28
Premarketing Applications of
Pharmacoepidemiology
HARRY A. GUESS
Merck Research Laboratories, Uest Point, PA, and University of North Carolina, Chapel Hill, NC, USA
INTRODUCTION
Uhile most of the interest in pharmacoepidemiol-
ogy has centered on its role in the evaluation of
drug safety after marketing,1−3 epidemiologic
methods are also useful in evaluating safety
concerns that arise prior to market approval.
Much of the premarketing patient exposure to
many pharmaceutical products occurs in studies
conducted without a control group. Often these
studies are open extensions of randomized trials
in which all patients are offered the option of
receiving active treatment upon completion of
the blinded phase of the studies. Other sources
of premarketing patient exposure without a
control group are Treatment IND (Investigational
New Drug) Studies4 and Compassionate Use
Pro- grams.5 These programs are intended to
give physicians the opportunity to use
investigational drugs to treat patients with life-
threatening ill- nesses which are not adequately
controlled with approved drugs and for which the
investigational drugs are considered likely to be
beneficial. Typically, such patients are more
seriously ill than
Phavmacoepidemiology, Third Edition. Edited by B. L. Strom.
g 2000 John Uiley & Sons, Ltd.
those in the randomized clinical trials designed
to evaluate efficacy. Use of epidemiologic
methods can sometimes be useful in
determining whether rare adverse events in
noncomparative clinical studies are occurring at
rates above those expected among similar
patients not exposed to the study drug. In
addition, retrospective analyses of non-
comparative clinical trials can help identify risk
factors for specific adverse events and thereby
contribute to safer use of the study drug.
Premarketing applications of pharmacoepide-
miology are similar in many ways to analyses of
postmarketing surveillance studies (phase 4 stu-
dies). Two important ways in which
premarketing and postmarketing applications of
pharmaco- epidemiology differ are that (i) safety
questions arising in clinical trials typically require
answers in a matter of days rather than weeks or
months, and
(ii) the threshold for either the manufacturer or a
regulatory agency to decide to halt human
exposure to a drug is much lower before market
approval than after approval.
In this chapter, we will first review the clinical
and regulatory context for epidemiologic evalua-
45 PHARMACOEPIDEMIOLOGY
0
tions of drug safety during premarketing clinical
studies. Next, we will briefly mention information
technology requirements and statistical
methods. Finally, we will discuss some
interesting case studies in which the analyses
resulted in publica- tions. Because premarketing
applications of phar- macoepidemiology
normally require access to premarketing, as yet
unpublished, clinical data, such studies are of
greatest importance and interest to those in the
pharmaceutical industry or regulatory agencies,
and the perspective taken in this chapter reflects
this. However, the topic should also be of
interest to others in the field of
pharmacoepidemiology, including academic
researchers and clinical investigators.
CLINICAL PROBLEMS TO BE
ADDRESSED BY
PHARMACOEPIDEMIOLOGY
RESEARCH
The need for premarketing epidemiologic assess-
ments of drug safety may arise in several situa-
tions, including evaluation of serious adverse
events in noncomparative clinical trials, preparing
integrated summaries of safety for a New Drug
Application (NDA) or international marketing
application, and answering inquiries from regula-
tory agencies or advisory committees. For clinical
trials conducted under US Food and Drug
Administration (FDA) regulations for Investiga-
tional New Drugs (IND), the Code of Federal
Regulations (21 CFR 312.32) currently (January
1999) mandates reporting to FDA and all partici-
pating investigators within 15 calendar days any
adverse experience which simultaneously meets the
three conditions„ (i) ‘‘serious,’’ (ii) ‘‘unexpected,’’
and (iii) having a reasonable possibility of having
been caused by the study drug.6 If, in addition to
the above three conditions, the event is also fatal
or life threatening, a report to FDA by telephone
or facsimile transmission within seven calendar
days is required (6). The publication of these
regulations in the Federal Register6 noted that
they were issued to implement definitions, report-
ing periods, formats, and standards as recom-
mended by the International Conference on
Harmonisation of Technical Requirements for
Registration of Pharmaceuticals for Human Use
(ICH)7, 8 and by the Uorld Health Organization’s
Council for International Organizations of Med-
ical Sciences (CIOMS).9, 10 This is part of an
ongoing process of international standardization
of definitions and reporting procedures used in
drug safety evaluation.
According to current regulations,6 a ‘‘serious’’
adverse drug experience is any adverse drug
experience occurring at any dose that results in
any of the following outcomes„ death, a life-
threatening adverse drug experience, inpatient
hospitalization or prolongation of existing hospi-
talization, a persistent or significant disability/
incapacity, or a congenital anomaly/birth defect.
Important medical events that may not result in
death, be life threatening, or require hospitaliza-
tion may be considered a serious adverse drug
experience when, based upon appropriate medical
judgment, they may jeopardize the patient or
subject and may require medical or surgical
intervention to prevent one of the outcomes listed
in this definition. Examples of such medical events
include allergic bronchospasm requiring intensive
treatment in an emergency room or at home, blood
dyscrasias or convulsions that do not result in
inpatient hospitalization, or the development of
drug dependency or drug abuse.
An ‘‘unexpected’’ adverse drug experience is
any adverse drug experience, the specificity or
severity of which is not consistent with the
current investigator brochure; or, if an
investigator bro- chure is not required or
available, the specificity or severity of which is
not consistent with the risk information described
in the general investiga- tional plan or elsewhere
in the current application, as amended. In
reporting IND safety reports, the regulations
indicate that the sponsor shall identify all safety
reports previously filed with the IND concerning
a similar adverse experience and shall analyze
the significance of the adverse experience in
light of the previous similar reports.
