3 Statistical Approaches To Analysis of Small Clinical Trials - Small Clinical Trials - Issues and Challenges - The National Academies Press

26/03/2019 3 Statistical Approaches to Analysis of Small Clinical Trials | Small Clinical Trials: Issues and Challenges | The National
| The National Academies Press
Small Clinical Trials: Issues and Challenges
Small Clinical Trials: Issues and Challenges (2001)

×
Chapter: 3 Statistical Approaches to Analysis of Small
Clinical Clinical
trials are Trials
used to elucidate the most
 Buy Paperback | $50.00

appropriate preventive, diagnostic, or treatment
options for individuals with a given medical condition.
Visit NAP.edu/10766 to get more information about this book,Perhaps
to buy it the mostor
in print, essential feature
to download ofaafree
it as clinical
PDF.trial is
that it aims to use results based on a limited sample of

Buy Ebook | $39.99 research participants to see if the intervention is safe
and e ective or if it is comparable to a comparison
treatment. Sample size is a crucial component of any
Page 60 MyNAP members save
clinical trial. A trial with a small number of research
participants is more prone to variability and carries a
10% online. considerable risk of failing to demonstrate the
3 Login or Register to e ectiveness of a given intervention when one really is
save! present. This may occur in phase I (safety and
Statistical Approaches to Analysis of Small pharmacologic pro les), II (pilot e cacy evaluation),
and III (extensive assessment of safety and e cacy)
ClinicalTrials Download Free PDF trials. Although phase I and II studies may have smaller
sample sizes, they usually have adequate statistical
power, which is the committee's de nition of a "large"
A necessary companion to well-designed trial.clinical
Sometimes trial
a trialiswith
itseight
appropriate sta-
participants may
have adequate statistical power, statistical power being
tistical analysis. Assuming that a clinical trial will produce data that could
the probability of rejecting the null hypothesis when
reveal differences in effects between two or moreisinterventions,
the hypothesis false. statistical
analyses are used to determine whether such differences are real or are
Small Clinical Trials assesses the current
due to chance. Data analysis for small clinical trialsand
methodologies intheparticular
appropriate must befor the
situations
focused. In the context of a small clinical conduct
trial,of it
clinical trials with small
is especially sample sizes.
important This
for
report assesses the published literature on various
researchers to make a clear distinctionstrategies
between such preliminary evidence
as (1) meta-analysis to combine and
con rmatory data analysis. When the sample disparate population
information from isseveral
small,studies
it is including
im-
Bayesian techniques as in the con dence pro le
portant to gather considerable preliminary method evidence
and (2) other onalternatives
relatedsuch subjects
as assessing
before the trial is conducted to de ne the size needed to determine a criti-
therapeutic results in a single treated population (e.g.,
astronauts) by sequentially measuring whether the
cal effect. It may be that statistical hypothesis testing is premature. Thus,
intervention is falling above or below a preestablished
testing of a null hypothesis might be particularly challenging
probability outcome range and in the predesigned
meeting context
speci cations as opposed to incremental
of a small clinical trial. Thus, in some cases it might be important to focus
improvement.
on evidence rather than to test a hypothesis (Royall, 1997). This is because a
small clinical trial is less likely to be self-contained, providing all of the
necessary evidence to effectively test a particular hypothesis. Instead, it
might be necessary to summarize all of the evidence from the trial and
combine it with other evidence available from other trials or laboratory
studies. A single large clinical trial is often insuf cient to answer a bio-
medical research question, and it is even more unlikely that a single small
https://www.nap.edu/read/10078/chapter/5#89 1/47
26/03/2019 3 Statistical Approaches to Analysis of Small Clinical Trials | Small Clinical Trials: Issues and Challenges | The National Academies Press
×
clinical trial can do so. Thus, analyses of data must consider the limitations
Small
ofClinical
the Trials: Issues and Challenges
Clinical trials are used to elucidate the most
Page 61  Buy Paperback | $50.00

Perhaps the most essential feature of a clinical trial is
3-1
BOX Buy Ebook | $39.99research participants to see if the intervention is safe
Some Statistical Approaches to Analysis of
MyNAP members save
Small Clinical Trials
10% online.
considerable risk of failing to demonstrate the
Login or Register to e ectiveness of a given intervention when one really is
Sequential analysis
pharmacologic pro les), II (pilot e cacy evaluation),
Hierarchical models and III (extensive assessment of safety and e cacy)
Download Free PDF trials. Although phase I and II studies may have smaller
Bayesian analysis
trial. Sometimes a trial with eight participants may
Decision analysis have adequate statistical power, statistical power being
Statistical prediction the hypothesis is false.

Meta-analysis and other alternatives
methodologies and the appropriate situations for the
conduct of clinical trials with small sample sizes. This
Risk-based allocation report assesses the published literature on various
strategies such as (1) meta-analysis to combine
disparate information from several studies including
method and (2) other alternatives such as assessing
data at hand and their context in comparison with
astronauts) those ofmeasuring
by sequentially other similar or
whether the
related studies. intervention is falling above or below a preestablished
probability outcome range and meeting predesigned
Since data analysis for small clinical trials inevitably involves a number of
improvement.
assumptions, it is logical that several different statistical analysis be con-
ducted. If these analysis give consistent results under different assump-
tions, one can be more con dent that the results are not due to unwar-
ranted assumptions. In general, certain types of analysis (see Box 3-1) are
more amenable to small studies. Each is brie y described in the sections
that follow.
×
SEQUENTIAL ANALYSIS
Sequential analysis refers to an analysis of the data as they accumulate,
with a view toward stopping the study as soon as the results become sta-
tistically compelling. This is in contrastClinical
to a sequential design
trials are used to (see
elucidate Chapter
the most

Buy Paperback
2), in which the probability | $50.00
appropriate
that a participant preventive, diagnostic,
is assigned or treatment
to a particular in-
tervention is changed depending on the accumulating
Perhaps results.
the most essential Inofsequential
feature a clinical trial is
analysis the probabilty of assignment tothat anit intervention
aims to use resultsisbased on a limited
constant sample of
across
the study. 
Sequential analysis
MyNAP methods
members were
save rst used in the context of industrial
quality control in online.
10% the late 1920s (Dodgeconsiderable
and Romig, 1929). The use of se-
risk of failing to demonstrate the
Logininorclinical
quential analysis Registertrials
to has been extensively
e ectiveness of a given described
intervention by
when Ar-
one really is
mitage (1975),save!
Heitjan (1997), and Whitehead present. This may occur in phase I (safety and
(1999). Brie y, the data are an-
alyzed as the results for each participant andare obtained.
III (extensive Afterofeach
assessment safetyobserva-
and e cacy)
tion, the decision isDownload
made to Free PDF trials.
(1) continue the Although
studyphase I and II studies
by enrolling may have smaller
additional
participants, (2) stop the study with thepower,conclusionwhich is thethat there de
committee's is anition
statisti-
of a "large"
cally signi cant difference between the treatments, or (3) stop the studymay
trial. Sometimes a trial with eight participants
and conclude that there is not a statistically signi ofcant
the probability rejectingdifference between
the null hypothesis when
the the hypothesis is false.

Page 62 strategies such as (1) meta-analysis to combine
interventions. The boundaries for the decision-making processsuch
method and (2) other alternatives areascon-
assessing
structed by using considerations of power and size needed to determine an
effect size similar to those used to determine sample size (see, for example
Whitehead [1999]). Commercially available software
probability outcome can beand
range used to con-
meeting predesigned
struct the boundaries.
improvement.
In sequential analysis, the nal sample size is not known at the beginning
of the study. On average, sequential analysis will lead to a smaller average
sample size than that in an equivalently powered study with a xed-sam-
ple-size design. This is a major advantage to sequential analysis and is a
reason that it should be given consideration when one is planning and ana-
lyzing a small clinical trial. For example, take the case study of sickle cell
disease introduced in Chapter 1 and consider the analysis of the clinical
×
design problem introduced in Box 1-4 as an example of sequential analysis (
Small
BoxClinical
3-2). Trials: Issues and Challenges
Data from a clinical trial accumulate gradually over a period of time that
can extend to months or even years. Thus, results for patients recruited
early in the study are available for interpretation while patients
diagnostic, are still be-

Buy Paperback | $50.00
ing recruited and allocated to treatment.
appropriate preventive,
options
Thisforfeature
individualsallows
with a given
or treatment
themedical condition.
emerging
evidence to be used to decide when to that stop thetostudy.
it aims In particular,
use results it may
based on a limited sample of
be desirable to stopBuy
theEbook
study if| a$39.99
clearresearch
treatment difference is apparent,
 participants to see if the intervention is safe
thereby avoiding the allocation of further patients to the less successful
therapy. Investigators may also want toclinical
stoptrial.
a study that
A trial with no longer
a small number ofhas
research
MyNAP members save
much chance10% of demonstrating
online. a treatment difference (Whitehead, 1992,
1997). Login or Register to e ectiveness of a given intervention when one really is
For example, consider the analysis of an pharmacologic
intervention pro les), II (pilot e cacy evaluation),
(countermeasure) to
prevent the loss of Download
bone mineral density
Free PDF in Although
trials. sequentially treated
phase I and groups
II studies may have ofsmaller
astronauts resulting from their exposure sample sizes, they usually during
to microgravity have adequate
space statistical
travel ( Figure 3-1). The performance index is the bone
trial. Sometimes a trial mineral density (in
with eight participants may
grams per square centimeter) of the calcaneus. S refers to success, where being
have adequate statistical power, statistical power p
is the probability of success and p* is the cumulative mean. F refers to fail-
the hypothesis is false.
ure, where q is the probability of failure and q* is the cumulative mean. The
con dence intervals for p and q are obtained after each space mission, that
is, for p, (P1, P2), and for q, (q1, q2). The sequential accumulation
conduct of clinical trials with small of datasizes.
sample thenThis
allows one to accept the countermeasure if p is greater than p* and q2 is
strategies1 such as (1) meta-analysis to combine
less than q* or reject the countermeasure if p2information
disparate from p*
is less than q1 is including
or studies
several
greater than q*. Performance indices will be acceptable
method when success
and (2) other alternatives S, a
such as assessing
gain or mild loss, occurs on at least 75 percent
therapeutic(p* = 0.75)
results of the
in a single cases
treated (as- (e.g.,
population
tronaut missions) and when F, severe bone mineral density loss, occurs in
no more than 5 percent (q* = 0.05) of the cases.outcome
probability Unacceptable performance
range and meeting predesigned
indices occur with less than a 75 percent success rate or more than a 5
improvement.
percent failure rate. As the number of performance indices increases, level
1 performance crite-
Page 63
×
BOX 3-2
Clinical Trial for Treatment of Sickle Cell
Disease
Sickle Cell disease is a red blood cell (RBC) disorder
that affects 1 in 200 African Americans. Fifty percent
options for individuals with a givenof
medical condition.
individuals living with sickle cell disease die before ageof a clinical trial is
Perhaps the most essential feature
40. The most common complications includetostroke,
 Buy Ebook | $39.99 research participants
e ective or if it is
see if the intervention is safe
renal failure, and chronic severe pain. Patients who to a comparison
and comparable
have a stroke are predisposedclinical
to having another
trial. A trial one.
with a small number of research
MyNAP members save
10%donor
Mixed online. and host stem cell chimerism
considerable (e.g.,tothe
risk of failing demonstrate the
recipient patient has stem cells of her or his own origin
and also those from the transplant donor)
pharmacologic pro is curative
les), II (pilot e cacy evaluation),
for sickle cell disease. Only 20 percent of donor RBC
production (and 80 percent of recipient
sample sizes, theyRBC produc-
usually have adequate statistical
power, which is the
tion) is required to cure the abnormality. committee's de nition of a "large"
Conditioning
of the recipient is required forhavethe transplanted
adequate bonestatistical power being
statistical power,
marrow stem cells to becometheestablished.
probability of rejecting
The degreethe null hypothesis when
of HLA (human leukocyte antigen) mismatch as well as
Small Clinical
the sensitization state (i.e., chronic Trials assesses
transfusion the current
immu-
nizes the recipient) in uences howofmuch
conduct clinical conditioning
trials with small sample sizes. This
is required to establish 20 percent donorthe
report assesses chimerism.
published literature on various
In patients who have an HLA-identical donor and who
have not been heavily transfused,
method200 centigrays
and (2) (cGy)
other alternatives such as assessing
of total body irradiation (TBI) is suf cient to establish
donor engraftment (establishintervention
a cure). This dose
is falling aboveof
or irra-
below a preestablished
diation has been shown to beprobability outcome range
well tolerated. and meeting predesigned
In heavily
transfused recipients who are HLA mismatched, more
improvement.
conditioning will probably be required. The optimal
dose of TBI for this cohort has not been established.
The focus of this study is to establish the optimum
dose of TBI to achieve 20 percent donor cells
(chimerism) in patients enrolled in the protocol.
How many patients must be enrolled per cohort to ob-

tain durable bone marrow stem cell establishment (en-
×
graftment)? Patients are monitored monthly for the
Small Clinical level
Trials:ofIssues
donor and Challenges
chimerism. Engraftment can be consid-
ered durable if 20 percent donor chimerism is present
at ≥6 months. When can TBI dose escalation be imple-
mented? How many patients Clinical
are required pertogroup
trials are used elucidate the most
before anBuy 
Paperback
increase in dose| can
$50.00
appropriate
be made? preventive, diagnostic, or treatment
Cohort Number of subjectsresearch
needed TBI dose
to see (cGy)
 Buy Ebook | $39.99 participants if the intervention is safe
A MyNAPto members save
be determined (see below)is more
participants 200prone to variability and carries a
B save! "
present. This may occur in phase I (safety and
250
C " 300
D " the probability of 350rejecting the null hypothesis when

