MI250: Introduction To Bayesian PK-PD Modeling & Simulation

MI250: Introduction to Bayesian PK-PD Modeling &
Simulation
Practical use of WinBUGS, BUGSModelLibrary and R for
PK-PD applications
Bill Gillespie and other Metrum Institute scientists
Metrum Institute
c 2010 Bayesian PK/PD M&S Using WinBUGS 1 / 173
Getting Started Course Introduction
Course Introduction
MI250: Introduction to Bayesian PK-PD Modeling & Simulation:
Practical use of WinBUGS, BUGSModelLibrary and R for PK-PD
applications provides an introduction to Bayesian modeling and the
practical use of WinBUGS for pharmacometric applications. The
course duration and content is equivalent to a single semester 3 credit
course at a typical institution of higher learning. Each weeks topic will
consist of a lecture (two hours) followed by a hands-on lab (one hour).
The general plan will be as follows:
Lectures will be on Mondays at 2 PM EDT.
Hands-on labs will be on Thursdays at 2 PM EDT (in some cases,
the lecture may nish during the rst part of the lab on Thursday).
Getting Started Student Expectations and Requirements for Certicate
Student Expectations and Requirements for Certicate
All students are expected to attempt the hands-on exercises prior
to the Thursday lab. Instructors will not grade homework
assignments, but will review solutions with the entire class on a
weekly basis.
A midterm take-home exam will be assigned at the midpoint of the
course.
A nal take-home exam will be assigned at the end of the course
(due one week after it is posted).
Students will be required to complete and submit a modeling
project before the end of the course. This project will be based on
a real-world (or similar) problem, and will include components of
data assembly, model development and evaluation, and a brief
report. More details to follow...
Course grade will be based on the midterm (25%), nal (25%) and
modeling project (50%)
Getting Started Course Content Management Site
Course Content Management Site
All students should already have an account to access the main course
website. Heres the link: (http://training.metruminstitute.org). Postings to this
site will automatically generate an email message to your own e-mail
accounts. This site is intended to be the primary repository for all course
resources including:
News Forum: Here youll nd updates about class schedule,
assignments, etc.
Discussion Forum: Direct your questions about course content to
instructors or other students here. You can contribute to ongoing
discussions or start a new thread.
Technical Support: Use this forum to submit technical support tickets
for problems with any of the Metrum Institute web resources.
Links to the Metrum Institute Web Server (SIMI), and GoToWebinar
webcast registration form.
Course Materials: Youll nd course notes, examples, and links to
recorded lectures under each weekly class heading.
Getting Started About the Course Materials
About the Course Materials
Each week, you will nd course materials such as examples and problems on the
main course website (http://training.metruminstitute.org). The initial course
materials, including the software we will use throughout the course, are contained in
the (rather large) zip le named MI250USB.zip. For the course we recommend you
run the software and course examples directly from a USB ash drive via the
shortcuts provided. Install the contents of MI250USB.zip according to:
1
Download MI250USB.zip to your computer.
2
Unzip MI250USB.zip. This will create a folder named MI250USB.
3
Obtain and insert a USB ash drive with 1GB capacity.
4
Copy the contents of MI250USB (not the MI250 folder itself) to the ash drive.
5
Check the software installation by double-clicking on the shortcuts named R
2.11.1.cmd, WinBUGS14.cmd and BlackBoxWinBUGS.cmd. In each case
the corresponding program (R, WinBUGS and BlackBox Component Builder)
should be successfully launched. If not, please report it via the Technical
Support forum on the course website.
Getting Started Course Outline
Course Outline
The following course outline was originally developed for a 3 day
workshop. All of the listed topics and examples will be covered, but you
should anticipate changes to better t the 1 semester webcast format,
e.g., reordering of topics, subdividing tasks within the hands-on
examples, and adding new topics and examples.
Course Outline
Overview of the current and potential role of Bayesian methods in
clinical drug development
Introduction to Bayesian statistical principles and methods
Bayes Rule
Bayesian modeling & inference process
Likelihood principle
Computation for Bayesian modeling
Key challenge of Bayesian modeling and inference:
high-dimensional multiple integration
General computational approach: posterior simulation
Brief intro to Markov chain Monte Carlo simulation
WinBUGS basics
What is it? How do I get it? How do I run it?
