Statistical Heresy ? Response Adaptive Randomization in Clinical - - PowerPoint PPT Presentation

Response Adaptive Randomization in Clinical Trials

Mi-Ok Kim

Associate Professor of Pediatrics

• Div. of Biostatistics and Epidemiology

CCHMC, UC College of Medicine

Supported by CCTST Method Grant June 17th, 2011

? Statistical Heresy

When info. is available,

• Shall we use it?

– Yes.

• What about if the information is from an
• ngoing clinical trial and we consider using

the information to change some aspects of the trial under way?

– No.

A Dilemma Faced by Dr. Chmielowski

• Mr. McLaughlin: the experimental drug

stopped the growth of the tumor.

• Mr. Ryan: chemotherapy, priori known

ineffective, could not hold back the tumors.

• Mr. Ryan is highly likely to benefit from the

experimental drug yet would not be allowed to switch as it would muddy the trial’s results.

Conflicts of Convent. Design w Ethics

• Why wouldn’t Mr. Ryan be allowed to cross-
• ver to the experimental drug?
• Why hadn’t Mr. Ryan be given a greater than

50:50 chance of being assigned to the experimental drug?

Outline

• Response Adaptive Design

– Frequentist/Bayesian Approach

• Early Immunomodulator Trt Use in Pediatric

Ulcerative Colitis Patients

– Motivation – Logistical & Statistical Issues – Simulation Results

• Conclusion

Response Adaptive Randomization (RAR)

• Skews alloca. prob. away from equal alloca.
• ver the course of a trial to favor the better or

best performing trt arm adaptively based on

• resp. data accrued thus far w/o undermining

the validity and integrity of the trial

• Allocation prob. are adapted “by design”, not
• n “ad hoc” basis.
• Same inference procedure works as with a

fixed (non-adaptive) design

Randomized Play-the-Winner

Success Failure Success

Doubly Biased Coin Design

• Pick a target allocation

– E.g) Minimizing the expected total # of failures

• True unknown.
• Use estimates based on available data

successively over the course of a trial.

C T T

p p p 

C T p

p ,

How to compare different designs?

Suppose a target allocation probability is given & the sample size is fixed: The power increases as the variance of the sample allocation ratio gets smaller (Hu & Rosenberger, 2003)

• Ped. U Colitis Pts
• Current trt regimens are far from optimal.

– Up to 45% on corticosteroid (CS) 1yr after Dx. – Up to 26% receiving colectomy within 5yrs post Dx – No guidance as to who shall receive IM therapy, not 5-ASA monotheray, the least toxic UC drug

• PROTECT: observational study that aims to

– Est. the success rate of standardized therapeutic protocol – Develop a prediction model

Early IM Trt Use in Ped. U Colitis Pts

Enrollment High Likelihood Group Early IM Control Early IM Control Low Likelihood Group Steroid Free Remission at 1 yr Success 2ndary Outcome: Remission by day 30

Doubly Biased Coin Design

• Pick a target allocation

– Urn model

• Estimate the unknown based on

available data

% 3 . 58 ) 1 ( ) 1 ( 1     

C T C

p p p

C T p

p ,

Issues in Implementing an RAR Design

• 1. Delay in the response
• 2. Heterogenous pt population
• Okay.
• DABCD (Duan and Hu, 2007), Urn model (Bai &

Hu, 1999, 2005), Drop-the-loser rule (Zhang, et. al., 2007)

• Okay. Update when resp. become available.
• DABCD (Hu et al., 2007), Urn model (Bai et al.,

2002; Hu and Zhang, 2005), Drop-the-loser rule (Zhang, et. al., 2007)

Issues in Implementing an RAR Design

• 1. Delay in the response
• 2. Heterogenous pt population
• Use the short-term Seconndary endpoint as a

strata variable.

• Use the Kaplan-Meier estimator to incorporate

the delayed (or unavailable) responses & to update based on all available data.

