Bayesian Decision Procedures For Binary and Continuous Bivariate Dose-Escalation Studies

PHARMACEUTICAL STATISTICS
Pharmaceut. Statist. 2006; 5: 125133

Published online in Wiley InterScience
(www.interscience.wiley.com) DOI: 10.1002/pst.222
Bayesian decision procedures for binary

and continuous bivariate dose-escalation
studies
Yinghui Zhou1,*,y, John Whitehead1, Elisa Bonvini2 and John
W. Stevens3
1
Medical and Pharmaceutical Statistics Research Unit, The University of Reading, UK

Section of Applied Statistics, School of Biological Sciences, The University of Reading,
UK
3
AstraZeneca R&D Charnwood, Loughborough, UK
2
In this paper, Bayesian decision procedures are developed for dose-escalation studies based on binary
measures of undesirable events and continuous measures of therapeutic benet. The methods
generalize earlier approaches where undesirable events and therapeutic benet are both binary. A
logistic regression model is used to model the binary responses, while a linear regression model is used
to model the continuous responses. Prior distributions for the unknown model parameters are
suggested. A gain function is discussed and an optional safety constraint is included. Copyright #
2006 John Wiley & Sons, Ltd.
Keywords: Bayesian decision procedure; dose escalation; logistic regression; linear regression;
phase I clinical trial
1. INTRODUCTION
In phase I clinical trials, most formal doseescalation procedures have been developed in the
context of oncology [15]. In that setting, the
objective is usually to determine the maximum
tolerated dose (MTD). Conventionally, the MTD
*Correspondence to: Y. Zhou, Medical and Pharmaceutical
Statistics Research Unit, The University of Reading, PO Box
240, Earley Gate, Reading RG6 6FN, UK.
y
E-mail: y.zhou@reading.ac.uk
Copyright # 2006 John Wiley & Sons, Ltd.
is dened as a statistic arising from the trial results,

such as the highest dose administered for which no
more than one subject in six experienced toxicity.
In statistical accounts of dose-escalation studies,
the MTD is typically taken to be the dose at which
the probability of toxicity reaches some specied
value, such as 0.20 or 0.33. Models describing the
doseresponse relationship are usually monotonically increasing as doses increase. The Continual
Reassessment Method of OQuigley et al. [6] uses a
one-parameter model; essentially xing the slope
126 Y. Zhou et al.
of a regression line, while the Bayesian decisiontheoretic approach [7] uses a two-parameter
regression model.
Conventional phase I cancer studies are designed to assess drugs which kill cancer cells. These
are called cytotoxic agents and they have high
toxicity rates. Data concerning ecacy are not
obtained at the same time as the observation of
toxicities in this type of study. Recently, attention
has been moving away from cytotoxic drugs to less
toxic therapy: cytostatic drugs [8]. These agents do
not kill cancer cells. They work by blocking
receptors on the surface of the cancer cells. When
these receptors are triggered, the cell begins to
divide into two new cells. So, by blocking them,
the treatment is stopping the cancer from growing.
It is more important in assessing cytostatic therapy
to take into account the potential benet of
treatments along with the undesirable outcomes.
A desirable outcome (DO) in that context might be
a benecial therapeutic response, or more likely a
surrogate or a biomarker normally associated with
eventual benet. Also, in therapeutic areas other
than cancer, where dose escalation is typically
conducted in healthy volunteers, the bivariate
situation is more common, and a DO might be
some pharmacodynamic measure believed to
reect the intended mechanism of action of the
drug.
The literature relating to dose-escalation designs
for bivariate responses has been limited until very
recently. In recent years, Gooley et al. [9], Thall
and Russell [10], OQuigley et al. [11], Ishizuka
and Ohashi [12], Braun [13] and Loke et al. [14]
have all investigated bivariate outcome models.
Both desirable and undesirable responses were
limited to binary outcomes in these papers, and the
applications discussed were mainly to bone marrow transplantation. Most of these methods were
intensively discussed in [15], where a Bayesian
approach based on logistic regression models is
presented.
Whitehead et al. [15] considered dose-escalation
studies in which two responses are observed on
each subject. One of these was referred to as a dose
limiting event (DLE), and the other as a DO. A
DLE is an event that would give the investigators,
or the safety committee responsible for conduct of