One traditional approach to this problem
methodologies andisthe
toappropriate
identify situations for the
an acceptable engraftment rate and to then identify
the number of subjects required to ensure
strategies such as (1)that the
meta-analysis to combine
disparate information
con dence interval for the true proportion is suf - from several studies including
ciently narrow to be protective of human
method health.
and (2) other For such as assessing
alternatives
example, if the desired engraftment
therapeuticrate isin95
results percent,
a single treated population (e.g.,
19 subjects will provide a 95 percent con dence inter-
val with a width of 10 percentprobability
(i.e., 0.85 to 1.00).
outcome range Ifandfor a predesigned
meeting
particular application, this interval is too wide, a width
improvement.
of 5 percent can be obtained with 73 subjects (0.90 to
1.00).
On the basis of these results, should 73 subjects be re-

quired for each TBI dose group? Is a total of 292 pa-
tients really needed for all dose groups? The answer is
that a much smaller total number of patients is re-
quired by invoking a simple sequential testing strategy.
×
For example, assume that the study begins with three
Small Clinical patients
Trials: Issues
in theand Challenges
lowest-dose group and it is observed
that none of the patients are cured. On the basis of a
binomial distribution and by use of a target engraft-
ment proportion of 0.95, the Clinical
probability
trials are that zero
used to of the most
elucidate
three engraftments
Buy Paperback | established
will be $50.00
when the diagnostic,
true or treatment
population Perhaps the most essential feature of a clinical trial is
 Buy Ebook | $39.99 research participants to see if the intervention is safe

MyNAP members save
Page 64 10% online. considerable risk of failing to demonstrate the
proportion is 0.95 is approximately 1 in 10,000. Similar-
ly, the cumulative probabilitysample
of one orthey
sizes, fewer cures
usually is
have adequate statistical
less than 15 percent. As such,power,
afterwhich
onlyis three
the committee's
patients de nition of a "large"
are tested, considerable information
have adequate regarding
statistical power, statistical power being
whether the true cure rate isthe
95probability
percentofor moretheisnull
rejecting al-hypothesis when
ready available. Following this simple sequential strate-
gy, one would test each doseSmall(beginning with
Clinical Trials thethe
assesses low- current
est dose) with a small number of patients (e.g., three
patients) and increase to the report
next assesses
dose level if the re-
the published literature on various
sults of the screening trial indicate that the probability
of cure for the targeted proportion (e.g., 0.95
Bayesian techniques as percent)
in the con dence pro le
is small. In the current example, one would clearly in-
crease the dose if zero of three patients
astronauts) was cured
by sequentially and whether the
measuring
would most likely increase the intervention
dose toisthe fallingnext
abovelevel
or below a preestablished
even if one or two patients werespeci cured.
cations asIf, in thisto ex-
opposed incremental
ample, all three patients engrafted, one would then
improvement.
test either 19 or 73 patients (depending on the desired

width of the con dence interval) and determine a con-
dence interval for the true engraftment rate with the
desired level of precision. If the upper con dence limit
is less than the targeted engraftment rate, then one
would proceed to the next highest TBI dose level and
repeat the test.

ria can be set; for example, S is equal to a gain or no worse than 1 percent
loss of bone mineral density relative to that at baseline. Indeterminate (I) is
equal to a moderate loss of 1 to 2 percent from that at the baseline. F is
×
equal to the severe loss of 2 percent orClinical
moretrials
from are used to elucidate the most
that at the baseline

(Feiveson, 2000). See Box 1-2 for an alternate design
options for discussion
individuals with a given of thiscondition.
medical case
study. Perhaps the most essential feature of a clinical trial is
The use of study stopping

Buy Ebook | $39.99
(cessation)
research participants to see if the intervention is safe
rules
and e that
ectiveare
or if based on successive
it is comparable to a comparison
examinations of accumulating data maytreatment.
cause dif Sample culties
size is abecause of theof any
crucial component
MyNAP
need to reconcile members
such saverules with the standard approach to sta-
stopping participants is more prone to variability and carries a
10%
tistical analysis online.
used for the analysis ofconsiderable
data fromrisk most clinical
of failing trials. This
to demonstrate the
standard approach is known as the “frequentist approach.” In this approach
the analysis takes a form that is dependent on thepro
pharmacologic study les), IIdesign. When
(pilot e cacy such
evaluation),
analyses assume a design in which all data are simultaneously available, it is
called a “ xed-sample analysis.” If the data
sample from
sizes, athey
clinical
usually havetrialadequate
are notstatistical
ex-
amined until the end of the study, thenpower, which is the committee's
a xed-sample analysisdeisnition valid.of aIn
"large"
comparison, if the data are examined inhave a way thatstatistical
adequate mightpower, lead statistical
to earlypowerces-being
sation of the study or to some other change of design,
the probability then
of rejecting theanullxed-sample
hypothesis when
analysis will not be valid. The lack of validity is a matter of degree: if early
cessation or a change of design is an extremely
Small Clinicalremote possibility,
Trials assesses the currentthen
xed-sample methods will be approximately valid (Whitehead, 1992, 1997).
For example, in a randomized clinical trial for investigation
strategies of thetoeffect
such as (1) meta-analysis combineof
a selenium nutritional supplement on the prevention of skin cancer, it is
determined that plasma selenium levelsmethod
are notand rising
(2) other as expected
alternatives such in some
as assessing
patients in the supplemented group, indicating a possible noncompliance
problem. In this case, the failure of some subjects
intervention to receive
is falling above or the
belowprescribed
a preestablished
amount of selenium supplement wouldprobability
have ledoutcome
to a loss rangeofand meetingto
power predesigned
de-
tect a signi cant bene t, if one was present. One could then initiate a
improvement.
prestudy
Page 65

×

 Buy Ebook | $39.99research participants to see if the intervention is safe

MyNAP members save

improvement.
~ enlarge ~
FIGURE 3-1 Parameters for a clinical trial with a sequential design for pre-
vention of loss of bone mineral density in astronauts. A. Group sample
sizes available for clinical study. B. Establishment of repeated con dence
intervals for a clinical intervention for prevention of loss of bone mineral
×
density for determination of the success (S) or failure (F) of the
intervention.
SOURCE: Feiveson (2000).
 Buy Paperback | $50.00appropriate preventive, diagnostic, or treatment

Page 66
MyNAP members save
BOX 3-3 participants is more prone to variability and carries a
Login or Register to
Sequential Testing with Limited Resources
e ectiveness of a given intervention when one really is
As an illustration of sequential testing in small clinical
studies, consider
DownloadtheFreeinnovative
PDF trials.approach
Although phase to Iforensic
and II studies may have smaller
drug testing proposed by Hedayat, Izenman, and Zhang
(1996). Suppose that N units such as pills aor
trial. Sometimes trialtablets
with eightor
participants may
squares of lysergic acid diethylamide (LSD) are ob-
tained during an arrest and one would like
the hypothesis to deter-
is false.
mine the minimal number that would have to be
screened to state with 95 percent con dence
methodologies that at situations for the
and the appropriate
conduct
least N1 of the total N samples will beof clinical trials with
positive. Tosmall sample sizes. This
solve
the problem, de ne m as thestrategies
expected suchnumber of neg-to combine
as (1) meta-analysis
disparate
ative units in the initial random information
sample from several
of n units and studies including
X as the observed number of method
negative units
and (2) otherin a sample
alternatives such as assessing
of size n. Typically, the forensic scientist assumes that population (e.g.,
therapeutic results in a single treated
m is equal to 0, n samples areintervention
collected, and the actual
is falling above or below a preestablished
number of negative samples (X) is determined.
probability outcome rangeNext, and meeting predesigned
de ne k as the minimum number of positive drug sam-
improvement.
ples that are needed to achieve a conviction in the
case. One wishes to test with a con dence of 100(1 −
Î±, where Î± is the probability of committing a type 1
error) percent that N1 ≥ k. The question is: what is the
smallest sample size n needed?
Hedayat and co-workers showed that the problem can

be described in terms of the inequality
×
maxN1<kProb[X ≤ m | N1] ≤ Î±,
which is equivalent to
maxN1≤k−1 Prob[X ≤ m | N1] ≤ Î±,

 Buyed
and is satis Paperback
by | $50.00
Prob[X ≤ m | N1 = k − 1] ≤ Î±. that it aims to use results based on a limited sample of
This is a cumulative probability of the hypergeometric
distribution, that can be expressed asA trial with a small number of research
clinical trial.
MyNAP members save
~ enlarge ~
treatment period in which potential noncompliers
strategies such ascould be identi
(1) meta-analysis ed and
to combine
eliminated from the study before randomization (Jennison
disparate information and Turnbull,
from several studies including
1983). method and (2) other alternatives such as assessing
Another reason for early examination of study results
astronauts) is tomeasuring
by sequentially check the as- the
whether
sumptions made when designing the trial. For example, in an experiment
where the primary response variable isspeci
quantitative, the sample
cations as opposed size is of-
to incremental
improvement.
ten set assuming this variable to be normally distributed with a certain
variance. For binary response data, sample size calculations rely on an as-
sumed value for the background incidence rate; for time-to-event data
when individuals enter the trial at staggered intervals, an estimate of the
subject accrual rate is important in determining the appropriate accrual
period. An early interim
×
Page 67
For example, assume that the total number of units

under investigation is 150 andClinical
suppose that
trials are usedone wantsthe most
to elucidate
 Buy Paperback
to claim with 95 percent con | $50.00
dence that the number
of positive units, N1, is at least 135. If
Perhaps theone
mostassumes thatof a clinical trial is
essential feature
there will be no negative units in the initial sample (i.e.,

m = 0) thenBuyone
Ebook | $39.99
can begin
with
and an initial
e ective or ifsample of 25,to a comparison
it is comparable
using the inequality given above.treatment.
TheSample size is a crucial component of any
investigators
MyNAP
draw members
a random save of 25, and if no negative units
sample participants is more prone to variability and carries a
10% online. considerable risk ofwith
failing95 to demonstrate the
are found, the investigators can conclude per-
cent con dence that the totalpresent.
number of positive units
save! This may occur in phase I (safety and
(N1) is greater than 135. Note pharmacologic
that if one pro actually ob-e cacy evaluation),
les), II (pilot
serves X to be equal to 1 negative unit, one then can
determine what value of N1 issample
feasible
sizes,or
theyrecompute n.
usually have adequate statistical
For example, say that one observes X is equal
trial. Sometimes to 2
a trial with neg-
eight participants may
ative units among 25 initial samples. With 95 percent
con dence one can claim that k equal to 118 positive
units will be found. Alternatively, if one requires N1 to
be ≥ 135 positive units, one can increaseand
methodologies n tothe61 sam- situations for the
appropriate
ples by drawing an additionalconduct
61 - 25 = 36 random
of clinical trials with small sample sizes. This
samples. strategies such as (1) meta-analysis to combine
A useful example in clinical trials is the
Bayesian comparison
techniques ofdence
as in the con a pro le
new drug with a standard drug for the treatment of a
rare disease. For example, it may be known
astronauts) that the
by sequentially measuring whether the
intervention is falling
rate of response to an existing drug is 80 percent; above or below a preestablished
however, the drug has serious side
speci effects.
cations A new
as opposed drug
to incremental
without the side effect pro le of the old drug has been
improvement.
developed, but it is not known whether it is equally ef-

cacious. Power computations revealed that 150 sub-
jects are required to document that the response rate
is at least 90 percent with 95 percent con dence (i.e.,
at least 135 of 150 patients respond). Unfortunately, 150
subjects are not available. Using the strategy developed
by Hedayat and colleagues (1996), one can examine 25
×
patients, and if they all respond, then one can con-
Small Clinical clude
Trials:with
Issues and Challenges
95 percent con dence that the total num-
ber of responders is at least 135 among the 150 patients
that the investigators would have liked to test. There
are numerous applications ofClinical
this type ofused
trials are sequential
to elucidate the most
testing strategy 
appropriate
in small clinical trials.preventive, diagnostic, or treatment

