Enrichment Strategies for Clinical Trials to Support Approval of Human Drugs and Biological Products: Guidance for Industry

Administrative guidance, not enacted legislation. This captured publisher edition preserves the listed document and attachments; later developments are not independently consolidated.

Enrichment Strategies for
Clinical Trials to Support
Determination of
Effectiveness of Human Drugs
and Biological Products
Guidance for Industry
U.S. Department of Health and Human Services
Food and Drug Administration
Center for Drug Evaluation and Research (CDER)
Center for Biologics Evaluation and Research (CBER)
March 2019
Clinical/Medical

Enrichment Strategies for
Clinical Trials to Support
Determination of
Effectiveness of Human Drugs
and Biological Products
Guidance for Industry
Additional copies are available from:
Office of Communications, Division of Drug Information
Center for Drug Evaluation and Research
Food and Drug Administration
10001 New Hampshire Ave., Hillandale Bldg., 4th Floor
Silver Spring, MD 20993-0002
Phone: 855-543-3784 or 301-796-3400; Fax: 301-431-6353; Email: druginfo@fda.hhs.gov
https://www.fda.gov/Drugs/GuidanceComplianceRegulatoryInformation/Guidances/default.htm
and/or
Office of Communication, Outreach, and Development
Center for Biologics Evaluation and Research
Food and Drug Administration
10903 New Hampshire Ave., Bldg. 71, rm. 3128
Silver Spring, MD 20993-0002
Phone: 800-835-4709 or 240-402-8010; Email: ocod@fda.hhs.gov
https://www.fda.gov/BiologicsBloodVaccines/GuidanceComplianceRegulatoryInformation/Guidances/default.htm
U.S. Department of Health and Human Services
Food and Drug Administration
Center for Drug Evaluation and Research (CDER)
Center for Biologics Evaluation and Research (CBER)
March 2019
Clinical/Medical

i
TABLE OF CONTENTS
I. 	INTRODUCTION............................................................................................................. 1
II. 	BACKGROUND ............................................................................................................... 2
III. 	DECREASING VARIABILITY ...................................................................................... 3
A. 	Encouraging Adherence ................................................................................................................ 4
B. 	Decreasing Placebo Responses and Spontaneous Improvement ............................................... 4
IV. 	PROGNOSTIC ENRICHMENT STRATEGIES — IDENTIFYING HIGH-RISK
PATIENTS......................................................................................................................... 5
A. 	Experience With Prognostic Enrichment Strategies .................................................................. 6
1. CV Studies ........................................................................................................................................ 6
2. Oncology Studies ............................................................................................................................. 7
3. Pulmonary Studies ........................................................................................................................... 7
4. Neurology Studies ............................................................................................................................ 8
B. 	Potential Strategies for Prognostic Enrichment .......................................................................... 8
1. CV Studies ........................................................................................................................................ 8
2. Oncology Studies ............................................................................................................................. 8
a. 	Prostate cancer .......................................................................................................................... 8
b. 	Breast cancer ............................................................................................................................. 8
V. 	PREDICTIVE ENRICHMENT — IDENTIFYING MORE-RESPONSIVE
PATIENTS......................................................................................................................... 9
A. 	Increased Efficiency or Feasibility ............................................................................................. 11
B. 	Enhanced Benefit-Risk Relationship.......................................................................................... 12
C. 	Approaches to Predictive Enrichment ....................................................................................... 12
1. Empiric Strategies.......................................................................................................................... 13
a. 	Open-label single-arm trial followed by randomization ......................................................... 13
b. 	An individual’s history of response to a treatment class......................................................... 14
c. 	Factors identified in results from previous studies.................................................................. 15
2. Pathophysiological Strategies ....................................................................................................... 15
a. 	Metabolism of the test drug..................................................................................................... 16
b. 	Effect on tumor metabolism.................................................................................................... 16
c. 	Proteomic and genetic markers with known pathophysiologic effect ..................................... 16
3. Empirical Genomic Strategies ....................................................................................................... 17
4. Randomized Withdrawal Studies ................................................................................................... 18
5. Studies in Nonresponders or Patients Intolerant to Other Therapy .............................................. 20
a. 	Studies in nonresponders......................................................................................................... 21
b. 	Study in intolerants: ARBs in patients who cough on lisinopril ............................................ 23
VI. 	ENRICHMENT STUDY DESIGN AND OTHER CONSIDERATIONS ................. 24
A. 	General Considerations ............................................................................................................... 24
B. 	Which Populations to Study ........................................................................................................ 25
1. Studying Marker-Positive Patients Only........................................................................................ 26
2. Studying Both Marker-Positive and -Negative Patients ................................................................ 27
C. 	Type I Error Rate Control for Enriched Study Subpopulations ............................................. 29

ii
D. 	Adaptive Enrichment................................................................................................................... 29
E. 	Cautions in Interpretation .......................................................................................................... 31
VII. 	ENRICHMENT — REGULATORY ISSUES ............................................................. 32
A. 	Summary — The Decision to Use an Enrichment Strategy ..................................................... 32
1. Does the Enrichment Strategy Identify the Patients to Whom the Drug Should Be Given? .......... 32
2. Might the Drug Be Useful in a Broader Population Than Was Studied? ...................................... 33
B. 	Study of Marker-Negative Patients ............................................................................................ 34
C. 	Labeling ........................................................................................................................................ 36
APPENDIX: ADDITIONAL GUIDANCE RELATED TO ENRICHMENT...................... 37
REFERENCES............................................................................................................................ 38

Contains Nonbinding Recommendations
1
Enrichment Strategies for Clinical Trials to Support Determination
of Effectiveness of Human Drugs and Biological Products
Guidance for Industry1
This guidance represents the current thinking of the Food and Drug Administration (FDA or Agency) on
this topic. It does not establish any rights for any person and is not binding on FDA or the public. You
can use an alternative approach if it satisfies the requirements of the applicable statutes and regulations.
To discuss an alternative approach, contact the FDA office responsible for this guidance as listed on the
title page.
I. 	INTRODUCTION
The purpose of this guidance is to assist industry in developing enrichment strategies that can be
used in clinical investigations intended to demonstrate the effectiveness of drug and biological
products. Enrichment is the prospective use of any patient characteristic to select a study
population in which detection of a drug effect (if one is in fact present) is more likely than it
would be in an unselected population. Although this guidance focuses on enrichment directed at
improving the ability of a study to detect a drug’s effectiveness, similar strategies can be used in
safety assessments.
The enrichment strategies described in this guidance are intended to increase the efficiency of
drug development and support precision medicine, i.e., tailoring treatments to those patients who
will benefit based on clinical laboratory, genomic, and proteomic factors. This guidance also
discusses design options for enrichment strategies and discusses the interpretation of the results
of studies that use enrichment strategies.
In general, FDA’s guidance documents do not establish legally enforceable responsibilities.
Instead, guidances describe the Agency’s current thinking on a topic and should be viewed only
as recommendations, unless specific regulatory or statutory requirements are cited. The use of
the word should in Agency guidances means that something is suggested or recommended, but
not required.
1 This guidance was developed by the Center for Drug Evaluation and Research in coordination with the Center for
Biologics Evaluation and Research at the Food and Drug Administration. Although the principles and examples
discussed in this guidance relate primarily to the safety and effectiveness of drugs and biologic products many of
these principles also apply to studies for other medical products, including devices. FDA encourages a sponsor that
is considering applying these principles to the study of a device to discuss such a proposal with the Center for
Devices and Radiological Health, specifically with the organizational unit responsible for that product area.

Contains Nonbinding Recommendations
2
II. 	BACKGROUND
Sponsors of investigational drug products use a variety of strategies to enrich the study
population by selecting a subset of patients in which the potential effect of a drug can more
readily be demonstrated.
Three broad categories of enrichment strategies as listed below are addressed in this guidance:
(1) Strategies to decrease variability — These include choosing patients with baseline
measurements of a disease or a biomarker characterizing the disease in a narrow range
(decreased interpatient variability) and excluding patients whose disease or symptoms
improve spontaneously or whose measurements are highly variable (decreased
intrapatient variability). The decreased variability provided by these strategies would
increase study power (see section III., Decreasing Variability).
(2) Prognostic enrichment strategies — These include choosing patients with a greater
likelihood of having a disease-related endpoint event (for event-driven studies) or a
substantial worsening in condition (for continuous measurement endpoints) (see section
IV., Prognostic Enrichment Strategies — Identifying High-Risk Patients). These
strategies would increase the absolute effect difference between groups but would not be
expected to alter relative effect.
(3) Predictive enrichment strategies — These include choosing patients who are more likely
to respond to the drug treatment than other patients with the condition being treated.
Such selection can lead to a larger effect size (both absolute and relative) and can permit
use of a smaller study population. Selection of patients could be based on a specific
aspect of a patient’s physiology, a biomarker, or a disease characteristic that is related in
some manner to the study drug’s mechanism. Patient selection could also be empiric
(e.g., the patient has previously appeared to respond to a drug in the same class) (see
section V., Predictive Enrichment — Identifying More-Responsive Patients).
Enrichment characteristics can be dichotomous (e.g., sex, presence of genetic marker(s), a
concomitant illness) or continuous (e.g., age, blood pressure (BP)). Studies using the latter
would ordinarily dichotomize the continuous variable (e.g., BP over 160 systolic) or examine
several different cut-offs (e.g., BP over 140, over 160).
The enrichment strategies described in this guidance are discussed primarily in the context of
randomized controlled trials (but could also be relevant to other designs, such as single-arm
studies or historically (externally) controlled trials). In almost all cases, the strategies for patient
selection are prospectively planned and fixed prior to study initiation (with a few exceptions for
adaptive strategies to be noted later). These strategies, therefore, generally do not compromise
the statistical validity of the trials or the meaningfulness of the conclusions reached for the
population actually studied. The illustrative examples and design options described have been
used in the past, but they should not be regarded as an exclusive list or a limitation. FDA
encourages the development of additional approaches as experience with these strategies grows.

Contains Nonbinding Recommendations
3
The principal concerns with the use of enrichment strategies relate to the generalizability and
applicability of the study results. When considering use of an enrichment design, sponsors
should consider whether the enrichment strategy could be used in practice to identify the patients
to whom the drug should be given and whether the drug might be useful in a broader population
(see section VII.A., Summary — The Decision to Use an Enrichment Strategy). The extent to
which patients should be studied who do not meet the selection criteria for enrichment (see
section VII.B., Study of Marker-Negative Patients) is therefore a critical consideration. In
addition, in the setting of predictive enrichment strategies, the accuracy of the measurements
used to identify the enrichment population and the sensitivity and specificity of the enrichment
criteria used to distinguish treatment responders and nonresponders are also critical issues.
III. 	DECREASING VARIABILITY
Approaches to increasing study power (the ability of a clinical trial to demonstrate a treatment
effect if one is present) by decreasing heterogeneity (nondrug-related variability) are widely
practiced. The following strategies are useful and generally accepted ways to decrease
variability:
• 	Defining entry criteria carefully to ensure that enrolled patients actually have the disease
that is being studied.
• 	Training investigators to adhere to protocol-specified entry definitions and criteria.
• 	Identifying and selecting patients likely to adhere to treatment to decrease variability in
drug exposure. Note: Excluding poor compliers identified after randomization in the
analysis is not acceptable because such patients are not likely to be a random sample of
the study population and because adherence itself has been linked to outcome, even
adherence to a placebo treatment (Coronary Drug Project Research Group 1980).
• 	Using placebo lead-in periods before randomization to eliminate patients who improve
spontaneously or have large placebo responses.
• 	Decreasing intrapatient variability by enrolling only patients who give consistent baseline
values (e.g., for BP measurements, treadmill exercise tests, pulmonary function tests, or
patient-reported outcome measures).
• 	Excluding patients taking drugs that are pharmacologically similar to, or that could
interact with, the study drug.
• 	Excluding patients unlikely to tolerate the drug.
• 	Excluding patients likely to drop out for nonmedical reasons (e.g., because they have
difficulty getting to the study site).