the study, cause for concern and indicate that it
may be unwise to proceed to a higher dose.
Toxicities are included in this category. A Bayesian decision approach was proposed to determine
a therapeutic window of acceptable doses that
have a suciently large probability of a DO
accompanied by a small enough probability
of a DLE. An example was given based on a
dose-escalation study of an experimental antifactor Xa compound in healthy volunteers. Such
drugs have potential benet in the treatment of
thromboembolic disorders. A DLE could be
dened as a clinically important increase in the
clotting factors APPT or PT. Anti-factor Xa
activity measurements in plasma can be used as
potential therapeutic benet. Any recorded adverse event was taken to constitute a DLE. The
original continuous measures of benet were
dichotomised in order to t in with their method.
A further investigation [16] used intensive simulations to evaluate the Bayesian approach, based on
features of a another recently completed phase I
study, this time of a compound with potential
benet to patients suering from inammatory
diseases of the lung. No specic adverse events
were anticipated, but all were monitored and any
might count as a DLE. Potential therapeutic
benet was assessed indirectly using an ex vivo
assay that measured the inhibition of leukotriene
B4 (LTB4). Inhibition of LTB4, indicates a reduction in the production of the white blood cells
responsible for inammation, and this continuous
measure can be taken as the DO: in [16] it was
dichotomized depending on whether subjects
achieved 90% inhibition of LTB4 over 24 h.
Simulation results demonstrated that the Bayesian
designs with pessimistic priors provide safe
dose-escalation and accurate estimates of the
model.
In many situations, benet will be continuous
rather than binary, whereas it will often be
appropriate to represent a DLE as a binary
outcome. Therefore, in this paper, we extend the
Bayesian approach of Whitehead et al. [15,16] so
that the bivariate outcomes are binary DLEs and
continuous DOs. In doing so, we introduce an

DOI: 10.1002/pst
Binary and continuous bivariate bayesian dose-escalation approach
allowance for the correlations between repeated

observations of DOs on the same subject, although
we have not been able to do the same for the
binary DLEs. This point will be discussed again in
Section 6. The logistic regression model and the
linear mixed regression model that underlie the
method of this paper are described in the next
section. The latter model is based on the approach
developed in [1719] in connection with doseescalation studies in which subjects are repeatedly
dosed, and pharmacokinetic measurements relating to the pattern of the concentration of the drug
in plasma during the hours following its administration are observed. The linear mixed regression
model used here is based on [19], where renements of the approach taken in the previous
papers were included. Prior distributions imposed
on the unknown model parameters are described
in Section 3. The gain function and constraints on
dose escalation are discussed in Section 4 and a
simulated example given in Section 5.
2. REGRESSION MODELS FOR

BIVARIATE RESPONSES
The subjects in this trial will be referred to as Si,
i 1; . . . ; n and the dosing periods as Pj, j
1; . . . ; k: The dose administered to Si in Pj will be
denoted by dij, for those combinations of i and j
for which an active dose is administered. The
probability of a DLE occurring is denoted by pDLE
d where d is the dose administered. A logistic
regression model is used in which the probability
of a DLE occurring for subject Si in period Pj is
given by
expb1 b2 log dij
1
pDLE dij
1 expb1 b2 log dij
The continuous DO response observed on
subject Si in period Pj will be denoted by yij. A
model is proposed for its conditional value, given
that the subject suers no DLE in that period. This
conditional model is a linear mixed eects regression model:
yij y1 y2 log dij si eij
127
where si is a random subject eect, taken to be