An epidemiologic approach to the preparation
of an IND safety report can occasionally be
useful when the report concerns a rare adverse
event for which appropriate comparative
incidence data are available. For example
cancer incidence and
mortality data by age, gender, and race are
available online from a number of countries. 11−13
To produce an epidemiologic assessment it is
necessary to perform, review, and report the
analysis in time to meet the 15 calendar day
regulatory reporting deadline. This naturally puts
a constraint on how comprehensive such an
assessment can be.
The FDA has made available a number of
guidelines to assist manufacturers in preparing
NDAs and FDA medical officers in reviewing
them. A recent draft guideline for medical
reviewers outlines the goals of a safety review of
an NDA„14
In general, the goals of a safety review are (1)
to identify important adverse events that are
causally related to the use of the drug, (2) to
estimate incidence for those events, and (3) to
identify factors that predict the occurrence of
those events, e.g., patient factors such as age,
gender, race, comorbid illnesses, and drug
factors such as dose, plasma level, duration of
exposure, concomitant medications. The safety
review is useful not only in making a
risk/benefit decision, but also in drafting
labeling for a drug that is going to be approved.
The guideline also emphasizes the difference
between the formal, prespecified efficacy
evalua- tion and the more exploratory
approaches used in safety evaluation in pre-
approval clinical trials„
Approaches useful for evaluating the safety of a
drug under development generally differ
substan- tially from those useful in evaluating its
effective- ness. In fact, most of the studies in
phases 2 − 3 of a development program focus
on establishing a drug’s effectiveness. In
designing these trials, critical efficacy endpoints
are identified in ad- vance and sample sizes are
estimated to permit an adequate test of the null
hypothesis. For the most part, phase 2 − 3 trials
are not designed to test hypotheses about
safety. In fact, the safety end- points are
generally not known prior to the conduct of
these trials, and for many of the observed safety
outcomes, one can assume that the available
studies are underpowered. The usual approach
is to screen broadly and sensitively for adverse
events, and it is hoped that this approach
should reveal the common adverse event profile
of a new drug and detect some of the less
common and more serious adverse events
associated with
the drug as well. Uhile hypothesis testing for
effectiveness outcomes generally is done within
individual studies, it usually is not appropriate to
proceed with hypothesis testing procedures for
safety outcomes. Rather, the approach to safety
data may be viewed more as exploration and
estimation.
This discussion of the goals and approaches of a
pre-approval safety evaluation makes it clear that
epidemiologic reasoning and methods are clearly
useful. The need to evaluate data from all trials,
with different kinds of patient population, differing
designs, and differing durations makes
epidemiolo- gic approaches especially appropriate.
The approach required is one that recognizes the
limited nature of the data and avoids
overinterpretation in either direction.
METHODOLOGIC PROBLEMS TO BE
ADDRESSED BY
PHARMACOEPIDEMIOLOGY
RESEARCH
EPIDEMIOLOSIC APPROACHES TO
ASSESSIUS CAUSALITY OF ADVERSE
EVEUTS IU PRE-APPROVAL CLIUICAL
TRIALS
Clinical investigators and clinical monitors con-
tinually review reports of adverse events
occurring in pre-approval clinical trials.
Investigators are asked to provide a clinical
assessment as to the causality of each event.
Such assessments are a necessary part of
safety monitoring in clinical research, although
they are known to be subjective and imprecise14
(see also Chapter 32). Criteria to help guide
causality assessments have been pub- lished by
FDA.15 These criteria are most useful when there
are well defined differences in clinical features of
drug related and non-drug-related cases of the
adverse event. However, for serious, uncommon
adverse events where the clinical features of the
drug related cases could be similar to those of
non-drug-related cases, it can some- times be
helpful to supplement clinical causality
assessments of individual cases by
epidemiologic
 assessment criteria based on comparisons
between groups of patients.
The epidemiologic literature provides several
sets of criteria for helping to decide whether an
empirical association is likely to be causal. The
best known criteria are those proposed in 1965
by Bradford-Hill to help evaluate evidence linking
cigarette smoking with lung cancer16 (see also
Chapter 2). The nine Bradford-Hill criteria are
discussed briefly below as they relate to
evaluation of adverse experiences in
premarketing clinical trials.
Strength of Association
This is commonly quantified in terms of a suitably
adjusted hazard function ratio or risk ratio
(relative risk), rather than a p-value. In general,
the farther the ratio is from unity, the less likely it
could be entirely attributable to imbalances in risk
factors between groups. An exception to this
occurs when the ratio is based on very small
numbers or highly influenced by only a few cases.
In addition, it is worth noting that with the large
number of different kinds of adverse events often
seen in large clinical trials, it is quite likely that
some risk ratios far from unity will occur by
chance alone. Multiple comparisons not only
distort p-values, but can also bias risk ratios and
their confidence intervals away from unity. This
happens because screening many different adverse
event terms and selecting those with very low p-
values inherently selects for relative risks that are
biased away from unity.17 Since the confidence
interval is centered about the observed relative
risk, it will also be biased away from unity. This
type of selection bias, a form of regression to the
mean, is common to all programs of screening for
unusually large or small values and is well known
in the statistics literature. One example where this
appears to have occurred is in some
observational studies of vasectomy and
prostate cancer.17, 18 Because pre-approval
reviews of drug safety use the same set of data
for identifying potentially drug related events and
for providing preliminary estimates of the risk, the
relative risk estimates can
be biased away from unity.
It is also important to recognize that the
abzence of any association between a drug and
any given adverse event has to be judged in the
context of the limited amount of patient exposure
in pre-approval clinical trials. This topic is
reviewed in ICH Guide- line E1A, which
addresses the extent of population exposure
needed to assess clinical safety for drugs
intended for long term treatment of non-life-
threatening conditions.7, 8 The guideline notes
that
It is expected that short-term event rates (cumu-
lative 3-month incidence of about 1%) will be
well characterized. Events where the rate of
occurrence changes over a longer period of time
may need to be characterized depending on
their severity and importance to the risk−benefit
assessment of the drug. The safety evaluation
during clinical drug development is not expected
to characterize rare adverse events, for
example, those occurring in less than 1 in 1000
patients.