Contains Nonbinding Recommendations
4
• 	Excluding patients with comorbid illness that would make completing the treatment
period unlikely.
Two of the strategies described above — encouraging adherence and reducing the number of
spontaneous improvers or placebo responders — are further discussed in sections III.A.,
Encouraging Adherence, and III.B., Decreasing Placebo Responses and Spontaneous
Improvement.
Some strategies to decrease variability can result in studies that provide too little information
about the full range of patients who will receive a drug in clinical practice, such as the elderly,
patients with multiple illnesses, and patients taking multiple drug therapies. It is not clear that
concomitant illnesses that do not affect survival or other endpoint measurements or concomitant
drugs unrelated to a test drug really do interfere with assessment of a treatment effect.
Therefore, the implications of using these strategies should be carefully considered before they
are used and should be balanced against the need for information in critical patient subgroups.
A. 	Encouraging Adherence
Practices that have become standard for ensuring adherence include: encouraging adherence by
making patients aware of the conditions and demands of the trial, avoiding overly rapid titration
of drugs that could cause intolerable early adverse reactions, using adherence prompts and alert
systems, and counting pills (or using smart bottles to monitor drug use) so that nonadherent
patients can be encouraged to perform better. On occasion, more protocol-specific efforts have
been used to identify and enroll good adherers into clinical trials. For example:
• 	In the Veterans Administration Cooperative hypertension studies of the late 1960s and
early 1970s, prospective patients were given placebo tablets containing riboflavin during
a single-blind prerandomization period. Patients’ urine was then examined for
fluorescence; and randomized only patients whose urine fluoresced (evidence that the
patients had been taking the riboflavin tablets) on two consecutive visits during a 2- to 4-
month observation period were randomized (Veterans Administration Cooperative Study
Group on Antihypertensive Agents 1967, 1970).
• 	The Physicians’ Health Study used an 18-week, prerandomization placebo run-in during
which patients (all physicians) self-reported whether they were taking the drug as
specified in the protocol (Steering Committee of the Physicians’ Health Study Research
Group 1989). About one-third of the screened patients were not randomized because of
self-reported poor adherence. Adherence during the randomized study was reported as a
very satisfactory 90% over the 5 years of the study, greatly increasing its power (Lang et
al. 1991).
B. 	Decreasing Placebo Responses and Spontaneous Improvement
In placebo-controlled trials of drugs for symptomatic conditions (e.g., depression, anxiety,
angina) or laboratory/vital sign abnormalities (e.g., dyslipidemia, hypertension), use of a single-
blind, placebo lead-in period and exclusion of placebo responders (i.e., randomize only patients

Contains Nonbinding Recommendations
5
whose signs or symptoms remain above some threshold value) are relatively common. This is
done to identify and not randomize patients who would have had an improvement for a reason
other than a response to the test treatment (e.g., spontaneous improvement, a placebo response,
an effect of expectations on observations) that resolved or reduced a patient’s symptoms or signs,
making the patient less likely to show a response to treatment. Also, many signs and symptoms
vary spontaneously; therefore, initial screening values that would support enrollment may
represent random highs of the disease course that would be followed by regression to the mean,
leaving the patient with mild to no symptoms and little opportunity to no opportunity to benefit
from a drug.
IV. 	PROGNOSTIC ENRICHMENT STRATEGIES — IDENTIFYING HIGH-RISK
PATIENTS
Prognostic enrichment strategies are designed to increase the proportion of patients likely to have
a particular disease-related endpoint event or substantial worsening in condition.
Event-based studies: For any given desired power in an event-based study, a study to lower the
rate of cardiovascular (CV) events or tumor recurrence, the appropriate sample size will depend
on effect size and the event rate in the control group. If the patients enrolled have a high event
rate in the course of the study, the power of a study to detect any given level of relative risk
reduction will increase. A wide variety of prognostic indicators has been used to identify
patients with a greater likelihood of having the outcome event of interest (or a large change in a
continuous measure of interest, e.g., worsening of symptoms). These prognostic indicators
include clinical and laboratory measures, medical history, genomic, and proteomic measures,
among others. In many cases patients at high risk for events are chosen for the initial outcome
study of a drug, and if successful, larger studies in lower risk patients are conducted later.
Progression-based studies: Prognostic enrichment strategies are also potentially applicable to the
study of drugs intended to delay progression of a variety of diseases, such as Alzheimer’s
disease, Parkinson’s disease, rheumatoid arthritis, and multiple sclerosis, and other conditions for
which patients likely to have more rapid progression could be selected. A prognostic marker
may also be predictive (see section V., Predictive Enrichment — Identifying More-Responsive
Patients). That is, the more rapidly progressing patients could be less responsive to treatment
(i.e., rapid progression would be a negative predictor of response) or more responsive to
treatment, as illustrated in the angiotensin converting enzyme (ACE) inhibitor case described in
section IV.A.1., CV Studies.
For any given desired power in an event-based study, the appropriate sample size will depend on
effect size and the event rate in the control group. Prognostic enrichment does not affect the
relative risk reduction but will increase the number of events in a shorter time period, generally
allowing for a smaller sample size. For example, reduction of mortality from 10% to 5% in a
high-risk population is the same relative effect as a reduction from 1% to 0.5% in a lower risk
population, but the reductions in absolute risk are 5% and 0.5%, respectively. The sample size
needed would be smaller in the high-risk population. The increased absolute effect could also
enhance the benefit-risk relationship in the studied population.

Contains Nonbinding Recommendations
6
A. 	Experience With Prognostic Enrichment Strategies
1. 	CV Studies
In CV disease, the severity of the illness being studied and other factors that can indicate
increased risk have been used to identify patients at greater risk for CV events, considerably
reducing the sample sizes needed to show an effect in outcome studies. Examples include a
history of recent myocardial infarction (MI) or stroke; the presence of concomitant illness such
as diabetes or hypertension; and certain blood markers, such as very high low-density lipoprotein
(LDL) cholesterol, low high-density lipoprotein (HDL) cholesterol and high C-reactive protein
(CRP). Outcome studies using ACE inhibitors in heart failure (HF) and 5-hydroxy-3-
methylglutaryl-coenzyme A (HMG CoA) reductase inhibitors in hyperlipidemia (the enalapril
and statin trials, respectively) illustrate this approach.
In the enalapril trials, mortality reduction and decreases in morbid events (such as
hospitalization) were first assessed in a very ill HF population of New York Heart Association
Class IV patients (Cooperative North Scandinavian Enalapril Survival Study (CONSENSUS)),
then in less ill patients (Studies of Left Ventricular Dysfunction (SOLVD) treatment trial), and
eventually in asymptomatic patients (SOLVD prevention trial) (CONSENSUS Study Group
1987; SOLVD Investigators 1991, 1992). In the later studies, composite endpoints were needed
because the number of early deaths was too low to allow a mortality effect to be demonstrated;
event rates were about 15% at 1 year on placebo in the SOLVD treatment trial and about 5% in
the SOLVD prevention trial, far lower than the 44% 6-month mortality in CONSENSUS in the
placebo group. The very high early mortality in CONSENSUS, together with the large effect
size (40% reduction in mortality), allowed demonstration of a survival effect in just 253 very ill
Class IV patients, while the studies in less ill patients required sample sizes of 2,000 to 4,000
patients. The higher risk patients enrolled as a result of prognostic enrichment showed, as
expected, a larger absolute effect size, but relative effect size was also greater in the more ill
patients, suggesting that severity was also a predictive marker (see section V., Predictive
Enrichment — Identifying More-Responsive Patients).
More recently, patients with a ventricular ejection fraction less than 35% to 40% (normal 55–
65%) and levels of plasma B-type natriuretic peptide (BNP) or N-terminal pro-BNP above
specific levels were enrolled in a study of the neprilysin inhibitor sacubitril in HF because these
markers were thought to reflect HF severity. The study showed a significant effect on the
combined endpoint of CV death plus HF hospitalization, which occurred in the control group at
26.5% over about 2 years (McMurray et al. 2014).
A similar strategy of identifying people at very high risk was used in the statin trials. The early
CV outcome trials with statins were able to evaluate the effects of the drugs on mortality because
the patients enrolled in the trials had a history of heart disease and very elevated cholesterol
levels, an indicator of patients whose mortality risk was substantial (Pedersen et al. 1994). As
the benefit of statins became established in high-risk patient populations, patients with less
marked LDL cholesterol elevations and without known coronary artery disease were enrolled in
subsequent CV outcome trials. These populations had not yet been shown to benefit from LDL

Contains Nonbinding Recommendations
7
cholesterol lowering and, therefore, could still be ethically studied; patients in those trials were
identified as high risk because of some other illness (e.g., type 2 diabetes mellitus) or risk factor
(e.g., low HDL cholesterol, elevated high-sensitivity CRP). As the population’s risk became
lower, sample sizes increased considerably, but prognostic factors made the studies possible. For
example, in the JUPITER (Justification for the Use of Statins in Primary Prevention: An
Intervention Trials Evaluating Rosuvastatin) study (n=17,802), a statin was shown to have an
effect on outcome in patients with LDL cholesterol levels that were considered “normal,” but
who were at higher CV risk based on factors other than LDL cholesterol, including age, one
additional CV risk factor, and a high-sensitivity CRP greater than or equal to 2 mg/L (Ridker et
al. 2008). As the magnitude of risk declined in these study populations, relying on composite
endpoints often became necessary because the mortality rate was too low to allow a mortality
trial of reasonable size.
Choosing patients at relatively high risk of CV events can also be critical for safety studies to be
able to rule out a given level of CV risk with a reasonably feasible study size. This approach is
now recommended for the development of new antidiabetic treatments2 and has been a
consideration in the design of studies to evaluate the CV risk of nonsteroidal anti-inflammatory
drugs.
2. 	Oncology Studies
Studies intended to show reduced or delayed occurrence or recurrence of cancer are clearly more
likely to be successful in patients identified as high risk for such events (e.g., by genetic or other
characteristics). For example, adjuvant therapy studies of tamoxifen showed that the drug not
only delayed development of metastases in patients with breast cancer but also reduced the risk
of contralateral tumors (new primary tumors) in this high-risk group (high risk because the
patients previously had breast cancer). Tamoxifen was then studied in 13,000 high-risk women
(calculated using the Gail model) without a prior diagnosis of breast cancer who were followed
for 4 years (National Surgical Adjuvant Breast and Bowel Project (NSABP) P-1) (Fisher et al.
1998). The study showed a 44% relative reduction in risk of invasive breast cancer, and FDA
approved tamoxifen for reducing the risk of breast cancer in high-risk individuals identified
using the Gail model calculator. A study in patients at lower risk would have required a
substantially larger sample size. For example, a study of patients with a risk that was 25% of the
risk of the NSABP P-1 study population would have needed about 20,000 patients to detect an
effect of the size observed in the NSABP P-1 study with 90% power.
3. 	Pulmonary Studies
History of a recent exacerbation in patients with chronic obstructive pulmonary disease (COPD)
is generally thought to predict the likelihood of exacerbations in the subsequent year, and for this
reason, most COPD clinical trials use history of exacerbation as an inclusion criterion (e.g., only
2 See the guidance for industry Diabetes Mellitus — Evaluating Cardiovascular Risk in New Antidiabetic Therapies
to Treat Type 2 Diabetes (December 2008). We update guidances periodically. To make sure you have the most
recent version of a guidance, check the FDA guidance web page at
https://www.fda.gov/RegulatoryInformation/Guidances/default.htm.

Contains Nonbinding Recommendations
8
patients with at least 1 exacerbation in the past 12 months are enrolled). FDA has also qualified
plasma fibrinogen as a prognostic biomarker for all-cause mortality and COPD exacerbations,
improving the power of COPD clinical trials, even when history of exacerbation is also used for
enrichment.3
4. 	Neurology Studies
Studies of drugs intended to decrease exacerbations of multiple sclerosis have generally required
patients to have had a recent exacerbation (one in the last year or two in the last 2 years) or
specified magnetic resonance imaging findings before entry.
B. 	Potential Strategies for Prognostic Enrichment
Examples of additional strategies that may prove useful for prognostic enrichment are described
below. Whether these strategies are useful as enrichment tools is not yet established.
1. 	CV Studies
There may be additional approaches to identifying high-risk CV patients. A report in 2005 noted
that a higher resting heart rate, a small increase in exercise heart rate, and delayed recovery of
heart rate in a population of asymptomatic working men between the ages of 42 and 53 years
were all strong predictors of sudden death, suggesting potential enrichment strategies in studies
of drugs to prevent sudden death (Jouven et al. 2005). The potential for risk prediction based on
genetic factors has been examined, as has the predictive value of coronary artery calcium score
(Ripatti et al. 2010; Polonsky et al. 2010; Ioannidis and Tzoulaki 2010).
2. 	Oncology Studies
a. 	Prostate cancer
In men with localized prostate cancer following radical prostatectomy, high prostate-specific
antigen (PSA) velocity (PSA increase greater than 2 ng/mL during prior year) has been reported
to strongly predict prostate cancer recurrence and mortality over a 10-year period (D’Amico et
al. 2004). Prognostic markers such as PSA velocity, if validated in future studies, could be used
to identify high-risk patients. Studies of adjuvant treatment for prostate cancer would likely be
better able to detect an effect on survival if patients with a high risk of death were enrolled.
b. 	Breast cancer
Many investigators have reported gene expression profiles that appear to predict the likelihood of
breast cancer recurrence after surgery. For an adjuvant therapy trial to be successful in showing
a reduction in tumor recurrence and increased survival, selection of a population with a high rate
of recurrence and poor survival is critical. In a report on the use of five different gene expression
3 See the guidance for industry Qualification of Biomarker — Plasma Fibrinogen in Studies Examining
Exacerbations and/or All-Cause Mortality in Patients With Chronic Obstructive Pulmonary Disease (September
2016).

