Title: Optimal MultiStage Phase II Design With Sequential Testing of Hypotheses Within Each Stage
1Optimal Multi-Stage Phase II Design With
Sequential Testing of Hypotheses Within Each
Stage
- S. Poulopoulou1, 2, U. Dafni1, M. Kateri2, CT.
Yiannoutsos3, D. Karlis4 - 1University of Athens, Athens, Greece
- 2University of Piraeus, Piraeus, Greece
- 3Indiana University, Indiana, USA
- 4University of Economics and Business, Athens,
Greece - Email spoulopo_at_unipi.gr
2Phase II Trials
- The objective of a Phase II trial is to decide if
a particular therapeutic regimen is effective
enough to warrant further study (Phase III). - The hypothesis to be tested is
- ?O p ? pO versus ?A p ? pA,
pAgtpO - The conditions to be satisfied
- Preject HO HO true Ptype I error ? a
(false positive) - Preject HA HA true Ptype II error ? ß
(false negative)
3Simons Two Stage Design (1989)
- A two stage design, testing two complementary
events is proposed. - The following hypotheses are tested
- the null hypothesis ?O p ? pO in the first stage
and if rejected - the alternative hypothesis ?A p ? pA, (pAgtpO) in
the second and last stage - Conclusion on the alternative hypothesis requires
additional patients (going to the next stage)
4Multi-stage Designs
- In a Multi-stage Design we consider
- i1,..,k stages, N total patients
- n1,,nk, where n1nkN
- Si the sum of responses observed up to i-th
stage - ai acceptance point in i-th stage
- ri rejection point in i-th stage
- Decision rule
- If Si ? ai , stop and reject ?A p ? pA
- If Si ? ri ,stop and reject ?O p ? pO
- If ai ? Si ? ri , continue to the stage i1
- In the last stage rk ak 1
- Average Sample Number
- ASN(p)
5Multi-stage Designs (cont.)
- According to Schultz et al (1973)
- The probability that Sim is given by
- Ci(m,p)
- where ?max(ai - 11, m-ni) and Fmin(ri - 1,m).
- POi P(accept HO at i-th stage)
- PAi P( reject HO at i-th stage)
- Fleming (1982) proposed a method of determining
appropriate acceptance and rejection points based
on the asymptotic normal distribution and the
general formulation by Schultz.
6Multi-stage Designs Points of Interest
- The hypothesis of a multi-stage design consists
of two non-complementary events. - The conclusion that a particular therapy is
effective enough to warrant further study is
based only on rejecting the hypothesis p?pO (and
not on accepting the hypothesis p?pA) - (Schultz et al (1973), Fleming (1982), Simon
(1989) and Chen (1997) etc) - Ambiguity pointed out by Storer (1992)
- the evaluation of type I and II errors and
- the choice of the appropriate practical decision
at the end of the study.
7Three Outcome Design Storer (1992), Sargent et
al(2001)
- To overcome these problems a three outcome
design was proposed - Decision rule
- If Si ? ai ,stop sampling and reject ?A p ? pA
- If Si ? ri ,stop sampling and reject ?O p ? pO
- If ai ? Si ? ri ,continue to stage i1
- In the last stage k
- If Sk ? ak , stop sampling and reject ?A p ? pA
- If Sk ? rk ,stop sampling and reject ?O p ? pO
- If ak ? Sk ? rk , reject neither ?O nor ?A and
decide that the trial has been inconclusive. - The additional constraint in the three outcome
design increases the sample size. - Probability calculations are based on the
asymptotic normal distribution.
8Proposal Sequential Testing of Two Hypotheses
within the Same Stage
- In order to address the testing of
non-complementary events at the same stage, we
propose a class of designs in which we test two
hypotheses sequentially at each stage (i1,,k) - First hypothesis
- ?O1 p ? pO versus ?A1 p gt pO
- If we reject ?O1, then we test the second
hypothesis - ?O2 p ? pA versus ?A2 p lt pA
9Characteristics of the Proposed Optimal
Sequential Testing (OST) Procedure
- Two distinct hypotheses tested sequentially
within the same stage - Use of Exact Binomial distribution instead of
asymptotic Normal Distribution - Optimal design minimizing ASN with respect to
conditions on a and ß levels - Confidence interval estimation for p if
inconclusive at the last stage - Sample size chosen to satisfy predetermined
precision for the confidence interval of p
10Decision Rule
- For each stage (except the last), i1,,k -1
- If Si ? ai, stop sampling and conclude the
therapy is harmful - If Si gt ai, proceed to test the second hypothesis
- ?o2 p ? pA
- If Si ? ri, stop sampling and conclude the
therapy is effective - If Si lt ri, continue to stage i1
11Decision Rule (cont.)