analysis can reveal inaccurate assumptions treatment.in time
Sample for
sizeadjustments to beof any
is a crucial component
made to the design
MyNAP (Jennison
members saveand Turnbull, 1983).
Sequential methods
Login or typically
Register tolead to savings
e ectivenessin sample
of a givensize, time,when
intervention andonecost
really is
compared with save! present. This procedures
those for standard xed-sample may occur in phase I (safety
( Box and
3-3).
However, continuous monitoring is notand always practical.
III (extensive assessment of safety and e cacy)
HIERARCHICAL MODELS sample sizes, they usually have adequate statistical
Hierarchical models can be quite usefulhave in adequate
the context of power,
statistical smallstatistical
clinicalpower
tri-being
als in two regards. First, hierarchical models provide a natural framework
for combining information from a series of small clinical trials conducted
Page 68 disparate information from several studies including
within ecological units (e.g., space missions or clinics). In the case where
the data are complete, in which the same response
astronauts) measure
by sequentially is available
measuring whetherforthe
each individual, hierarchical models provide a more rigorous solution than
meta-analysis, in that there is no reason to cations
speci use effect magnitudes
as opposed as the
to incremental
improvement.
unit of observation. Note, however, that a price must be paid (i.e., the total
sample size must be increased) to reconstruct a larger trial out of a series
of smaller trials. Second, hierarchical models also provide a foundation for
analysis of longitudinal studies, which are necessary for increasing the
power of research involving small clinical trials. By repeatedly obtaining
data for the same subject over time as part of a study of a single treatment
or a crossover study, the total number of subjects required in the trial is
×
reduced. The reduction in the sample size number is proportional to the
Small Clinical
degree Trials: Issues and
of independence Challenges
of the repeated measurements.
A common theme in medical research is two-stage sampling, that is, sam-

pling of responses within experimental units (e.g., patients) and sampling of
experimental units within populations.appropriate
For example, in prospective longi-

tudinal studies patients are repeatedly options
sampled
preventive,
for individuals
and assessed
diagnostic, or treatment
with a given
in medical
termscondition.
of a
variety of endpoints such as mental andthat physical
it aims to use levels ofbased
results functioning
on a limitedor in of
sample
terms of the response BuyofEbook
one or|more biological systems to ifone or more is safe
 $39.99research participants to see the intervention
forms of treatment. These patients are treatment.
in turn sampled from a population,
Sample size is a crucial component of any
often strati ed on the basis of treatment delivery,
clinical for
trial. A trial withexample, in aofclinic,
a small number research
MyNAP members save
in a hospital, 10%
or during
online.space missions.considerable
Like all biological and behavioral
characteristics, theoroutcome
Login measureseexhibit
Register to ectiveness individual differences.
of a given intervention when oneIn-really is
save! be interested in not present.
vestigators should only the Thismean
may occur in phase I (safety
response pattern and but
also the distribution of these response and patterns (e.g.,
III (extensive time trends)
assessment of safety in
andthe
e cacy)
Download
population of patients. Freethen
One can PDF address
trials. Although phase I andor
the number II studies may have of
proportion smaller
patients who are functioning more or less power, positively
which is the at a speci dec nition
committee's rate.ofOne
a "large"
can then describe the treatment-outcome relationship
trial. Sometimes noteight
a trial with as aparticipants
xed law may
but as a family of laws, the parameters the of which describe the individual
probability of rejecting the null hypothesis when
biobehavioral tendencies of the subjects thein the population
hypothesis is false. (Bock, 1983).
This view of biological and behavioral research may lead to Bayesian meth-
ods of data analysis. The relevant distributions
methodologies existandobjectively
the appropriateand can for
situations bethe
investigated empirically.
In medical research, a typical example of two-stage
disparate sampling
information is the
from several longi-
studies including
Bayesian
tudinal clinical trial, in which patients are techniques
randomly as in the con
assigned to dence pro le
different
treatments and are repeatedly evaluated over the
therapeutic course
results of the
in a single study.
treated De-(e.g.,
population
astronauts)
spite recent advances in statistical methods forby sequentially measuring
longitudinal research,whether
thethe
cost of medical research is not always commensurate withand
probability outcome range the quality
meeting of
predesigned
the analyses. Reports of such studies often
speci consist of littletomore
cations as opposed than an
incremental
improvement.
end-
Page 69
point analysis in which measurements only for those participants who have
completed the study are considered in the analysis or the last available
×
measurement for each participant is carried forward as if all participants
Small Clinical
had, Trials:
in fact, Issuesthe
completed andstudy.
Challenges
In the rst example of a “completer-
only” analysis, the available sample at the end of the study may have little
similarity to the sample initially randomized. There is some improvement
in the case of carrying the last observation Clinicalforward. However,
trials are used to elucidate participants
the most
treated in the analysisBuyas 
Paperback
if they have | $50.00
appropriate
had identical preventive,
exposuresdiagnostic,
to or
thetreatment
drug
may have quite different exposures in reality Perhaps the or most
theiressential
experiences
feature of awhile
clinical re-
trial is
ceiving the drug may be complicated bythat other
it aimsfactors thatbased
to use results led ontoatheir
limited sample of
withdrawal from the Buy

study
Ebookbut |that
$39.99research participants to see if the intervention is safe
areand
ignored in the analysis. Both cases
e ective or if it is comparable to a comparison
lead to dramatic losses of statistical power since
treatment. the size
Sample measurements made
is a crucial component of any
on the intermediate occasions
MyNAP members are simply discarded. In these studies a re-
save
10% online.
view of the typical level of intraindividual variability
considerable risk ofof responses
failing should
to demonstrate the
raise serious questions regarding reliance on any single measurement.
To illustrate the problem, consider the andfollowing example.
III (extensive assessment Suppose a longi-
of safety and e cacy)
tudinal randomizedDownload Free is
clinical trial PDF trials. Although
conducted phase Ithe
to study and effects
II studies may
of ahave
par- smaller
ticular therapeutic intervention (countermeasure)
power, which is the oncommittee's
bone mineral density
de nition of a "large"
measurements taken at multiple pointstrial. in time during
Sometimes a trialthe
with course of a may
eight participants
space mission. At the end of the study, the
theprobability
data comprise a le of bone
of rejecting the null hypothesis when
mineral density measurements for eachthepatienthypothesis (astronaut)
is false. in each treat-
ment group. In addition to the usual completer or end-point analysis, a
data analyst might compute means for methodologies
each week and andthemight t separately
appropriate situations for the
for each group a linear or curvilinear trend line that shows average bone
mineral density loss per week. A more sophisticated
strategies such as (1)analyst might
meta-analysis t the
to combine
disparate information from several
line using some variant of the Potthoff-Roy procedure, although this would studies including
require complete and similarly time-structured
method and data
(2) otherfor all subjects
alternatives such as(Bock,
assessing
1979). therapeutic results in a single treated population (e.g.,
Despite the question of whether bone mineral
probability density
outcome rangemeasurements are
and meeting predesigned
related to the ability of an astronaut tospeci
function
cationsin space, to
as opposed most objection-
incremental
improvement.
able is the representation of the mean trend in the population as a biologi-
cal relationship acting within individual subjects. The analysis might pur-
port that as any astronaut uses a countermeasure he or she will decrease
the effect of life in a weightless environment on bone mineral density loss
at some xed rate (e.g., 0.1 percent per week). This is a gross oversimpli -
cation. The account is somewhat improved by reporting of mean trends for
important subgroups: astronauts of various ages, males and females, and so
on. Even then, within such groups some patients will respond more to a
×
given countermeasure, some will respond less, and the responses of others
Small Clinical
will Trials:atIssues
not change and
all. Like all Challenges
biological characteristics, there are individual
differ-

Buy Paperback | $50.00appropriate preventive, diagnostic, or treatment
Page 70 Perhaps the most essential feature of a clinical trial is
ences in response trends.
Buy Ebook | $39.99
Therefore, research participants to see if the intervention is safe
both the mean trend and the distrib-
ution of trends in the population of patients
treatment.areSample
of interest. Onecomponent
size is a crucial can then of any
speak of the number or proportion clinical trial. A trial with a small number of research
MyNAP members save of patients who respond to a clinically
acceptable degree and the rates at which
10% online. their biological
considerable risk of failingstatus changes
to demonstrate the
over time. Login or Register to e ectiveness of a given intervention when one really is
In a longitudinal study, repeated observations are nested within individuals
and the hierarchical model isFree
Download used to incorporate
PDF the effects
trials. Although phase of intrasub-
I and II studies may have smaller
ject correlation on estimates of uncertainty (i.e., standard errors and con -
dence intervals) and tests of hypotheses for
trial. the xed
Sometimes effects
a trial with eightorparticipants
structural may
parameters (e.g., differential treatment ef cacy) in the model. Note that hi-being
have adequate statistical power, statistical power
erarchical models are equally useful in the thehypothesis
contextis of clustered data, in
false.
which participants are nested within groups (e.g., different studies or space
missions), and the sharing of this similar environment
methodologies and theinduces a correlation
conduct
among the responses of participants within strata.of clinical trials with small sample sizes. This
Analysis of this type of data (under the disparate
assumptions that
information a subset
from of the
several studies re-
including
gression parameters has a distribution Bayesian
in the population
techniques as inof theparticipants
con dence pro le
and that the model residuals have a distribution in the population of re-
sponses within participants and also inastronauts)
the population of participants)
by sequentially measuring whether be-the
longs to the class of statistical analytical models called:
mixed model (Elston and Grizzle, 1962; Longford, 1987);
improvement.
regression with randomly dispersed parameters (Rosenberg, 1973);
exchangeability between multiple regressions (Lindley and Smith,

1972);
two-stage stochastic regression, (Fearn, 1975);
James-Stein estimation (James and Stein, 1961);

×
variance component models (Dempster, Rubin, and Tsutakawa, 1981;
Small Clinical Trials:
Harville, Issues and Challenges
1977);
random coef cient models (DeLeeuw and Kreft, 1986);
Clinical
hierarchical linear models (Bryk and trials are used to
Raudenbush, elucidate the most
1987);
multilevel models (Goldstein, 1986); and the most essential feature of a clinical trial is
Perhaps
random-effectBuy
regression
Ebook |models (Laird participants
and Ware, 1982).
 $39.99research to see if the intervention is safe
Along with the seminal articles that have described these statistical mod-
MyNAP members save
els, several book-length texts that further describe
participants is morethese
pronemethods
to variability have
and carries a
been published (Bock, 1989; Bryk and Raudenbush, 1992; Diggle, Liang, and
Zeger, 1994; Goldstein,
save! 1995;Jones, 1993; Lindsey,
present. This may1993; Longford,
occur 1993).
in phase I (safety and For
the most part, these treatments are based pharmacologic pro les), II (pilot e that
on the assumptions cacy evaluation),
the
residual effects areDownload
normallyFreedistributed with
PDF trials. zerophase
Although means I andand a covariance
II studies may have smaller
Page 71 the hypothesis is false.