Contains Nonbinding Recommendations
9
profiling approaches in a nonrandomized 285-patient sample treated with local therapy,
tamoxifen, tamoxifen plus chemotherapy, or chemotherapy alone, Fan et al. (2006) found that
four of the five approaches for classifying patients had a striking ability to predict large
differences in both recurrence and mortality (as illustrated in Figure 1) showing the difference
between patients with good and poor 70-gene profiles. It is apparent that showing improvement
in the low relapse population would require studies of enormous size. Studies in high-risk
patients can be much smaller.
Figure 1: Relapse Rates and Survival in Breast Cancer Patients Based on 70-Gene Profile *
* Fan C et al., 2006, Concordance Among Gene-Expression-Based Predictors for Breast Cancer, N Engl J Med;
355(6): doi: 10.1056/NEJMoa052933.
FDA has cleared MammaPrint (an in vitro diagnostic test using the gene expression profile of
fresh breast cancer tissue samples to assess a patient’s risk for distant metastasis) as a prognostic
test for certain breast cancer patients. As noted, use of such a diagnostic test represents a
potential enrichment strategy for adjuvant trials to identify a population at higher risk for
recurrence.
Women with a deleterious BRCA 1 or 2 mutation have a lifetime incidence of breast cancer and
ovarian cancer of 60% and 15% to 40%, respectively, compared to a risk of 12% and 1.4%,
respectively, in women without a BRCA mutation.4 Selecting women with such markers for a
primary prevention trial in breast cancer or ovarian cancer would increase the likelihood of
cancer events during the trial, thereby permitting a smaller sample size and a shorter study.
V. 	PREDICTIVE ENRICHMENT — IDENTIFYING MORE-RESPONSIVE
PATIENTS
There are many possible ways to identify patients more likely to respond to a particular
intervention, and these strategies have long been used in clinical trials when selection of patients
has been based on a specific aspect of pathophysiology, past history of response, or a disease
characteristic that is related in some manner to the study drug’s mechanism of action (e.g.,
genomic or proteomic factor). For example:
4 See the National Cancer Institute web page BRCA1 and BRCA2: Cancer Risk and Genetic Testing at
https://www.cancer.gov/about-cancer/causes-prevention/genetics/brca-fact-sheet.

Contains Nonbinding Recommendations
10
• 	Congestive heart failure (CHF) can result from either systolic or diastolic dysfunction.
Presumably, a population with systolic dysfunction would be more likely to respond in a
study of an inotrope that would increase ventricular contraction.
• 	High renin status predicts a greater antihypertensive response to beta-blockers, ACE
inhibitors, and angiotensin II receptor blockers (ARBs), all of which work by inhibiting
aspects of the renin angiotensin system. A population with high renin hypertension
would be more responsive than a general hypertension population in studies of drugs in
these classes.
• 	Antibacterial drug effects are best evaluated in patients who are infected with an
organism that is potentially responsive to the antibacterial drug. Most commonly,
patients are randomized before the specific infecting organism is known, but only those
patients with the type of organism targeted by the test antibacterial are evaluated for
effectiveness in the primary analysis. The category of organism is a baseline
characteristic, even though the assessment occurs post-randomization.
• 	An initial screening for response — a biomarker measurement (e.g., radiographic
response, reduction of ventricular premature beats (VPBs)), early clinical response, or
full-fledged clinical response — in an open-label prerandomization period can be used to
identify a responder population that would then be randomized in the controlled study.
This approach is of particular value when responders constitute only a small fraction of
the overall population to be treated.
• 	A population of nonresponders to a different drug can be randomized to the new drug or
to the drug they did not respond to. The comparison is enriched with respect to the active
control comparison because the population is expected to have a poor response to the
original drug compared to the test drug. These designs are not appropriate when
effectiveness is critical to survival or another irreversible outcome or the intolerance is
serious or life-threatening.
• 	Protein or genetic markers related to a drug’s mechanism of action can be used to identify
potential responders. Examples include use of human epidermal growth factor receptor 2
(HER2) overexpression in breast cancer to indicate responsiveness to trastuzumab (a
monoclonal antibody that targets HER2) for breast cancer; tumor surface epidermal
growth factor receptor (EGFR) measurements or mutations indicating responsiveness to
EGFR tyrosine kinase inhibitors for lung cancer, and cystic fibrosis transmembrane
conductance regulator mutation type indicating responsiveness to ivacaftor for cystic
fibrosis.
• 	A protein or genetic marker shown to predict response, even without a documented
mechanism of action.
Identifying a responder population (i.e., a subset of the overall population with a larger than
average response to treatment) and studying this population in a clinical trial can provide two

Contains Nonbinding Recommendations
11
major advantages: (1) increased study efficiency or feasibility and (2) an enhanced benefit-risk
relationship for the patients in the subset compared to the overall population.
A. 	Increased Efficiency or Feasibility
Identification of a population with a high rate of response or a larger response greatly increases
the chance that a study of an effective drug will be able to detect a treatment effect, if one exists,
and allows a study to demonstrate this effect with a smaller sample size than would be needed
for a study in an unselected population. The strategy can be particularly useful for early
effectiveness studies because it can provide clinical proof of concept for later studies. When the
treatment responder population constitutes only a small fraction of all patients, for example, 20%
(a common situation in oncology settings), enrichment can permit a showing of effectiveness
when a study in an unselected population may have difficulty showing any effect.
The prevalence of the enrichment marker and the relative effectiveness of the drug in the marker-
positive and marker-negative populations will determine how much the enrichment strategy can
reduce the sample size needed to adequately power a study. Table 1 illustrates how sample size
ratios — the ratio of the number of subjects needed in an unselected population versus the
number needed if only the marker-positive population is studied — change with varying
prevalence of marker-positive patients and different magnitudes of treatment effect in marker-
negative patients (treatment effect in marker-negative patients of either 0% or 50% of the effect
in marker-positive patients). Table 1 assumes the classification of patients into positive versus
negative is 100% accurate.
Table 1: Sample Size Ratios as a Function of the Prevalence of Marker-Positive Patients
Prevalence of Marker-
Positive Patients
Treatment Effect
in Marker-Negative Patients
(% of Marker-Positive Response)
100%
75%
50%
25%
0% 	50%
Sample Size Ratio 	Sample Size Ratio
1.0
1.8
4
16
1.0
1.3
1.8
2.6
In general, the lower the prevalence of marker-positive patients in the unselected population and
the smaller the treatment effect in the marker-negative population, the more the sample size can
be reduced in a study of marker-positive patients compared to a study in an unselected
population. For example, when the prevalence of marker-positive patients in a population is only
25% and no treatment effect is expected in the 75% of patients who are marker-negative, the
required sample size in a study of an unselected population would be 16 times the sample size
needed for a study that included only marker-positive patients. Simon and Maitournam (2004)
have presented a more detailed description of these results and the conditions under which the
results were obtained.

Contains Nonbinding Recommendations
12
B. 	Enhanced Benefit-Risk Relationship
Identification of a responder population can enhance the benefit-risk relationship of a drug by
avoiding exposure and potential toxicity in patients who would not benefit from the drug. For
drugs with significant toxicity and a low overall response rate in the overall population with the
medical condition to be treated — factors that could deter further development — identifying a
responder population could make a risk more acceptable and facilitate continued development
and approval. For example, the significant survival advantage (approximately 5 months) seen
with trastuzumab in the treatment of metastatic breast cancer in patients with HER2
overexpressing tumors (about 25% of breast cancers) ultimately supported use of the drug in the
marker-selected population despite the significant cardiotoxicity that emerged (see further
discussion in section V.C.2.c., Proteomic and genetic markers with known pathophysiologic
effect). The much smaller effect (less than 2 months) that would have been observed in an
unselected population (i.e., everyone with metastatic breast cancer), and the fact that only about
one-fourth of patients would have benefited, might have made approval difficult to support in the
face of the observed cardiotoxicity of the drug.
Identifying a more responsive population does not necessarily indicate that no benefit exists for
the remaining population. It is therefore generally desirable to have some data in the nonselected
(nonenrichment) population to determine whether that population responds less well or indeed
does not respond at all. These data also can provide an assessment of safety in the nonselected
population in the event that such patients are exposed after approval. The data in the nonselected
population need not be obtained in the controlled trials supporting effectiveness but could be
obtained in earlier studies showing absence of an effect (clinical or biomarker) in a clinical trial,
absence of a critical pharmacologic effect, or even lack of an effect in pertinent nonclinical
studies. A strong mechanistic rationale can make study of the nonenriched population
unnecessary (e.g., study of effects in an infection caused by an organism clearly known to be
resistant; study in a genetically determined disease such as cystic fibrosis, where some patients
do not have the specific genetic variant affected by the treatment).
A trial intended to provide evidence of effectiveness to support approval could include a broad
range of patients but be prospectively designed to evaluate in its primary analysis the effect in
the enriched population subset. This is a standard (and unavoidable) approach when the baseline
characteristic can only be determined after randomization (e.g., the infectious organism, tumor
characteristic), but the approach (preferably with stratified randomization) is also valuable in
other settings to gather some information on the marker-negative population.
C. 	Approaches to Predictive Enrichment
The discussion below considers five predictive enrichment strategies: (1) empiric strategies; (2)
pathophysiologic strategies; (3) empiric genomic strategies; (4) randomized withdrawal studies;
and (5) studies in nonresponders or patients intolerant to other therapy.

Contains Nonbinding Recommendations
13
1. 	Empiric Strategies
With an empiric strategy, the selection of likely responders for a study is not based on any
understanding of the basis for differences in response between patients with or without a
particular characteristic, but on observations of a response during screening periods or prior
experience with the drug or related drugs.
a. 	Open-label single-arm trial followed by randomization
A straightforward enrichment strategy, in cases when a treatment response can be identified
shortly after treatment initiation, is to give the investigational drug to all patients; identify
apparent responders, either on the planned study endpoint or on a biomarker or other short-term
response thought to be predictive of clinical response; withdraw the treatment; and then
randomize only responders into a placebo-controlled trial. This strategy is particularly useful
when the rate is low. This strategy has been discussed in the past and was used throughout the
1970s to develop new antiarrhythmic drugs (Temple 1994; Roden et al. 1980). Patients were
titrated on the investigational drug until they had an acceptable reduction of VPBs. Only the
responding patients were then randomized into placebo-controlled trials, often fixed-dose, dose-
response studies, and sometimes with active controls as well. This approach is useful only when
the response to the drug is of relatively short duration after drug withdrawal. The approach
would not be acceptable if withdrawal would be dangerous to the patient.
The Cardiac Arrhythmia Suppression Trial (CAST) — a study of the mortality effect of
suppressing VPBs in patients with recent acute myocardial infarctions (AMIs) and at least six
VPBs per hour — is one of the best-known studies conducted in apparent responders (Echt et al.
1991). Patients with greater than 10 VPBs per hour after an AMI were known to have a fourfold
increase in the rate of sudden death. Previous failed attempts to show survival benefits with
antiarrhythmics had been criticized because of the low rate of VPB suppression achieved. The
CAST used an open-label screening period to identify responders to two drugs, encainide and
flecainide, which were shown to be very effective in suppressing VPBs in a previous study
(Cardiac Arrhythmia Pilot Study Investigators 1988). Only patients who had at least a 70% VPB
reduction were randomized. Unfortunately, despite the enrichment effort, these antiarrhythmic
drugs did not decrease mortality but instead more than doubled it. This result reflects either the
inadequacy of the surrogate endpoint of VPB reduction as a predictor of an effect on mortality
or, more likely given the adverse effect on survival, an off-target pro-arrhythmic effect of the test
drugs. The result did not, however, reflect a problem with the study design. The enrichment
design yielded a study capable of showing an effect of VPB suppression and allowed clear
interpretation of the study, which showed, contrary to expectations, that even in VPB responders
the drugs were not helpful and, indeed, were harmful.
Use of an initial open-label phase without a control group does raise some concerns that need to
be considered. For example, the deaths that occurred during the screening period for CAST
(which were not unexpected given the recent infarction) were difficult to interpret in an open,
uncontrolled setting where all patients received the active drug. In CAST II (ethmosin versus
placebo), the initial screen for VPB suppression used a randomized comparison of drug to
placebo, with the responders then randomized into the placebo-controlled trial (CAST-II

Contains Nonbinding Recommendations
14
Investigators 1992). This strategy showed that the drug used in the screen was itself lethal (19
deaths on ethmosin versus 1 on placebo), and the study was stopped early.
A similar problem was described in outcome trials of carvedilol in the treatment of chronic HF
(Pablos-Méndez et al. 1998). These studies, unlike CAST, clearly showed a benefit of treatment.
In two large studies, some patients were excluded during a run-in period because they could not
tolerate carvedilol. Some of those patients died. The dropout rates during the subsequent
controlled trials were undoubtedly decreased by the screening procedure that excluded patients
intolerant to beta-blockers, and the results made carvedilol seem better tolerated than it would
actually have been in patients starting therapy. The randomized comparisons and the benefits
demonstrated are fully valid in these trials for the populations studied, but the benefits and risks
facing unselected patients, who would be treated in clinical use of the drug, may be different
from those benefits and risks observed in the clinical trial, requiring close attention to the
screening period results. These results would need to be described in the drug’s labeling.
There are many other outcome study settings in which it would be possible to select patients
more likely to benefit from treatment. Patients with a lipid abnormality might be given the
planned treatment in a screening period to evaluate their biochemical responses. For the
randomized trial, only patients with a response of a certain size might be randomized, giving a
greater mean effect on the lipid level and, presumably, a larger effect on outcome. That
approach could be useful in an early outcome trial, but it would also be possible to randomize a
broader population stratified by such an initial response with the intent of making the primary
study endpoint the result in the high-response subgroup while also gaining some information
about the less responsive group. Again, the response in such a selected group would not describe
the response in an unselected population.
Active, open screening for empiric responders is particularly advantageous when a population
includes subsets with potentially different responses to interventions that are not identifiable
before treatment based on genetic or other pathophysiologic assessments. Although it is hard to
know in advance when this is true, certain difficult-to-study conditions, such as irritable bowel
syndrome or fibromyalgia, might be candidates for this approach (Temple 1994).
The overall strategy (open trial followed by randomization of responders) is an efficient way to
document effectiveness, but it cannot be used prospectively to identify the responder population
when the drug is used in clinical practice. In some cases, however, an early response could be
used to determine who should stay on the drug, which is usually how symptomatic treatments are
used in practice.
b. 	An individual’s history of response to a treatment class
Information about prior responses to a drug in a pharmacologic class, if available, can be used to
identify potential responders for a study of a new member of that class. As is the case with an
open-label trial followed by randomization, use of patient history of response to a drug class can
greatly increase the efficiency of a trial in demonstrating effectiveness. In most cases, however,
it will not help identify the population to be treated in clinical practice.