normally distributed with mean 0 and variance t2,
and eij is a random error, taken to be independent
of si and normally distributed with mean 0 and
variance s2. The within-subject correlation following from equation (2) is equal to t2 =s2 t2 : Its
value is taken to be xed at 0.6. The policy of
xing the value of the correlation is discussed and
justied in [19], where it is shown that attempts to
allow for uncertainty concerning its value using a
prior distribution can reduce the accuracy of
estimation more than imposing a wrong value.
The value of 0.6 has been found to be suitable in
retrospective analyses of data in other doseescalation settings: if sucient data exist then
such investigations would help in the choice of a
suitable value in the therapeutic area of interest.
As indicated in Section 1, model (1) does not allow
for correlations between the binary DLE outcomes
within a subject. Neither model (1) or (2) include
period eects. In practice, within the carefully
controlled conditions of phase I studies, no
substantial period eects would be anticipated. It
would be inappropriate to divert some of the very
limited information available in a phase I study to
the estimation of period eects that are likely to be
small and strongly confounded with dose eects.
3. PRIOR INFORMATION FOR THE

UNKNOWN PARAMETERS
Bayesian decision procedures operate by making
the decision on what doses should be recommended for the next cohort of subjects based on
the prior or the current posterior, the data
observed so far and the selection criteria. Before
dose escalation begins, we express our prior belief
concerning the parameters b1, b2, y1, y2 and s. We
use the conjugate prior introduced by Tsutukawa
[20] for the parameters b1 and b2, which has
previously been used in [14,15]. For i 21; 0, pi
pDLE di is considered to have a prior beta
distribution with parameters ai and bi. The two
beta distributions are treated as independent of
one another. The resulting joint prior density of b1

DOI: 10.1002/pst
128 Y. Zhou et al.
and b2 is
h0 b1 ; b2
0
Y
i1

pi
d1
log

Bai ; bi
d0
pai i 1
bi
where pi pDLE di is related to b1 and b2

according to (1).
A normal-gamma conjugate prior density is
used for the parameters h y1 ; y2 and s2:
hjs2 Nl0 ; s2 Q1
0 ;
s2 Ga1 s0 ; t0
4. OPTIMAL AND SAFE DOSE

ESCALATION
where Ga1 denotes an inverse gamma distribution, and l0, Q0, s0 and t0 are to be determined by
the investigators.
The specied prior information can be thought
of in the form of pseudo-data. For pDLE(d), these
consist of ni ai bi pseudo-subjects treated at
dose di, of whom ai suer DLEs, i 1; 0: The
value ni represents the strength of the prior
opinion. The values of ai and bi i 1; 0 should
all be chosen to be positive, as will be natural when
choosing them using pseudo-data. This will ensure
that the prior distribution, and all subsequent
posterior distributions, will be unimodal so that
unique modal estimates of the parameters b1 and
b2 can be found. In the ctitious world of pseudosubjects, the numbers of toxicities, ai, can be set to
non-integer values. To set the normal-gamma
prior, governing the DO, rst predict responses
y01 and y02 corresponding to two log dose levels 01
and 02. The components m01 and m02 of l0 are
taken to be the intercept and slope of the line
joining these two points. Let y0 y01 ; y02 T ; X0 be
the matrix with rows (1, 01) and (1, 02). Then
l0 X1
0 y0 ; and Q0 is the matrix satisfying
s2 Q0 1 varh; that is Q0 XT0 X0 : The values
for s0 and t0 of the inverse Gamma distribution are
more dicult to set. Some imposed values will give
a starting point.
Once data have been observed, posterior distributions can be deduced. As priors are of
conjugate form, posterior distributions will have
the same form of distribution as the priors. In this
paper, we shall be concerned with prior and
posterior modal estimates of parameters b1 and
b2, h and s2. Posterior modal estimates are easy to
obtain and can make simulations quicker than full
Bayes estimates based on expectations. Also
modal estimates can be obtained by adding the