Consistency
The original wording was, ‘‘Has [the association]
been repeatedly observed by different persons,
in different places, circumstances and times.’’ This
is a useful criterion for assessing adverse
experiences in a program of several clinical trials.
Results that show a consistently elevated
incidence on drug across studies are generally
more convincing than those in which the
elevated risk is largely due to one study. It has
been noted by FDA, however, that an apparent
lack of consistency among trials may simply
reflect differences in trial design, making the
event less likely in some trials than others. For
example, in a combined analysis of several
studies in an NDA, a reassuringly low incidence
of phototoxicity was seen. Examination of
individual studies revealed that in one study the
rate of phototoxicity was substantial— a finding
that had been obscured by combining the data.
This study was the only outpatient study and
thus these were the only patients who had an
opportunity to have the event by virtue of having
sun exposure.14
TemporaFity
In both epidemiologic and clinical
assessments of causality, it is important to
distinguish between
events having onset before a drug was started
and those having onset after drug therapy was
started. Especially in the evaluation of adverse
experiences from studies without a comparison
group, it sometimes occurs that early symptoms of
a disease which is present but not yet
recognized lead a patient to be prescribed a
drug, which then appears to be the cause of the
disease when it is eventually diagnosed. This
has been called ‘‘protopathic bias’’ and is a
special case of the broader concept of
‘‘confounding by indication’’ (see also Chapter
34). A classic example of this form of bias was
seen in uncontrolled postmarketing surveillance
studies of cimetidine, where a higher than
expected incidence of gastric carcinoma was
found among users than among nonusers. It is
likely that many of the cancers were present but
undiagnosed at the time the cimetidine was
started. Subsequent studies showed that
elevations in gastric cancer risk diminished with
duration of followup, returning to baseline with
long term use.19
Another way in which the concept of
temporality plays a role in the evaluation of
adverse experiences is whether the timing of the
reaction in relation to duration of exposure is
consistent with the proposed mechanism. Thus,
an elevated incidence of cancer in patients who
had been taking a drug for several years would
be of more concern than would an elevated
incidence in the first year of therapy. Timing
plays a major role in the evaluation of adverse
events that are thought to be due to immune
mechanisms (e.g., anaphylaxis, angioedema,
hemo-
lytic anemia, serum-sickness), altered
metabolism, or drug interactions.15 A classic
example where timing provided highly convincing
evidence of causality was in the evaluation of
Guillain-Barre´ syndrome in association with the
‘‘swine-flu’’ vaccine20 (see Figure 28.1).
Dose– Response
Adverse effects caused by exaggerated pharmaco-
logical actions of a drug are often dose
dependent. Examples include hypotension
resulting from the use of antihypertensive drugs
and gastrointestinal hemorrhage from
nonsteroidal anti-inflammatory drugs. For such
adverse events it is especially important to
characterize how the incidence of the event
varies with dose in different patient popula-
tions. This is not only important in assessing
causality and quantifying incidence, but also in
understanding the mechanism and providing
guidance to clinicians. The case of renal failure
in patients with congestive heart failure treated
with the angiotensin converting enzyme (ACE)
inhibi- tor enalapril provides an excellent
example of where a life saving drug was initially
thought by some commentators to be inherently
nephrotoxic, when the actualproblem was that
the starting dose in patients with severe
congestive heart failure was too high. Uhile ACE
inhibitors improve survival in patients with
congestive heart failure, too high a dose can
shut off renal function, because angio- tensin-II
is part of the compensatory process for
Figwre 28.1. Onset of SwiFFain-Barre´ syndrome in reFation to day of vaccination.16 Reprodwced with permission.
maintaining adequate renal perfusion pressure
in the face of low cardiac output. This example
illustrates the importance of understanding both
pathophysiology and dose−response relationships
in evaluating drug safety.
ExperimentaF Evidence
Evidence from animal experiments or from pre-
vious human clinical trials can be especially
important in helping to interpret adverse experi-
ences. Uell designed animal studies can be
especially helpful in determining the extent to
which animal results are applicable to humans.
On the other hand, poorly designed animal
studies may produce misleading results, which
can require carefully designed further
experiments to correct.
BioFogicaF PFawsibiFity
In the original statement of biological plausibility
as one of the considerations in judging causality,
Bradford-Hill noted
It will be helpful if the causation we suspect is
biologically plausible. But this is a feature I am
convinced we cannot demand. Uhat is biologi-
cally plausible depends on the biological knowl-
edge of the day.
He went on to note, among other examples,
that the role of rubella in causing congenital
malformations was initially doubted on the basis
of presumed lack of biological plausibility. 16 As
important as we consider biological plausibility to
be, it is equally important to realize that it can
mislead in either direction. It is worth noting,
however, that some epidemiologists have
expressed the view that too little attention is
currently paid to biological plausibility in reviews
of epidemiologic evidence.21
Coherence
The postulated cause-and-effect interpretation
should not seriously conflict with what is known
of the natural history and biology of the event. In
this regard, biological and laboratory evidence
may strengthen a causalinterpretation, but lack
of it cannot be used to nullify one.
AnaFogy
Reasoning by analogy often serves as a basis
for having a lower threshold for judging an
adverse event to be causally related to one drug
in a class when the same effect is considered to
be causally related to another drug in the class.
Such reasoning is even more problematic than
reasoning based on biological plausibility and
this criterion has been criticized as being
potentially misleading. It is worth noting,
however, that certain types of adverse event are
commonly enough associated with drugs to be
worth anticipating as topics which regulatory
agencies are likely to ask to be specifically
addressed. The FDA medical reviewer guidance
document14 states that
There is an expanding checklist of adverse
events that we now routinely think about when
reviewing new drugs, e.g., sedation,
anticholinergic effects, vasodilator effects, QT
prolongation, tachycardia, beta agonist effects,
hepatotoxicity, hematological effects, such as
neutropenia. All of these standard concerns
should be specifically addressed within the
appropriate body systems.