WinBUGS demo: Linear regression
Course Outline
Introduce PK/PD modeling case study to be used throughout the
course
Hands-on example 1: Simple nonlinear regression, e.g., a PK/PD
model relating a single exposure metric to a single continuous PD
outcome
Topics in Bayesian model development using WinBUGS I
Using WinBUGS scripts
R tools for running WinBUGS and analyzing MCMC simulations
Assessing convergence
Programming hierarchical models (aka mixed effect or population
models)
Course Outline
Hands-on example 2: Nonlinear mixed effects, e.g., a PK/PD
model relating observed drug concentrations to continuous PD
measurements at the same time.
Topics in Bayesian model development using WinBUGS II
Model evaluation and comparison
BUGSModelLibrary for pharmacometric modeling: Introduction &
demonstration
Dealing with censored data in WinBUGS
BQL data
Hands-on example 3: Population PK.
Course Outline
Topics in Bayesian model development using WinBUGS III
Informative prior distributions in clinical pharmacology applications
Using the cut function in WinBUGS
Programming models in terms of differential equations
Hands-on example 4: Population PK-PD using an indirect action
model
Additional topics & closing discussion
Selected published examples of Bayesian applications
Considerations in deciding whether to use:
Bayesian or maximum likelihood methods
WinBUGS or NONMEM or ?
What didnt we cover?
Licensing
Licensing
Your use of the materials provided in this course is subject to the terms
of use described in the courseware license agreement in the Appendix.
Bayesian modeling Core notions
Bayesian modeling
Bayesian principles and methods provide a coherent framework
for:
Quantifying uncertainty,
Making inferences in the presence of that uncertainty.
It is also the basis for formal approaches to incremental model
building, parameter estimation and other statistical inference as
knowledge and data are accumulated.
The two core notions that distinguish Bayesian analysis are:
Unknown quantities are viewed as random variables, i.e., they are
described in terms of probability distributions.
Bayes rule provides a formal mechanism for combining prior
knowledge and new data.
Bayesian modeling Core notions
Bayesian approach to uncertainty
Uncertainty is described in terms of probability:
Can assign a quantitative description of uncertainty in the model
parameters and evaluate how it affects the uncertainty in predicted
outcomes.
Probability can be used to quantitatively represent uncertainty
based on either objective evidence/data or subjective expert
opinion or belief.
0 5 10 15 20 25
0
.
0
0
0
.
0
5
0
.
1
0
0
.
1
5
CL (L/h)
p
r
o
b
a
b
i
l
i
t
y

d
e
n
s
i
t
y
Pr(CL > 15) = 0.0524
Role of Bayesian methods in clinical drug development
Current and potential role of Bayesian methods in
clinical drug development
PK and PK/PD modeling
Model development: Bayesian hierarchical models
Model evaluation/selection
Synthesis of new data with prior data and knowledge
Quantitative treatment of uncertainty in simulations of drug
treatment and clinical trial simulations
Bayesian decision analysis for optimization of:
Clinical trial design and analysis
Development programs
Treatment regimens
Adaptive trial designs
Motivating scenario 1
You need to characterize the risk to human workers posed by
VOC-X in the air.
VOC-X is a volatile organic compound known to cause a range of
adverse health effects.
Because of its toxicity we cannot perform controlled studies
involving the administration of VOC-X to humans.
As a result we have no direct knowledge about TK, the
relationships between atmospheric exposure, systemic exposure,
and toxicity in humans.
We do have such information for 3 other species.
Limited information is available for closely related compounds.
How would you synthesize this information to make quantitative
inferences about the probable ranges of systemic exposure and
toxicity in humans?
How would you characterize the uncertainty in your inferences?
You need to recommend a dose or dose range for drug X that
optimizes the trade-offs between benets and risks.
Phase II has been completed.