Proposed Method

• 1. “Standard” Method: Primary Only / Primary +

Secondary

• 2. K-M Method: Primary Only / Primary +

Secondary Heterogeneous delays are okay Simulation – Use the Ped. IBD Collaborative Research Group Registry (n=353)

50 100 150 200 0.45 0.50 0.55 0.60

Low Likelihood Group, Long Delay

Patient number in the order of entry Mean % Patients assigned to the TRT group

"Standard" method K-M method

50 100 150 200 0.45 0.50 0.55 0.60

High Disk, Short Delay, n=228

Patient number in the order of entry Percentage of patients assigned to the treatment group

Standard method Proposed method Low likelihood Group, Short Delay

K-M method

1 9 19 31 43 55 67 79 91 104 119 134 149 164 179 194 209 0.2 0.3 0.4 0.5 0.6 0.7 0.8

"Standard" Method, Long Delay

Patient number in the order of entry % Patients assigned to the TRT group

Primary only

1 9 19 31 43 55 67 79 91 104 119 134 149 164 179 194 209 0.2 0.3 0.4 0.5 0.6 0.7 0.8

"Standard" Method, Long Delay

Patient number in the order of entry % Patients assigned to the TRT group 1 9 19 31 43 55 67 79 91 104 119 134 149 164 179 194 209 0.2 0.3 0.4 0.5 0.6 0.7 0.8

Primary only Primary+Secondary

Results (2000 simulated data replicates)

N ≈ 228 (0.5 vs 0.3) Fixed Primary Only Primary + Secondary “Standard” K-M “Standard” K-M 90.8 (80.7%) 93.1 (80.3%) 94.7 (80.7%) 92.9 (78.8%) 94.8 (79.5%) 99.7 (91.6%) 103.4 (88.7%) 101.8 (94.0%) 104.7 (93.5%)

Bayesian Approach

• Prior knowledge about parameters + Data

= Posterior knowledge about parameters

• Prior for rate of success stratified by the

secondary endpoint.

• Require more extensive pre-trial research to

appropriately design the approach with acceptable operating characteristics

Bayesian Approach

• When the priors are appropriately specified,

may perform better.

When Bayesian approach may help?

• When there exist sufficient priori info. on

which to base a relatively strongly informative prior – may bring substantially greater gain.

• Continuous assessment of trt effects is
• natural. Easier to incorporate early stopping

rules and multiple hypotheses.

• Testing a drug in genetically defined many sub-

patient populations

• Testing many drugs with limited resources

Conclusion

• Response adaptive randomization is a well-

established randomization method that increases pt benefit without undermining the validity or integrity of clinical trials.

• RAR can be applied for delayed responses,

while maintaining the benefits of the adaptive design.

• Bayesian approach can be more beneficial

• Ms. Chunyan Liu (DBE)
• Dr. Jack J. Lee (MD Anderson)
• Dr. Feifang HU (Univ. of Virginia)
• Dr. Lee Denson (Direct Inflammatory Bowel

Disease Center at CCHMC)

• the Ped. IBD Collaborative Research Group
• Dr. Lili Ding (DBE) – Adaptive Dose Finding
• Ms. Yangqing Hu – Covariate Adaptive

Randomization Design

References

• Hu, F. and Rosenberger, W.F. (2003). Optimality, variability, power:

Evaluating response-adaptive randomization procedures for treatment

• comparisons. Journal of the American Statistical Association, 98, 671-

678.

• Hu, F. F., L. X. Zhang, et al. (2008). Doubly adaptive biased coin

designs with delayed responses. Canadian Journal of Statistics-Revue Canadienne De Statistique 36(4): 541-559.

• Bai, Z., Hu, F. and Rosenberger, W.F. (2002). Asymptotic properties of

adaptive designs for clinical trials with delayed response. Ann. Statist. Vol 30, No 1, 122-139.

• Hu, F. and Zhang, L.X. (2004). Asymptotic normality of adaptive

designs with delayed response. Bernoulli. 10, 447-463.

• Zhang, L.-X., Chan, W.S., Cheung, S.H., Hu, F.

A generalized drop-the-loser urn for clinical trials with delayed

• responses. (2007) Statistica Sinica, 17 (1), pp. 387-409.

• Duan, L. L. and F. F. Hu (2009). Doubly adaptive biased coin designs

with heterogeneous responses. Journal of Statistical Planning and Inference 139(9): 3220-3230.

• Bai, Z. D. and Hu, Feifang (1999) Asymptotic theorems for urn models

with nonhomogeneous generating matrices. Stochastic Processes and their applications, Vol. 80, 87-101.

• Bai, Z. D. and F. F. Hu (2005). Asymptotics in randomized URN
• models. Annals of Applied Probability 15(1B): 914-940.
• Hu, F. and W. F. Rosenberger (2006). The theory of response-adaptive

randomization in clinical trials. Hoboken, N.J., Wiley-Interscience.

References