Contains Nonbinding Recommendations
15
A study enriched with prior responders to a pharmacologic class can be useful in demonstrating
effectiveness at the proof-of-concept stage. This design may be particularly advantageous for
randomized, fixed-dose, dose-response studies (the preferred dose-response study design as
described in the ICH guidance for industry E4 Dose-Response Information to Support Drug
Registration (November 1994)). A responder population provides a larger overall treatment
effect and, therefore, a steeper dose-response curve, which generally allows for easier
interpretation of the curve (identifying the steep area and plateau of the curve) and more precise
characterization of dose-response, especially for doses providing near-maximum effects. For
example, a dose-response study of indapamide in known responders to diuretics demonstrated
mean decreases of 29/12 mmHg (systolic/diastolic) for the 2.5 mg dose and 37/15 mmHg for the
5 mg dose, an increase in effect with the 5 mg dose that was considerably larger than that seen in
studies of unselected patients, where 2.5 mg and 5 mg gave similar results.
c. 	Factors identified in results from previous studies
Analyses of results of previous studies can sometimes point to a substantially greater effect in
a specific subset of the overall population and provide a basis for studying that subset in a
subsequent study, either as the sole population studied or as the identified primary endpoint
subset in a study of a broader population. For example, isosorbide dinitrate/hydralazine
hydrochloride, a treatment for severe HF, was approved on the basis of a placebo-controlled
study in 1050 patients carried out entirely in self-identified blacks (the African-American
Heart Failure Trial) (Temple and Stockbridge 2007; Taylor et al. 2004). The selection of a
black population was based on two previous studies (the Vasodilator-Heart Failure Trials (V-
HeFT) I and II) of an isosorbide dinitrate/hydralazine hydrochloride combination versus
placebo in a racially mixed population that strongly suggested effectiveness in blacks (Cohn
et al. 1986; Cohn et al. 1991). In those studies, the combination had not shown an overall
benefit, but post hoc analyses revealed a nominally significant effect in black patients in V-
HeFT I and apparent equivalence to enalapril in V-HeFT II. In contrast, there was little or no
effect of the combination in whites in V-HeFT I and nominally significant inferiority of the
combination to enalapril in whites in V-HeFT II. The replication of the observed effect in
blacks was strong, with only a suggestion of a modest effect in whites, perhaps a third of the
effect in blacks. A trial to establish this small effect in a white population would have
required 16,000 patients. The product was approved for “self-identified blacks” only.
2. 	Pathophysiological Strategies
Pathophysiological strategies involve selection of likely responders to a drug based on a patient’s
individual physiology or on the assessment of disease pathophysiology that suggests that only
certain patient subgroups will respond to the mechanism of the drug or that certain subgroups
will respond better than others.
Indicators of an individual’s pathophysiology could include biomarkers (e.g., a specific mutation
that affects tumor proliferation), imaging findings, and possibly even demographic or clinical
characteristics that correlate with certain disease phenotypes (e.g., age and race may associate
with renin-angiotensin-aldosterone system activity in hypertension).

Contains Nonbinding Recommendations
16
a. 	Metabolism of the test drug
For a drug that acts through an active metabolite, as is the case for the antiplatelet drug
clopidogrel, patients may differ in their ability to metabolize the prodrug to its active metabolite.
Some patients may not form the active metabolite at all, and others may not make enough to
respond to the dose selected. Including these patients in a trial will dilute the overall drug effect
and can also lead to inefficient or inappropriate use of the drug in practice if the two subsets of
patients are not identified and treated differently. In some cases, adjusting (increasing) the dose
in the poor metabolizers would be possible, but patients who cannot make the active metabolite
at all should probably not be included in the trial or in the planned primary analysis. A closely
related approach is the assessment of uptake of the test drug by a tumor (Stroobants et al. 2003).
Historically, before treatment of thyroid tumors with I-131, a low dose was given to determine
whether the tumor did, in fact, take up iodine and the extent of uptake so that the needed dose
could be estimated.
b. 	Effect on tumor metabolism
Patients for a cancer trial can be selected by screening for an effect on a tumor metabolic
response, as assessed by a positron emission tomography (PET) scan. For example, response to
the tyrosine kinase inhibitors imatinib and sunitinib in patients with gastrointestinal stromal
tumors (GISTs) has been shown to correlate well with metabolic responses (decreased tumor
glucose utilization) assessed by 15F-fluorodeoxyglucose PET imaging (Prior et al. 2009). A trial
could enroll only the identified metabolic responders or could enroll all patients, stratified by
metabolic response, with the primary hypothesis to be tested being the treatment effect in the
metabolic responder stratum.
c. 	Proteomic and genetic markers with known pathophysiologic effect
Increasingly, cancer treatments are directed at enzymatic, hormonal, or other functions that are
tied to tumor cell surface or intracellular receptors and enzymes. The following examples
illustrate use of proteomic markers, or genetic markers that are linked to a proteomic marker, that
are known to be essential for the activity of the drug.
• 	Trastuzumab was developed to bind to the HER2 receptor, which is present on normal
and malignant cells but is overexpressed in about 25% of breast cancers. Binding of
trastuzumab to the HER2 receptor blocks receptor-mediated growth-stimulating
intracellular signaling, decreasing cellular repair after chemotherapy and radiation
therapy and also increasing apoptosis. In activity-estimating trials, antitumor activity in
patients with lower levels of HER2 receptor expression (1+ by immunohistochemical
staining) was minimal, so that efficacy trials in patients with metastatic disease were
limited to patients with HER2 2+ or 3+ overexpression.
In the treatment of metastatic disease, when added to either of two background regimens,
trastuzumab was estimated to have increased median survival by about 5 months, about
three to four times the effect that would have been expected in an unselected population,
assuming no response (which a modest amount of testing showed was the case) in the

Contains Nonbinding Recommendations
17
HER2-negative patients. Enrichment thus allowed a modest-size study to show a striking
effect and directed treatment to the population that could benefit. In addition, because the
drug was shown to be moderately cardiotoxic in the metastatic breast cancer trials, the
focus on potential responders (i.e., patients with HER2 receptor overexpression) was
considered critical when designing adjuvant studies (Smith et al. 2007).
• 	Evidence from the metastatic breast cancer setting has demonstrated that the likelihood of
response to endocrine therapy is related to the hormone receptor status of the tumor. For
example, when treated with tamoxifen, a selective estrogen receptor modulator, patients
whose tumors express both estrogen receptors (ERs) and progesterone receptors (PRs)
have a response rate of approximately 70%; patients whose tumors express either ER or
PR, but not both, have a 20% response rate; and patients whose tumors are ER and PR
negative have a response rate less than 5%. As a consequence, testing of all breast cancer
specimens to direct decisions regarding endocrine therapy, in both the early stage and the
advanced setting, has become the standard of care and would be expected in any trial of
endocrine therapy (Early Breast Cancer Trialists’ Collaborative Group 2005).
• 	A more recent illustration is the use of somatic mutations in the gene encoding the serine-
threonine protein kinase BRAF to identify potential responders to vemurafenib in
melanoma; 40% to 60% of all melanomas carry this activating mutation. In a study of 49
patients with melanoma, 11 of 16 patients with BRAF V600E mutation who received
vemurafenib had a tumor response, compared to 0 of 5 without the mutation (the
remaining 28 patients did not undergo BRAF mutation testing). The phase 3 trial
compared vemurafenib to dacarbazine in 675 patients with metastatic or unresectable
melanoma who had the BRAF V600E mutation. The trial was stopped after an interim
analysis showed a 63% reduction in the risk of death with vemurafenib. The confirmed
response rate was 48% for vemurafenib versus 5% for dacarbazine (Chapman et al.
2011).
The examples of pathophysiologic selection described above reflected, at least initially, variables
related to cellular receptors that could be described as proteomic variables, but that were in many
cases later identified as tumor genetic markers (e.g., EGFR and BRAF genetics). In such cases,
the genetic marker may sufficiently define the pathophysiologic state.
When proteomic and genetic markers are used in an enrichment strategy, adequate
characterization of the test for the marker is critical. An inaccurate assay will undermine an
enrichment effort if the study aims to demonstrate superiority or noninferiority of the test
treatment (Wang et al. 2011). It is also important to gain as much information as possible about
the marker-response relationship.
3. 	Empirical Genomic Strategies
Studies directed at subsets of patients with specific genomic patterns that appear to be associated
with outcomes (e.g., RNA expression profiles, single nucleotide polymorphism arrays) are
becoming increasingly common. Although most genomic markers (e.g., for a tumor surface
property) have been linked to a pathophysiologic property, this linkage is not essential. Use of a

Contains Nonbinding Recommendations
18
genomic marker could instead be an empiric strategy, identifying subsets of responders without
providing a pathophysiologic basis for the difference in response (i.e., before such a basis is
recognized).
Simon and coauthors, for example, in Freidlin and Simon (2005) and Freidlin, Jiang, et al.
(2010), have suggested that a trial population could be divided into two portions, with an
unblinded exploratory analysis of many different genetic markers to identify a predictive
classifier in the first portion. A confirmatory analysis would then be carried out in the
biomarker-defined subgroup in the remaining portion of the trial. Treatment effects would then
be evaluated in the overall population and the biomarker-defined subset from the remaining
portion, with appropriate control of the type I error rate ensured. Any such approach would need
scrupulous attention to maintaining the blind, perhaps by using an independent group to do the
biomarker analysis and should be thoroughly discussed with FDA in advance.
4. 	Randomized Withdrawal Studies
In a randomized withdrawal study, patients who have an apparent response to treatment in an
open-label period or in the treatment arm of a randomized trial are randomized to continued drug
treatment or to placebo treatment. Because such trials generally involve only patients who
appear to have responded, this is a study enriched with apparent responders, an empiric strategy.
The study evaluation can be based on signs or symptoms during a specified interval (e.g., BP,
angina rate), on recurrence of a condition that had been controlled by the drug (e.g., depression),
or on the fraction of patients developing a rate or severity of symptoms that exceeds some
specified limit (i.e., a failure criterion).
The randomized withdrawal design was initially proposed as a way to establish long-term
effectiveness of drugs in settings in which long-term use of a placebo would not be acceptable on
either ethical or practical grounds. Angina was the initial example, but this would also be true
for most psychiatric conditions, pain treatments, and antihypertensive drug treatments (Amery
and Dony 1975). A randomized withdrawal design in which the study population is on treatment
for an extended duration followed by blinded, randomized withdrawal of treatment for a short
duration can provide evidence of prolonged effectiveness with only brief exposure to the
placebo. The design can allow a patient to be removed from the study (for having reached an
endpoint) when the condition returns at a specified severity, avoiding long-term exposure to an
ineffective treatment.
The randomized withdrawal design can also be used as an initial trial to show effectiveness when
there is an existing population of patients in an open-label treatment setting (e.g., under an IND
or as an off-label use of an approved drug), as illustrated by the cases of nifedipine and gamma-
hydroxybutyrate (GHB).
The approval of nifedipine for vasospastic angina (the first drug approved for this condition)
illustrates the utility of this design. An open-label, historically controlled trial was considered
inadequate to support approval because the natural history of vasospastic angina was not well
established (Antman et al. 1980). A randomized withdrawal design (see Figure 2) was
conducted in patients already receiving the drug, with a primary endpoint of recurrence of severe