Dose-response information is available for key efcacy and
safety-related outcomes of drug X treatment.
How would you synthesize the cumulative quantitative evidence on
safety and efcacy in order to predict outcomes for various doses?
How would you quantify the uncertainty in those predictions?
How would you use those predictions in conjunction with patient
and prescriber opinions regarding acceptable risk relative to
benet?
You need to design a Phase II strategy for a new drug with an
unprecedented MOA for treating a life-threatening disorder for
which there is no effective treatment.
There is large uncertainty in whether the known pharmacological
effects of the drug will translate into therapeutic benet,
And even more uncertainty in dose-response for both efcacy and
safety-related responses.
As a result, a conventional dose nding trial using 3-4 active doses
has a high risk of failing to yield good dose-response information.
How would you design an efcient dose-nding and PoC strategy
that is likely to yield relatively conclusive inferences regarding
go/no-go and dose selection?
Introduction to Bayesian statistical principles and methods Bayesian approach to statistical inference
Introduction to Bayesian statistical principles and
methods
Bayesian approach to statistical inference
Model parameters and predictions described in terms of
probability distributions representing uncertainty
Results reect the combined evidence of data and prior
knowledge or belief
Focuses on estimation and inferences related to probabilities of
unknown quantities: parameters, future data, hypotheses.
Inferences are described directly in terms of probabilities:
Hypothesis test: Probability of no treatment effect given data
Interval estimation: Probability that a parameter value lies within an
interval given data
Introduction to Bayesian statistical principles and methods Bayesian applications
Bayesian applications
Decision analysis
Bayesian principles are well-suited to decision-making
Assess probable outcomes of different decisions via predictive
probabilities
Choose decision that maximizes average gain (minimizes average
loss)
Modeling and simulation
Bayesian modeling and simulation are practical options for many
applications due to advances in hardware, numerical methods and
software.
Now possible to model complex nonlinear systems (e.g., PK/PD)
and processes (e.g., clinical trials).
Introduction to Bayesian statistical principles and methods Bayesian inference
Bayesian inference
Bayes Rule
Bayes Rule is the basis for inference about model parameters () given
data (y) and prior knowledge about model parameters (p ()):
p (|y) =
p () p (y|)
p (y)
=
p () p (y|)
_
p () p (y|) d
p () p (y|)
The ps are probabilities or probability densities of the specied
random variables.
Bayesian modeling/inference process
1
Assess prior distribution p ()
viewed as random variables
Subjective
Ideally base on all available evidence/knowledge (or belief)
Or deliberately select a non-informative (or weakly informative)
prior (e.g., reference, vague or improper prior)
2
Construct a model for the data p (y|), also known as the
likelihood function when viewed as a function of .
3
Calculate posterior distribution p (|y).
Use for inferences regarding parameter values
4
Calculate posterior predictive distribution p (y
new
|y).
Use for inferences regarding future observations
p (y
new
|y) =
_
p (y
new
|) p (|y) d
A simple one parameter example
Estimating the mean of a normal distribution with a known variance
where the prior is also normal:
N
_
0
,
2
0
_
y| N
_
,
2
_
p (|y) p () p (y|) = p ()
n
i =1
p (y
i
|)
exp
_
1
2
2
0
(
0
)
2
_
n
i =1
exp
_
1
2
2
(y
i
)
2
_
exp
_
1
2
_
1
2
0
(
0
)
2
+
1
2
n
i =1
(y
i
)
2
__
|y N
_
n
,
2
n
_
n
=
1
2
0
0
+
n
2
y
1
2
0
+
n
2
and
2
n
=
1
1
2
0
+
n
2
A simple one parameter example
Posterior mean is a weighted average of prior mean (
0
) and the sample mean
(y).
Posterior precision
_
1
2
n
_
is a sum of the prior precision
_
1
2
0
_
and the data
precision
_
n
2
_
.
The prior can also be interpreted in terms of equivalent number of data points,
i.e., n
0
=

2
2
0
.
0 5 10 15 20 25 30
0
.
0
0
.
2
0
.
4
0
.