prior pseudo-data to the real data in commonly
available frequentist software, such as SAS PROC
LOGISTIC and PROC MIXED.
We now consider the situation where a schedule of

doses, d 1 5 5d m ; is available for administration to successive cohorts of subjects. The selection
criteria can, for example, be set to choose a dose
that will maximize therapeutic eects and minimize side eects. More specically, if a DLE
occurs, we gain nothing. If a DO is observed
without a DLE, our gain is taken to be the
magnitude y of the DO. The scale on which DOs
are measured must give positive values, and reect
an appropriate advantage over the zero gain
registered when a DLE occurs. Let y0 y if there
is no DLE, and 0 otherwise. Then the expected
value of y0 is
y1 y2 logdose
5
Ey0
1 expb1 b2 logdose
The selection criterion can then be as follows: give
the next cohort of subjects the dose at which Ey0
is the greatest.
It is important to emphasize that, in practical
applications, dose recommendations should
always be reviewed and may be modied before
administration by the safety committee overseeing
the study. In certain situations, extra caution
should be taken in order to prevent that many
subjects will be dosed at unacceptably high dose
levels. This extra caution can be expressed via gain
functions or simply via a safety constraint, for
example, not to administer any dose d for which
pDLE d > g
where g is some predetermined value, for example

0.2. To apply this during the trial, the intercept
and slope parameters are replaced with their
current modal estimates in evaluating the lefthand side of (6). If none of the available doses
satisfy the safety constraint, the study should be

DOI: 10.1002/pst
terminated. Otherwise, the maximum safe dose

for a subject is dened as the dose for which pDLE
d g:
5. AN ILLUSTRATIVE EXAMPLE
This example is based on the anti-factor Xa
compound described in [14]. All recorded adverse
events were taken to constitute a DLE. The data
listed in Table I are constructed to resemble data
from the anti-factor Xa study. The planned doses
were 10.5, 35, 87.5, 262.5, 700 and 1050 mg. The
Table I. Data constructed following the anti-factor Xa

study.
Cohort
Subject
Dose
(mg)
1
2
3
4
5
6
7
8
9
10
1
2
3
4
5
6
7
8
9
10
1
2
3
4
5
6
7
8
9
10
10.5
10.5
10.5
10.5
10.5
35.0
35.0
35.0
35.0
35.0
87.5
87.5
87.5
87.5
87.5
262.5
262.5
262.5
262.5
262.5
700.0
700.0
700.0
700.0
700.0
1050.0
1050.0
1050.0
1050.0
1050.0
DLEs
Anti-factor
Xa (units/
ML 10)
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
1
1
1
0
0
0
0
0
0
0
1
0
0
0.43
0.43
0.36
0.39
0.37
0.73
0.86
0.81
0.92
0.84
1.14
1.22
1.08
1.20
1.13
1.54
1.56
1.43
1.43
1.55
1.91
1.94
1.83
1.88
1.89
1.99
2.08
1.98
2.05
2.03
129
set of subjects treated together in a single period

are referred to as a cohort, so the cohort numbers
shown in Table I are also the period identiers. Six
cohorts of size ve are shown, patients randomly
assigned to placebo being ignored for this illustration.
The real study was conducted according to a
conventional design, although the investigators
did discuss prior information and the denitions of
DLEs and DOs with the authors before conducting the study. Dose escalation followed a predened pattern, with a safety committee receiving
data on DLEs and other relevant outcomes before
conrming the doses to be administered in each
round. The anti-factor Xa data, used in this
illustration as DOs, were not available until the
end of the study and so they did not inuence dose
escalation. It would have been feasible to make
them available to use for dose escalation if that
had been intended, although it would have
required changing current practice. The data that
were collected are not suitable for the illustration
here for various reasons, and so we have reconstructed idealized data of a similar magnitude and
combined them with the real DLE data. Practical
matters concerning the DO data are discussed in
Section 6.
The prior belief concerning the relationship
between DLEs and dose is set so that the
probability p10.5 of a DLE at dose 10.5 mg is
equal to 0.20, and so that this information is
equivalent to observing three pseudo-subjects of
whom 0.6 suer a DLE. Similarly, the probability
p700 of a DLE at dose 700 is equal to 0.33, this
information being equivalent to observing three
pseudo-subjects of whom one suers a DLE.
Formally, p10.5 and p700 are assigned independent
beta distributions:
p10:5 Beta0:6; 2:4;
p700 Beta1; 2
Therefore, the prior modal estimates of the