The document suggests that medical reviewers
perform a ‘‘review of body systems’’ to help
judge whether potential safety concerns within
each body system had been adequately
evaluated14
... within each body system, the reviewer should
comment on the adequacy of the development
program in evaluating the new drug with regard
to the body system in question. This gets at the
question of whether or not ‘‘all tests reasonably
applicable’’ were conducted to assess the
safety of the new drug. Uere all the appropriate
animal tests done? Uere all the appropriate
clinical tests done? Uere all potentially important
findings adequately explored, e.g., to what
extent was psychomotor impairment specifically
assessed in a drug that is sedating? If not, this
could be the basis for a non-approval action, or
alternatively, a phase 4 study for additional
testing ....
Thus, it is important to anticipate potential
mechanism based toxicity and to design the
drug
development program to address concerns that
can reasonably be anticipated and tested. This
requires understanding both the pharmacology
of the compound and its class and the natural
history and epidemiology of the disease being
treated.
Specificity
The finding that an adverse event has a very
specific presentation or is associated with a
specific histopathology can be useful evidence
in favor of causality. Absence of specificity in an
epidemiolo- gic study often manifests itself as
elevation in positive associations between an
exposure and a large number of unrelated
outcomes. Such a finding raises the possibility of
uncontrolled confounding or selection bias.
In summary, while the Bradford-Hill criteria
were originally proposed for use in interpreting
evidence from observational studies, they also
provide a good framework for evaluating
evidence from unplanned comparisons of
adverse experi- ences in pre-approval clinical
trials.
Another approach to assessing causality is
through Bayesian causality assessment, 22, 23 which
provides a framework for using clinical and
epidemiologic information to compute a prob-
ability that a specific drug caused a specific set
of events in a specific patient. In Bayesian
causality assessments the statement that a
given drug ‘‘D’’ cauzed an event ‘‘E’’ is defined to
mean that the event would not have happened
as and when it did, if drug ‘‘D’’ had not been
administered. Using this definition, a Bayesian
causality assess- ment begins by decomposing
the calculation of a probability of causation into a
number of subcalculations, some of which make
use of epidemiologic information and others of
which make use of pharmacologic and medical
informa- tion specific to the case. Thus, rather
than being alternatives to epidemiologic
assessments, Baye- sian causality assessments
vequive epidemiologic assessments as a
(somewhat hidden) part of a process which
yields a numerical probability of causation.
Bayesian analysis is discussed in more detail in
Chapter 32.
StatisticaF AnaFysis Methods
Uhile both cohort and case−control study
designs are commonly used in postmarketing
pharmacoe- pidemiology, premarketing studies
typically in- volve only cohort designs. One of the
most common problems in premarketing
pharmacoepi- demiology is to compare the
incidence of a given adverse event in the cohort
of patients exposed to the study drug to the
incidence in a suitably chosen cohort of
historicalcontrols. Here incidence is used in the
epidemiologic sense, where the numerator
refers to the number of events (counting only the
initial event in each patient) and the denominator
often refers to the total pevzon time at vizk during
exposure to the study drug, sometimes referred
to as incidence density. Person time at risk is
typically measured in pevzon dayz, where each
patient contributes one person day for each day
he or she is on the study drug and is at risk for
the given adverse event. The person time of
exposure for each patient is thus most
commonly measured from the day of first
exposure to the study drug, up to the earlier of (i)
the date of last followup for the patient or (ii) the
day that the patient first experienced the adverse
event. Uhen analyzing incidence it is important
to review the data for any evidence of temporal
trends in relation to start of therapy. Uhen such
trends are found, it may be necessary to
produce separate risk computations for different
windows of time on therapy. This is part of the
general topic of effect modification by time on
therapy, as discussed elsewhere24 (see also
Chapter 43).
For example, Suissa and colleagues recently
showed that when risk in relation to duration of
therapy was taken into account in comparing the
risk of venous thromboembolism among new
users of second and third generation
oralcontraceptives, the incidences in the two
groups were much closer than had been found
in previous analyses where this approach had
not been used.25 This phenom- enon has been
termed ‘‘effect modification by duration of
therapy’’24 or ‘‘depletion of suscepti- bles’’.26
Analyses of events in relation to person time
can be analyzed by Cox regression, Poisson
regression,27 or by stratified incidence density
computations28 using standard software packages, breast cancer in the exposed community that is
such as Stata or SAS . Uhen the sample sizes
not attributable to the known risk factors. The
are very small confidence intervals on the
corresponding quantity, IC(1—AR)C, is the in-
stratified relative risk estimates may be
cidence of breast cancer in the comparison
computed by exact methods.29 Analytic methods
communities that is not attributable to the known
and their limitations are further discussed in
risk factors. Hence, the ratio of these two
standard references.28, 30 Issues that often arise
quantities is the incidence ratio for breast cancer
include calculation of point and interval
not attributable to the known risk factors. This is
estimates of standardized morbidity or mortality
the risk ratio of interest for comparing risk in
ratios (SMRs) (30, p 65), comparing adjusted
the two communities, while adjusting for the
risk ratios for different groups of patients (30, pp
difference in prevalence of risk factors.
