| HOME | PEOPLE | PROJECTS | LINKS | CONTACT US |
Berry Consultants' areas of expertise include:
Some of our actual consultations are described below.
Many of the projects apply to a number of
different consultations, all having the same theme. We
have suppressed proprietary details.
PROJECT 1
We have developed numerous types
of Bayesian designs for medical device clinical trials. We
write the protocols, provide the
software necessary for carrying out the eventual analyses and we
carry out those analyses. We develop
innovative designs and analyses for comparing an
experimental device with historical
controls. Types of endpoints considered include dichotomous,
continuous and time to an event.
In some trials we have incorporated surrogate endpoints and early
endpoints for improving the inferential
efficiency. Some protocols have been to show superiority
and others have been to establish
equivalence. In our view it is economically essential to pay heed
to information as it accumulates
in a clinical trial. Therefore, all of these trials include sequential
decision-making. The Bayesian approach
is ideally suited for designing sequential designs and
making interim decisions. Types
of decisions made during these trials include (i) stopping early
because of positive results, (ii)
stopping early for futility, (iii) extending the trial beyond the originally
planned maximum sample size because
the available results are equivocal, (iv) dropping an arm,
and (v) modifying a device. In
making decisions we use predictive probabilities of future
observations, another aspect for
which the Bayesian approach is ideally suited. These predictive
probabilities apply for additional
follow-up of patients already in the trial and also for patients who
have yet to be accrued. The resulting
trial designs are efficient in several ways. For example,
incorporating the possibility of
stopping early or continuing beyond the planned maximum sample
size provides greater statistical
power while at the same time reducing the average number of
patients.
PROJECT 2
We have developed Bayesian models
that incorporate both dose-effect and historical information
about comparitor drugs (including
their doses) in evaluating an experimental drug. Two registration
trials included a comparitor drug
(at one or two doses) and placebo as controls. The comparitor
had been investigated by itself
and in comparison with other agents and placebo in a number of
earlier trials. We exploited these
earlier trials to enhance the strength of the comparison of the
experimental drug with the comparitor,
and also that of the experimental drug with placebo.
Reactions from the company: "Your
Bayesian report gives just the kind of analysis we should be
doing more regularly here to get
as much information from our phase III trials as possible." "Your
Bayesian analysis is really useful
for the clinical team."
We developed a decision analysis
model for the expected shortage of year 2000 influenza vaccine.
We analyzed historical data sets
for the relationship between protection with the vaccine and
subsequent influenza. We used Bayesian
models for the relationship between dose and efficacy,
and simulation models to compare
different dosing schemes in view of the projected shortage of
vaccine. We also modeled vaccine
efficacy as it depends on dose. Our analysis provides a
structure for vaccination schedule
should there be an influenza pandemic.
We provided expertise in the development
of Bayesian neural networks in medical decision
making. We developed methodology
for optimal sequential decision making, creating a structure
for the optimal sequence of questions
in the context of diagnosis. We developed a justification for
Bayesian networks in the context
of medical decision making.
We have developed models for the
evaluation of a medical therapy identified as being effective in a
subgroup of patients. We developed
Bayesian mixture models for binomial outcomes and also for
survival that borrow information
from historical subgroups. We have use these models to design a
sequential confirmatory clinical
trial, as in the next project.
We have developed designs of confirmatory
clinical trials when there is evidence that a device or
drug is effective but only in a
subgroup of patients. We developed Bayesian mixture models that use
information available from a previous
trial: "borrowing strength." The amount of borrowing
depends on the similarity of the
data from the confirmatory trial and the historical trial. If the results
of the confirmatory trial are similar
to the historical results in the subgroup of interest then there is
more borrowing, and if the results
in the confirmatory study differ then little or no information is
borrowed. This recognizes the pitfalls
of subgroup analyses while not ignoring the critical role they
play in discovery and in delivering
good medicine.
We have developed analyses addressing
the sequential decision to stop or continue a Phase III
clinical trial. We incorporate
the ability to drop doses from the study based on the accumulating
evidence about efficacy and safety.
An economic decision analysis considers potential profit, the
cost of conducting the trial, and
the duration of the trial to find the optimal design.
We have developed decision analyses
addressing the question of whether to conduct a phase III
confirmatory study. In one case
we developed a Bayesian hierarchical model considering the
available literature studies. We
modeled the expected profit based on the efficacy of the treatment
and considered the cost of various
Phase III designs. We write the statistical software to find the
optimal phase III design--which
may be "drop development"--depending on the various inputs.
We have designed trials that blur
the distinction between phases II and III of drug development. In
one setting we exploited a possible
relationship between an early "phase II" endpoint and the
ultimate endpoint of survival to
build a seamless phase II/phase III design. The trial is initiated in a
single center. Depending on the
results from that center (and using predictive probabilities) the trial
either ends, continues in its phase
II mode or moves into phase III without halting accrual. In phase
III additional centers are added.
All the data from both phases are used in the eventual inferences
and for registration. Especially
important is the relationship between the early endpoint and survival
that is observed in the trial,
including whether this relationship depends on treatment. The result is
an enormous savings in sample size
and total trial duration, and this savings accrues whether the
drug is concluded to be effective
or not.
We have designed dose-finding studies
that exploit accumulating data in making the decision of
which dose to assign next. Traditional
dose-finding studies almost always end with the conclusion
that a different assignment of
doses or a different number of patients would have more efficiently
identified the important aspects
of the dose-response curve. Our designs avoid that flaw by looking
at the data as they accrue and
asking what is best to do next. The trial stops when the
dose-response curve has been sufficiently
well defined, where "sufficiently well" can be based on
a decision analysis. We use predictive
probabilities in assigning doses; when responses are not
immediately available we calculate
predictive probabilities by exploiting early endpoints and their
empirical correlation with the
final endpoint. Our designs have many benefits, including that they
learn about the standard deviation
of response as it depends on dose. If the standard deviation turns
out to be small then the study
sample size will be smaller. If it turns out to be large then the sample
size will be larger--or in the
worst case of very large standard deviation and only moderate
apparent drug effect, our algorithm
recommends stopping the trial because the drug's development
will be too costly. A clear benefit
of our designs is that they are very efficient in discovering when a
dose-response curve is flat, that
is, when the drug has no benefit.