parameters b1 and b2 are e
b10 1:774; b* 20
0:165: This prior setting, with the safety constraint
(6) for g 0:2; forces dose escalation to start from
the lowest available dose, 10.5 mg. Prior opinion
concerning the continuous DO responses (given no
DLE) suggests that y would be equal to 0.483 at

DOI: 10.1002/pst
130 Y. Zhou et al.
dose 10.5 mg and to 1.840 at dose 700 mg. For the

inverse gamma prior for variance, we impose the
values 1 and 0.002 for s0 and t0, respectively. In
formal terms, the prior of having a DO given no
DLE has the structure:
!
y1 2
s
y2
!
! !
0:277
0:330 0:039
N
;
s2 ;
0:323
0:039 0:009
s2 Ga1 1; 0:002
The choice t0 0:002 is made so that the prior

will convey only weak information. The data in
Table I are used as a rst illustration of the
procedure. As Table I represents data collected by
escalating doses one step at a time without
reference to the Bayesian procedure, the recommended doses are not those actually used: in fact
they are higher. The Bayesian procedure reacted to
the four DLEs observed in cohort 4, by reducing
the recommendation to dose 87.5 mg, although the
investigators proceeded as these DLEs were mild
and judged not to be causally related to the study
drug. This point will be discussed in Section 6.
Table II gives the doses that were recommended by
the Bayesian decision procedure. Figure 1 shows
the tted relationship between the probability of a
DLE and log dose based on the prior and
posterior modal estimates of intercept and slope
and on the frequentist maximum likelihood
estimates of those parameters. In a similar way,
Figure 2 shows the tted relationship between
E(y | no DLE) and log dose. The real data have
shown that there is less risk of a DLE than
Table II. Doses recommended via the Bayesian approach.

After cohort
Actual doses
Recommended doses
0
1
2
3
4
5
10.5
35.0
87.5
262.5
700.0
1050.0
10.5
87.5
262.5
262.5
87.5
87.5
0
-1 0 1 2 3 4 5
6 7 8 9 10 11 12 13 14 15 16 17 18 19
log dose
Prior
Data
Posterior
Figure 1. The prior and posterior most likely relationships between the probability of a DLE and log dose,
together with the model tted by maximum likelihood to
all the data.
7
6
5
4
3
2
1
0
-1
-1 0 1 2 3 4 5
6 7 8 9 10 11 12 13 14 15 16 17 18 19
log dose
Posterior
Data
Prior
Figure 2. The prior and posterior most likely relationships between E(DO | no DLE) and log dose, together
with the model tted by maximum likelihood
to all the data.
supposed by the pessimistic prior. Figure 1 shows

the response to this: the estimated risk is reduced,
although some of the caution inherent in the prior
remains. Figure 2 shows little reaction to the data:
in this illustration, we had no genuine prior
information and so constructed the prior a posteriori! The posterior modal estimates of the parameters b1 and b2 are e
b1 3:410; b* 2 0:371:

DOI: 10.1002/pst
The posterior model for the DOs given no DLE

has the structure:
!
y1 2
s
y2
! !
!
0:0029 0:0004
0:417
s2 ;
;
N
0:0004 0:0001
0:350
s2 Ga1 13:5; 0:026
The nal dose recommended by the procedure to

be safe was 87.5 mg: it maximized the gain subject
to the safety constraint. Although 1050 mg was
associated with greater gain, it was still deemed
unsafe.
We now make a comparison by running the
dose-escalation procedure according to the Bayesian method. We need data from subject/dose
combinations that were not used in practice, and
so have simulated the data from the following
models, tted by maximum likelihood to the data
of Table I,
exp4:276 0:501 log dij
pDLE dij
9
1 exp4:276 0:501 log dij
periods and the nal recommended safe dose for a

new untested subject was 262.5 mg, although
700 mg maximized the gain. The posterior modal
estimates of the parameters b1 and b2 are
e
b1 5:6459; b* 2 0:7049: The posterior model
for the DOs given no DLE has the structure
(Table III):
!
y1 2
s
y2
!
! !
0:367
0:0040 0:0007
N
;
s2 ;
0:343
0:0007 0:0001
s2 Ga1 13:5; 0:019
10
where si N0; 0:0010; and eij N0; 0:0008:

Note that the tted value of the correlation, r, is
0.56. This is close to, but not equal to, the imposed
value of 0.6. As the user of the method will not
know the truth about r, we simulate the situation
in which its true value is 0.56 while the investigator
imposes 0.6.
As before, the procedure started from the lowest
available dose, 10.5 mg. For the second period in
which subjects 610 received their rst dose, the
gain function indicated that the top dose would
give the greatest gain. However, the recommended
dose was 87.5 mg because of the safety constraint.
In any event, the 35 mg dose was skipped using
this procedure. For subjects 15 returning in the
third period, 1050 mg again gave the greatest gain
but 262.5 mg was the highest safe dose to use. Two
DLEs encountered in this period meant that
87.5 mg was the highest safe dose for periods 4
and 5. The procedure stopped after six treatment
11
Table III. A simulated example to demonstrate dose

escalation via the Bayesian approach.
Cohort
Subject
Dose
(mg)
1
2
3
4
5
6
7
8
9
10
1
2
3
4
5
6
7
8
9
10
1
2
3
4
5
6
7
8
9
10
10.5
10.5
10.5
10.5
10.5
87.5
87.5
87.5
87.5
87.5
262.5
262.5
262.5
262.5
262.5
87.5
87.5
87.5
87.5
87.5
87.5
87.5
87.5
87.5
87.5
262.5
262.5
262.5
262.5
262.5
and
yij 0:435 0:356 log dij si eij
131
DLEs
Anti-factor
Xa (units/
ML 10)
0
0
0
0
0
0
0
0
0
0
0
0
1
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0.40
0.42
0.48
0.44
0.40
1.18
1.12
1.11
1.14
1.12
1.52
1.52
1.56
1.57
1.58
1.24
1.19
1.13
1.18
1.26
1.14
1.15
1.25
1.18
1.15
1.64
1.49
1.54
1.57
1.59

DOI: 10.1002/pst
132 Y. Zhou et al.
6. DISCUSSION
This paper extends the Bayesian decision-theoretic
approach to dose escalation from studies yielding
binary bivariate outcomes to the case of bivariate
outcomes where one is binary and the other
continuous. It should have broader applications
than the previous method. The dosing recommendation from the Bayesian approach should serve as
a guide to the clinical investigators or to the trial
safety committee, who will have the power to
override the recommendations if necessary.
The procedure described is based on as simple a
model as we have been able to construct. Such
a model would be inappropriate for the analysis of
a substantial amount of data from a completed
study. However, we wish to use the model before
any data have been collected at all, and then to
revise it having seen observations from around ve
subjects, and then from 10, and so on. There will
be far too little data early on to support more
elaborate models, and so here we have embraced
simplicity. In particular, we have assumed that the
DLEs are independent and so we do not allow for
their within-subject correlation. It appears dicult
to make such an allowance, a practical diculty
being that subjects will often be removed from the
study if a DLE occurs, limiting the data available
for estimating correlations as well as introducing
bias.
In the anti-factor Xa study, the DLEs observed
at 262.5 mg did not cause the subjects to be
removed, nor was the escalation of doses interrupted. Although the denition of a DLE had been
discussed with investigators prior to the start of
the study, the denition encompassed all adverse
events and proved to be too wide. The events
observed concerned nausea, and were neither
considered to be serious nor to be related to the
study drug. This result underlines both the
importance of correctly dening a DLE and of
allowing the overruling of the Bayesian dosing
recommendations. A DLE should be an event that
would cause serious concern about whether dose
escalation should continue. Thus the safety committee would have overruled these recommendations and escalated the dose. In fact, they might
alternatively have chosen to redene the DLE at