106 − 109), and testing for heterogeneity
The novelty of the approach lies in using the
(representing potential effect modification) and
prevalence of risk factors in only the cases to
dose−response trend (30, p 110). As with all
estimate the attributable risk fraction, AR. The
epidemiologic studies, it is essential to approach
method for doing this was outlined by Bruzzi and
these analyses with recognition of the biological
colleagues (33, equation (6)) and involves using
and clinical aspects of the problem, as well as
published information on relative risks
with an understanding of the meaning and
associated with known risk factors to compute
limitations of the formal computational methods.
the values of AR for the exposed community
Examples of applications of these methods are
and the control community. Here information
given in a later section.
from published sources is used to compute a
Since pre-marketing epidemiologic evaluations
relative risk for each case based on all known
typically have to be performed under tight time
risk factors for that case, relative to some
constraints, it is sometimes necessary to use pub-
reference group. Of course, the same reference
lished data to help estimate risks. One method
group and the relative risk informa- tion must be
which has been used in environmental
used for the cases in the exposed community
epidemiology, but does not appear to have been
and the cases in the control commu- nities. One
used yet in pharma- coepidemiology, involves
major drawback of this approach is that no
using risk factor preva- lence in cases together
methods for expressing the statisticaluncertainty of
with relative risks from published studies to
the resulting incidence ratio have been published.
produce adjusted incidence ratios.31 Using this
This approach would appear to be mainly useful
method, one can compute risk comparisons with
in noncomparative drug studies to provide a
an externalcontrolgroup in which the prevalence
rough idea whether what appeared to be an
of risk factors differs from those in the cases that
elevated incidence of an adverse event could be
have occurred in the clinical study.
attributed to a greater prevalence of risk factors
Dean and colleagues31 presented this method
in the exposed group, relative to the prevalence
and used it to determine whether the observed
in some historical control group. For example,
elevation in breast cancer incidence in one com-
the method would appear to be of some value
munity with a contaminated water supply was
in evaluating adverse events among a group of
higher than in surrounding communities, after
very ill patients in a compassionate use
adjusting for differences in prevalence of known
program. The advantage of only requiring risk
risk factors for breast cancer. The method of
factor data from the cases is that sometimes
adjustment makes use of the population attribu-
these are the only patients for whom
table risk fraction (AR) for each population, i.e.,
information on important risk factors for
the fraction of risk in each population attributable
the adverse events is available.
to the known risk factors (32, p 163). The
complementary fraction, 1— AR, is the fraction
of risk that is not attributable to the known risk Information TechnoFogy Reqwirements
factors. If IE is the total incidence in the exposed
Effective premarketing pharmacoepidemiology re-
community, then IE(1 — AR)E is the incidence of
quires considerable advanced planning to make
sure that information can be assembled, types of laboratory data on all patients exposed to
analyzed, summarized, reviewed, and reported a
in hours or days. For historical data to be useful,
it must be either available in sufficient detail from
published sources or from existing datasets that
can be accessed, analyzed, and checked for
errors within a day or two. This can only be done
if one can anticipate at least some types of
problems likely to arise and have appropriate
sources of epidemiolo- gic information readily
accessible.
In addition, the clinical trials data management
organization has to be able to produce accurate,
suitably detailed patient exposure information
from many different trials in a number of different
countries. This is more difficult than it might
initially appear to be. Because of regulatory
report- ing requirements, adverse experience
information (the numerator) will typically be
current to within a few days, while the patient
exposure information (the denominator) may be
weeks behind and may not yet have essential
covariate information and demographics in
accurately retrievable form.
Getting new drugs approved in a timely
manner can depend on the ability to answer
safety questions rapidly and accurately, and that
can depend on the way in which data
management is staffed, organized, and
equipped. One of the rate- limiting factors in
securing timely approval is the manpower
needed to process worldwide clinical trial data
on thousands of patients from many different
countries accurately. This entails produ- cing
both standard efficacy and safety analyses and
special ad hoc safety analyses required for
responding to questions raised by reviewers at
drug regulatory agencies. The data
management problems posed by the need to
answer ad hoc safety questions rapidly are
different from those posed by the need for the
formal statistical analyses of efficacy required for
drug approval. To be successful, the data
management organiza- tion and systems must
meet both challenges.
As potential capabilities of data management
systems increase, the potential benefit from
using them effectively increases dramatically, as
does the potential gap in productivity from using
them ineffectively.34 For example, a request by a
regulatory agency for an analysis of certain
study drug for more than two months might important to read the actual written question and
take only a day or two for a response if the the surrounding
data from all worldwide studies were stored
so that a query could readily be run against
the entire database. However, the query
could take weeks and consume much more
scarce manpower critically needed for other
tasks if„ the specific data requested were
stored differently by different countries; the
way the data were stored made custom
programming necessary; a decision had
been made not to have centralized computer
storage of the specific data needed; the
laboratory methods used in different
countries were different and cross-translation
files had not been centrally stored; there
were problems with getting data entered,
checked, and available for analysis in a
timely manner; or there were other
organizationalimpediments to producing a
unified worldwide tabulation and analysis.
During the course of regulatory review of a
new drug many standard and ad hoc analyses
must be performed for different regulatory
agencies in different countries. The
difference between effective and somewhat
less effective clinical data management and
analysis can easily translate into differences
of several weeks in drug approval time. Such
differences can translate into substantial lost
revenue for each new drug.

CURRENTLY AVAILABLE SOLUTIONS

Epidemiologic investigation of pre-marketing

drug safety problems is a collaborative
undertaking that requires the ability not only
to respond to inquires promptly and
accurately, but also to recognize and answer
questions that should have been asked but
were not. Sometimes the initial request from
the clinical or regulatory group in a company
to the epidemiology group asks for a specific
tabulation or calculation, which may not in
fact represent the most appropriate way to
approach the problem. Requests transmitted
through several intermedi- aries often change
in meaning, so that what finally reaches the
person responsible for the analysis may be
quite different from what was originally
asked. Uhen the originalrequest is a written
inquiry (e.g., from a regulatory agency), it is
context before attempting to answer a several
paraphrased version of it.
As with all epidemiologic and statistical con-
sulting, the best way to approach a request for
consultation on a potential safety problem with a
premarketing investigationaldrug is to first
under- stand the broader context surrounding
the im- mediate inquiry. This includes gaining at
least a basic understanding of the
pharmacology, me- chanism of action, preclinical
toxicity profile, and clinical safety profile of the
compound. It also entails understanding the
characteristics and co- morbidities of the patient
population in which the drug has been studied.
