplicity adjustment - No further priority order for endpoints within
each category
7Trial Setting (3)
- Traditional procedures may be lack of power.
- For example for a trial with two primary
and two secondary endpoints-a total of 8 active
to control comparisons, if a Hochberg procedure
is used and if the lower dose shows no effect on
the two secondary endpoints, the higher dose on
the primary endpoints will be assessed at most at
level of ?/3. - Since no monotonic order in dose response is
assumed, stepdown trend tests are not applicable.
- More specific gatekeeping procedures may have
higher power.
8Procedures for a Two-dimensional Problem Two
doses and multiple endpoints
- Hij is the null hypothesis for comparing the i
th dose to the control on the j th endpoint, and
pij is the corresponding p-value, i1, 2 and j1,
2, , g. (Dose 1 the higher and primary dose) - Modified Bonferroni-Closed Procedure (2.1)
- Step1. Reject all Hij if all pij ?, i1,2,
j1,, g. - Otherwise,
- Step 2. Reject H1j if p1j ?/g
- Reject H2j if p1j ?/g and p2j
?/g
9Modified Hochberg-Bonferroni Procedure (2.2)
- Step 1. Reject all Hij if all pij ?, i1,2,
j1,, g. - Otherwise,
- Step 2. Use Hochberg procedure for the higher
dose at ?2?/3. - Step 3. Let l of endpoints of no effect for
the higher dose. If l0, use - a Hochberg procedure for the lower
dose at level ?. Otherwise, - Step 4
- a) Use Hochberg procedure for the lower dose
at ? with all endpoints - b) Assume no effect on the l (gt0) endpoints
for the lower dose - c) Use Hochberg procedure for the lower dose
on the other (g-l) - endpoints at level ?/3 then
- d) Claim the effect for the lower dose on an
endpoint if the - effect is significant at both a) and c).
10Weighted Modified Bonferroni-Closed Procedure
(2.3)
- Step 1. Reject all Hij if all pij?
- Otherwise,
- Step 2. Reject H1j if p1j ?wj
- reject H2j if p1j ?wj and p2j
?wj . - where wj 0 and Swj1
- Applicable to a three-dimensional problem by
assigning more weights on the primary and less
weights on the secondary endpoints. However,
different from PAAS, no need to split the
significance level if all pij ?.
11Properties of Procedures (2.n)
- Reject all the hypotheses if all p-values are ?.
- Step down from the higher dose to the lower dose
to form closed procedures. - Theoretically or based on simulation, provide
strong control on the FWE rate under appropriate
conditions. - for (2.1) and (2.2) the condition for the
Simes test - for (2.3) positive regression dependency
condition for - the weighted Simes test.
12Procedures for the Three-dimensional Problem
- Hij is the null hypothesis for comparing the i th
dose to the control on the j th primary endpoint,
i1, 2 and j1,, gp - Sik is the null hypothesis for comparing the i th
dose to the control on the k th secondary
endpoint, i1, 2 and k1,, gs. - Following Westfall and Krishen (2001),
- Gatekeeping Procedure (3.1) for gp1 and gs1
- (a) H11 ? (b) H21, S11 ? (c) S21
- level a for each .
13Generalized Gatekeeping Procedure (3.2)
- (a) H1j (b) H2j,
S1k One-dimensional - (c) Sik
Two-dimensional - One-dimensional procedure at (a) and (b), and
two-dimensional procedure at (c). - Reject H1j (H2j) if it should be rejected at
(a) ((b)) - Reject S1k if it should be rejected at both (b)
and (c). - Reject S2k if all H2js have been rejected, S1k
has been - rejected at (b) and (c), and S2k should be
rejected at (c).
14Other Procedures for the Three-dimensional Problem
- PAAS Hij two-dimensional at level ap and
- Sik two-dimensional at level as
(apas a) - Primary-Secondary Gatekeeping
- (a) Hij two-dimensional
- ?
- (b) Sik two-dimensional
- each at level a
- Procedure (2.2) or others for the
two-dimensional Problems
15Numerical Examples(RRejection, NRNo
Rejection)
- 4 Procedures are compared to the regular Hochberg
Procedure - P1PAAS with P(2.2) for primary and secondary
endpoints - P2P(2.3) with weight wp for primary and ws for
secondary - P3P(3.2) with P(2.1) for Step (c)
- P4P(3.2) with P(2.2) for Step (c)
16Numerical Example 1
- gp1, gs3, ap.0375, as.0125, wp3/6 and
ws1/6
H11 H12 S11 S12 S13 S21 S22 S23
P-value .005 .018 .006 .014 .070 .012 .018 .100
P1 R R NR NR NR NR NR NR
P2 R R R NR NR NR NR NR
P3 R R R R NR R NR NR
P4 R R R R NR NR NR NR
HB R NR R NR NR NR NR NR
17Numerical Example 2
- gp2, gs2, ap.0375, as.0125, wp3/8 and
ws1/8
H11 H12 H21 H22 S11 S12 S21 S22
P-value .001 .014 .012 .045 .006 .033 .009 .130
P1 R R R NR NR NR NR NR
P2 R R R NR R NR NR NR
P3 R R R R R NR R NR
P4 R R R R R R R NR
HB R NR NR NR R NR NR NR
18Real Data Example
- A randomized trial to compare mitoxantrone 12
mg/m2 (n64) or 5 mg/m2 (n64) to placebo (n60)
on patients with worsening relapsing remitting or
secondary progressive multiple sclerosis (Hartung
et al. 2002). - Here, two primary and two secondary endpoints for
which p-values for both doses are available
EDSS, ambulation index, with relapses and
admitted to hospital.