6
d
e
n
s
i
t
y
true mean
posterior
prior
posterior
n = 1
Introduction to Bayesian statistical principles and methods The likelihood principle
The likelihood principle
The likelihood function p (y|) contains all the information in the
observed data y from an experiment relevant to inferences about the
parameter .
Follows directly from Bayes Rule and Bayesian inference.
Basically it says that inferences should be based on what is
observed, not on what might have been observed.
Key distinction from frequentist principles and methods.
Has important practical consequences for clinical trial design and
analysis.
Removes many frequentist restrictions/penalties on:
Interim analyses.
Interim design changes, e.g., adaptive dose assignment.
Caution: How the data is collected is still relevant. For example
randomization is still a necessary element of experimental design for
assuring exchangeability and validity of causal inference.
Introduction to Bayesian statistical principles and methods Comparison of Bayesian and frequentist inference
Comparison of Bayesian and frequentist inference
Probability concepts
Bayesian: Probability of parameters (& new data) given observed
data
Frequentist: Probability of observed data given parameters
Evidence used
Bayesian: All available (relevant) information/knowledge
Conceptually appealing but hard work to do correctly/honestly
Frequentist: Specic to experiment
Experimental results upon which inferences depend
Bayesian: Observed data
Frequentist: Also depends on the probabilities of data that were
possible but not observed
Experimental design exibility
Bayesian: May adapt trial design as evidence accumulates
Sample size need not be pre-specied
Interim analyses may be conducted anytime and at any frequency
Number and timing has no effect on the analyses
Explicit cost/risk/benet analysis may be applied to optimize trial
design as it progresses
Stop at any time for any reason
Assign experimental treatments to optimize trial outcome
Frequentist: Inferences depend on trial design and require that a
pre-specied design be followed
Interim analyses possible but more restricted
Must be pre-specied
Number and timing affects the analyses
Decision-making
Bayesian: Formal decision analysis conceptually simple: maximize
expected gain (minimize expected loss)
Frequentist: Formal decision analysis possible but conceptually and
operationally difcult; historically avoided
Computation for Bayesian modeling The bad news
So whats the bad news?
Youve a selected suitable distributions for p () and p (y|).
is a vector of 6 parameters;
y is a vector of 200 observations.
You would like to estimate the expected value of a function of ,
i.e., E (f () |y).
You know that p (|y) p (y|) p () so
E (f () |y) =
_
1
_
2
_
3
_
4
_
5
_
6
f () p (|y) d =
_
1
_
2
_
3
_
4
_
5
_
6
f () p (y|) p () d
_
1
_
2
_
3
_
4
_
5
_
6
p (y|) p () d
Hmmmm. How do you do that?
Computation for Bayesian modeling The bad news
So whats the bad news?
If youre lucky those integrals have known analytic solutions, but
that is rarely true for PK/PD modeling applications.
For integrals in fewer dimensions, say up to 3, a numerical
quadrature method might be practical, but 6 dimensions is not so
feasible.
Now imagine the computational requirements for hierarchical
models, e.g., population PK models, with individual-specic
parameters in the hundreds!!
Computation for Bayesian modeling Posterior simulation
But theres hope: posterior simulation
What if you could simulate samples of from the joint posterior
distribution?
Then you could estimate E (f () |y) by the arithmetic mean:
E (f () |y)
1
n
n
i =1
f (
i
)
But how do you simulate samples from a high dimensional joint
posterior distribution?
Markov chain Monte Carlo (MCMC) simulation via WinBUGS is
the approach we will explore today.
Computation for Bayesian modeling Posterior simulation
Inferences from posterior simulations
Posterior simulation yields vectors of parameters and/or
predictions from a joint posterior distribution
Marginal distributions
To describe the posterior distribution of any scalar function of the
parameters apply the function to each simulated vector. The
empirical distribution of those values approximates the posterior
distribution.
The marginal distribution of any single parameter is just a special
case of that approach.
Inferences are usually based on moments, probabilities or
percentiles from marginal posterior distributions. They are readily
estimated from the corresponding sample statistics for the
simulated values.