Finally, it requires knowing how and in what
context the question arose. A review of several
case studies may be helpful in conveying these
concepts.
One example, which illustrates the interplay
between clinical pharmacology and
epidemiology, involved seizures in seriously ill
hospitalized patients with systemic gram-
negative infections being treated with the Ø-
lactam antibiotic, imipe- nem/cilastatin. During
initial randomized clinical trials few seizures
were reported, either with imipenem/cilastatin or
with control antibiotics. In subsequent
noncomparative studies an increas- ing number
of seizures were noted, often in patients with
predisposing factors, such as com- promised
renal function that could alter drug metabolism
and affect serum levels of the anti- biotic. An
association of seizure risk with anti- biotic level
was biologically plausible, in light of known
epileptogenic properties of Ø-lactam anti- biotics.
At the time these studies were conducted,
laboratory techniques to measure serum levels
had not yet become widely available and so
serum levels on these patients were rarely
known.
To study seizure risk in relation to serum level,
clinical pharmacology studies were reviewed to
determine how serum levels varied with dose,
body weight, gender, age, and renal function. An
equation was developed to predict serum level
as a function of these parameters. The predicted
serum levels were then used as one variable in
an analysis of seizure risk in the noncomparative
clinical trial patients. It was found that risk of
seizure was strongly and independently related
to predicted serum level, after controlling for
non-drug-related seizure risk factors. 35 At the decision was made to
same time, however, the other factors found
to be related to seizure risk in patients
receiving imipenem/cilastatin were also found
to be risk factors for seizures in patients who
had not received imipenem/cilastatin. These
factors in- cluded a history of seizures, central
nervous system insults, and renal
impairment.36 Age was not found predictive of
seizures when adjustment for the above
factors was made.
This study illustrates the concept of
phavmaco- epidemiology in a study where
methods of clinical pharmacology and
epidemiology were both brought into play to
investigate and solve a problem that arose
during the course of premar- keting clinical
trials. It also illustrates the impor- tant point
that merely quantifying risk and identifying
patients at increased risk would not have been
sufficient. Uhat was needed was to identify
measures to help reduce the risk and to help
educate physicians on the need for dosage
adjustments. The investigation into seizure
risk led to improved prescribing information
with better recommendations for dosage
adjustments in the presence of impaired renal
function.
An illustration of statistical methods for
phar- macoepidemiology cohort studies is
provided by a study comparing the observed
incidence of Guillain-Barre´ syndrome among
recipients of a plasma-derived hepatitis-B
vaccine with that which would be expected in
the general population, controlling for age in a
stratified analysis.37 Although this study
involved postmarketing data, it provides a
good illustration of the type of cohort study
methods typically used in pre-marketing
analyses. During the first three years of
marketed use of the vaccine, nine cases of
Guillain-Barre´ syndrome were reported among
vaccinees. Six of the nine cases occurred
within six weeks after the first dose and the
remaining three occurred within eight weeks
after either the second or third dose. (The
vaccine was administered in three injections,
with the second occurring one month after the
first and the third occurring six months after
the first.) Based on knowledge of the temporal
pattern of risk of Guillain-Barre´ syndrome with
other vac- cines and on the serological
response of the hepatitis-B vaccine, a
consider four analyses based on two time
windows (six weeks or eight weeks) applied to
either the first dose only or to all three doses.
The total number of vaccinees was estimated
from sales data and the age distribution was
estimated from surveys. Comparison data were
obtained on the age- specific incidence of
Guillain-Barre´ syndrome in severalstudies,
including a community based study of Guillain-
Barre´ syndrome among residents of Olmsted
County, MN. Because the number of cases in
vaccinees was so small, the confidence intervals
about the relative risks were calculated by an
exact binomial method.29 The age adjusted
relative risks in the four analyses ranged from
the lowest value of 1.7 (95% confidence interval„
0.8 − 3.6), based on nine cases occurring in an
at-risk intervalof 8 weeks after each of the three
doses, to the highest value of 3.6 (95%
confidence interval„
1.3 − 8.3), based on six cases occurring in an
at-risk interval of six weeks after the first dose
only (37, Table 6). Although the numerator
(number of cases) in the latter analysis was less
than in the former, the denominator (person time
at risk) was also less, and the age adjusted
relative risk was higher.
Another example involved the question of
whether herpes zoster might be expected to
occur more frequently following vaccination with
a live- virus varicella (chickenpox) vaccine than
following natural varicella. It was found that the
incidence and clinical characteristics of herpes
zoster in childhood following natural varicella
were not well enough known to permit any
meaningful compar- ison with cases of childhood
herpes zoster follow- ing vaccination with the
varicella vaccine. This led to an epidemiologic
study to document the incidence and clinical
features of herpes zoster following natural
varicella.38 It was found that herpes zoster in
children was both much more common than had
been previously thought and also was much
milder than herpes zoster in adults. Using these
data as a point of comparison, subsequent
studies suggested that the incidence of herpes
zoster in childhood following vaccina- tion did
not appear to be more common with the vaccine
than with natural varicella.39
In this example, what was needed was more
information about the incidence and clinical
spectrum of the adverse event following natural
varicella. This illustrates the point that to quantify
risk associated with a therapeutic or prophylactic
intervention, it is often necessary to develop
additional information about the natural course
of the disease itself.40 Such studies need to be
started very early in drug or vaccine
development in order to be able to have
information available when needed to address
safety concerns that arise in clinical studies.
A third example evaluated the risk of
angioede- ma in relation to angiotensin
converting enzyme inhibitors, based on data
from three large studies involving about 12 000
patients each.41 Previous studies had suggested
that angioedema was more common early in
therapy than later in therapy. This hypothesis was
tested by examining incidence as a function of
time on therapy. It was found that the incidence
in the first week of therapy was about one case
per 3000 patients per week; thereafter the
incidence was about 14-fold lower, with no
evidence of any temporal trend in incidence
beyond the first week of therapy (Figure 28.2).
