challenges of non-inferiority trial designs r. sridhara, ph.d
TRANSCRIPT
Challenges of Non-Inferiority Trial Designs
R. Sridhara, Ph.D.
12/6/06 Pediatric Advisory Committe 2
Terminology Superiority = Drug T better than Placebo (P) or
T better than Active Control (C) C may or may not have known efficacy
Non-inferiority (NI) = T not much less effective than C ‘Was Not Different’ or ‘Similar’
C must have known efficacy
12/6/06 Pediatric Advisory Committe 3
Clinical Trials to Demonstrate Efficacy
Superiority Claim First choice Direct evidence
Non-inferiority Claim Last choice Indirect evidence Interpretation could be misleading
12/6/06 Pediatric Advisory Committe 4
Non-inferiority Claim Implies:
1. The effect of T and C are close2. T and C may be both equally
beneficial orT and C may be both equally not beneficial
12/6/06 Pediatric Advisory Committe 5
Assumptions for NI consideration
C has an effect (compared to placebo, P) Therefore, C can not be another
experimental therapy in NI trial Can reliably estimate C effect size * C effect size compared to P Control (C) effect in the future
study population will be same as in the historical population.
12/6/06 Pediatric Advisory Committe 6
Non-inferiority Trial Design Considerations
Endpoint - Overall Survival - Progression-free Survival - Response Rate
Control Effect
% Retention
12/6/06 Pediatric Advisory Committe 7
Non-inferiority or Superiority
Challenges with Non-inferiority Trial How well we know the effect of the
control How much of the control effect can we
afford to give up At least ‘’% retention of the control
effect necessary to make sure that the new treatment is significantly better than placebo
12/6/06 Pediatric Advisory Committe 8
Superiority trial designs with OS as the primary endpoint
Null Hypothesis, H0: HR(T/C) = 1 vs.
Alternative, H1: HR (T/C) < 1 if T is better than C
12/6/06 Pediatric Advisory Committe 9
Non-inferiority trial designs with OS as the primary endpoint
H0: HR(T/C) ≥ M vs. H1: HR (T/C) = 1 if the two are expected to be
similar, Or H1: HR(T/C) = 0.95, for example, if the new
treatment (T) is expected to be slightly betterM is the margin determined based on the active
control (C) effect size estimated from historical trials and percentage of the effect to be retained
HR > 1 implies, T is worse than C
12/6/06 Pediatric Advisory Committe 10
Percent Retention and Estimated Active Control Effect Size
Non-inferiority ≈ “not much less effective” ‘X’ is the effect size of the active control;
example: Point estimate of HR (C/P) = 0.5 (or HR(P/C) = 2.0) implies an estimate of active control effect size was a 50% reduction in the risk of event (ex: death)
Percent retention is a percentage of ‘X’ that is retained; example: a 50% retention of the 50% effect size is 25% effect size, i.e., the putative HR(T/P) = 0.75 (or HR(P/T) = 1.33)
12/6/06 Pediatric Advisory Committe 11
Estimates of True Control Effect
Lower 95% C.L.
( << 0.025)
Lower % C.L.
( = 0.025)
Point Estimate
( >> 0.025)
12/6/06 Pediatric Advisory Committe 12
Methods for estimating active control effect size: No method is ideal and no particular method is endorsed by the Agency
All methods assume that the control effect has not changed over time
Some methods do not consider the variation between and within studies
Other methods incorporate variation between and within studies
12/6/06 Pediatric Advisory Committe 13
Hypothetical Example of a NI Trial with Overall Survival Endpoint
Suppose HR(P/C) = 2.0 (95% CI:
1.9, 2.1)
Then, Lower 70% Confidence limit = 1.97
12/6/06 Pediatric Advisory Committe 14
Hypothetical example: Estimates of True Control Effect
Lower 95% C.L. of HR = 1.9
Lower 70% C.L. of HR: 1.97
Point Estimate
HR(P/C) = 2.0
12/6/06 Pediatric Advisory Committe 15
Hypothetical Example of a NI Trial with Overall Survival Endpoint
Suppose we want to retain 50% of 1.97, i.e., HR(P/T) = 1.49, and we can accrue 100 patients per unit time
If H1: HR(T/C) = 1, then N = 407 events or 1045 patients will be required
If H1: HR(T/C) = 0.95, then N = 290 events or 836 patients
12/6/06 Pediatric Advisory Committe 16
Sample Size Depends On:
How good the historical data is: One randomized study – large
confidence interval – more uncertainty about the effect size
Many randomized studies – smaller confidence interval – unless there were fundamental differences in the conduct of the studies: population, dose, etc.
12/6/06 Pediatric Advisory Committe 17
Summary - 1
Superiority trials provide direct evidence New drug can be compared to placebo or
control (established or not) Non-inferiority trials provide indirect
evidence New drug must be compared to
established control Interpretation can be misleading
12/6/06 Pediatric Advisory Committe 18
Summary - 2
For NI trial consideration, active control effect must be well characterized Able to estimate effect size Control effect is same now as before
NI trials are generally large Sample sizes for a NI trial is dependent on
the estimate of control effect, population, % retention and the alternative hypothesis
12/6/06 Pediatric Advisory Committe 19
Summary - 3
In considering NI trial, any potential loss of efficacy must be weighed against risk-benefit ratio
Failed superiority does not imply non-inferiority