this point, and reclassify the subjects as not having
suered DLEs. The revised Bayesian dosing recommendations might then have been acceptable.
A Bayesian procedure, operated in a phase I trial,
should be implemented exibly and imaginatively.
Some changes of practice would have been
necessary in order to implement this procedure.
It is necessary to know individual responses and
the doses at which they occur in order to make
dosing recommendations for the next period. This
means that DO data, such as factor Xa measures,
need to be available quickly. Currently, such values
are collected only when the study is complete. Some
of the real anti-factor Xa values were below the
limit of detection. In order to implement the
procedure it would be necessary either to use more
accurate measurement or to allow for missing
values. Our purpose here is to present the principle
of the method: further elaborations would include
dealing with continuous distributions other than
the normal and making allowance for censored or
missing values. Finally, the data cannot be blind, or
at least someone (preferably the safety committee)
must be able to relate the responses to the subject
indentiers and the doses received. Such practical
issues need to be considered carefully when the
method is applied.
In subsequent work, the authors wish to
investigate further issues identied through attempts to implement this procedure. These include
dealing with DO distributions that are non-normal
and those which are left-censored due to the
occurrence of values below the limit of detection.
Non-linear relationships between DO and log dose
are also of interest, usually the form of nonlinearity would need to be anticipated before data
collection began, although normal-dynamic models might provide a exible approach [21]. It will
also be important to create software to conduct
multiple simulations of proposed designs under a
variety of scenarios in order to assess designs so
that satisfactory choices can be made of the prior
distributions, the cohort size, the number of
available doses and so on. Such investigations
are described for the situation in which DOs are
binary in Reference [16].

DOI: 10.1002/pst
The authors recognize that other statistical

issues could be explored, such as proper allowance
for the correlation of binary DLEs, the inclusion
of period eects and the use of dierent forms of
prior and dierent techniques for their elicitation.
However, these appear to be more subtle questions, perhaps worthy of study if and when the
methodology receives widespread use in its present
form. The use of non-conjugate priors becomes
worthwhile only if prior knowledge is suciently
detailed to defy expression within the conjugate
form. Having specied a non-conjugate prior,
computation of the posterior requires computerintensive methods such as the MCMC approach,
and indeed the MCMC approach has itself
prompted new thinking about the elicitation of
priors [22]. The time required to run MCMC
algorithms becomes an issue, even in modern
computing environments, when multiple simulation runs are to be completed in order to select a
design. Sophisticated approaches to the elicitation
of priors, such as the formal synthesis of opinions
from multiple experts will become important once
clinical pharmacologists understand this method
more completely and are better able to appreciate
what is being asked of them.
8.
9.
10.
11.
12.
13.
14.
15.
16.
REFERENCES
1. Edler L. Overview of phase I trials. In Handbook of