BOX 3-4 conduct of clinical trials with small sample sizes. This
Power Consideration for Space Mission
Clinical Trials Bayesian techniques as in the con dence pro le
A natural application for hierarchical
therapeutic regression mod-
results in a single treated population (e.g.,
els is the problem in which astronauts are nested with-
in space missions and the intervention (e.g.,range
probability outcome the and
pres-meeting predesigned
speci cations as opposed
ence or the absence of a particular countermeasure) is to incremental
improvement.
randomly assigned at the level of the space mission. To
illustrate the problem, assume that one is interested in
detecting a difference of a 0.5 standard deviation unit
between control and experimental conditions by a
one-tailed test. In addition, assume that ve astro-
nauts are available per space mission and that the in-
traspace mission correlation is 0.2.
×
Assuming a Type I error rate of 5 percent (i.e., 95 per-
Small Clinical cent
Trials: Issues
con andhow
dence), Challenges
many space missions are re-
quired to have 80 percent statistical power of detec-
tion of a difference? Using statistical power computa-
tions for the clustered t distribution
Clinical trials(Hsieh,
are used1988) onethe most
to elucidate
nds thatBuy 
Paperback
detection | $50.00
appropriate
of a difference of preventive,
0.5 standard diagnostic,
de-or treatment
viation unit with 80 percent power Perhaps the would requirefeature
most essential for of a clinical trial is
each condition (i.e., the control that versus
it aims to the experimen-
use results based on a limited sample of
tal condition) 
Buy Ebook
18 space| $39.99 research participants to see if the intervention is safe
missions, each with 5 subjects,
or a total of 180 astronauts.Note thatSample
treatment. if thesizeeffect sizecomponent of any
is a crucial
is increased to a difference
MyNAP members save of 1 standard deviation unit,
10%may
which online.
be acceptable for considerable
bone mineral risk of density
failing to demonstrate the
measurement data, the number of space missions is
reduced to 5 per condition, for a total ofpro
pharmacologic 50les),
astronauts.
II (pilot e cacy evaluation),
In a longitudinal study (i.e., repeated evaluation ofofas-
and III (extensive assessment safety and e cacy)
tronauts during their tour of sampleduty),sizes,statistical
they usually power
have adequate statistical
computations become more complex power, whichbecause they can
is the committee's de nition of a "large"
involve both random effects and residual autocorrela-
tions. (The interested reader the is referred
probability ofto the paper
rejecting the null hypothesis when
by Hedeker, Gibbons, and Waternaux [1999]).
strategies such
matrix in all participants, and that the random as (1) are
effects meta-analysis
normallyto combine
dis-
tributed with zero means and covariance matrix.
Bayesian Recent
techniques as inreview articles
the con dence pro le
summarize the use of hierarchical models in biostatistics and health ser-
vices research (Gibbons, 2000; Gibbonsastronauts)
and Hedeker, 2000). Some statisti-
cal details of the general linear hierarchical regression
intervention model
is falling above are aprovided
or below preestablished
in Appendix A. The case study presented in Box 3-4 provides an example of
how hierarchical models can be used toimprovement.
aid in the design and analysis of
small clinical trials.
BAYESIAN ANALYSIS
The majority of statistical techniques that clinical investigators encounter
are of the frequentist school and are characterized by signi cance levels,
con dence intervals, and concern over the bias of estimates (Jennison and
Turnbull, 1983). The Bayesian philosophy of statistical inference however is
×
fundamentally different from that underlying the frequentist approach
(Malakoff, Issues
1999; Thall, and Challenges
2000). In certain types of investigations Bayesian
Page 72 
Buy Paperback | $50.00appropriate preventive, diagnostic, or treatment
analysis can lead to practical methods that that it are
aims similar to those
to use results based onused bysample
a limited sta- of
tisticians who use the 
Buyfrequentist
Ebook | $39.99 research participants to see if the intervention is safe
approach.
The Bayesian approach has a subjectiveclinicalelement. It focuses
trial. A trial with a smallon an unknown
number of research
MyNAP members save
parameter value q, which measures theparticipants
effect ofisthe moreexperimental
prone to variabilitytreat-
and carries a
ment. BeforeLogin
designing a study
or Register to or collecting any of
e ectiveness data, the
a given investigator
intervention when one ac-
really is
quires all available
save! information about the activities
present. This mayof both
occur the experimen-
in phase I (safety and
tal and the control treatments. This provides some information about the
possible value of Î¸.Download Free PDF trials. Although phase I and II studies may have smaller
The Bayesian approach is based on the power, which is the committee's de nition of a "large"
supposition that the investigator's
opinion can be expressed in the form ofhave a value
adequate P(Î¸ ≤ power,
forstatistical x) forstatistical
every x be-being
power
tween − ∞ and ∞. Here P(Î¸ ≤ x) represents the probability of rejecting the
the probability nullÎ¸hypothesis
that is less when
than or equal to x. The probability is not frequentist: it does not represent
Smallor
the proportion of times that Î¸ is less than Clinical
equal Trials
toassesses the current
x. Instead, P(Î¸ ≤ x)
represents how likely the investigator thinks
conduct it to betrials
of clinical thatwith
Î¸ is less
small than
sample or This
sizes.
equal to x. The investigator is allowed toreport assesses
think onlythe inpublished
terms of literature on various
functions
P(Î¸ ≤ x) which rise from 0 at x = − ∞ to disparate
1 at x =information
∞. Thus fromP(Î¸ several
≤ x) de nesincluding
studies a
probability distribution for Î¸1, which will be called
Bayesian theassubjective
techniques distribu-
tion of Î¸. Notice how deep the division therapeutic
betweenresults
the frequentist and
in a single treated the (e.g.,
population
Bayesian goes: even the notion of probability receives
astronauts) a different
by sequentially measuringinterpre-
whether the
tation (Jennison and Turnbull, 1983, p. 203).
Thus, before the investigator has observed any data, a subjective distribu-
improvement.
tion of Î¸ can be formulated from the experiences and knowledge gained by
others. At this stage, the subjective distribution can be called the prior dis-
tribution of Î¸. After data are collected, these will in uence and change
opinions about Î¸. The assessment of where q lies may change (re ected by
a change in the location of the subjective distribution), and uncertainty
about its value should decrease (re ected by a decrease in the spread of
this subjective distribution). The combination of observed data and prior
×
opinion is governed by Bayes's theorem, which provides an automatic up-
Small Clinical
date of theTrials: Issues and
investigator's Challenges
subjective opinion. The theorem then speci es a
new subjective distribution for Î¸, called a posterior distribution (Jennison
and Turnbull, 1983).
The attraction of the Bayesian approach lies in its simplicity of concept and

appropriate
options forand
the directness of its conclusions. Its exibility
preventive,
individuals
diagnostic,
lack with
or treatment
a given medical
of concern for condition.
in-
terim inspections are especially valuable
thatin sequential
it aims clinical
to use results trials.
based on The
a limited sample of
main problem withBuy the Bayesian approach, however, lies in ifthe idea of a

Ebook | $39.99 research participants to see the intervention
is safe
subjective distribution. treatment. Sample size is a crucial component of any

MyNAP members save
Subjective opinions are a legitimate part of personal
participants is moreinferences. A small
prone to variability in- a
and carries
vestigating team might be in suf cient agreement to share the same prior
distribution but it
save! is less likely that all members
present. This of
maythe team
occur willI (safety
in phase holdandthe
same prior opinions and some members pharmacologic pro les), II (pilot
will be reluctant e cacy evaluation),
to accept an
analysis based in part on opinions
Download that trials.
Free PDF theyAlthough
do notphaseshare.
I andAn alternative
II studies may have smaller
possibility is for investi- sample sizes, they usually have adequate statistical
Page 73
gators to adopt a prior distribution representing only
conduct of clinical vague
trials subjective
with small sample sizes. This
opinion, which is quickly overwhelmedreport assesses the published
by information from the literature
data.on The
various
latter suggestion leads to analyses which are similar
disparate informationtofrom
frequentist infer-
several studies including
ences, but it would appear to lose the spirit
Bayesianoftechniques
the Bayesianas in theapproach.
con dence proIf le
the prior distribution is not a true representation of subjective opinion
then neither is the posterior (Jennison astronauts)
and Turnbull, 1983, measuring
by sequentially p. 204). whether the
More generally, the Bayesian approach probability
has the outcome
following range and meeting predesigned
advantages:
improvement.
Problem formulation. Many problems, such as inferences or decision
making based on small amounts of data, are easy to formulate and
solve by Bayesian methods.
Sequential analysis. Because the posterior distribution can be updat-

ed repeatedly, using each successive posterior distribution as the pri-
or distribution for the next update, it is the natural paradigm for se-
quential decision making.
×
Meta-analysis. Bayesian hierarchical models provide a natural frame-
Small Clinical
workTrials: Issues and
for combining Challenges
information from different sources. This is often
referred to as “meta-analysis” in the context of clinical trials, but the
methods are quite broadly applicable.
Prediction. An especially useful tool is the predictive probability of a

appropriate
options
future event. This allows one to make
preventive, diagnostic,
for individuals such
statements with a given
or treatment
medical condition.
as “Given that
an astronaut has not suffered bone mineral
that it aims to density
use resultsloss
basedduring thesample of
on a limited
rst year of a 2-year space|mission, the probability that heintervention
or she will

Buy Ebook $39.99 research participants to see if the
is safe
suffer bone mineral density loss during the second year is 25 percent.”
MyNAP members save
Communication. Bayesian models,participants
methods, andprone
is more inferences are
to variability andoften
carries a
easier to communicate to nonstatisticians. This is because most peo-
ple thinksave!
and behave like Bayesians, whether
present. This mayoroccur
notintheyphaseunderstand
I (safety and
pharmacologic
or are even aware of the formal paradigm. The proposterior
les), II (pilot edistribution
cacy evaluation),
provides a framework
Downloadfor describing
Free PDF trials.and communicating
Although phase I and II studies one's
maycon-have smaller
sample sizes, they usually
clusions in a variety of ways that make sense to nonstatisticians. Al- have adequate statistical
though the details are not presented trial. here,
SometimesBayesian
a trial with methods (Thall,
2000; Thall and Sung, 1998; Thall and Russell, 1998; Thall, Simon, and being
Estey, 1995; Thall, Simon, and Shen, 2000; White-head and Brunier,
1995) can be applied in most of the design and analysis situations de-
scribed in this report and in manymethodologies
cases will be extremely useful for
and the appropriate situations for the
the analysis of results of small clinical
conduct trials.
of clinical trials with small sample sizes. This
DECISION ANALYSIS disparate information from several studies including
Decision analysis is a modeling technique method
thatand (2) other alternatives
systematically such as assessing
considers all
possible management options for a problem (Hillner
astronauts) and Centor,
by sequentially measuring1987).
whetherItthe
intervention
uses probabilities and utilities to explicitly de ne is falling above orThe
decisions. belowcomputa-
a preestablished
improvement.
Page 74
tional methods allow one to evaluate the importance of any variable in the
decision-masking process. Sensitivity analysis describes the process of re-
calculating the analysis as one changes a variable through a series of plau-
×
sible values. The steps to be taken in decision analysis are outlined in Table
3-1.
As mentioned in Chapter 2, one can use decision analysis as an aid in the

experimental design process. If one models a clinical situation a priori, one
can test the importance of a single value in making the decision intreatment
ques-

appropriate
options
tion. Performance of a sensitivity analysis before
preventive,
for individuals
a study
diagnostic, or
withisadesigned
given medicalto condition.
provide an understanding of the in uencethat itof a given
aims valuebased
to use results on the decision.
on a limited sample of
Such analyses can determine the best use of a small clinical
to see if trial. This pre-

Buy Ebook | $39.99 research participants the intervention
is safe
analysis allows one to focus data collection on important variables (see Box
3-5). clinical trial. A trial with a small number of research
MyNAP members save
10% online.
The other major advantage of decision considerable
analysis occursrisk of failing to demonstrate the
after data collection.
If one assumes that the sample size is inadequate
save! present. This may and therefore
occur that and
in phase I (safety the
pharmacologic
con dence intervals on the effect in question are pro
wide, les),one
II (pilot
maye cacy evaluation),
still have a
clinical situation for which a Free
Download decision
PDF istrials.
required.
Although One
phase might havemay
I and II studies to have
make smaller
decisions under conditions of uncertainty, despite a desire to increase the have adequate statistical
certainty. The use of decision analysis can make explicit
trial. Sometimes a trial with the uncertain
eight participantsde- may
cision, even informing the level of con dence in the decision. A 1990 Insti-being
tute of Medicine report states: it is thistheexibility of decision analysis that
hypothesis is false.
gives it the potential to help set priorities for clinical investigation and ef-
fective transfer of research ndings to methodologies
clinical practice (Institute of Medi-
and the appropriate situations for the
cine, 1990). The formulation of a decision analytical
conduct of clinical model
trials withhelps investiga-
small sample sizes. This
tors consider which health outcomes are important and how important
they are to one another. Decision analysis alsoinformation
disparate facilitates from consideration of
several studies including
the potential marginal bene t of a newBayesian techniques
intervention byasforcing
in the con compar-
dence pro le
isons with other alternatives or “fallback positions.”
therapeutic resultsCombining
in a single treated several
population (e.g.,
methodologies, such as astronauts) by sequentially measuring whether the
TABLE 3-1 Steps in a Decision Analysis
improvement.
Frame the question
Establish a time horizon
Structure the decision tree (choices, chances, outcomes)
Assess the probabilities
Assess the utilities

×
Evaluate the decision tree (expected value)
Perform sensitivity analysis
Interpret the results

Iterate 
SOURCE: Pauker (2000). that it aims to use results based on a limited sample of

BOX 3-5 pharmacologic pro les), II (pilot e cacy evaluation),
Using Decision Analysis to Prevent
Osteoporosis in Space sample sizes, they usually have adequate statistical
Consider a decision analysis that takes the
have adequate following
into consideration: the probability of rejecting the null hypothesis when
a long space mission accelerates

Small Clinical bone mineral
Trials assesses the current
density loss, methodologies and the appropriate situations for the
bone mineral density loss can produce
strategies fractures to combine
such as (1) meta-analysis
now or in the future, disparate information from several studies including
fractures produce disabilities now and disabilities
in the future, and astronauts) by sequentially measuring whether the
a proposed special program may have ef cacy in
ameliorating bone mineral density loss, but may
improvement.
have side effects.
Determine the expectation by evaluating a decision

tree ( Figure 3.2). A decision tree can be set up for ei-
ther a surrogate measure of bone loss (e.g., bone min-
eral density loss) or for an actual disabling outcome
(e.g., bone fracture, as illustrated in Figure 3.3).
×
Then by making a number of assumptions, such as
probability of fracture (p) = 0.2,
ef cacy of special program (e) = 0.3,

 Buy Paperback
probability | $50.00
of side effectsappropriate
(s) = 0.05, preventive, diagnostic, or treatment
quality of life (LT) after fracture
that it aims (qFx) = 0.85,
to use results based on a limited sample of

quality of life (LT) after side
and e effects
ective or if(qSE) = 0.999,to a comparison
it is comparable
life expectancy
MyNAP members save of astronaut (LE) = 35 y, where y is
10% online.
years, considerable risk of failing to demonstrate the
short-term morbidity fracture (stmFx) = 0.3 y,
short-term
Download Free PDF trials.effects
morbidity side Although(stmSE)
phase I and =II0.1 y, may have smaller
studies
and sample sizes, they usually have adequate statistical
short-term morbidity special program (stmSP) =
0.04 y the probability of rejecting the null hypothesis when
one can assign values at the chance nodes to pick the

best option (see Figure 3.4). methodologies and the appropriate situations for the
improvement.

