Professional Documents
Culture Documents
Bayesian Decision Procedures For Binary and Continuous Bivariate Dose-Escalation Studies
Bayesian Decision Procedures For Binary and Continuous Bivariate Dose-Escalation Studies
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
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,
127
and b2 is
h0 b1 ; b2
0
Y
i1
pi
d1
log
Bai ; bi
d0
pai i 1
bi
s2 Ga1 s0 ; t0
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
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
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
p700 Beta1; 2
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.
10
11
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
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
8.
9.
10.
11.
12.
13.
14.
15.
16.
REFERENCES
17.
18.
19.
20.
21.
22.
133