Bayesian Methods for Making

Inferences about Rare Diseases in

Pediatric Populations
Laura A Thompson, PhD.
Division of Biostatistics/OSB
Center for Devices and Radiological Health
Food and Drug Administration

1
 Outline
• Special problems with Studying Rare Diseases in
Pediatric Populations.
• Bayesian Methods for solutions
– Zero Numerator Problem with Rare Events
– Borrowing Strength from Similar Studies to Boost
Sample Size
• Forthcoming Pediatric Extrapolation Draft Guidance
– Bayesian Adaptive Designs for Shorter Trials
• Summary
2
Special Problems with Studying Rare Diseases
in Pediatric Populations
• The pediatric population available for clinical trials is
limited even when the condition/disease is not rare.
– Informed consent might be more difficult in pediatrics.
– Finding an appropriate control could be difficult.
– Problematic: results more prone to variability and studies lack
power
• Rare conditions or events may not occur in a finite
collected sample of pediatric patients.
– Problematic: Estimating an event rate is difficult with no
events

3
 Overview of Bayesian Approach

• The Bayesian approach describes a method for

learning from evidence as it accumulates.
• The method combines prior information with
current study information on an endpoint of
interest (e.g., adverse event rate from using a
device) in order to form conclusions about the
endpoint.
• Prior information typically comes from results
of previous studies. 4
 Overview of Bayesian Approach

• Often, prior information can be used to help

estimate rare event rates and gain power for
small populations.
• In short, a way to combine the past (prior) with
the present (current study) to make decisions
about the future (posterior conclusions).
• FDA “Guidance for the Use of Bayesian Statistics in
Medical Device Trials” released in final form
February, 2010. 5
 Special Problem #1
Rare conditions or events may not occur in a
finite collected sample of patients.

6
 Zero Numerator Problem
Example based on Chen & McGee (2008)

• A standard test or device has been shown to

cause a serious reaction in about 15 of every
10,000 patients exposed to it (0.0015). A new
improved test/device was used on 167 patients
and none of them reported having the reaction.

• What can we say about the probability of a

serious reaction for the new test/device? Is it
really 0%?
7
 Zero Numerator Problem
Example based on Chen & McGee (2008)

• “Rule of three” estimate of the upper bound of a 95%

confidence interval is a conservative approximation:
3/n = 3/167 = 0.018
• Approximation holds better with larger n.
• We would like a point estimate of the occurrence rate
too.
• Bayesian methods can obtain this (even with small
samples), as well as uncertainty intervals with direct
probability interpretations.
8
 Actual Submission (Zero Numerator):
EssureTM System for Permanent Birth Control
• SSED:
http://www.accessdata.fda.gov/cdrh_docs/pdf2/P020014b.pdf
• Micro-Insert that occludes the fallopian tubes
• Zero pregnancies were observed in pivotal study
(n=632). However, because no birth control is 100%
effective, an estimate of a 0% fertility rate at 12 months
appears inaccurate.
• Bayesian Statistics/Models can help so that the estimate
is not 0% when that is unrealistic.
9
 Bayesian Estimate of Rare Event Rate

• Prior distribution placed on p, the probability of

experiencing the event.
• Examples of prior distributions:
– Prior mean is equal to the standard rate (e.g.,
0.0015), and there is a 95% chance that p is less than
0.0075.
– “Vague” Uniform prior distribution (equal
probability that p falls anywhere between 0 and 1.0)
– Hierarchical model: common method used in
CDRH
10
 Bayesian Estimate of Rare Event Rate

• Posterior Estimates (from posterior distribution)

– Posterior mean rate is not 0%, but something more
realistic and satisfying.
• (Chen & McGee Example) The posterior mean is
0.00022, which is much less than 0.0015.
• (Uniform prior) The posterior mean is 0.0016.
– Posterior probability statements can be made:
• (Chen & McGee Example) There is 96% posterior
probability that the rate is lower than the standard rate of
0.0015.
• (Uniform prior) There is 39% posterior probability that
the rate is lower than the standard rate of 0.0015.
11
 Special Problem #2:
The pediatric population available for clinical
trials is limited.

Bayesian Methods can be used to gain power

by combining prior studies with a current
study.

12
 Boost Sample size by “borrowing strength”
(information) from prior studies

• By borrowing from appropriate prior

information, the same decision might be reached
with a smaller (recruitment) sample size.
– The extent of borrowing depends on the similarity
of previous studies with the current study.
– If prior study results are different from current
study result, then borrowing strength weakens (and
can go to zero).

13
 Bayesian Hierarchical Models
• “Borrow strength” from prior studies similar to a
current study on an endpoint of interest.
– Effective sample size boost: we borrow information provided
by subjects in the prior studies
– We don’t know how much we will borrow until the current
data become available.

• The model lets the current and prior studies determine how
much to borrow.

14
 Assumption of Exchangeability is Required
for the Hierarchical Model
• Exchangeability of studies means that knowing a result
would not divulge which study it came from. (Are the
studies comparable?)

• Ideally, it is decided upon before seeing any study results

(even the prior study results).

• To decide whether exchangeability of prior and current

studies can be assumed, we need clinical input.

15
 Assumption of Exchangeability is Required
for the Hierarchical Model

• To decide whether exchangeability of prior and current

studies can be assumed, we need clinical input.