×


MyNAP members save
~ enlarge ~ the hypothesis is false.
FIGURE 3-2 Decision analysis expectation. SOURCE: Pauker (2000).
Page 76 disparate information from several studies including
improvement.

×


MyNAP members save
~ enlarge ~ power, which is the committee's de nition of a "large"
trial.osteoporotic
FIGURE 3-3 Decision tree for preventing Sometimes a trialfractures
with eight participants
in space. may
SOURCE: Pauker (2000).

improvement.

×


MyNAP members save

~ enlarge ~ disparate information from several studies including
FIGURE 3-4 Assigning values at chanceBayesian
nodestechniques
to pick the as in best
the conoption
dence pro le
(clinical intervention) for preventing osteoporotic
method and (2) fractures in space.
other alternatives such as assessing
SOURCE: Pauker (2000). therapeutic results in a single treated population (e.g.,
Page 77 improvement.
decision analysis, with a sequential clinical trials approach potentially of-

fers additional improvements in the means of determining the ef cacy of a
therapeutic intervention in small trial populations.
Although decision analysis does not address the questions raised by small
clinical trials, it can allow a better trial design to be used and interpretation
of the results of such trials.
×
Decision analytical models can combine data from diverse sources
Small Clinical Trials: Issues
and examine and Challenges
interactions.
Decision analytical models are most powerfully used to answer the

question “What if?” by sensitivity analyses.

Decision analytical models can examine
options forthe impactwith
individuals ofamorbidity
given medical and
condition.
effects on quality of life because they can integrate many attributestrial
Perhaps the most essential feature of a clinical in is
a utility structure.

Decision analyses might be used sequentially
treatment. Sample insize
small ongoing
is a crucial trials,of any
component
in whichMyNAP
the results for every
members save additional patient might guide the use
10% online.
of the model for subsequent patients.
save!functions such a beta functions
Probability present. This maycanoccur in phase
provide I (safety
such and
automat-
ic updating of distributions in a model as moreassessment
and III (extensive patients'ofexperiences
safety and e cacy)
Download
are revealed (Pauker, Free PDF trials. Although phase I and II studies may have smaller
2000).
STATISTICAL PREDICTION trial. Sometimes a trial with eight participants may
When the number of control samples isthepotentially large and the number of
experimental samples is small and is obtained sequentially from a series of
clusters with small sample sizes (e.g., space missions), traditional compar-
isons of the aggregate means or medians mayofbe
conduct of limited
clinical value.
trials with small In those
sample sizes. This
cases, one can view the problem not as a classical hypothesis testing prob-
lem but as a problem of statistical prediction.
disparate Conceptualized
information from several instudies
that way,
including
the problem is one of deriving a limit orBayesian
intervaltechniques as in the con dence pro le
on the basis of the control
distribution that will include the mean therapeutic
or median forinall
results or a treated
a single subset of the (e.g.,
population
experimental cluster samples. For example,astronauts)
onebymay sequentially
wish to measuring
compare whether
thethe
median bone mineral density loss in 5 astronauts in each
probability outcome rangeofand ve future
meeting predesigned
space missions (i.e., a total of 25 astronauts clustered in groups of 5 each)
improvement.
with the distribution of bone mineral density loss in controls over a similar
period of time on Earth or alternatively with that for a control group of as-
tronauts who are in a weightless environment (e.g., the International Space
Station) but who are not taking part in a particular countermeasure pro-
gram. As the number of cluster samples increases, con dence in the deci-
×
Page 78
sion rule also increases. In the following, a general nonparametric ap-
proach to this problem is developed, and its use is illustrated with the
problem of testing for bone mineral density loss during space missions. Al-
though more general than parametric alternatives, a lossdiagnostic,
of statistical pow-

er is associated with the nonparametricoptions for individuals
approach. with a given
Parametric
or treatment
medical condition.
alternatives
(normal, lognormal, and Poisson distributions)
that it aimsare presented
to use results basedinonAppendix A of
a limited sample
and can be used whenBuythe observed data are consistent
participants to with one of these

Ebook | $39.99research see if the intervention
is safe
distributions. treatment. Sample size is a crucial component of any

MyNAP members save
The prediction problem involves construction
participantsof
is a limit
more ortointerval
prone variability that will a
and carries
contain one or more new measurements drawn from that same distribu-
tion with a given
save!level of con dence. As an example,
present. in environmental
This may occur in phase I (safety and
monitoring problems one may be interested pharmacologic pro les), II (pilot
in determining e cacy evaluation),
whether a sin-
gle new measurement (or theFree
Download PDFof n
mean new
trials. measurements)
Although obtained
phase I and II studies may have from
smaller
sample sizes, they usually have
an onsite monitoring location is consistent with background levels as char- adequate statistical
acterized by a series of n measurements obtained
trial. Sometimesfrom off-site
a trial with (i.e., back-
ground) monitoring locations. have adequate statistical power, statistical power being
If the new measurement(s) lies within the interval (or below [above] the up-
per [lower] limit), then one can conclude that
Small theTrials
Clinical measurement from the
assesses the current
on-site monitoring location is consistent with the background measure-
ment and is therefore not affected by activities at the
report assesses site from
the published which
literature the
on various
measurement was obtained. By contrast, if the new measurement(s) lies
outside of the interval, one can conclude that techniques
Bayesian it is inconsistent
as in the con with
dencethe
pro le
background measurement and may potentially have been affected byassessing
method and (2) other alternatives such as the
activities at the site (e.g., disposal of waste or some
astronauts) industrial
by sequentially process).
measuring whether the
One can imagine that as the number ofprobability
future measurements (i.e., new
outcome range and meeting predesigned
monitoring locations and number of constituents
improvement.
to be examined) gets
large, the prediction interval must expand so that the joint probability of
any one of those comparisons by chance done is small, say 5 percent. Of
course, this results in a loss of statistical power. To this end, Gibbons
(1987b) and Davis and McNichols (1987) (see Gibbons [1994] for a review)
suggested that the new measurements be tested sequentially so that a
smaller and more environmentally protective limit can be used. The basic
idea is that in the presence of an initial value that exceeds the background
level in an on-site monitoring location (initial exceedance), another sample
×
for independent veri cation of the level should be obtained. A true ex-
Small Clinical is
ceedance Trials: Issues
indicated and
only if Challenges
both the initial level and the veri cation re-
sample exceed the limit (or are outside the interval). There are many varia-
tions of this sequential strategy in which more than one additional sample
(resample) may be obtained. The net result Clinicalistrials
that areaused
much smaller
to elucidate thepredic-
most
tion limit can be used 
Buysequentially
Paperbackcompared
| $50.00
appropriate with preventive, diagnostic,
the limit that or treatment
would be
used if the statistical prediction decision was the
Perhaps based on the feature
most essential resultofof a single
a clinical trial is
comparison, leading to a dra- that it aims to use results based on a limited sample of
 research participants to see if the intervention is safe

Buy Ebook | $39.99
MyNAP members save
Page 79
matic increase in statistical power. In fact, this strategy is now used almost
exclusively in environmental monitoring andprograms
III (extensive in the United
assessment States
Download
(Davis, 1993; Gibbons, Free PDF
1994, 1996;
trials. Although phase I and II studies may have smaller
Environmental Protection Agency, 1992).
This idea can be directly adapted to thetrial.
problem
Sometimes ofa loss of eight
trial with bone mineralmay
participants
have adequate statistical power,
density in astronauts, particularly with respect to the design and analysis statistical power being
of data from a series of small clinical trials (e.g., space
the hypothesis missions, each con-
is false.
sisting of a small number of astronauts) and in which a potentially large
number of outcomes are simultaneously assessed.and
methodologies Totheprovide a foundation,
consider the case in which a study has n control subjects (e.g., astronauts
on the International Space Station or instrategies
a simulated environment,
such as (1) meta-analysis to butcombine
without countermeasures) and a series of p replicate experimental cohorts
(e.g., space missions), each of size ni (e.g., ni =and
method
5 astronauts in each of p = 5
(2) other alternatives such as assessing
the n control
space missions). The objective is to usetherapeutic results inmeasurements to de-(e.g.,
a single treated population
rive an upper (lower) bound for a subset (e.g., 50 percent) of the n experi-
intervention is falling above or below ai preestablished
mental subjects in at least one of the p probability
experimentaloutcomesubject
range andcohorts (e.g.,
meeting predesigned
space missions). improvement.
Given the previous characterization of the problem and the questionable

distributional form of the outcomes of multiple countermeasures, a natural
approach to the solution of this problem is to proceed nonparametrically.
For a particular outcome (e.g., bone mineral density), de ne an upper pre-
diction limit as the uth largest control measurement among the n control
subjects. If u is equal to n, the prediction limit is the largest control mea-
surement for that particular outcome measure or endpoint. If u is equal to
×
n − 1 then the prediction limit is the second largest control measurement
Small
forClinical Trials: Issues
that outcome andA Challenges
measure. natural advantage of using u < n is that it pro-
vides an automatic adjustment for outliers, in that the largest n − u values
are removed. Note, however, that the larger the difference between u and
n the lower the overall con dence, if everything else
Clinical trials are istokept
used equal.
elucidate the most

options
Now consider the experimental subjects. that nwith
for individuals
Assume a given medical condition.
experimental sub- i
jects (e.g., astronauts who are subjectedthat toitexperimental
aims to use resultscountermeasures)

exist in each of p experimental
Buy Ebook |subject research participants to
$39.99cohorts (e.g., space missions). see Letisssafe
if the intervention
i
be the number of subjects required to be contained
treatment. Samplewithin the interval
size is a crucial componentfor of any
cohort i. For MyNAP
example, if ni is save
members
clinical trial. A trial with a small
equal to 5 and one wishes to have the median number of research
value for cohort10%ionline.
be below the upper prediction
considerable risk limit, then
of failing si is equalthe
to demonstrate to 3.
An effect of theLogin or Register tointervention
experimental e ectiveness of a given intervention
on a particular outcome when one really is
mea-
sure is declared only if the sith largest measurement
pharmacologic pro (e.g., the emedian)
les), II (pilot lies
cacy evaluation),
outside of the prediction interval (or above and III [below]
(extensive assessment
the prediction of safety limit
and e in cacy)
the one-sided case) in all p experimental subject
sample cohorts.
sizes, they usually have adequate statistical
The questions of interest are as follows:trial. Sometimes a trial with eight participants may

Page 80 methodologies and the appropriate situations for the
1. What is the probability of a chance exceedance
strategies in all p experimental
such as (1) meta-analysis to combine
subject cohorts for different values of n, u, ni, si, and p?
2. How is this probability affected therapeutic
by various numbers
results in a singleof outcome
treated population (e.g.,
measures (i.e., k)? astronauts) by sequentially measuring whether the
3. What is the power to detect a real difference between control and
experimental conditions for a given statistical strategy?
improvement.
A drawback to this method is that the control group is typically not a con-
current control group. Thus, if other conditions, in addition to the inter-
vention being evaluated, are changed, it will not be possible to determine if
the changes are in fact due to the experimental condition.
Speci c details regarding implementation of the approach and a general

methodology for answering these questions is presented in Appendix A and
×
is illustrated in Box 3-6.
The use of statistical prediction limits described here represents a par-
adigm shift in the way in which small clinical studies are designed and ana-
lyzed. The method involves characterization of the distribution of control
measurements and the use of parameters for the controldiagnostic,
distribution to

appropriate
options forsamples
draw inferences from a series of more limited
preventive,
individuals ofwithexperimental
or treatment
a given medical condition.
measurements. This is a classical problem that itin statistical
aims to use resultsprediction and de-
parts from the more commonly used paradigm of hypothesis testing. The

Buy Ebook | $39.99 research participants to see if the intervention
is safe
methodology described here is applicable to virtually any problem in which

the number of potential endpoints is largeclinicaland trial.the
A trialnumber
with a smallofnumber
available of research
MyNAP members save
subjects is small.
10% In a recent work by Gibbons
online. and colleagues (submitted
for publication), a similar
Login approach
or Register to wase developed
ectiveness of ato compare
given interventionrelatively
when one really is
small numbers save!
of experimental tissues present.
to a larger This may occur in phase
number I (safety and
of control tis-
sues in terms of potentially thousands of andgene expression
III (extensive assessment levels obtained
Download Free
from nucleic acid microarrays. ToPDF trials.ease
provide Although of phase I and II studies
application, they may have smaller
devel-
oped a “probability calculator” that computes
power, which conis thedence levels
committee's deandnitionstatis-
of a "large"
tical power for any set of values of n, nitrial.
, p, Sometimes
u, i, si, and k.with
a trial (The probability
calculator is freely available at www.uic.edu/labs/biostat/ and
the probability of rejecting the null is usefulwhen
hypothesis for
both design and analysis of small clinical
thestudies.)