This study illustrates several important points.
First, it is essential to establish a clear case
definition in any epidemiologic study. Failure to
do this can lead to continued confusion about
what is being counted. The epidemiologic case
definition may be different from whatever case
definition is used for regulatory reporting. (The
latter definition is typically maximally inclusive,
while the former may be more restrictive.) In this
study, the clinical case definition was
established in consultation with clinical experts
in the diagnosis and treatment of angioedema.
Using this case definition, all case reports that
might have represented angioedema were re-
reviewed, classi- fied, abstracted, and used in
computing incidence. Second, the study
illustrates the importance of examining risk in
relation to time on therapy.24−26 Any temporal
patterns detected will be of clinical relevance.
Further, failure to account for temporal
differences can invalidate calculation of relative
risks. Finally, this study shows how to combine
epidemiologic analyses with a clinical review of
case reports so as not only to quantify the
incidence and temporal pattern of occurrence of
Figwre 28.2. Angioedema associated with angiotensin converting enzyme inhibition. Incidence in reFation to
time on drwg. The Fines show the 95% confidence intervaFs on the observed incidence rates. The centraF bar
on each Fine represents the observed incidence rate. The abbreviations for the three stwdies on which these
rates were based are PMS = postmarketing swrveiFFance stwdy, CDSP = cFinicaF deveFopment stwdy program,
and PEM = prescription event monitoring stwdy. These are expFained in the pwbFication from which these data
were taken, with permission.37

an adverse event, but also to describe its clinical
features and spectrum of severity.
THE FUTURE
Premarketing and postmarketing applications of
pharmacoepidemiology differ in the speed of
response required and in the fact that the
threshold for halting human exposure to a drug
is lower before market approval than after
market ap- proval. In addition, premarketing
pharmacoepide- miology uses cohort study
designs almost exclusively, while postmarketing
pharmacoepide- miology often uses
case−control study designs as well. Safety
monitoring in clinical trials can be improved by
anticipating potential questions and organizing
data resources in advance to answer them as
rapidly as possible. Having a responsive data
management system is as essential to pre-
marketing applications of
pharmacoepidemiology as it is to the entire
process of drug development. Epidemiologic
planning should be regarded as one of the first
steps in drug development, rather than as
something to be considered at or after market
approval. This requires effort to build and main-
tain a repository of disease incidence and
natural history data from governmental surveys,
public- use data tapes of governmental studies,
publica- tions, and other sources.42 It also
requires under- taking epidemiologic studies far
enough in advance of clinical trials to be able to
contribute to clinical trial planning and analysis.
Finally, it requires that the epidemiologic group
maintain an awareness of pertinent research
findings in areas relevant to the diseases under
study and to the types of measurement and
analytic technique likely to be needed.
REFERENCES
1. Lawson DH. Pharmacoepidemiology„ a new disci-
pline. Bv Wed J 1984; 289„ 940 − 1.
2. Tilson HH. Pharmacoepidemiology„ the future.
Dvug Intell Clin Phavm 1988; 22„ 416 − 21.
3. Guess HA. Adverse and beneficial effects in
pharmacoepidemiology. In„ Armitage P, Colton T,
eds, Encyclopedia of Bioztatizticz. New York„ Uiley,
1998; 3332 − 8.
 4. Mattison N. The FDA’s treatment IND„ current
controversies. Phavmaceut Wed 1988; 3„ 159 − 71.
5. Calandra GB, Garelik JP, Hohler PT, Brown HB.
Problems and benefits of an antibiotic compassio-
nate therapy program. Rer Infect Diz 1987; 9„ 1095
− 101.
6. Yedeval Regiztev. Expedited Safety Reporting
Requirements for New Drugs and Biologics. 7
October, 1997, 62 (194)„ 52 237 − 53. Available
at
http„]]www.access.gpo.gov]sudocs]aces]aces140.html.
7. International Conference on Harmonisation of
Technical Requirements for Registration of Phar-
maceuticals for Human Use (ICH). Available at„
http„]]www.ifpma.org/ich1.html. Accessed 30 Janu-
ary, 1999.
8. ICH Topics and Guidelines. Available at„ http„]]
www.ifpma.org]ich5.html. Accessed 30 January,
1999.
9. Council for International Organizations of Medical
Sciences (CIOMS). Available at
http„]]www.who.int] dsa]cat98]zcioms8.htm.
Accessed 30 January, 1999.
10. CIOMS Uorking Groups I−IV. Available at„ http„]]
www.who.int]dsa]cat98]zcioms8.htm. Accessed 30
January, 1999.
11. National Cancer Institute. Suvreillance, Epidemiol-
ogy and End Rezultz. Available at„ http„]]www-
seer.ims.nci.nih.gov]. Accessed 25 November, 1998.
12. Rotterdam Cancer Registry. Cancev Regiztviez on the
Eeb. Available at„ http„]]www.ikr.nl]can- regs.htm.
Accessed 25 November, 1998.
13. International Agency for Research on Cancer.
Cancev in the Euvopean Jnion. Available at„ http„]]
www-dep.iarc.fr]dataava]eucandat.htm. Accessed
25 November, 1998.
14. US Food and Drug Administration, Center for Drug
Evaluation and Research. Conducting a Safety Reriew
of a New Pvoduct Application and Pvepaving a Repovt on the
Reriew. Available at„ http„]]www.fda.gov]cder]
guidance]index.htm. Accessed 22 November,
1998.
15. Food and Drug Administration Staff College.
Clinical Zhevapeuticz and the Recognition of Dvug- Induced
Dizeaze. MedUatch Continuing Education Article,
June 1995. Available at„ http„]]www.fda.-
gov]medwatch]articles]dig]rcontent.htm. Accessed
24 November, 1998.
16. Hill AB. The environment and disease„
association or causation. Pvoc R Soc Wed 1965;
58„ 295 − 300.