- For the last stage k, stop sampling and
- If Sk ? ak, conclude the therapy is harmful
- If Sk gt ak, proceed to test the hypothesis ?02 p
? pA - If Sk ? rk, conclude the therapy is effective
- If Sk lt rk, conclude the therapy is not harmful
but also not as effective as originally targeted
(polt p ltpA) and - Estimate Confidence interval for p
12First Hypothesis HO1
- For testing the hypothesis
- ?O1 p ? po versus ?A1 p gt pO
- We have two complementary events
- Probability of rejecting ?A1
- A1(p)
- Probability of rejecting ?O1
- R1(p) 1-A1(p)1-
- Error Probabilities
- a1 sup P(reject HO1 p pO) R1(pO)
- ß1 P(reject HA1 p gt pO) A1(pA) i.e., at ppA
13Second Hypothesis HO2
- For testing the hypothesis
- ?O2 p ? pA versus ?A2 p lt pA
- We have two complementary events
- Probability of rejecting ?A2
-
-
14Second Hypothesis Ho2 (cont.)
- Probability of rejecting ?O2 R2(p) R1(p) -
A2(p) - Error Probabilities
- a2 sup P(reject HO2 p pA) R2(pA) R1(pA)
- A2(pA) - ß2 P(reject HA2 p lt pA) A2(po) , i.e., at
ppo - Average Sample Number for the total procedure
- ASN(p)
15Optimal Sequential Testing procedure
- Following Simon, we define an Optimal Sequential
Testing procedure that satisfies the constraint
parameters for the errors and minimizes the
Average Sample Number under Ho (ASN(po)). - For finding the Optimal Sequential Testing design
we use the Simulated Annealing Method.
16Algorithm Outline Simulation Annealing Method
- For given n1,, nk, pO and pA
- Step 1 - initialization For prob0 and
temperature1, choose a start value for ai and ri
that satisfy the constraints and calculate the
ASN(pO). - Step 2 For i1 to 15000
- Step 2.1 Generate random numbers xi and yi from
discrete uniform distribution U-1,1. - Step 2.2 Calculate new critical points
ai(new) ai xi and ri(new) ri yi - Step 2.3 If new critical points satisfy the
constrains find the difference
diffASN(po)new-ASN(po). - If difflt0 then ai ai(new) and riri(new)
- If diffgt0 then generate random number u from
U(0,1) and calculate probexp(-diff/temperature) - If ultprob then ai ai(new) and riri(new)
17Design Parameters for Two- and Three-stage
Optimal Sequential Testing (OST) Procedure
- For pO0.05 , pA0.20 with a2lt0.05 and ß1, ß2lt0.20
18Comparison of Flemings Design with the Optimal
Sequential Testing Procedure
- For pO0.05 , pA0.20 with a2lt0.05 and ß1,
ß2lt0.20
19Comparison of Storers Three Outcome Design with
the Optimal Sequential Testing Procedure
20Required Sample Size for a 10 Precision for a
95 Confidence Interval for p
- If the conclusion at the last stage is that (polt
p ltpA) - the therapy is not harmful
- but also not as effective as originally targeted
- i.e., inconclusive region
21Summary of Results for Proposed Design
- Comparable ai and ri to Flemings procedure
- Optimal ASN(pO), comparable or lower than
Flemings procedure - Ability to control the error rates when reaching
conclusions - reject a therapy as harmful when it has response
probability p at most p? (type II error ß1) - accept a therapy as effective when it has
response probability p at least pA, (type II
error ß2) - reject both ??1 and ??2, i.e., p?ltpltpA (type I
error a2) - Confidence interval estimation for p within 10
precision feasible with proposed sample sizes
22References
- Schultz, JR., Nichol, FR., Elfring, GL. and Weed,
SD. (1973). Multiple stage procedures for drug
screening. Biometrics 29, 293-300. - Fleming, TR. (1982). One-Sample Multiple Testing
Procedure for Phase II Clinical Trials.
Biometrics 38, 143-151. - Storer, BE. (1992). A class of phase II designs
with three possible outcomes. Biometrics 48,
5560. - Simon, R. (1989). Optimal two-stage designs for
phase II clinical trials. Controlled Clinical
Trials 10, 1-10. - Chen, TT. (1997). Optimal three-stage designs for
phase II cancer clinical trials. Statistics in
Medicine 16, 2701-2711. - Sargent, DJ., Chan, V. and Goldberg, MD. (2001) A
three-outcome desugn for phase II clinical
trials. Controlled Clinical Trials 22, 117-125.