META-ANALYSIS: SYNTHESIS OF RESULTS OF methodologies and the appropriate situations for the
INDEPENDENT STUDIES report assesses the published literature on various
Meta-analysis refers to a set of statistical procedures used to summarize
empirical research in the literature ( Table
method3-2). Although
and (2) the concept
other alternatives of
such as assessing
combining the results of many studies has its origins in the early 1900s
agricultural experiments, Glass in 1976 intervention
coined the term
is falling to mean
above or below“the analy-
a preestablished
sis of
improvement.
Page 81
BOX 3-6
×
Case Study of Bone Mineral Density Loss
Small Clinical During
Trials: Issues andMissions
Space Challenges
Space travel in low Earth orbit or beyond Earth's orbit

exposes individuals (astronauts or trials
Clinical cosmonauts) to en-the most
are used to elucidate

vironmentalBuystresses
Paperback (e.g.,|microgravity
$50.00
and cosmic diagnostic,
ra-or treatment
diation) that, if unabated, could result
Perhaps in radiation-in-
the most essential feature of a clinical trial is
duced physiological damage that or marked
it aims to usephysiological
results based on a limited sample of

adaptation Buy Ebook | $39.99
(microgravity-induced
shifts in calcium and
bone metabolism) that could treatment.
be deleterious
Sample size oris even fa-
a crucial component of any
tal MyNAP
duringmembers
space travel,
save on landing on another planet,
10% online.
or after the return to Earth. Based on the
considerable risk ofpreceding
failing to demonstrate the
Login or Register to e ectiveness of
discussion of statistical prediction, one can consider a given intervention when one really is
the details of design and analysis of a potential
pharmacologic pro les), II study
(pilot e of
cacy evaluation),
and
bone mineral density loss in astronauts.III (extensive assessment of safety and e cacy)
Download Free PDF
trials. Although phase I and II studies may have smaller
Assume that there are n control
power, astronauts in eitherdea nition of a "large"
which is the committee's
simulated environment or ontrial.
theSometimes
International Space
a trial with eight participants may
Station not taking part in a countermeasure program,
or perhaps serving as matched the control
hypothesisastronauts
is false. on
Earth and ni experimental subjects (e.g., astronauts
subjected to experimental countermeasures)
methodologies and thein each ofsituations for the
appropriate
p space missions. Let si be the number of subiects re-
quired to be contained withinstrategies
the interval for
such as (1) space to combine
meta-analysis
mission i. For example, if ni isBayesian
equa to 5 and one wishes
techniques as in the con dence pro le
to have the median value in space
methodmission i below
and (2) other thesuch as assessing
alternatives
upper prediction limit, then sastronauts)
i is equalbyto 3. An effect of
sequentially measuring whether the
the experimental intervention on a particular
intervention outcome
measure is declared only if the s th largest measure-
specii cations as opposed to incremental
ment (e.g., the median) lies outside of the prediction
improvement.
interval (or above [below] the prediction limit in the

one-sided case) in all p space missions. Statistical de-
tails for both parametric and nonparametric solutions
to this problem are presented in Appendix A.
Returning to the example, suppose that a series of 20

control astronauts on Earth are monitored for the
same period of time that 5 astronauts on a single space
×
mission are evaluated for the effects of a series of
Small Clinical countermeasures
Trials: Issues andon Challenges
bone mineral density loss. The
question is whether the countermeasures are suf -
cient to eliminate the effect of the space mission on
bone mineral density loss, such that
Clinical the
trials are bone
used tomineral
elucidate the most

Buy Paperback
density measurements for| the
$50.00
appropriate
experimental preventive, diagnostic, or treatment
astronauts
are consistent with the bonePerhapsmineral the density
most essentialmeasure-
feature of a clinical trial is
ments for the control astronauts. To this
that it aims to useend,
resultsconsider
a comparison 
Buy Ebook research participants to see if the intervention is safe
| $39.99 of 20 control measure-
of the maximum and e ective or if it is comparable to a comparison
ments (n = u = 20) with the mediantreatment. far a single
Sample size is space
mission
MyNAP(pmembers ni equal to ve experimental as-
= 1 ) withsave
10% online.
tronauts (i.e., si = 3) for a single outcome
considerable risk measure
of failing to demonstrate the
(Case A). Using the previous equations, one obtains an
overall con dence level of 99.6 percent,pro
pharmacologic indicating
les), II (pilotan
e cacy evaluation),
extremely low probability that the experimental medi-
an will be above the largest control bone
sample sizes, theymineral
usually have den-
adequate statistical
sity measurement by chance power,
alone. which is the committee's de nition of a "large"
One can do better,
however. Instead of selectinghavetheadequate
most extreme control
the probability
measurement, take the 18th largest of rejecting the null hypothesis when
measurement
(Case B). In this case, the con dence is 96 percent and
Small Clinical
one has a more powerful decision rule.Trials
Now, assesses
considerthe current
the effects of multiple endpoints
conduct (Case C). trials
of clinical With with k small
equalsample sizes. This
to 10 endpoints and the prediction limit de
report assesses ned as literature
the published the on various
18th largest control measurement, the overall con -
dence level for the experiment is reduced
Bayesian techniques toas68 per-
cent. To counteract this effect (Case D), one can add a
second space mission (p = 2),astronauts)
each with ni equal measuring
by sequentially to ve whether the
astronauts, and the overall con dence is increased
back to 96 percent. If one had instead
speci cationsconsidered p
as opposed to incremental
improvement.
equal to four space missions (Case E), a con dence of
94 percent would be achieved by setting the prediction
limit to the 15th largest control measurement, again
increasing the statistical power of the decision rule.
How does one select from among the various strate-

gies described in the simple example described above?
The answer is to select the strategy that has reason-
×
able con dence (i.e., a low rate of false-positive re-
Small Clinical sults,
Trials:e.g.,
5 percent) and that has the maximum statis-
tical power for a desired effect size. To this end, one
can evaluate the power of the test to detect a true dif-
ference between the control Clinical
grouptrials
andarethe
usedexperi-
mental group 
by simulation. appropriate preventive, diagnostic, or treatment

analyses.” Meta-analysis is widely
Login or Register to used einectiveness
education (seeintervention
of a given Box 3-7),when psychol-
one really is
save!
ogy, and the medical present. This may occur
sciences (e.g., in evidence-based in phase I (safety
medicine) and and has
frequently been used to study the ef cacies of different
and III (extensive treatments
assessment of safety and e cacy)
(Hedges and Olkin,Download
1985). Free PDF trials. Although phase I and II studies may have smaller
A meta-analysis can summarize an entire set of research in the literature, a
sample from a large population of studies, have or somestatistical
adequate de ned subset
power, of stud-
statistical power being
ies (e.g., published studies or n-of-1 studies). The degree to which the re-
sults of a synthesis can be generalized depends in part on the nature of the
set of studies. In general, meta-analysisSmall Clinical Trials assesses the current
serves as a useful tool to answer
questions for which single trials were underpowered or with
conduct of clinical trials notsmall
designed to This
sample sizes.
address. More speci cally, the following report
areassesses
bene the published
ts of literature on various
meta-analysis:
It can provide a way to combine the results
Bayesian of studies
techniques with
as in the different
con dence pro le
designs (within reason) when similar research questions are of assessing
method and (2) other alternatives such as
interest. astronauts) by sequentially measuring whether the
It uses a common outcome metricprobability
when studies vary and
outcome range in the ways
meeting in
predesigned
which outcomes are measured.
improvement.
It accounts for differences in precision, typically by weighting in pro-

portion to sample size.
Its indices are based on suf cient statistics.
It can examine between-study differences in results (heterogeneity).
×
It can examine the relationship of study outcomes to study features (Beck-
Small
er,Clinical
2000). Trials: Issues and Challenges
TABLE 3-2 Key Points in the Conduct of Meta-Analyses
Clinicaloften
Systematic reviews of study ndings trials are
useused to elucidatestatistical
complex the most

methods to synthesize and interpret
options data from individual
for individuals studies,
with a given medical condition.
Perhaps the most essential feature
and an understanding of their basic principles is important in inter- of a clinical trial is
preting their results.
Buy Ebook | $39.99
Quantitative synthesis cannot replace
treatment. sound
Sampleclinical reasoning;
size is a crucial component of any
combining
MyNAP poor-quality
members save or overly biased data that do not make
10%
sense is online.
likely to produce unreliable results.
save!
When appropriate, combining datapresent.
from Thisvarious
may occurstudies
in phase Ito
(safety
obtainand a
common estimate can increase the and statistical power for
III (extensive assessment the discov-
Download
ery of treatment Freeand
ef cacy PDFcantrials. Althoughthe
increase phase I and II studies
precision of may
the have smaller
estimate. power, which is the committee's de nition of a "large"
Sensitivity analyses should be performed
have adequate tostatistical
determine power,the robust-
statistical power being
ness of conclusions. the hypothesis is false.
Patients, clinical settings, and treatment

Small Clinicalresponses
Trials assessesare
the expected
current to
vary across trials that have studied the same problem. Insight into
reasons for the heterogeneity of report
trial assesses
resultsthe may oftenliterature
published be as im-
on various
portant as or even more important than producing aggregate
results. Bayesian techniques as in the con dence pro le
SOURCE: Lau, Ioannidis, and Schmid therapeutic results in a single treated population (e.g.,
(1997).
Page 83 improvement.
BOX 3-7
Combining n-of-1 Studies in Meta-
Analysis: Results from Research in Special
Education
×
Researchers in special education are often concerned
Small Clinical with
Trials:individualized
treatments for behavior disorders
or with low-incidence disabilities and disorders. Sin-
gle-case research designs are quite common. Study
designs typically involve a baseline period
Clinical trials are usedfollowed bythe most
to elucidate
a treatment 
Buyperiod
Paperback | $50.00
appropriate
and possibly follow-up.preventive, diagnostic, or treatment
Multiple
measures are usually obtained fromthe
Perhaps each
most case during
essential feature of a clinical trial is
the baseline and treatment. Crossover
that it aims to treatment
use results based de-on a limited sample of
signs areBuy 
Ebook | used
occasionally $39.99research participants to see if the intervention is safe
and usually involve only no
baseline-treatment cycles. Meta-analysis
treatment. Samplehas been
size is ap-
plied
MyNAPsincemembers
the 1980s to summarize these case-study
save
10% online.
designs. considerable risk of failing to demonstrate the
save! the methods proposed
However, present. This may
have beenoccur in phase I (safety and
controver-
sial and the statistical properties
and III of the methods
(extensive assessmenthave
not beenDownload
rigorouslyFree PDF Three
studied. trials. Although phase I andhave
approaches II studies may have smaller
been used to measure effects. power, which is the committee's de nition of a "large"
(1) Single-case effect size.have
Some researchers
adequate have
used an index similar tothe the effect size, computed
for n > 1 studies as the standardized difference
between group means: methodologies and the appropriate situations for the
~ enlarge ~ astronauts) by sequentially measuring whether the
YÌ„treatment is the subject's mean
speci score during
cations as opposed treat-
to incremental
improvement.
ment, YÌ„baseline is the mean before treatment, and SY
pooled is obtained by pooling intrasubject variation
across the two time periods.
(2) Percentage of nonoverlapping data index (PND)

(Scruggs, Mastropieri, and Castro, 1987). The per-
centage of nonoverlapping data index is also
based on the idea of examining the data from the
×
baseline and treatment periods of the case study.
TheIssues
indexand Challenges
is the percentage of datum values ob-
served during treatment that exceed the highest
baseline data value.
(3) Regression approaches (Center, Skiba, diagnostic,
and
appropriate
options for
preventive,
individuals
Casey, 1986). The researcher estimates the treat- with a given
or treatment
medical condition.
ment effect and separate effects
that it aims toofusetime during
baseline (t = 1 to n ) and treatment (t = n to ifn)
 Buy Ebook | $39.99
a research participants to
a see the intervention is safe
phases via treatment. Sample size is a crucial component of any
MyNAP members save
Yi = b0 + b1Xi + b2t + b3Xi (t − nparticipants
a) + et is more prone to variability and carries a
The effects of interest, say, b1present.
for X Thisor bmay3 for X (t), are
save! occur in phase I (safety and
then evaluated via incremental F tests, which
pharmacologic pro les),are
II (pilot e cacy evaluation),
transformed and summarized.
A relevant question is: when does a meta-analysis of small
have adequate statistical studies
power, rule
statistical outbeing
power
the need for a large trial? One investigation showed that the results of
smaller trials are usually compatible with the results of larger trials, al-
though large studies may produce a more precise answer to a particular
question when the treatment effect is not large
conduct but is
of clinical clinically
trials important
with small sample sizes. This
(Cappelleri, Ioannidis, Schmid, et al., 1996). When the small studies are
replicates of each other—as, for example, in collaborative
disparate information fromlaboratory
several studiesor
including
clinical studies Bayesian techniques as in the con dence pro le
Page 84 speci cations as opposed to incremental
improvement.
or when there has been a concerted effort to corroborate a single small

study that has produced an unexpected result—a meta-analysis may be
conclusive if the combined statistical power is suf cient. Even when small
studies are replicates of one another, however, the population to which
they refer may be very narrow. In addition, when the small studies differ
too much, the populations may be too broad to be of much use (Flournoy
and Olkin, 1995). Some have suggested that the use of meta-analysis to
×
predict the results of future studies is important but would require a de-
Small Clinical
sign formatTrials: Issues and
not currently Challenges
used (Flournoy and Olkin, 1995).
Meta-analysis involves the designation of an effect size and a method of

analysis. In the case of proportions, some of the effect sizes used are risk
differences, risk ratios, odds ratios, number needed to treat, variance-sta-

appropriate
options for
preventive,
individuals
diagnostic,
with
bilized risk differences, and differences between expected and observed a given
or treatment
medical condition.
outcomes. For continuous outcomes, the thatstandardized mean
it aims to use results baseddifference or of
on a limited sample
correlations are common measures. The technical aspects
to see of these proce-

Buy Ebook | $39.99 research participants if the intervention
is safe
dures have been developed by Hedges and Olkin (1985).