– CDRH clinicians and engineers compare previous studies

with proposed study for similarity in relevant factors,
including
device used patient population
protocol inclusion/exclusion criteria
prognostic factors patient management
proximity operator training/experience

16
 Pediatric Medical Device Safety &
Improvement Act (PMDSIA) 2007
• To improve the process for the development of needed pediatric
medical devices.
• Allows determination of a pediatric indication for a medical device,
using adult data, if:
– Similar Course of Disease or Condition, or
– Similar Effect of Device
• “Extrapolation” of a device’s effect or safety may be made:
– From adults to pediatric patients
– Between pediatric subpopulations
• Can potentially be made for approvals and clearances (PMAs,
HDEs, 510Ks), as well as during the IDE stage.
17
 Draft Guidance Document
“Extrapolation of Data for Pediatric Uses of
Medical Devices”
• General Factors for Consideration for Extrapolation:
1. Similarity of Adult Population/Response Data with future
Pediatric Response Data
– Will there be differences in device characteristics, disease process, or
patient characteristics that will likely make responses to treatment
with device different for the pediatric population than adults?
2. Quality of Adult Data
– How were the data collected, assigned to treatments? (Recent final
CDRH Guidance)
• The higher the similarity and quality, the more likely
extrapolation will be appropriate for regulatory
submissions. If both are low, we cannot rely on adult
data for pediatric indication. 18
 Are Adult and Pediatric Studies
Exchangeable?
• Obvious Differences in physiology
• Study Conduct Differences
– Enrollment might differ between adult and pediatric
studies.
– Informed consent might differ between adult and
pediatric studies.
– Treatment or handling in the trial might differ
between adult and pediatric studies.
• With these dissimilarities, how can we still
borrow from adult studies? 19
 Three-level Hierarchical Model Structure:
Studies within Patient Populations are Exchangeable
Patient Populations

Level 3: Adults Pediatrics

Level 2: Study 1 Study 2 Study 3 Future Study

Level 1: y1,…,yn1 y1,…,yn2 y1,…,yn3 y1,…

Level 1: Patients (y) exchangeable within studies

Level 2: Studies exchangeable within patient populations.
Level 3: Patient populations are exchangeable. 20
 Conditional Exchangeabilty
• Important for pediatrics: Growth or size of
the patient might influence effectiveness of the
device.
• If the covariate is measured in all studies, we can
assume exchangeability across populations,
conditional on this covariate, and hence borrow
strength from adults to pediatrics.

21
 Hypothetical Example:
SlimFix Device for Weight Loss
Single Arm Study
Average Excess Weight Loss in Percentages

Adult Study 1 Adult Study 2 Adolescent Study

(n=250) (n=150) (n=20)
41% 34% 20%

22
 Borrowing Across Studies
0.6 Adjusting for a covariate
Percent Excess Weight Loss

Adolescent
0.4

Adult 1
Adult 2
0.2
0.0

0.2 0.4 0.6 0.8 1.0 1.2

Baseline “Size” variable

23
No Borrowing from Adult Studies
Adjusting for “Baseline Size”
Population Study Posterior Mean Percent
Excess Weight loss (SD)
Adolescent Study 3 (n=20) 22.8% (5.1%)
Baseline
“Size”=0.85

Population Study Posterior Mean Percent

Excess Weight loss (SD)
Adolescent Study 3 (n=20) 19.8% (3.1%)
Baseline
“Size”=0.60

24
 Borrowing from Adult Studies
Population Study Posterior Mean Percent
Excess Weight loss (SD)
Adults Study 1 (n=250) 38.7% (0.7%)
“Size”=0.85 Study 2 (n=150) 33.0% (0.9%)
Adolescent Study 3 (n=20) 24.3% (3.0%)
“Size”=0.85

Population Study Posterior Mean Percent

Excess Weight loss (SD)
Adults Study 1 (n=250) 32.2% (1.7%)
“Size”=0.60 Study 2 (n=150) 27.4% (1.9%)
Adolescent Study 3 (n=20) 20.0% (2.2%)
“Size”=0.60
25
 Borrowing from Adult Studies

Effective Sample Size in Pediatric Study (when “Size” = 0.85) = 58:

38 subjects’ worth of information was borrowed from the adult studies
(out of a possible 250 + 150 = 400)

Effective Sample Size in Pediatric Study (when “Size” = 0.60) = 40:

20 subjects’ worth of information was borrowed from the adult studies
(out of a possible 250 + 150 = 400)

26
 Adaptive/Flexible Designs
• Trial designs that allow modifications during the course of a trial
without negatively impacting false positive error rate.
• Adaptations are performed at an interim look, based on revised
estimates of variance and/or treatment effect, or external information.
• Examples
– Change criteria for entry into trial
– Dropping/Adding an arm
– Change randomization ratio
– Sample size re-estimation
– Stop early for effectiveness or futility
• Specific adaptations should be pre-specified in order to be
carried out without complications/concerns from regulators .
• Interim looks should be performed by an independent third party

27
 Bayesian adaptive sample size using
predictive probability

• Predictive Distribution describes what the

unobserved outcomes for future patients
(enrolled or not yet accrued) will be midcourse
in a trial, given the observed patients’ data.
• This distribution provides the predictive
probability of trial success before all patients
finish the trial.
28
 Bayesian Predictive Probability
• Might be used to predict a clinical outcome from a
valid surrogate.
• Might be used to stop a trial early (for success or
futility).
• Might be used to stop accrual of patients into the
trial.
• Key point: Often lead to shorter trials or smaller
trials.

29
 Summary Statements
• Bayesian Methods can handle difficulties with
studying rare conditions in pediatric populations.
– More realistic estimates of rare event rates
– Borrow strength from adult data to make decisions
about device performance in pediatrics. (Adult
clinical data may be available from previous
marketing applications).
• Adaptive Designs and Predictive Probability
may shorten lengthy trials.
30