Contains Nonbinding Recommendations
19
vasospastic angina leading to study withdrawal. A total of 28 patients participated in the study.
One-third of the patients randomized to placebo withdrew early, as compared to no withdrawals
in patients randomized to nifedipine (see Table 2).
Figure 2: Nifedipine Randomized Withdrawal Trial in Vasospastic Angina
Table 2: Results of Nifedipine Randomized Withdrawal Study
Nifedipine
(n=13)
Placebo
(n=15)
Early withdrawal 	0 	5*
Early withdrawal or AMI** 	0 	6*
* Statistically significant at p ≤ 0.05
** Acute myocardial infarction
Another example in which patients already using a drug were studied involved GHB (sodium
oxybate), which was approved for treatment of cataplexy on the basis of a single placebo-
controlled study of conventional design and a second, small, randomized withdrawal study in 55
long-term (7 to 44 months) users randomized to 2 weeks of continued treatment with GHB or
placebo. The second study produced a clinically and statistically impressive result (p< 0.001, as
shown in Table 3) and needed little time for recruitment.
Table 3: Randomized Withdrawal Study of GHB* in Cataplexy
Median Attacks/2 Weeks
Treatment Group 	Baseline 	Change in Rate
Placebo (n=29)
GHB (n=26)
4.0
1.9
+21.0
0
* GHB = gamma-hydroxybutyrate
By randomizing patients to different doses, the randomized withdrawal design can also be used
to obtain long-term dose-response data. For example, this design was used to demonstrate
effectiveness of a single weekly dose of fluoxetine in preventing recurrence of depression;
Randomization
4 weeks
Nifedipine
4 weeks
Placebo
Open-label
Nifedipine
Single-blind
Nifedipine
2 weeks

Contains Nonbinding Recommendations
20
patients on 20 mg per day were randomized to a placebo, fluoxetine 70 mg per week as a single
dose, or continued 20 mg per day. Both fluoxetine groups were superior to the placebo group in
reducing the rate of recurrence. This design has become the standard for studies to demonstrate
the ability of psychotropic drugs to prevent recurrent depression, psychosis, and anxiety.
5. 	Studies in Nonresponders or Patients Intolerant to Other Therapy
A comparative study can be enriched by selection of patients who failed to respond to an existing
drug or who were intolerant of that drug. Although nonresponsive or treatment-intolerant
patients are not more likely than an unselected population of patients to respond to or tolerate the
new drug and would not enrich a placebo-controlled trial, they would generally be less likely to
respond to or tolerate the existing comparator drug, giving the test drug a potential advantage if it
was in fact more effective or better tolerated. Because patients in a trial sometimes respond to a
drug to which they had previously failed to respond, in most cases studies in nonresponders are
informative for the between-drug comparison of effectiveness only if patients are randomized to
both the new and failed drug (i.e., not simply placed onto the new drug in a single-arm study or
randomized to new drug versus placebo). This approach can provide important information to
practitioners; it is critical to know whether another member of a pharmacologic class or a
member of a different class can be useful in patients who fail on previous treatments. A drug’s
effectiveness in nonresponders can be a critical component of a risk-benefit assessment (i.e., it
can allow approval of a drug with a substantial risk). It should be appreciated that for life-
threatening diseases that are progressing, it may not be ethical to randomize to the failed
treatment. The approach may be useful in two settings:
(1) Where the drug has a different mechanism of action from the previous treatments, it may
be more effective in nonresponders to those treatments even if it is not more effective in
an unselected population.
(2) Where the treatment effect of a new drug is moderately superior to the existing drug in an
unselected population, but where a very large study would be needed to show superiority
if the study included unselected patients, many of whom would respond to the less
effective drug. For example, if the new drug response rate is 90% and the existing drug
response rate is 80%, a study with 90% power to detect that 10% difference would
require about 600 patients. In contrast, if only nonresponders to the existing drug were
randomized (20% of the patients treated with the existing drug), few would respond to
the existing drug; but if half of those patients responded to the new drug, the difference
would be detectable with fewer than 40 of the nonresponders.
Note: In neither case would showing an advantage for the new drug in nonresponders to existing
therapy establish superiority of the new drug in an unselected population.

Contains Nonbinding Recommendations
21
a. 	Studies in nonresponders
Several past studies in nonresponders illustrate the value of this study design.
Captopril
To support a claim in severe hypertension unresponsive to other agents, a study was designed to
evaluate patients who had not responded to standard triple therapy (propranolol 320 mg,
hydrochlorothiazide 100 mg, and hydralazine 200 mg). Those patients, who had diastolic
pressure that was severely elevated, were observed for 1 to 2 weeks on the same regimen (triple
therapy lead-in), and if their diastolic pressure did not exceed a defined limit, the patients were
randomized to the same standard triple therapy they had failed on, or to captopril, with a 2:1
captopril to triple therapy randomization ratio. The number of responders (diastolic pressure less
than 90 mmHg or a decrease by at least 10 mmHg) clearly favored captopril in this difficult-to-
treat population (see Table 4).
Table 4: Results of the Captopril Severe Hypertension Trial for the Group Randomized to
Captopril or Triple Therapy
Time Period 	DBP* Findings 	Captopril
(n=66)
Triple Therapy
(n=30)
Week 4 	Normalized DBP ≤ 90 mmHg 	21 (32%) 	5 (17%)
Reduction in DBP ≥ 10 mmHg 	8 (12%) 	3 (10%)
Week 8 	Normalized DBP ≤ 90 mmHg 	22 (33%) 	4 (13%)
Reduction in DBP ≥ 10 mmHg 	14 (21%) 	3 (10%)
* DBP = diastolic blood pressure
Note: Approximately 25% of the triple therapy nonresponders actually responded (diastolic
blood pressure less than 90 or 10 mmHg fall) to the previously failed therapy in the new trial.
This finding reinforces the need for randomization to the new and reportedly failed therapy in a
study of nonresponders. A study that had merely switched patients from the failed therapy to the
new one and found 25% responders might have been interpreted as showing an effect of the new
drug in the nonresponders to the failed therapy when, in fact, the response to the new drug was
no better than the response to the failed therapy would have been in a new study. The finding of
an effect in nonresponders helped overcome concerns about agranulocytosis seen with high
doses of captopril.
Clozapine
Clozapine is an antipsychotic drug associated with serious toxicity, a greater than 1% rate of
potentially fatal agranulocytosis. For clozapine to be approved, the sponsor had to show that the
drug offered a clear advantage over safer alternatives. To show this, a study was conducted in
hospitalized schizophrenic patients with a history of poor response to neuroleptics who, in
addition, had failed to respond to 6 weeks of treatment with haloperidol. These patients were
randomized to 4 weeks of treatment with clozapine or chlorpromazine plus benztropine. The
results showed a striking advantage for clozapine on Clinical Global Impression and the Brief
Psychiatric Rating Scale standard measures in antipsychotic drug trials (see Table 5). Despite its

Contains Nonbinding Recommendations
22
serious risk, clozapine was approved for use in patients not responding to other antipsychotic
drugs.
Table 5: Results of Clozapine Study in Nonresponders to Standard Psychotropic Drugs
Response (%)
Measure 	Clozapine 	Chlorpromazine
CGI* (decrease > 1) 	71 	37
BPRS* items (decrease > 1)
Conceptual disorganization
Suspiciousness
Hallucinations
Thought content
60
64
59
15
39**
42**
51
2**
CGI and BPRS 	15 	2**
* CGI = Clinical Global Impression; BPRS = Brief Psychiatric Rating Scale
** p < 0.05
Rofecoxib
Individual patients are often believed to respond differently to different nonsteroidal anti-
inflammatory drugs. To examine this belief, a controlled trial was conducted in which
osteoarthritis patients identified as nonresponders to celecoxib were randomized to celecoxib or
rofecoxib (Temple 2012). In fact, no difference was observed between the treatments (see
Figure 3).

Contains Nonbinding Recommendations
23
Figure 3: Study Comparing Rofecoxib to Celecoxib in Celecoxib Nonresponders*
* Temple R, 2012, A Regulator’s View of Comparative Effectiveness Research, Clin Trials, 9:56–65.
Considerable and prompt improvement in pain was reported in both groups. A baseline-
controlled, single-arm trial of rofecoxib would have led to a clearly erroneous conclusion that
rofecoxib was effective in celecoxib nonresponders and even a placebo-controlled trial of
rofecoxib in this population might have shown an effect that would have been incorrectly
interpreted as an effect in celecoxib nonresponders because whether they would have responded
to celecoxib in the controlled trial setting would not be known.
b. 	Study in intolerants: ARBs in patients who cough on lisinopril
Studies of the tolerability of a new drug in patients who do not tolerate a previous treatment are
also informative and efficient. Comparative studies in an unselected population could provide
some information on relative tolerability, but a very large study would be needed to show small
differences. For example, if the true rates of cough for an ACE inhibitor and an ARB were 5%
and 1%, respectively, a study with 90% power to show a difference in an unselected population
would need about 800 patients. In contrast, a study in patients known to cough on ACE
inhibitors would need fewer than 20 patients, if, for example, the cough rate were greater than
90% in the ACE inhibitor arm and 20% in the ARB arm.

Contains Nonbinding Recommendations
24
This approach was used in a study of 84 elderly hypertensive patients with history of coughing
on an ACE inhibitor (Chan et al. 1997). Cough was confirmed by rechallenge with the ACE
inhibitor lisinopril, administered for up to 8 weeks, which had to cause at least moderate
coughing for patients to continue in the study. Lisinopril was then withdrawn for 4 weeks, and
coughing had to disappear. The patients were then randomized to losartan 50 mg, lisinopril 10
mg, or metolazone (diuretic active control that does not induce coughing) for 10 weeks. The
study achieved a very persuasive result with this small population (see Table 6), a result cited in
losartan’s labeling.
Table 6: Comparison of Coughing Rates With ACE Inhibitor, ARB,
and an Additional Active (Non-Cough-Inducing) Control
Lisinopril
(n=28)
Losartan
(n=28)
Metolazone
(n=28)
Any cough 	97% 	18%* 	21%*
* p < 0.001, lisinopril vs metolazone and losartan
As was the case in studying nonresponders to previous therapy, randomization to the poorly
tolerated previous therapy and the new drug is critical to reach a conclusion that the new drug
has superior safety because the adverse reactions do not always reappear when a treatment is
repeated. For the same reason, FDA recommends that sponsors should include a placebo group
(or, as in the above example, an active drug clearly lacking the adverse effect) to be certain that
the adverse reactions were indeed reproduced in the previous treatment group. This study design
is not feasible if the adverse reaction is dangerous to the patient.
VI. 	ENRICHMENT STUDY DESIGN AND OTHER CONSIDERATIONS
A. 	General Considerations
An enrichment design should be explicitly described in the protocol and final report and should
fully detail the enrichment maneuvers and their effect on the interpretation of results (see section
VI.E., Cautions in Interpretation).
Some enrichment strategies depend on a screening measurement for selecting the enriched
patient population. For both prognostic and predictive markers, understanding the accuracy and
performance characteristics of the test used to identify patients or marker-defined subgroups for
enrichment is important. Note that the question of who to include in the study and what patients
to analyze in the primary analysis are distinct issues and may differ for predictive and prognostic
enrichment.
For prognostic enrichment, the ideal screening measurement to be used for patient selection will
have good sensitivity and specificity. That is, it will reliably identify patients with a greater
likelihood of having disease-related endpoint events or substantial worsening in their conditions
(high-risk group) and reliably exclude patients in the lower risk group, both of which will serve
to enhance the trial’s efficiency. If the screening measurement’s specificity is poor, many

Contains Nonbinding Recommendations
25
patients who will not have clinical worsening will be selected, and the effect of enrichment will
be diluted. If sensitivity is poor, patients who are at high risk will be missed in the screening
process, delaying enrollment into the study. Note, however, that although exclusion of lower
risk patients will increase absolute effect size, allowing a smaller trial, the relative effect is
expected to be similar in higher and lower risk patients. This is in contrast to predictive
enrichment, where patients without the enrichment factor would be expected to have a lower
relative effect size.
Uncertainties about the performance characteristics of an enrichment strategy based on a
prognostic marker have different implications for superiority and noninferiority studies. For a
superiority study, a prognostic enrichment strategy with poor sensitivity and/or specificity can
increase the sample size or duration required to observe the needed number of endpoint events.
Although this will reduce the efficiency gained through enrichment, it will not lead to a false
conclusion of effectiveness. For a study intended to demonstrate noninferiority, however, the
use of a prognostic enrichment strategy with poor ability to select subjects for entry into the
study could have the effect of increasing the chances of reaching a false noninferiority
conclusion because the control effect size would be smaller than expected from historical
experience (Wang et al. 2011). In general, patients in a noninferiority study should be selected
according to criteria similar to the past studies of the active control.
For predictive enrichment, a screening measurement of a biomarker with both high sensitivity
and specificity for identifying responders is also desirable. A screening measurement cut-off can
be set to cast a wide net, i.e., with high sensitivity giving a high probability of identifying and
treating patients who are more likely to respond to the treatment (higher sensitivity), at the risk
of also treating more nonresponder patients (lower specificity). Alternatively, the measurement
cut-off can be set to identify and exclude from treatment patients who might not respond (higher
specificity), at the risk of also excluding more responder patients. These trade-offs can be
explored using receiver operating characteristic analyses.
The performance characteristics of enrichment strategies based on predictive markers also have
different implications for superiority and noninferiority studies. For a superiority study,
predictive markers with poor sensitivity and specificity will lead to increased sample size but
will not increase the chance of a false conclusion of efficacy. For a noninferiority study, the
effect on the type I error rate is more complex and depends on whether the marker is pertinent to
both treatments or only one treatment; again, however, patients should generally be selected
according to criteria used in past studies of the active control.
B. 	Which Populations to Study
As will be described in more detail below, trials can be designed to: (1) include only patients
with the enrichment factor; or (2) include patients with and without the enrichment factor, but
with an intent to analyze only the patients with the enrichment factor as one of the primary study
hypotheses. This is a potentially critical step in the predictive enrichment setting where patients
without the factor are not expected to respond, thereby diluting the effect size if included.
Studies including both populations need not include a wholly unselected population of patients
with the disease to be treated but can designate separate sample sizes for patients with and