MyNAP members save
Meta-analysis sometimes refers to the participants
entire process of synthesizing
is more prone to variability andthe
carries a
results of independent studies, including the collection of studies, coding,
abstracting, and so
save! on, as well as the statistical
present. Thisanalysis.
may occur inHowever, some
phase I (safety and re-
searchers use the term to refer to only pharmacologic
the statistical pro les), II (pilot ewhich
portion, cacy evaluation),
in-
cludes methods such as the analysis
Download Free PDF of trials.
variance,
Althoughregression, Bayesian
phase I and II studies may have smaller
analysis and multivariate analysis. The con dence pro le method (CPM), have adequate statistical
another form of meta-analysis (Eddy, Hasselblad,
trial. Sometimesand a trialShacter, 1992) adopts
with eight participants may
the rst de nition of meta-analysis and attempts to deal with all the issuesbeing
in the process, such as alternative designs, outcomes, and biases, as well as
the statistical analysis, which is Bayesian. Methods of analysis used for
CPM include analysis of variance, regression, nonparametric analysis, and
Bayesian analysis. The CPM analysis approach
conduct ofdiffers from
clinical trials with other meta-
small sample sizes. This
analysis techniques based on classical statistics in that it provides marginal
probability distributions for the parameters
disparate ofinformation
interestfrom andseveral
if an studies
integratedincluding
Bayesian techniques
approach is used, a joint probability distribution for allasthe in the con dence pro le
parameters.
More common meta-analysis procedures provide
therapeutic a point
results estimate
in a single for one(e.g.,
treated population
or more effect sizes together with con astronauts) by sequentially
dence intervals formeasuring whether the
the estimates.
Although exact con dence intervals can be obtained
probability outcomeusingrange and numerical inte-
meeting predesigned
gration, large sample approximations often provide
speci cations suf ciently
as opposed accurate
to incremental
improvement.
results even when the sample sizes are small.
Some have suggested that those who use meta-analysis should go beyond
the point estimates and con dence intervals that represent the aggregate
ndings of a meta-analysis and look carefully at the studies that were in-
cluded to evaluate the consistency of their results. When the results are
×
Page 85
largely on the same side of the “no-difference” line, one may have more
con dence in the results of a meta-analysis (LeLorier, Gregoire, Benhad-
dad, et al., 1997).

Sometimes small studies (including n-of-1
optionsstudies) are omitted
for individuals with a givenfrom medicalmeta-
condition.
analyses (Sandborn, McLeod, and Jewell, 1999). Others, however, view trial is
Perhaps the most essential feature of a clinical
meta-analysis as a remedy or as a means to increase power
to see ifrelative to the

Buy Ebook | $39.99 research participants
power of individual small studies in a research domain (Kleiber and Harper,
and e ective or if it is comparable
the intervention
to a comparison
is safe

1999). Because those who perform meta-analyses typically weight the re-
MyNAP members save
sults in proportion to sample size, smallparticipants
sample issizes more pronehavetoless of anand
variability effect
carries a
on the results than larger ones. A synthesis based mainly on small sample
sizes will produce
save! summary results with moreThis
present. uncertainty
may occur in phase (larger standard
I (safety and
pharmacologic pro
errors and wider con dence intervals) than a synthesis based on studies les), II (pilot e cacy evaluation),
with larger sample Download
sizes. Thus, a cumulative meta-analysis
Free PDF trials. Although phase I and IIrequires
studies mayahave stop- smaller
ping procedure that allows one to say that a treatment is or is not effective have adequate statistical
(Olkin, 1996). trial. Sometimes a trial with eight participants may
When the combined trials are a homogeneous setofdesigned
the probability rejecting theto nullanswer
hypothesis thewhen
same question for the same population, the use of a xed-effects model, in
which the estimated treatment effects Small
varyClinical
across studies
Trials assessesonly as a result
the current
of random error, is appropriate (Lau, Ioannidis, and Schmid, 1998). To as-
sess homogeneity, heterogeneity is often tested
report onthe
assesses the basis literature
published of the chi-
on various
square distribution, although this lacks power. If heterogeneity is detected,
the traditional approach is to abort theBayesian
meta-analysis
techniques as orintotheuse
con random-
dence pro le
method and (2) other alternatives
effects models. Random-effects models assume that no single treatment such as assessing
effect exists, but each study has a different truebyeffect,
astronauts) with
sequentially all treatment
effects derived from a population of such truths assumed to followpreestablished
intervention is falling above or below a a nor-
mal distribution (Lau, Ioannidis, and Schmid, 1998) (see section on
Hiearchical Models and Appendix A). Neither
improvement.xed-effects nor random-ef-
fects models are entirely satisfactory because they either oversimplify or
fail to explain heterogeneity. Meta-regressions of effect sizes affected by
control rates have been used to develop reasons for observed heterogene-
ity and to attempt to identify signi cant relations between the treatment
effect and the covariates of interest; however, a signi cant association in
regression analysis does not prove causality. Because heterogeneity can be
a problem in the interpretation of a meta-analysis, an empirical study (En-
×
gels, Terrin, Barza, et al., 2000) showed that, in general, random-effects
Small Clinical
models forTrials: Issuesand
odds ratios andrisk
Challenges
differences yielded similar results. The
same was true for xed-effects models. Random-effects models were more
conservative both for risk differences and for odds ratios. When studies are
homogeneous it appears that there is consistency
Clinical trials areof results
used when
to elucidate risk dif-
the most
ferences or odds ratios 
Buy Paperback
are used and| consistency
$50.00
of re- diagnostic, or treatment

Page 86 treatment. Sample size is a crucial component of any
MyNAP members save
suits when random-effects
10% online. or xed-effects models
considerable risk ofare used.
failing Differences
to demonstrate the
Login or Register to
appear when heterogeneity is present (Engels, Terrin, Barza, et al., 2000).really is
e ectiveness of a given intervention when one
The use of an individual subject's data rather than summary data from each
study can circumvent ecological
Download Free fallacies.
PDF trials.SuchAlthoughanalyses
phase I and can provide
II studies maxi-
may have smaller
mum information about covariates to which heterogeneity can be ascribed
and allow for a time-to-event analysis (Lau, Ioannidis,
trial. Sometimes a trialand Schmid,
with eight 1998).
participants may
Like large-scale clinical trials, meta-analyses cannot always show how in- being
dividuals should be treated, even if theytheare usefulisfor
hypothesis false.estimation of a pop-
ulation effect. Patients may respond differently to a treatment. To address
this diversity, meta-analysis can rely onmethodologies
response-surface modelssituations
and the appropriate to sum-for the
marize evidence along multiple covariates of interest. A reliable meta- This
conduct of clinical trials with small sample sizes.
analysis requires consistent, high-quality reporting
strategies such as of the primary
(1) meta-analysis data
to combine
from individual studies. disparate information from several studies including
Meta-analysis is a retrospective analytical method, the results of which will
be based primarily on the rigor of the technique
astronauts) by(the trial designs)
sequentially and the
quality of the trials being pooled. Cumulative meta-analysis can help deter-
mine when additional studies are needed speciand canasimprove
cations opposed tothe predictabili-
incremental
improvement.
ty of previous small trials (Villar, Carroli, and Belizan, 1995). Several work-
shops have produced a set of guidelines for the reporting of meta-analysis
of randomized clinical trials (the Quality of Reporting of Meta-Analysis
group statement [Moher, Cook, Eastwood, et al., 1999], the Consolidated
Standard of Reporting Trials conference statement [Begg, Cho, Eastwood,
et al., 1996], and the Meta-Analysis of Observational Studies in Epidemiolo-
gy group statement on meta-analysis of observational studies [Stroup,
Berlin, Morton, et al., 2000]).
×
RISK-BASED ALLOCATION
Empirical Bayes methods are needed for analysis of experiments with risk-
based allocation for two reasons. First, the natural heterogeneity from sub-
ject to subject requires some accounting for trials
Clinical randomare used effects; and
to elucidate thesecond,
most

Buy Paperback
the differential selection | $50.00
of groups due appropriate preventive, diagnostic,
to the risk-based allocation or treatment
is han-
dled perfectly by the “u-v” method introduced
Perhaps theby mostHerbert
essential E. Robbins.
feature The
of a clinical trial is
u-v method of estimation capitalizes onthat it aims togeneral
certain use resultsproperties
of dis-
tributions such as the 
BuyPoisson
Ebook or| $39.99research participants to see if the intervention is safe
normal
anddistribution
e ective or if it isthat hold under
comparable arbi-
to a comparison
trary and unknown mixtures of parameters, treatment.thus allowing
Sample size is a for the
crucial existence
component of any
MyNAPAt
of random effects. members
the same save
time, the u-v method allows estimation of
averages under10%a online.
wide family of restrictions on the
considerable risk sample
of failing tospace, such
demonstrate theas
restriction to high-risk or low-risk subjects, thus addressing the risk-based
alloca- pharmacologic pro les), II (pilot e cacy evaluation),
Page 87 have adequate statistical power, statistical power being
tion design the hypothesisare
feature. These ideas and approaches is false.
considered in greater
detail in Appendix A. Small Clinical Trials assesses the current
Another example from Finkelstein, Levin, conduct
andofRobbins
clinical trials with small
(1996b) samplein
given sizes.
Box This
3-8 illustrates the application of risk-based allocation
strategies such as (1) to a trial studying
meta-analysis to combine
the occurrence of opportunistic infections in very sick AIDS patients. including
disparate information from several studies This
example was taken from an actual randomized
method andtrial,
(2) otherACTG Protocol
alternatives such as002,
assessing
which tested the ef cacy of low-dose versus high-dose zidovudine (AZT). (e.g.,
therapeutic results in a single treated population
Survival time was the primary endpointintervention
of the clinical trial, but for the pur-
pose of illustrating risk-based allocation, Finkelstein
probability outcomeand rangecolleagues fo-
cused on the secondary endpoint of opportunistic infections. They studied
improvement.
the rate of such infections per year of follow-up time with an experimental
low dose of AZT that they hoped was better tolerated by patients and
which would thereby improve the therapeutic ef cacy of the treatment.
SUMMARY
Because the choice of a study design for a small clinical trial is constrained
by size, the power and effectiveness of such studies may be diminished,
×
but these need not be completely lost. Small clinical trials frequently need
Small
toClinical Trials:
be viewed Issues
as part of aand Challenges
process of continuing data collection; thus, the
objectives of a small clinical trial should be understood in that context. For
example, a small clinical trial often guides the design of a subsequent trial.
Therefore, a key question will be what information from
Clinical trials are used to the current
elucidate the most trial
will be of greatest value 
Buy Paperback
in designing| the
$50.00
appropriate
next one? preventive, diagnostic,
In small or treatment
clinical trials of
drugs, for example, the most importantPerhapsresultthe might be to provide
most essential feature of ainforma-
clinical trial is
tion on the type of postmarketing surveillance
that it aims that
to useshould follow.