Computation for Bayesian modeling Markov chain Monte Carlo (MCMC) simulation
Tools for Bayesian modeling & inference
Recent development of more exible methods and computational
tools, e.g., Markov chain Monte Carlo (MCMC) simulation.
Now possible to model complex nonlinear systems (e.g., PK/PD)
and processes (e.g., clinical trials).
Markov chain Monte Carlo (MCMC) simulation
Simulation methods for performing high dimensional integration
required for Bayesian modeling.
Simulate random variables from the posterior distributions of
interest, e.g., means and variances of the population distributions of
PK/PD parameters.
Inferences follow directly from the distributions of simulated
parameter values
Point estimates from mean, median or mode.
Posterior intervals from percentiles, e.g., 95% interval from the 2.5
and 97.5 percentiles.
Markov Chain Monte Carlo (MCMC) simulation
Involves random draws from approximate distributions and then
correcting those draws to better approximate the joint posterior.
The samples are drawn sequentially so that each draw depends
on the previous one, thus forming a Markov chain.
Eventually the Markov chain converges to a stationary distribution
that is the joint posterior distribution.
Algorithms for MCMC include:
Metropolis-Hastings algorithm
Gibbs sampling
MCMC samples are serially correlated:
Inferences based on MCMC require more samples than would be
required for independent samples
Practical consequences:
Use only samples drawn after convergence is achieved, i.e.,
discard samples from a burn-in phase.
Draw more samples than you would for independent random draws.
Gibbs sampling
In most cases a convergent Markov chain may be constructed by
progressively sampling from the univariate full conditional
distributions:
i
1
p
_
1
|
i 1
2
,
i 1
3
, . . . ,
i 1
n
, y
_
i
2
p
_
2
|
i
1
,
i 1
3
, . . . ,
i 1
n
, y
_
.
.
.
i
n
p
_
n
|
i
1
,
i
2
, . . . ,
i
n1
, y
_
This reduces the multivariate posterior sampling problem to a
sequence of more manageable univariate sampling problems.
Gibbs sampling example
Data: single observation from a bivariate normal distribution:
y = (y
1
, y
2
) N (, )
Unknown mean: = (
1
,
2
)
Known covariance matrix: =
_
1
1
_
Improper uniform prior on
The conditional posterior distributions of
1
and
2
are:
1
|
2
, y N
_
y
1
+ (
2
y
2
) , 1
2
_
2
|
1
, y N
_
y
2
+ (
1
y
1
) , 1
2
_
y = (0, 0) and = 0.8
For this simple case Gibbs sampling can be done with the
following short R function (See R script gibbs.example.R):
gi bbs1 < f unct i on ( nsamp, y , sigma2 , mu. i n i t ) {
## Gibbs sampl er f or bi v ar i at e normal mean wi t h
## known covar i ance mat r i x ( sigma2 ) , an i mproper
## uni f or m p r i o r and gi ven one dat a poi nt ( y )
mu < mat r i x ( doubl e ( 2nsamp) , ncol =2)
mu[ 1 , ] < mu. i n i t
f or ( i i n 2: nsamp) {
mu[ i , 1 ] < rnorm . cond ( 1 , 1 , y , sigma2 , mu[ i 1 , ] )
mu[ i , 2 ] < rnorm . cond ( 1 , 2 , y , sigma2 , mu[ i , ] )
}
mu
}
4 chains with different starting points
10 samples per chain 1000 samples per chain
4 2 0 2 4
2
0
2
4
2
q
q
q
q
4 2 0 2 4
2
0
2
4
2
q
q
q
q
WinBUGS basics What is it?
WinBUGS
What is it?
BUGS = Bayesian analysis Using Gibbs Sampling
Versatile program for Bayesian modeling using Markov Chain Monte Carlo
(MCMC) simulation
Both graphical and text model specication
Data format is R-like.
Uses Gibbs sampling with sampling from the conditional distributions according
to:
Continuous distributions
Conjugate Direct sampling using standard algorithms
Log-concave Derivative-free adaptive rejection sampling
Restricted range Slice sampling
Unrestricted range Metropolis
Discrete distributions
Finite upper bound Inversion
Shifted Poisson Direct sampling using standard algorithms
WinBUGS basics What is it?