statistics in clinical oncology, Crowley J (ed.). Marcel
Dekker: New York, 2001; pp. 134.
2. OQuigley J. Dose-nding designs using continual
reassessment method. In Handbook of statistics in
clinical oncology, Crowley J (ed.). Marcel Dekker:
New York, 2001; pp. 3572.
3. Whitehead J, Zhou Y, Stallard N, Todd S, Whitehead A. Learning from previous responses in phase
I dose-escalation studies. British Journal of Clinical
Pharmacology 2001; 52:17.
4. Rosenberger WF, Haines LM. Competing designs
for phase I clinical trials: a review. Statistics in
Medicine 2002; 21:27572770.
5. Zhou Y. Choice of designs and doses for early phase
trials. Fundamental and Clinical Pharmacology 2004;
18:373378.
6. OQuigley J, Pepe M, Fisher L. Continual reassessment method: a practical design for phase I clinical
trials in cancer. Biometrics 1990; 46:3348.
7. Whitehead J, Williamson D. An evaluation of
Bayesian decision procedures for dose-nding
17.
18.
19.
20.
21.
22.
133
studies. Journal of Biopharmaceutical Medicine

1998; 8:445467.
Cancer Research UK website, http://www.cancerhelp.org.uk/help/default.asp?page=5078
Gooley TA, Martin PJ, Fisher LD, Pettinger M.
Simulation as a design tool for phase I/II clinical
trials: an example from bone-marrow transplantation. Controlled Clinical Trials 1994; 15:450462.
Thall PF, Russell KE. A strategy for dose-nding
and safety monitoring based on ecacy and adverse
outcomes in phase I/II clinical trials. Biometrics
1998; 54:251264.
OQuigley J, Hughes MD, Fenton T. Dose-nding
designs for HIV studies. Biometrics 2001; 57:
10181029.
Ishizuka N, Ohashi Y. The continual reassessment
method and its applications: a Bayesian methodology for phase I cancer clinical trials. Statistics in
Medicine 2001; 20:26612681.
Braun TM. The bivariate continual reassessment
method: extending the CRM to phase I trials of two
competing outcomes. Controlled Clinical Trials
2002; 23:240256.
Loke YC, Tan SB, Cai Y, Machin D. A Bayesian
dose nding design for dual endpoint phase I trials.
Statistics in Medicine 2006; 25:322.
Whitehead J, Zhou Y, Stevens J, Blakey G, Price J,
Leadbetter J. Bayesian decision procedures for doseescalation based on evidence of undesirable events
and therapeutic benet. Statistics in Medicine 2006;
25:3753.
Whitehead J, Zhou Y, Stevens J, Blakey G. An
evaluation of a Bayesian method of dose-escalation
based on bivariate binary responses. Journal of
Biopharmaceutical Statistics 2004; 14:969983.
Patterson S, Francis S, Ireson M, Webber D,
Whitehead J. A novel Bayesian decision procedure
for early phase dose nding studies. Journal of
Biopharmaceutical Statistics 1999; 9:583597.
Whitehead J, Zhou Y, Patterson S, Webber D,
Francis S. Easy-to-implement Bayesian methods for
dose-escalation studies in healthy volunteers. Biostatistics 2001; 2:4761.
Whitehead J, Zhou Y, Mander A, Ritchie S, Sabin
A, Wright A. An evaluation of Bayesian designs for
dose-escalation studies in healthy volunteers. Statistics in Medicine 2006; 25:433445.
Tsutakawa RK. Bayesian inference for bioassay.
Technical Report 52, Mathematical Sciences, University of Missouri, Columbia, 1975.
West M, Harrison PJ. Bayesian forecasting and
dynamic models. Springer: New York, NY 1997.
Hahn ED. Re-examining informative prior elicitation through the lens of Markov chain Monte Carlo
methods. Journal of the Royal Statistical Society,
Series A 2006; 169:3748.

DOI: 10.1002/pst

Bayesian Decision Procedures For Binary and Continuous Bivariate Dose-Escalation Studies

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

Bayesian Decision Procedures For Binary and Continuous Bivariate Dose-Escalation Studies

Uploaded by

Copyright:

Available Formats

PHARMACEUTICAL STATISTICS

Pharmaceut. Statist. 2006; 5: 125133