Buy Ebook | $39.99
A major fundamental question is the qualitatively different goals that one
might have when studying very few people. clinical The
trial. Amain example
trial with hereofisresearch
a small number de-
MyNAP members save
termination of best treatment that participants
theonline.
10% allows astronauts is more prone to variability and carries a
to avoid bone
mineral density loss.
Login Such research
or Register to could have many
e ectiveness goals.
of a given One goal
intervention when would
one really is
be to providesave! present. Thisthat
information on this phenomenon may occur
is mostin phase I (safety
likely and
to be
correct in some universal sense; that is,and the knowledge
III (extensive and estimates
assessment of safety and earecacy)
as unbiased and as Download
precise asFree PDF trials.
possible. A second Althoughgoal phase I and II studies
might be to maytreathave
thesmaller
most astronauts in the manner that was most
power, likely
which is theto be optimal
committee's for of
de nition each
a "large"
individual. These are profoundly different trial. goals
Sometimes that would
a trial have
with eight to be both
participants may
articulated and discussed before any trial designs could be considered.
One can nd the components of a goal the discussion
hypothesis isin some of the descrip-
false.
tions of individual designs, but the discussion of goals is not identi ed as

part of a methodologies and the appropriate situations for the
Page 88 method and (2) other alternatives such as assessing
BOX 3-8 probability outcome range and meeting predesigned
Illustration of a Clinical Trial on improvement.
Opportunistic Infections Using Risk-Based

Allocation Analysis
A total of 512 subjects were randomly assigned in ACTG
Protacol 002 to evaluate the therapeutic ef cacy of an
experimental low dose of AZT (500 mg/day) versus the
high dose of AZT (1,500 mg/day) that was the standard
×
dose at the time of the trial. A total of 254 patients
Small Clinical were
Trials:randomized
Issues andto Challenges
the low-dose experimental group
and 258 were randomized to the high-dose standard
treatment group. Although all patients in the original
trial were seriously immunode cient,
Clinical theused
trials are focus here is
on the treatment 
effect amongappropriate
the subgrouppreventive,of diagnostic,
253 pa- or treatment
tients at highest risk, de nedPerhaps
as those with
the most initialfeature
essential CD4- of a clinical trial is
cell counts less than or equalthatto it60 aimsper microliter
to use results based ofon a limited sample of
blood. InBuy 
this Ebook | $39.99
subgroup,
the number of opportunistic
infections among the 125 patientstreatment. in the
Sample high-dose
size is a crucial component of any
(standard) treatment
MyNAP members group was observed to be 296
save
with10%
66,online.
186 days of follow-up, for an opportunistic
considerable in-
fection rate of 1.632 per year. Among the 128 high-risk
patients randomized to the low-dose
pharmacologic (experimental)
pro les), II (pilot e cacy evaluation),
and III (extensive
treatment group, the number of opportunistic assessment of safety and e cacy)
infec-
tions was 262 with 75,591 days of follow-up,
sample for have
sizes, they usually an op-adequate statistical
portunistic infection rate of 1.265
power, per whichyear. The ratiodeofnition of a "large"
is the committee's
the rate for the control grouphave to adequate
the treatment group
is 1.632/ 1.265 = 1.290, with athestandard
probabilityerror of ±0.109.
of rejecting the null hypothesis when
Thus, the “gold standard” (randomized) estimate of the
low-dose effect on the high-risk patients
Small Clinical Trialsis that itthere-
assesses current
duces their rate of opportunistic infections by about
22.5 percent (1−1/1.290 = 0.225) reportrelative
assessesto thethat for literature
published the on various
higher-dose group.
If the trial had used risk-based allocation
method with
and (2) other all of the
alternatives such as assessing
high-risk patients receiving the experimental low dose population (e.g.,
therapeutic results in a single treated
and all of the lower-risk patients (withisCD4-cell
intervention falling abovecounts
>60 per microliter) receiving probability
the standard
outcomehigh rangedose,
the effect of the standard dose on the high-risk pa-
improvement.
tients would have been estimated instead of being di-
rectly observed. This is done by rst tting a model for
the rate of opportunistic infections under standard
treatment with the data for the lower-risk patients.
Previous data suggest that the annual rate of oppor-
tunistic infection under the standard dose can be
modeled by the exponential function R(X) = A exp (BX),
where X is the CD4-cell count per microliter at the
×
start of the trial, and A and B are constants to be esti-
Small Clinical mated
Trials: Issues
from theand Challenges
trial data for the lower-risk patients.
The rate R(X) is the expected number of opportunistic
infections per year of survival per patient for those
with initial CD4-cell count X.Clinical
A Poisson trials areregression
used to elucidate the most
model was Buyused,
Paperback | $50.00
which assumesappropriate
that the preventive,
number diagnostic,
of or treatment
opportunistic infections occurring
Perhaps the in amostgiven time
essential pe- of a clinical trial is
feature
riod t under standard treatment that ithasaimsatoPoisson
use resultsdistrib-
ution withBuy
mean 
Ebook | $39.99
tR(X).
By using the data for the 133
lower-risk patients who received treatment. theSample
standard size is dose,
theMyNAP
maximum-likelihood
members save estimates of the model para-
10% online.
meters are A = 0.541 and B = −0.00155.
considerable risk The model
of failing esti-
to demonstrate the
mates that with a CD4-cell count of, for example, 60
per microliter, the opportunistic infection
pharmacologic pro rate
les), II would
(pilot e cacy evaluation),
and III (extensive
be 1.452 per year, whereas with a CD4-cell count of assessment of 10
safety and e cacy)
per microliter it would be 1.526 samplepersizes,year. theyThe
usuallytotol
have ex-
adequate statistical
pected number of opportunistic power, infections for the de nition of a "large"
which is the committee's
high-risk patients under standard treatment is the sum
of the model expectations over all 128 high-risk
the probability of rejecting the pa-null hypothesis when
tients (who in fact received the low dose). That sum is
340.46, whereas the actual numberSmall Clinical is 262. The esti-
Trials assesses the current
mated rate ratio among the high-risk patients is thus
340.46/262 = 1.2995 (with a standard
report assesses error of approxi-
the published literature on various
mately ±0.147 after adjusting for overdispersion). Under
risk-based allocation, then, the estimated
Bayesian techniques low-dose
as in the con ef-dence pro le
fect on the high-risk patients is a reduction in the rate as assessing
method and (2) other alternatives such
of opportunistic infection of astronauts)
23.0 percent (1 − 1/1.2995
= 0.230), close to the randomized estimate of 22.5orper-
intervention is falling above below a preestablished
cent. In the estimation of thespeciratecations
ratio,asthe use of risk-
opposed to incremental
based allocation and an appropriate
improvement. model generated
results that are virtually indistinguishable from those
generated by the randomized clinical trial.
SOURCE: Finkelstein, Levin, and Robbins (1996b).
×
Page 89
conceptual framework that would go into choosing which class of trials to
be used.
It is quite likely that there could be veryClinical trials are used

substantial to elucidate the about
disagreement most

those goals. The rst might lead one tooptions
include every subject
for individuals in medical
with a given a de ned condition.
Perhaps the most essential
time period (e.g., 10 missions) in one grand experimental protocol. The sec- feature of a clinical trial is
ond might lead one to identify a subgroup of individuals who would be the

Buy Ebook | $39.99 research
e
participants to see
initial experimental subjects and whose results would be applied to the re-
and ective or if it is
if the
comparable
intervention
to a comparison
is safe

mainder of the subjects. On the other hand, it might lead to a series of in-
MyNAP members save
tensive metabolic studies for each individual,
participantsincluding,
is more prone perhaps, n-of-1
to variability and carries a
type trials, which might be best for the individualization of therapy but not
for the production
save! of generalizable knowledge.
present. This may occur in phase I (safety and
and III (extensive
Situations may arise in which it is impossible assessment
to answer of safety and
a question e cacy)
with any
con dence. In those cases, the best that onesizes,
sample cantheydo usually
is usehavetheadequate
information
statistical
to develop new research questions. In other cases,
power, which it committee's
is the may be necessary
de nition of ato
"large"
answer the question as best as possiblehave
because a major, possibly irre-
adequate statistical power, statistical power being
versible decision must be made. In thosethe cases, multiple,
probability of rejectingcorroborative
the null hypothesis when
analyses might boost con dence in the ndings.
RECOMMENDATIONS methodologies and the appropriate situations for the

Early consideration of possible statistical analyses
strategies such as should be an integral
(1) meta-analysis to combine
part of the study design. Once the data are collected, alternative statistical
analyses should be used to bolster con method
dence andin(2)the
other interpretation
alternatives such asofassessing
re-
sults. For example, if one is performing a Bayesian analysis, a non-Bayesian(e.g.,
therapeutic results in a single treated population
analysis should also be performed, and intervention
vice versa; similar
is falling abovecross-validation
of other techniques should also be considered.
improvement.
RECOMMENDATION: Perform corroborative statistical analyses. Given
the greater uncertainties inherent in small clinical trials, several alter-
native statistical analyses should be performed to evaluate the consis-
tency and robustness of the results of a small clinical trial.
The use of alternative statistical analyses might help identify the more sen-
sitive variables and the key interactions in applying heterogeneous results
across trials or in trying to make generalizations across trials. In small clin-
ical trials, more so than in large clinical trials, one must be particularly
cautious about recognizing individual variability among subjects in terms
of their biology and health care preferences, and administrative variability
in terms of what can be done from one setting to another. The diminished
power of studies with small sample sizes might mean that the
generalizability
Page 90
of the ndings might not be a possibility in the short-term, if at all. Thus,

caution should be exercised in the interpretation of the results from small
clinical trials.
RECOMMENDATION: Exercise caution in interpretation. One should ex-

ercise caution in the interpretation of the results of small clinical trials
before attempting to extrapolate or generalize those results.
The National Academies of Sciences, Engineering, and Medicine

500 Fifth St., NW | Washington, DC 20001
© 2019 National Academy of Sciences. All rights reserved.

3 Statistical Approaches To Analysis of Small Clinical Trials - Small Clinical Trials - Issues and Challenges - The National Academies Press

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

3 Statistical Approaches To Analysis of Small Clinical Trials - Small Clinical Trials - Issues and Challenges - The National Academies Press

Uploaded by

Copyright:

Available Formats

26/03/2019 3 Statistical Approaches to Analysis of Small Clinical Trials | Small Clinical Trials: Issues and Challenges | The National

| The National Academies Press

Small Clinical Trials: Issues and Challenges

Small Clinical Trials: Issues and Challenges (2001)

 Buy Paperback | $50.00

Clinical trials are used to elucidate the most

Page 61  Buy Paperback | $50.00

Small Clinical Trials assesses the current

Small Clinical Trials assesses the current

How many patients must be enrolled per cohort to ob-

Small Clinical Trials assesses the current

On the basis of these results, should 73 subjects be re-

 Buy Ebook | $39.99 research participants to see if the intervention is safe

test either 19 or 73 patients (depending on the desired

Small Clinical Trials: Issues and Challenges

The use of study stopping

Small Clinical Trials: Issues and Challenges

 Buy Paperback | $50.00

 Buy Ebook | $39.99research participants to see if the intervention is safe

Small Clinical Trials assesses the current

SOURCE: Feiveson (2000).

Clinical trials are used to elucidate the most

 Buy Paperback | $50.00appropriate preventive, diagnostic, or treatment

Hedayat and co-workers showed that the problem can

maxN1≤k−1 Prob[X ≤ m | N1] ≤ Î±,

For example, assume that the total number of units

developed, but it is not known whether it is equally ef-

A common theme in medical research is two-stage sampling, that is, sam-

withdrawal from the Buy

Clinical trials are used to elucidate the most

regression with randomly dispersed parameters (Rosenberg, 1973);

exchangeability between multiple regressions (Lindley and Smith,

two-stage stochastic regression, (Fearn, 1975);

James-Stein estimation (James and Stein, 1961);

random coef cient models (DeLeeuw and Kreft, 1986);

Small Clinical Trials assesses the current

Clinical trials are used to elucidate the most

subjective distribution. treatment. Sample size is a crucial component of any

Sequential analysis. Because the posterior distribution can be updat-

As mentioned in Chapter 2, one can use decision analysis as an aid in the

Establish a time horizon

Structure the decision tree (choices, chances, outcomes)

Assess the probabilities

Assess the utilities

Interpret the results

a long space mission accelerates

Determine the expectation by evaluating a decision

ef cacy of special program (e) = 0.3,

 Buy Ebook | $39.99 research participants to see if the intervention is safe

one can assign values at the chance nodes to pick the

Small Clinical Trials: Issues and Challenges

 Buy Paperback | $50.00

 Buy Ebook | $39.99research participants to see if the intervention is safe

Small Clinical Trials: Issues and Challenges

Small Clinical Trials assesses the current

Small Clinical Trials: Issues and Challenges

 Buy Paperback | $50.00

 Buy Ebook | $39.99research participants to see if the intervention is safe

Small Clinical Trials assesses the current

decision analysis, with a sequential clinical trials approach potentially of-

Decision analytical models are most powerfully used to answer the

distributions. treatment. Sample size is a crucial component of any

 research participants to see if the intervention is safe