WinBUGS versions
WinBUGS 1.4.3
Windows application only available as executable code
Well-tested stable code. Probably the best version for production
use at this time.
No active development. Bug xes only.
OpenBUGS 3.0.8 (beta)
Probable successor to WinBUGS 1.4.*
Source code available
Primarily a Windows application, but a non-GUI version can be
ported to Unix/Linux systems with Intel/AMD processors
Active development. Currently a beta version.
Release of a stable version imminent.
WinBUGS basics How do I get it?
WinBUGS
How do I get it?
WinBUGS 1.4.3 is available at:
http://www.mrc-bsu.cam.ac.uk/bugs/winbugs/contents.shtml
Site also contains many links to other useful Bayesian sites
OpenBUGS 3.0.* is available at:
http://www.openbugs.info
WinBUGS basics How do I get it?
WinBUGS extensions/accessories used in this course
R2WinBUGS is a collection of R functions for running WinBUGS
and OpenBUGS from R
http://cran.r-project.org/src/contrib/Descriptions/
R2WinBUGS.html
http://www.stat.columbia.edu/
~
gelman/bugsR/
CODA is useful set of R functions for processing WinBUGS
output:
http:
//cran.r-project.org/src/contrib/Descriptions/coda.html
BUGSModelLibrary is a WinBUGS extension that includes PK
models, differential equation solvers and uses NONMEM data
conventions
http://bugsmodellibrary.googlecode.com
WinBUGS basics How do I run it?
WinBUGS demo: Linear regression
Linear model with weakly informative prior distributions for slope and
intercept:
y
i
N
_
y
i
,
2
_
y
i
= a + bx
i
Prior distributions:
a N (0, 10000)
b N (0, 10000)
Uniform(0, 10000)
for(i IN 1 : n)
y[i]
sigma
tau
yhat[i]
b a
Data (R format):
x = c(1,2,3,4,5,6,7,8,9,10)
y = c(5.19,6.56,9.19,8.09,7.6,7.08,6.74,9.3,8.98,11.5)
WinBUGS demo
WinBUGS topics to be introduced
Using the WinBUGS model language
Entering data and initial estimates
Generating MCMC samples
Assessing convergence
Fuzzy caterpillar criterion
Multiple chains with different initial estimates
Analyzing and interpreting the results
BUGS model specication language
Not a procedural language
Model specication similar to statistical literature conventions
Does not explicitly describe a sequence of calculations
Statement order is largely unimportant
Stochastic nodes a.k.a. probability distributions
Logical nodes a.k.a. deterministic functions
Very exible w.r.t. stochastic structure of a model
Lack of control structures (true loops & if-then-else) and control
over calculation order can make specication of complex
deterministic functions difcult (but there is a solution)
WinBUGS demo
WinBUGS model
y
i
N
_
y
i
,
2
_
y
i
= a+bx
i
Prior distributions:
a N (0, 10000)
b N (0, 10000)
Uniform(0, 10000)
stochastic
nodes
logical
nodes
&
&
&
&
&
&
&
&b
B
r
r
r
r
r
rj
~
&
&
&
&
&
&
&
&b
z
loop over
the data
$
$
$
$
$
$W
comments begin
with a #
model{
for ( i in 1:10){
# model for observed data (likelihood function)
y[ i ] dnorm(ymean[i],tau)
# simulation of new observations, i . e., samples from
# the posterior predictive distribution
ypred[i ] dnorm(ymean[i],tau)
# expected value of y[ i ] given x[ i ]
ymean[i] < a + bx[i]
}
# prior distributions
a dnorm(0,0.0001)
b dnorm(0,0.0001)
sigma dunif(0,10000)
tau < 1/(sigmasigma)
}
WinBUGS demo
Data and initial estimates
Formatting the data
Data format is like R list
List of scalars, vectors, matrices or arrays
Numbers onlyno character strings
Matrices/arrays: Unlike R, WinBUGS uses row major entry
list (
x = c (1,2,3,4,5,6,7,8,9,10) ,
y = c (5.19,6.56,9.19,8.09,7.6,7.08,6.74,9.3,8.98,11.5)
)
Specifying initial estimates
list (
a = 3,
b = 1,
sigma = 1.5
)
WinBUGS demo
WinBUGS steps to be demonstrated