17. Guess HA. Invited commentary„ vasectomy and
prostate cancer. Am J Epidemiol 1990; 132„ 1062 − 5.
18. Rosenberg L, Palmer JR, Zauber AG, Uarshauer
ME, Strom BL, Harlap S, Shapiro S. The relation
of vasectomy to the risk of cancer. Am J Epidemiol
1994; 14Є 431 − 8.
19. Ualker AM. Confounding by indication [editorial].
Epidemiology 1996; 7„ 335 − 6.
20. Schonberger LB, Bregman DJ, Sullivan-Bolyai JZ,
Heenlyside RA, Ziegler DU, Retailliau HF, Eddins
DL, Bryan JA„ Guillain-Barre´ syndrome following
vaccination in the NationalInfluenza Immunization
Program, United States, 1976 − 1977. Am J
Epide- miol 1979; 11Є 105 − 23.
21. Ueed DL, Hursting SD„ Biologic plausibility in
causal inference„ current method and practice.
Am J Epidemiol 1998; 147„ 415 − 25.
22. Lane DA. The Bayesian approach to causality
assessment„ an introduction. Dvug Inf J 1986; 2Є
455 − 61.
23. Naranjo CA, Lanctot HL, Lane DA. The Bayesian
differentialdiagnosis of neutropenia associated
with antiarrhythmic agents. J Clin Phavmacol 1990;
3Є 1120 − 7.
24. Guess HA. Behavior of the exposure odds ratio in
a case control study when the hazard function is
not constant over time. J Clin Epidemiol 1989; 42„
1179 − 84.
25. Suissa S, Blais L, Spitzer UO, Cusson J, Lewis M,
Heinemann L. First-time use of newer oral contra-
ceptives and the risk of thromboembolism. Contva-
ception 1997; 56„ 141 − 6.
26. Yola M, Lucien A„ Evidence of the depletion of
susceptibles effect in non-experimental pharmacoe-
pidemiologic research. J Clin Epidemiol 1994; 47„
731 − 7.
27. Frome EL, Checkoway H. Epidemiologic
programs for computers and calculators. Use of
Poisson regression models in estimating
incidence rates and ratios. Am J Epidemiol 1985;
121„ 309 − 23.
28. Greenland S, Rothman HJ. Introduction to
stratified analysis. In„ Rothman HJ, Greenland S,
eds, Wodevn Epidemiology, 2nd edn. Philadelphia,
PA„ Lippincott−Raven, 1998; Ch. 15, 253 − 79.
29. Guess HA, Thomas JE. A rapidly converging
algorithm for exact binomial confidence intervals
about the relative risk in follow-up studies with
stratified incidence density data. Epidemiology 1990;
1„ 75 − 7.
30. Breslow NE, Day NE. Statiztical Wethodz in Cancev
Rezeavch, Volume II„ the Dezign and Analyziz of Cohovt
Studiez, International Agency for Research on
Cancer Scientific Publication No. 82. Lyon„ Uorld
Health Organization, 1987.
31. Dean AG, Imrey HH, Dusich H, Hall UN.
Adjusting morbidity ratios in two communities
using risk factor prevalence in cases. Am J Epidemiol
1988; 127„ 654 − 62.
32. Hleinbaum DG, Hupper LL, Morgenstern H. Epide-
miologic Rezeavch„ Pvinciplez and Quantitatire Weth- odz.
Belmont, CA„ Lifetime Learning, 1982; 163.
33. Bruzzi P, Green SB, Byar DP, Brinton LA, Schairer
C. Estimating the population attributable risk for
multiple risk factors using case−control data. Am
J Epidemiol 1985; 122„ 904 − 14.
34. Guess HA. Uorldwide drug safety information
processing„ what’s ahead. Dvug Inf J 1989; 23„
169−70.
35. Calandra G, Lydick E, Carrigan J, Ueiss L, Guess
H. Factors predisposing to seizures in seriously ill
infected patients receiving antibiotics„ experience
with Imipenem/Cilastatin. Am J Wed 1988;
84„ 911 − 8.
36. Guess HA, Resseguie LJ, Melton LJ III,
Hurland LT, Lydick EG, Uilson UR. Factors
predictive of seizures among intensive care
unit patients with gram-negative infections.
Epilepzia 1990; 31„ 567 − 73.
37. Shaw FE, Graham DJ, Guess HA, Milstien JB,
Johnson JM, Schatz GC, Hadler SC, Huritsky JN,
Hiner EE, Bregman DJ, Maynard JE. Postmarket-
ing surveillance for neurological adverse events
reported after hepatitis B vaccination„ experience
of the first three years. Am J Epidemiol 1988; 127„
337 − 52.
38. Guess HA, Broughton DD, Melton LJ III,
Hurland LT. Epidemiology of herpes zoster in
children and
adolescents„ a population-based study. Pediatvicz
1985; 76„ 512 − 7.
39. Plotkin SA. Hell’s fire and varicella−vaccine safety
[editorial]. New Engl J Wed 1988; 318„ 573 − 5.
40. Guess HA, Stephenson UP, Sacks ST, Gardner
JS. Beyond pharmacoepidemiology„ the larger
role of epidemiology in drug development. J Clin
Epidemiol 1988; 41„ 995 − 6.
41. Slater EE, Merrill DD, Guess HA, Roylance PJ,
Cooper UD, Inman UHU, Ewan PU. Clinical
profile of angioedema associated with angiotensin
converting-enzyme inhibition. J Am Wed Azzoc
1988; 26Є 967 − 70.
42. Guess HA. The role of epidemiology in pharmaceu-
tical research, development, and marketing.
In„ Hartzema AG, Porta MS, Tilson HH, eds,
Phavma- coepidemiology„ an Intvoduction, 2nd edn.
Cincinnati, OH„ Uhitney, 1991; 321− 9.