Contains Nonbinding Recommendations
26
without the enrichment characteristic to allow collection of sufficient information to demonstrate
effectiveness in the enriched subgroup and to allow a reasonable estimate of effect in the
nonenriched group. Many design alternatives have been discussed in the literature (Mandrekar
and Sargent 2009; Freidlin, McShane, et al. 2010).
A critical question in all settings in which enrichment is used is therefore to what extent the
marker-negative population should be studied. In some cases, study of the general population
(one including the marker-negative population) would not be expected. For example, if
prognostic enrichment is used to ensure that there are sufficient events to make a trial feasible,
even if the treatment effect is thought to also be present in the lower risk population without the
marker (but at a lower absolute effect size), it may not be possible to design a trial that includes a
significant fraction of the marker-negative population without greatly increasing the sample size
— a strategy that may make the trial impractical and defeat a major purpose of prognostic
enrichment. Whether to use the drug in the unstudied marker-negative population would depend
on the particular circumstances. For example, the presence of significant toxicity could lead to
doubts about the advisability of using the drug in the lower risk population. The variability-
reducing factors discussed in section III., Decreasing Variability, would not ordinarily call for
study of the population lacking the enrichment factor (e.g., patients with poor adherence).
It is principally in the area of predictive enrichment, especially predictive enrichment using a
pathophysiological measurement or biomarker, that the question of studying the population
without the enrichment factor is most germane. Experience shows that the selected enrichment
factors often do not precisely dichotomize patients into subpopulations that will and will not
respond so that obtaining some information on the marker-negative population to assess
performance of the factor is usually desirable. However, even an imperfectly characterized
predictive marker can greatly increase the power and likelihood of success of a study. Moreover,
in treating serious and life-threatening illnesses (especially when alternative treatments exist),
using the test treatment in patients thought unlikely to respond raises critical ethical issues.
Efforts to use predictive enrichment thus offer a number of design choices. The study designs
illustrated below are fixed sample size designs that can be used with predictive enrichment
strategies (also see section VI.D., Adaptive Enrichment, concerning adaptive enrichment and
nonfixed sample size). The examples describe trials intended to show superiority of the test
treatment to a control (e.g., placebo, standard of care), but noninferiority studies would present
similar issues.
1. 	Studying Marker-Positive Patients Only
A study randomizing only marker-positive patients is shown in Figure 4. Because a study that
uses a marker-positive only population will provide no direct information about the marker-
negative population, its use should generally be limited to situations in which information about
the marker-negative population is not needed or is not feasible given the objectives of the study.
For example, if it appears clear based on mechanistic, nonclinical, or early clinical data that the
marker-negative patients will have no or minimal response or would be exposed to unreasonable
risk, inclusion of the marker-negative patients would, in most cases, not be justified.

Contains Nonbinding Recommendations
27
Figure 4: Study Design of Predictive Enrichment Trial With No Possible Effect in the
Marker-Negative Group
The study shown in Figure 4 would support an effectiveness claim in the enriched population,
but it would overstate the actual effectiveness for an unselected population, so that the fraction of
marker-positive patients would be important information. The study would provide no new
clinical evidence with respect to the marker-negative population and would not further
characterize the predictiveness of the marker because of a lack of ability to compare
effectiveness in marker positive and marker negative patients. Because no effectiveness or
safety information on the enrichment-marker-negative patients would exist, the selection process
should be fully described in labeling and in clinical practice patients would usually be assessed
for the enrichment marker before exposure. Moreover, because assessment of marker status is
critically important to determining whether the drug will be effective in patients, the test to
assess the enrichment marker that would be used after the drug’s approval would be an
established, FDA-cleared or -approved, laboratory test explicitly labeled for this purpose as a
companion diagnostic, although exceptions can be considered for a major advance in treatment.5
2. 	Studying Both Marker-Positive and -Negative Patients
FDA encourages inclusion of some predictive marker-negative patients in most trials intended to
provide primary effectiveness support, unless earlier studies have established that the marker-
negative patients do not respond or a strong mechanistic rationale makes it clear that they will
not respond. Significant toxicity of the test drug could reduce the level of the evidence needed to
conclude that the drug should not be studied in marker-negative patients. In general, the greater
the uncertainty about the marker cutoff and responsiveness of marker-negative patients, the more
important inclusion of a reasonable sample of marker-negative patients becomes. When
substantial incentive exists to use the drug in the marker-negative population (e.g., for serious
diseases with few alternative therapies), characterization of the response in the marker-negative
population is more important, especially if the drug has important safety concerns.
There are two cases to consider in studies that include both marker-positive and marker-negative
patients: (1) when the marker can be assessed before randomization; and (2) when the marker
can be assessed only after randomization. Figures 5 and 6 provide sample study designs for
these two cases (Freidlin, McShane, et al. 2010).
5 See the guidance for industry and FDA staff In Vitro Companion Diagnostic Devices.

Contains Nonbinding Recommendations
28
In the first case (see Figure 5), marker status is determined for all patients, and randomization is
stratified by marker status. The primary study objective would usually be a statistically rigorous
demonstration of the treatment effect in the marker-positive patients, and the study would be
powered for the effect in that group. The size of the marker-negative group would be determined
separately, and randomization of all marker-negative patients would not be necessary. Because
the treatment effect would be expected to be much smaller (if there were any effect) in the
marker-negative population, the size of the marker-negative population would usually be too
small to provide a definitive answer on the effect; however, the marker-negative patients would
provide at least some estimate of the effect in that population.
This design could also enable a risk-benefit assessment for the drug when used in the overall
population with the medical condition to be treated, which would be advantageous if some
exposure in marker-negative patients is anticipated in clinical practice (e.g., because the test is
not widely available). When substantial uncertainty exists about whether a marker is predictive
(i.e., whether it can select a population in which treatment is effective), the primary endpoint
could be the effect in the overall population, or study alpha could be divided between the two
endpoints (overall population and marker-positive population). In each of these scenarios, the
hypotheses should be clearly specified and control of the type I error rate should be addressed
(see section VI.C., Type I Error Rate Control for Enriched Study Subpopulations). In addition,
tests for heterogeneity of treatment effects (i.e., treatment by biomarker interactions) should be
prespecified.
Figure 5: Study Design of Predictive Enrichment Trial With Possible Effect in the
Marker-Negative Group
The second case (see Figure 6) is one in which a drug is expected to be effective only in the
marker-positive subset (e.g., only in patients with a targeted type of organism), but the drug must
nonetheless be given to all patients because the marker result is not available at randomization.
Having the primary study outcome be the effect in the marker-positive subset would still be
appropriate, but the risk-benefit assessment would reflect results in the entire population (i.e., the
population that would be exposed to treatment). In these cases, when the marker test results will
not be known before drug administration and when no patient management decisions will be
made on the basis of the test result (e.g., decisions to discontinue treatment in marker-negative
patients), FDA clearance or approval of the test contemporaneously with approval of the drug is
not needed.

Contains Nonbinding Recommendations
29
Figure 6: Study Design of Predictive Enrichment Trial Where All Patients Are
Randomized Because Marker Cannot Be Assessed Before Randomization of the
Marker-Negative Group
C. 	Type I Error Rate Control for Enriched Study Subpopulations
Generally, even if patients both with and without a predictive enrichment characteristic are
studied, the primary endpoint is expected to be driven by the result in the enriched subgroup.
For enrichment designs that enroll patients both with and without the enrichment characteristic,
the type I error rate for the study can be shared between a test conducted using only the enriched
subpopulation and a test conducted using the entire population. While allowing the assessment
in the enriched subgroup, this alpha allocation scheme also allows for assessment of the
treatment effect in the entire population when there may be some effect in patients without the
enrichment characteristic. Determining the required sample size needed to provide reasonable
power to test the different hypotheses while controlling the type I error rate (usually including a
prespecified order of testing or a multiple testing procedure allowing testing of both hypotheses)
is challenging. The interpretation of the study finding would have to take into consideration the
magnitude of the effect in the nonenriched group.
D. 	Adaptive Enrichment
Enriching a clinical trial using prognostic or predictive markers can pose a challenge if there is
uncertainty at the planning stage about the performance characteristics of the enrichment
strategy, such as incomplete information on the prevalence of the marker and/or the type and
strength of the marker-outcome relationship. This can be a particular problem for predictive
enrichment. A study design that incorporates planned adaptations to the enrichment strategy,
taking advantage of information gained on marker performance during the course of a clinical
trial, may be useful in meeting these challenges. For example, the prospective determination of
the sample size required to detect a particular treatment effect with sufficient power, either in the
overall population or in a marker-defined subpopulation, can be difficult when there are
uncertainties about the characteristics of the enrichment strategy, such as:
• 	The cutoff value of the biomarker used to identify patients in the marker-defined
subgroup
• 	The proportion of patients who are in the marker-positive subgroup
• 	The magnitude of the treatment effect in patients with and without the marker

Contains Nonbinding Recommendations
30
With a fixed sample enrichment design, such unknowns create uncertainty as to whether the
study will successfully meet its objectives. An adaptive design in which the sample size and
other design features can be modified in a prospectively planned manner to adapt to information
obtained during the course of the trial (e.g., the frequency of marker positivity) may enhance the
value of certain enrichment strategies. Similarly, determination that the enrichment factor has a
greater or lesser effect on response than anticipated or that the patients without the enrichment
factor have a higher or lower response than anticipated, could trigger planned changes in sample
size or entry criteria in response to the accumulating information. Such changes should be
prospectively planned and would often need appropriate type I error rate control to account for
interim, unblinded analyses of the accumulating data as well as analyses of multiple subgroups.
If the only change was increased sample size based on blinded, pooled results because the
prevalence of the marker-defined subgroup was lower than expected, there would be no need for
a type I error rate adjustment.
A few examples of adaptive enrichment trials have been reported in the literature. The
Investigation of Serial Studies to Predict Your Therapeutic Response With Imaging and
Molecular Analysis Phase 2 (I-SPY 2) is an example of a phase 2 screening trial using an
adaptive enrichment strategy. This is a clinical trial for women with newly diagnosed, locally
advanced breast cancer. The trial is designed to evaluate the effect of investigational drugs
added to standard chemotherapy compared to standard chemotherapy alone on pathological
complete response rates. The treatments being studied are neoadjuvant therapies, that is, drugs
administered before surgery. The treatment phase of this trial involves testing multiple
investigational drugs that are thought to target identified biomarkers of each patient’s tumor type.
The trial is an adaptive enrichment design in that biomarkers are used at baseline to identify
patients likely to respond to treatment, and the accumulating responses are then used to inform
treatment assignments for subsequent participating patients as the trial progresses. The results of
I-SPY 2 are intended to inform the design of future trials for establishing efficacy and safety of
those treatments that advance to phase 3 (Barker et al. 2009).
Although there are challenges with enriched study designs that involve preplanned changes to
the enrichment strategy based on information accrued during the trial (e.g., change sample size
after the start of the study), the following three examples illustrate potential adaptation strategies:
(1) In a study that includes both marker-positive and -negative patients, and with a primary
endpoint that is the effect in the marker-positive group, a planned interim look could
reveal, either on an early endpoint (e.g., imaging, pharmacodynamic biomarker, tumor
response rate) or later endpoint (e.g., progression-free survival), that the marker-negative
population has a much lower response than the marker-positive group. Additional
enrollment of marker-negative patients could be reduced or stopped entirely, provided an
adaptive enrichment strategy was planned and was appropriately accounted for in
statistical analyses.
(2) Planned interim analyses could suggest changing entry criteria to increase enrollment in
what may be a better responding subgroup. If such an adaptive enrichment strategy is
undertaken, appropriate control of the type I error rate may need to be demonstrated,
depending on the analyses used. As noted above, although type I error rate adjustments

Contains Nonbinding Recommendations
31
would be needed for interim unblinded assessments of efficacy, no additional adjustment
would be needed for changing the entry criteria per se, as long as all randomized patients
based on the originally planned sample size were included in the final analysis (Mehta et
al. 2009).
(3) A study could be designed to obtain more precise information on the performance
characteristics of the marker or other characteristics used for enrichment, for example, by
examining an early endpoint using several different biomarker cutoff values to determine
the optimal cutoff value. If a cutoff value proved too low or too high, it could be
changed. Such plans would need to be specified in advance, and appropriate type I error
rate control demonstrated.6
E. 	Cautions in Interpretation
Any use of an enrichment design should be explicit in the protocol and study report and should
fully detail the rationale for the design, the specific enrichment maneuvers, and their effects on
the interpretation of results. For example, if only half of the patients screened passed the entry
test, that should be noted, and the effect of this selection in terms of the expected response rate in
the overall population and on the generalizability of the results should be evaluated. The
importance of such descriptions is obvious for trials in which high-risk patients (prognostic
enrichment) and probable responders (predictive enrichment) have been selected, where the
description is critical to knowing to which patients the results apply, but such descriptions are
important for all types of enrichment studies. Given the potentially complex interpretation of
studies using enrichment designs, plans to use them should be discussed with FDA early in
development.
When enrichment depends on a proteomic or genetic biomarker, particularly if a test for the
biomarker is intended for use in practice to identify patients to be treated, the analytical validity
of the test is critical. In addition to assay validity, for any marker used to select patients, even a
familiar one, the biomarker’s clinical sensitivity, specificity, and positive and negative predictive
values should be well characterized. To the extent an enrichment strategy successfully identifies
patients with high event rates or high response rates and leads to a successful study, study results
could be said to speak for themselves (i.e., the randomized trial did show an effect; the event rate
was high enough) and certainly support the effectiveness of the drug in the population studied.
Again, however, the enrichment strategies should be clearly described in the study report and
labeling to indicate how the drug is to be used and to whom the results may apply (groups of
patients that are known to benefit or not to benefit and groups in which effect is not known).
Selection of the optimal predictive enrichment study design (specifically, whether to include
both marker-positive and -negative patients and whether to introduce adaptive elements) can be
difficult to determine when there is uncertainty about the properties of the enrichment marker.
Many publications have addressed these issues (Freidlin and Simon 2005; Mandrekar and
Sargent 2009; Freidlin, McShane, et al. 2010). One conclusion is that the greater the uncertainty
6 The draft guidance for industry Adaptive Designs for Clinical Trials of Drugs and Biologics (September 2018).
When final, this guidance will represent the FDA’s current thinking on this topic. For the most recent version of a
guidance, check the FDA guidance web page at https://www.fda.gov/RegulatoryInformation/Guidances/default.htm.