19Real Data Example (2)
- gp2, gs2, ap.040, as.010, wp4/10 and
ws1/10
H11 H12 H21 H22 S11 S12 S21 S22
P-value .0194 .0306 .0100 gt.050 .0206 .0024 .7150 .2031
P1 NR NR NR NR NR R NR NR
P2 R NR R NR NR R NR NR
P3 R R R NR R R NR NR
P4 R R R NR R R NR NR
HB NR NR NR NR NR R NR NR
20Procedures without Overall Strong Control
- Strong control is theoretically appealing.
However, the procedures are generally
conservative. - Strong control in certain key sub-families of
hypotheses may be sufficient. - Procedure ()
- (a) H1j H2j two-dimensional at
level a - ? ?
- (b) S1k S2k two-dimensional at
level a - Procedure () strongly controls the FWE rate
in the sub-families of individual doses,
individual endpoints, within the families of
primary endpoints and secondary endpoints. Also,
it weakly controls the overall FWE rate.
21Procedures without Overall Strong Control (2)
- Procedure () does not strongly control the
overall FWE rate. - When gp1 and gs1, Procedure () becomes
- H11 ? H21
- ? ?
each at level a - S11 ? and ? S21
- (pij is the p-value for Hij and qik is the
p-value for Sik) - The probability of rejecting H21S11 when H21S11
is true is - Sup Pr (p11 a)(p21 a) or (p11 a)(q11 a)
H21S11 - Sup Pr(p21a) or (q11a) H21S112a-a2
22Other Procedures without Strong Control
- For example Step-down Closed Procedure
- Step a. Apply a procedure for each dose vs. the
control separately at level ? for all primary and
secondary endpoints. - Step b. Reject H1j (or S1k) if it should be
rejected at Step a. Reject H2j (or S2k) if both
H1j (or S1k) H2j (or S2k) should be rejected at
Step a. - It controls the FWE rates in the strong sense
within sub-families of individual doses and
endpoints, but not the family of the primary
endpoints nor the family of the secondary
endpoints. It controls the overall FWE rate in
the weak sense.
23Discussion
- After Phase IIb dose finding study, Phase III
confirmatory trials may still have two doses of
the active treatment. The setting considered here
should have very broad application. - As shown by numerical examples (here and in our
manuscript), no procedure is uniformly more
powerful than the others. - All procedures need only individual p-values --
no need to calculate the adjusted p-values based
on different configurations of intersection
hypotheses. - One should think hard whether the overall strong
control is necessary for a particular trial.
24Major References
- Benjamini Y and Hochberg T. Multiple hypotheses
testing with weights. Scandinavian Journal of
Statistics 1997 24 407-418. - CPMP Points to Consider on Multiplicity Issues in
Clinical Trials. 2002. - DAgostino RB. Controlling alpha in a clinical
trial the case for secondary endpoints.
Statistics in Medicine 2000 19, 763-766. - Dmitrienko A, Offen WW and Westfall PH.
Getekeeping strategies for clinical trials that
do not require all primary effects to be
significant. Statistics in Medicine 2003 22
2387-2400. - Hartung HP, Gonsette R, Konig N, Kwiecinski H,
Guseo A, Morrissey SP, Krapf H, Zwingers T and
the Mitoxantrone in Multiple Sclerosis Study
Group. Mitoxantrone in progressive multiple
sclerosis a placebo-controlled, double-blind,
randomized, multicentre trial. Lancet 2002 360,
2018-2025. - Hochberg Y. A sharper Bonferroni procedure for
multiple tests of significance. Biometrika 1988
75, 800-802. - Macus R, Peritz E and Gabriel KR. On closed
testing procedures with special reference to
ordered analysis of variance. Biometrika 1976
64, 655-660. - Moye LA. Alpha calculus in clinical trials
considerations and commentary for the new
millennium. Statistics in Medicine 2000 19,
767-779. - Neer RM, Arnaud CD, Zanchetta JR, Prince R, Gaich
GA, Reginster JY, Hodsman AB, Eriksen EF,
Ish-Shalom S, Genant HK, Wang O and Mitlak BH.
Effect of parathyroid hormone (1-34) on fractures
and bone mineral density in postmenopausal women
with osteoporosis. The New England Journal of
Medicine 2001 344, 19, 1434-1441. - Quan H, Luo E, Capizzi T. Xun Chen, Lynn Wei, and
Bruce Binkowitz, Multiplicity adjustment for
multiple endpoints in clinical trials with
multiple doses of an active control. JSM
Proceedings 2003. - Samuel-Cahn, E. Is the Simes improved Bonferroni
procedure conservative? Biometrika 1996 83,
928-933. - Simes RJ. An improved Bonferroni procedure for
multiple tests of significance. Biometrika, 1986
63 655-660. - Treanor JJ, Hayden FG, Vrooman PS, Barbarash R,
Bettis R, Riff D, Singh S, Kinnersley N, Ward P
and Mills RG. Efficacy and safety of the oral
neuraminidase inhibitor oseltamivir in treating
acute influenza. JAMA 2000 283, 8, 1016-1024. - Westfall PH and Krishen A. Optimally weighted,
fixed sequence and gatekeeper multiple testing
procedures. Journal of Statistical Planning and
Inference 2001 99, 25-40.