Contains Nonbinding Recommendations
32
regarding the marker cutoff and responsiveness of marker-negative patients, the more sense it
makes to include a reasonable sample of marker-negative patients, perhaps using an adaptive
design to exclude such patients if they are seen not to respond. In general, stratification by
marker status would be sensible, especially when marker prevalence in the population studied is
relatively low.
VII. 	ENRICHMENT — REGULATORY ISSUES
A. 	Summary — The Decision to Use an Enrichment Strategy
The decision to use an enrichment design is largely left to the sponsor of the investigation, but
like the overall research and clinical communities, FDA is very interested in targeting treatments
to the patients who can most benefit from them (i.e., individualization or “precision medicine”).
FDA’s interests also include the adequacy of the study (Will it successfully assess effectiveness
in a defined population and, in so doing, support marketing approval?) as well as how study
findings can be described in drug labeling. As discussed above, a critical question is almost
always how much data will be needed in the off-target (nonenriched) population, particularly for
predictive enrichment strategies. These issues are critical aspects of a development program that
should be discussed with FDA early in development.
There are many reasons to use enrichment designs, including an enhanced benefit-risk
relationship if a population with an increased likelihood of response can be identified, and
efficiency in drug development, as smaller studies can often be used to demonstrate
effectiveness. Sponsors, however, should consider the following two critical regulatory issues
when contemplating the use of enrichment designs.
1. 	Does the Enrichment Strategy Identify the Patients to Whom the Drug Should Be
Given?
When patients with an increased likelihood of response can be defined before treatment by a
predictive marker (e.g., a pathophysiologic or genomic characteristic, a short-term screen such as
response to a test dose), a straightforward method is available for selecting patients for treatment.
In contrast, some empiric strategies that provide predictive enrichment (e.g., studying known
responders in a conventional study or in a randomized withdrawal study) can efficiently establish
the effectiveness of a drug in a subset of the population but provide no way for prescribers to
prospectively identify patients with a greater likelihood of response or to predict the magnitude
of response in an unselected patient. Although this type of untargeted treatment may seem
troubling (treatment of many to attain a response in only some), the reality is that this is
generally the case with treatments that are approved on the basis of conventional studies in a
nonenriched population, where there is typically a wide range of responses, including no effect at
all or even harm in some patients. However, it needs to be understood that the magnitude and/or
likelihood of a treatment response for an unselected patient could be substantially less than the
mean response observed in a clinical trial that employed an empiric enrichment strategy, and this
should be clear in labeling. When the prescriber is able to quickly gauge the effectiveness of a
drug in an individual patient (e.g., pain is relieved, cholesterol is reduced), the pretreatment

Contains Nonbinding Recommendations
33
ability to predict the likelihood of a drug response with accuracy may not be as critical as it
would be if response is delayed.
In some cases, enrichment cannot be used to select patients for study because the enrichment
factor is not known until after the treatment is initiated but is used to identify the subset of the
studied population to be analyzed — for example, the patients with a prerandomization sample
of tissue, sputum, etc. that shows (later, after treatment is started) the presence of a targeted type
of organism in studies of antimicrobial drugs. Again, the subset analysis documents
effectiveness, but the population entered into the trial and randomized to treatment groups, at
least initially, will be the unselected patients (i.e., a larger group than the population of potential
responders). Such situations are unavoidable, however, if treatment is urgent and must be
initiated before the enrichment test results are available.
Finally, knowing the precise utility of the enrichment strategy may not be critical to determining
that a drug has an effect (that is if the selected study population shows an effect, the drug was in
fact effective), but knowing the effect of enrichment and the difference in effect between
populations are important for therapeutic use of the drug. It is plainly undesirable to treat patients
who will not respond and to fail to treat potential responders because the enrichment marker has
poor performance characteristics or because cutoff points for the marker were poorly selected.
These problems notwithstanding, if the enrichment strategy enables a drug of value to be
developed and to be shown to be effective when disease and response variability would make
nonenriched studies unable or unlikely to succeed, there is clearly an important gain from use of
such strategies. Labeling will reflect limitations and concerns, but it seems clear that a drug
shown to be effective and safe in an enriched study should be available even if the responder
population is not identified as precisely as would be desirable.
2. 	Might the Drug Be Useful in a Broader Population Than Was Studied?
The data that should be obtained for the marker-negative patients are considered below, but it
can be anticipated that less information will be available about those patients and there will be
greater uncertainty as to their responses to the treatment. Studies in unselected patients (i.e., a
nonenriched population), the typical basis for drug approval, simply ignore the question of
identifying treatment responders and lead to treatment of many patients who will not benefit.
There would thus seem to be a gain from a process that seeks to establish the characteristics that
predict a drug response, rather than one that ignores the varied responses and overcomes the
variability by simply increasing sample sizes.
In general, then, FDA’s regulatory standards do not bar the Agency from approving drugs whose
effectiveness has been demonstrated primarily or even solely in enriched populations, and FDA
will seek to ensure truthful labeling that does not overstate the likelihood of a response, the
magnitude of the response, or the predictiveness of the enrichment factor. But the extent of data
that should be available on the nonenriched subgroup should always be considered.
Postmarketing commitments or requirements may be requested to better define the full extent of
a drug’s effect (including efficacy and safety studies and trials in a broader population).

Contains Nonbinding Recommendations
34
B. 	Study of Marker-Negative Patients
Well-controlled enrichment studies, if successful, provide evidence of effectiveness in the
population studied. In many cases, however, questions will remain about whether some effect of
the treatment occurs in marker-negative patients and the extent to which the treatment effect
should be characterized in this population. A related issue is whether labeling should discourage
use in marker-negative patients and, if so, how strongly such use should be discouraged. In
general, heterogeneity-reducing efforts raise few issues of this kind, but prognostic enrichment
strategies, and especially predictive enrichment strategies, do raise them. The solutions are not
always clear and may be circumstance specific.
In general, studies using prognostic enrichment have been accepted as a basis of approval
without a requirement to study broader populations. Nonetheless, when studies have shown that
a drug reduced the rate of serious or irreversible endpoints, showing an effect in a high-risk
population (e.g., high BP, high LDL with a history of MI, severe CHF), they have, historically,
been followed by later assessments of effects in lower risk patients. Subsequent studies in lower
risk populations have generally shown an effect, but the effect size has often been smaller, and
the study endpoints have sometimes changed (mortality in the high-risk studies versus composite
endpoints in the lower risk patients). Consequently, benefit-risk considerations may have
changed. FDA has generally accepted the results from prognostically enriched studies, approved
an indication based on the observed effect, and described the study, including the patient
population, in the CLINICAL STUDIES section of labeling, with any enrichment selection
criteria noted. In most cases, the specific patient population studied has been the indicated
population in the INDICATIONS AND USAGE section of labeling, in addition to its description
in the CLINICAL STUDIES section. Labeling should also note populations that were not
studied.
When predictive enrichment is used in drug development and the treatment represents an
important advance for the marker-positive group, delaying approval because of limited data in
the marker-negative group would generally be unreasonable. Nonetheless, in such cases,
potential effectiveness in the marker-negative group is of great interest, and questions will
therefore arise about whether there is a treatment effect, even if a smaller effect, in the marker-
negative patients and about the precision of the dividing line chosen to define marker positivity.
It must be appreciated that a study sized to show a treatment effect in the predictively enriched
population, even if there is a modest representation of marker-negative patients, will have
relatively little capacity to detect or rule out an effect in the marker-negative population.
Nevertheless, the marker-negative population will generally have at least some data, as the study
designs in Figures 5 and 6 indicate. In addition, information on possible differences in the
treatment effect based on the enrichment factor can sometimes be obtained from
pathophysiologic studies, nonclinical studies, in vitro clinical studies, or combinations of various
kinds of information.
Determining the need to characterize the treatment effect in the marker-negative population will
be based on potential benefit and risks in that population, including: (1) the nature of the
efficacy shown in the marker-positive population (which could range from reduction in

Contains Nonbinding Recommendations
35
symptoms to enhanced survival); (2) the risks of the drug, in particular risks that might be
present in the marker-negative population; (3) whether the effect of treatment would be apparent
to an individual patient; (4) the relative sizes of the marker-positive and -negative populations;
and (5) the desire to use the drug in the marker-negative population (the impetus is greater when
benefit is substantial and/or alternative treatments are limited).
When treatment benefit is clinically critical (e.g., leading to prolonged survival or prevention of
significant disability), physicians may want to prescribe the drug in the marker-negative
population, especially if treatment options are limited, and obtaining a reasonably reliable
assessment of the effect in the marker-negative group is especially important.
When treatment benefit is not clinically critical, it is important to consider the expected toxicity
in the marker-negative population, the relative size of the marker-positive and marker-negative
populations, and how much evidence there is that no treatment effect occurs in the marker-
negative population. The balance of these factors would determine both the need to discourage
or limit use through labeling and other measures and the need to characterize the treatment effect
in the marker-negative population.
When individual patients can determine whether they are deriving benefit from a drug (usually
for drugs used to reduce symptoms in the short term), patients can decide to discontinue a drug if
it fails to provide symptom benefit. In this situation, determining the treatment effect in the
marker-negative population is generally less important. This is because if the drug is used in the
marker negative population, the determination of whether or not to discontinue the drug is made
in a straightforward manner based upon the patient’s report. How reassuring this will be depends
substantially on the seriousness of the drug’s adverse effects.
For drugs used to reduce the risk of morbid outcomes, patients typically cannot determine
whether or not they will ever derive benefit, and such drugs will usually be used for long periods
of time. When the drug has the potential to cause significant harm in the marker-negative
population, a greater need to assess the treatment effect in that population arises. Also, the larger
the relative size of the marker-negative population, the more important characterization of the
benefit in that population becomes. For example, if only 10% of the population is marker-
positive, and marker-negative (90%) patients are not known to benefit from the drug, sponsors
should assess the benefit in the marker-negative population and communicate the findings.
Under these circumstances, the benefit of the treatment in the marker-positive population could
be outweighed by the overall risk in the marker-negative population, if this population is treated
extensively, and communicating both the risk and the lack of benefit in that population would be
important.
A number of considerations may support collection of less (or sometimes no) information on the
enrichment-factor-negative population:
• 	A clear pathophysiologic basis for concluding that the nonenriched population will not
respond (e.g., because the patients lack the molecular target of the drug, or because they
cannot convert a prodrug to its active metabolite). This type of support could be provided

Contains Nonbinding Recommendations
36
from nonclinical or clinical pharmacology and biomarker studies. In some cases, such
information could be obtained from the literature.
• 	Early clinical studies that show very marked differences in response between the
enrichment and nonenrichment populations.
Sponsors should discuss how much information to collect on the enrichment-factor-negative
population with FDA review staff.
C. 	Labeling
The use of enrichment designs will have implications for labeling, especially in the
INDICATIONS AND USAGE and CLINICAL STUDIES sections. Labeling should accurately
describe the enrichment strategies used, including any limitations or concerns they raise for
clinical use of the drug. Sponsors should discuss the potential effect of enrichment strategies on
labeling with FDA during drug development.

Contains Nonbinding Recommendations
37
APPENDIX:
ADDITIONAL GUIDANCE RELATED TO ENRICHMENT
FDA has issued a number of guidances that provide additional and related information about
clinical trial designs (including enrichment designs) and demonstrating effectiveness. See
especially the following guidances.1
• 	The draft guidance for industry Adaptive Designs for Clinical Trials of Drugs and
Biologics (September 2018)2 considers the case of enrichment approaches introduced
only after randomization and based on interim evaluations. Such a retrospective finding
would have to be carefully implemented and highly compelling to be accepted without
further study.
• 	The guidance for industry Clinical Pharmacogenomics: Premarket Evaluation in Early-
Phase Clinical Studies and Recommendations for Labeling (January 2013) focuses
particularly on use and evaluation of genomic strategies in early drug development and
highlights identification of enrichment options for later trials.
• 	The guidance for industry and FDA staff In Vitro Companion Diagnostic Devices
(August 2014) defines in vitro companion diagnostic devices that are essential for the
safe and effective use of their corresponding therapeutic products. The guidance
describes FDA’s policies for approval and clearance and for labeling companion
diagnostics contemporaneously with approval and labeling of the therapeutic product.
• 	The guidance for industry Providing Clinical Evidence of Effectiveness for Human Drug
and Biological Products (May 1998) describes the amount and type of evidence needed
to demonstrate effectiveness and is applicable to studies using enrichment designs.
1 We update guidances periodically. To make sure you have the most recent version of a guidance, check the FDA
guidance web page at https://www.fda.gov/RegulatoryInformation/Guidances/default.htm.
2 When final, this guidance will represent the FDA’s current thinking on this topic. For the most recent version of a
guidance, check the FDA guidance web page at https://www.fda.gov/RegulatoryInformation/Guidances/default.htm.

Contains Nonbinding Recommendations
38
REFERENCES
Antman E, Muller J, Goldberg S, MacAlpin R, Rubenfire M, Tabatznik B, Liang C, Hempler F,
Achuff S, Reichele N, Geltman E, Kerin N, Nelf R, and Paramawalt E, 1980, Nifedipine Therapy
for Coronary Artery Spasm, N Engl J Med, 302:1269–1273.
Amery W and Dony J, 1975, A Clinical Trial Design Avoiding Undue Placebo Treatment, J Clin
Pharmacol, 15:674–679.
Barker AD, Sigman CC, Kelloff GJ, Hylton NM, Berry DA, and Esserman LJ, 2009, I-SPY 2:
An Adaptive Breast Cancer Trial Design in the Setting of Neoadjuvant Chemotherapy, Clin
Pharmacol Ther, 86(1):97–100, doi: 10.1038/clpt.2009.68.
Cardiac Arrhythmia Pilot Study (CAPS) Investigators, 1988, Effects of Encainide, Flecainide,
Imipramine and Moricizine on Ventricular Arrhythmias During the Year After Acute Myocardial
Infarction. The CAPS, Am J Cardiol, 61:501–509.
Cardiac Arrhythmia Suppression Trial-II Investigators, 1992, Effect of Antiarrhythmic Agent
Moricizine on Survival After Myocardial Infarction, N Engl J Med, 327:227–233.
Chan P, Tomlinson B, Huang TY, Ko JT, Lin TS, and Lee YS, 1997, Double-Blind Comparison
of Losartan, Lisinopril, and Metolazone in Elderly Hypertensive Patients With Previous
Angiotensin-Converting Enzyme Inhibitor-Induced Cough, J Clin Pharmacol, 37:253–256.
Chapman P, Hauschild A, Robert C, et al., 2011, Improved Survival With Vemurafenib in
Melanoma With BRAF V600E Mutation, N Engl J Med, 364:2507–2516.
Cohn JN, Archibald DG, Ziesche S, Franciosa JA, Harston WE, Tristani FE, et al., 1986, Effect
of Vasodilator Therapy on Mortality in Chronic Congestive Heart Failure. Results of a Veterans
Administration Cooperative Study, N Engl J Med, 314:1547–1552.
Cohn JN, Johnson G, Ziesche S, Cobb F, Frances G, Tristani F, et al., 1991, A Comparison of
Enalapril With Hydralazine-Isosorbide Dinitrate in the Treatment of Chronic Congestive Heart
Failure, N Engl J Med, 325:303–310.
CONSENSUS Study Group, 1987, Effects of Enalapril on Mortality in Severe Congestive Heart
Failure. Results of the Cooperative North Scandinavian Enalapril Survival Study
(CONSENSUS), N Engl J Med, 316:1429–1435.
Coronary Drug Project Research Group, 1980, Influence of Adherence to Treatment and
Response of Cholesterol on Mortality in the Coronary Drug Project, N Engl J Med, 303:1038–
1041.
D’Amico AV, Chen MH, Roehl KA, and Catalona WI, 2004, Preoperative PSA Velocity and the
Risk of Death From Prostate Cancer After Radical Prostatectomy, N Engl J Med, 351:125–135.

Contains Nonbinding Recommendations
39
Early Breast Cancer Trialists’ Collaborative Group (EBCTCG), 2005, Effects of Chemotherapy
and Hormonal Therapy for Early Breast Cancer on Recurrence and 15-Year Survival: An
Overview of the Randomised Trials, Lancet, 365:1687–1717.
Echt DS, Liebson PR, Mitchell LB, Pears RW, Obias-Manno D, Barker AH, Arensberg D, Baller
A, Friedman L, Greene HL, Huther ML, Richardson DW, and the CAST Investigators, 1991,
Mortality and Morbidity in Patients Receiving Encainide, Flecainide, or Placebo: The Cardiac
Arrhythmia Suppression Trial, N Engl J Med, 324:781–788.
Fan C, Oh DS, Wessels L, Welgeh B, Nuyten DS, Nobel AB, van’t Veer LJ, and Perou CM,
2006, Concordance Among Gene-Expression-Based Predictors for Breast Cancer, N Engl J Med,
355:560–569.
Fisher B, Constantino JP, Wickerham DL, et al., 1998, Tamoxifen for Prevention of Breast
Cancer: Report of the National Surgical Adjuvant Breast and Bowel Project P-1 Study, J Natl
Cancer Inst, 90:1371–1388.
Freidlin B, Jiang W, and Simon R, 2010, The Cross-Validated Adaptive Signature Design, Clin
Cancer Res, 16:691–698.
Freidlin B, McShane LM, and Korn EL, 2010, Randomized Clinical Trials With Biomarkers:
Design Issues, J Natl Cancer Inst, 102:152–160.
Freidlin B and Simon R, 2005, Adaptive Signature Design: An Adaptive Clinical Trial Design
for Generating and Prospectively Testing a Gene Expression Signature for Sensitive Patients,
Clin Cancer Res, 11:7872–7878.
Ioannidis JP and Tzoulaki I, 2010, What Makes a Good Predictor? The Evidence Applied to
Coronary Artery Calcium Score, JAMA, 303:1646–1647.
Jouven X, Empana JF, Schwartz PJ, Desnos M, Cowben D, and Ducimetiere P, 2005, Heart-Rate
Profile During Exercise as a Predictor of Sudden Death, N Engl J Med, 352:1957–1958.
Lang JM, Buring JE, Rosner B, Cook N, and Hennekens CH, 1991, Estimating the Effect of the
Run-in on the Power of the Physicians’ Health Study, Stat Med, 10:1585–1593.
Mandrekar SJ and Sargent DJ, 2009, Clinical Trial Designs for Predictive Biomarker Validation:
Theoretical Considerations and Practical Challenges, J Clin Oncol, 27:4027–4034.
McMurray JVV, Packer M, Desai AS, Giong J, Lefkowitz MP, Rizkala AR, Rouleau JL, Shi
VC, Solomon SD, Swedberg K, and Zie MR, 2014, Angiotensin-Neprilysin Versus Enalapril in
Heart Failure, N Engl J Med, 371:993–1004.
Mehta C, Gao P, Bhatt DL, Harrington RA, Skerjanec S, and Ware J, 2009, Optimizing Trial
Design: Sequential, Adaptive and Enrichment Strategies, Circulation, 119:597–605.

Contains Nonbinding Recommendations
40
Pablos-Méndez A, Barr RG, and Shea S, 1998, Run-in Periods in Randomized Trials:
Implications for the Application of Results in Clinical Practice, JAMA, 279:222–225.
Pedersen TR, Olsson AG, Faergeman O, Kjekshus J, Wedel H, Berg K, Wilhelmsen L, Haghfelt
T, Thorgeirsson G, Pyörälä K, and Miettinen T, 1994, Randomised Trial of Cholesterol
Lowering in 4444 Patients With Coronary Heart Disease: The Scandinavian Simvastatin
Survival Study (4S), Lancet, 344:1383–1389.
Perez EA, Romond EH, Sumam VJ, et al., 2007, Updated Results of the Combined Analysis of
NCCTG N9831 and NSABP B-31 Adjuvant Chemotherapy With/Without Trastuzumab in
Patients With HER2-Positive Breast Cancer, J Clin Oncol, 25:65 (suppl 185, abst 512).
Polonsky TS, McClelland RL, Jorgenson NW, et al., 2010, Coronary Artery Calcium Score and
Risk Classification for Coronary Heart Disease Prediction, JAMA, 303:1610–1616.
Prior JO, Montemurro M, Orcurto MV, et al., 2009, Early Prediction of Response to Sunitinib
After Imatinib Failure by 18F-Fluorodeoxyglucose Positron Emission Tomography in Patients
With Gastrointestinal Stromal Tumor, J Clin Oncol, 27:439–445.
Ridker PM, Danielson E, Fonseca FA, et al., 2008, Rosuvastatin to Prevent Vascular Events in
Men and Women With Elevated C-Reactive Protein, N Engl J Med, 359:2195–2207.
Ripatti S, Tikkamen E, Orho-Melander M, et al., 2010, A Multi-Locus Genetic Risk Score for
Coronary Heart Disease: Case-Control and Prospective Cohort Analyses, Lancet, 376:1393–
1400.
Roden DM, Reele SB, Higgis SB, Mayol RF, Gammans RE, Oates JA, and Woosley RL, 1980,
Total Suppression of Ventricular Arrhythmias by Encainide, N Engl J Med, 302:877–882.
Shepherd J, Cobbe SM, Ford I, et al. for the West of Scotland Coronary Prevention Study Group,
1995, Prevention of Coronary Heart Disease With Pravastatin in Men With
Hypercholesterolemia, N Engl J Med, 333:1301–1307.
Simon R and Maitournam A, 2004, Evaluating the Efficiency of Targeted Designs for
Randomized Clinical Trials, Clin Cancer Res, 10:6759–6763.
Smith I, Procter M, Gelber RD, et al., 2007, 2-Year Follow-up of Trastuzumab After Adjuvant
Chemotherapy in HER2-Positive Breast Cancer: A Randomized Controlled Trial, Lancet,
369:29–36.
SOLVD Investigators, 1991, Effect of Enalapril on Survival in Patients With Reduced Left
Ventricular Ejection Fractions and Congestive Heart Failure, N Engl J Med, 325:293–302.
SOLVD Investigators, 1992, Effect of Enalapril on Mortality and the Development of Heart
Failure in Asymptomatic Patients With Reduced Left Ventricular Ejection Fractions, N Engl J
Med, 327:685–691.

Contains Nonbinding Recommendations
41
Steering Committee of the Physicians’ Health Study Research Group, 1989, Final Report on the
Aspirin Component of the Ongoing Physicians’ Health Study, N Engl J Med, 321:129–135.
Stroobants S, Goeminne J, Seegers M, et al., 2003, 18FDG-Positron Emission Tomography for
the Early Prediction of Response in Advanced Soft Tissue Sarcoma Treated With Imatinib
Mesylate (Glivec), Euro J Cancer, 39:2012–2020.
Taylor AL, Ziesche S, Yancy C, Carson P, D’Agostino R Jr, Ferdinand K, et al., 2004,
Combination of Isosorbide Dinitrate and Hydralazine in Blacks With Heart Failure, N Engl J
Med, 351:2049–2057.
Temple R, 1994, Special Study Designs: Early Escape, Enrichment Studies in Non-Responders,
Commun Stat Theory Methods, 23:499–531.
Temple, R, 2012, A Regulator’s View of Comparative Effectiveness Research, Clin Trials, 9:
56–65.
Temple R and Stockbridge N, 2007, BiDil for Heart Failure in Black Patients: The US Food and
Drug Administration Perspective. Ann Int Med, 146:57–62.
Veterans Administration Cooperative Study Group on Antihypertensive Agents, 1967, Effects of
Treatment on Morbidity in Hypertension: Results in Patients With Diastolic Blood Pressures
Averaging 115 Through 129 mmHg, JAMA, 202:1028–1034.
Veterans Administration Cooperative Study Group on Antihypertensive Agents, 1970, Effects of
Treatment of Morbidity in Hypertension II: Results in Patients With Diastolic Blood Pressure
Averaging 90 Through 114 mmHg, JAMA, 213:1143–1152.
Wang SJ, Hung HMJ, and O’Neill RT, 2011, Genomic Classifier for Patient Enrichment:
Misclassification and Type I Error Issues in Pharmacogenomics Non-Inferiority Trial, Stat
Biopharm Res, 3:310–319.
Wang SJ, O’Neill RT, and Hung HMJ, 2010, Statistical Considerations in Evaluating
Pharmacogenomics-Based Clinical Effect for Confirmatory Trials, Clin Trials, 7:525–536.