Open le(s) containing model, data and initial estimates
Compile the model (model specication tool)
Check model syntax
Load data
Compile
Load initial estimates
Generate initial values for any remaining un-initialized random variables
Specify the variables you want to save/monitor (sample monitoring tool)
Specify the number of sample draws and run the MCMC simulation (update tool)
Initial assessment of convergence by viewing sample histories (history button in
sample monitoring tool)
View summary statistics and density plots of marginal posterior distributions of
the sampled variables (density and stats buttons in sample monitoring tool)
View plots showing ts to the data (comparison tool)
Save MCMC samples to a le (coda button in sample monitoring tool)
WinBUGS
demo
WinBUGS demo
MCMC history plots for model parameters
a
iteration
1 2500 5000 7500 10000
-5.0
0.0
5.0
10.0
15.0
b
iteration
1 2500 5000 7500 10000
-1.0
0.0
1.0
2.0
sigma
iteration
1 2500 5000 7500 10000
0.0
5.0
10.0
15.0
WinBUGS demo
Summary statistics and density plots for posterior marginal
distributions of model parameters
node mean sd MC error 2.5% median 97.5% start sample
a 5.687 1.141 0.01127 3.375 5.682 7.899 501 9500
b 0.4225 0.1842 0.001956 0.06104 0.4221 0.7924 501 9500
sigma 1.601 0.5231 0.00797 0.9435 1.497 2.818 501 9500
a sample: 9500
-5.0 0.0 5.0 10.0
0.0
0.2
0.4
0.6
b sample: 9500
-1.0 0.0 1.0
0.0
1.0
2.0
3.0
sigma sample: 9500
0.0 5.0 10.0
0.0
0.5
1.0
1.5
WinBUGS demo
Model predictions (posterior median & 95% prediction intervals)
compared to observed data
model fit: ypred
0.0 2.5 5.0 7.5
0.0
5.0
10.0
15.0
Hands-on sessions
Fictional antithrombotic case study used for all
hands-on sessions: Clinical development of ME-2
We are developing a orally-administered direct Factor Xa inhibitor ME-2 for
indications including post-op prevention of VTEs and prevention of thrombotic
events in patients with atrial brillation. Early clinical development focuses on
post-op prevention because it can be done more quickly and with fewer
resources, and it is more likely to provide exposure-response information
relevant to all indications. Size, duration and variability (and cost) for atrial b
trials are all large, making dose-response studies prohibitive.
In Phase I we look at safety, PK and biomarker PD. Inhibition of factor Xa activity
is the primary biomarker we are using for both efcacy and bleeding liability. For
our example we will focus on inhibition of factor Xa activity and its relationship to
PK by analyzing the Phase I data. We will model the PK/PD relationship in a
couple different ways to illustrate both simple nonlinear regression and nonlinear
mixed effects regression.
Hands-on sessions
Fictional antithrombotic case study used for all
hands-on sessions: Clinical development of ME-2
Neutropenia was observed in some subjects receiving higher ME-2 doses in
Phase I. That data needs to be further analyzed to assess whether there is
sufcient probability of identifying a dose that is effective while still causing an
acceptably small degree of neutropenia. To facilitate this effort we will model the
relationship between neutrophil counts and drug exposure. We will adapt the
semi-mechanistic model proposed by Friberg and Karlsson, including use of
informative prior distributions for drug-independent parameters.
The clinical development plan includes a Phase IIa PoC trial comparing one
dose level of ME-2 to an active comparator (enoxaparin) in patients undergoing
total knee replacement. VTEs, bleeding events and PK are measured. We will
analyze the combined Phase I and Phase IIa PK data using a population PK
model.
Hands-on sessions
Antithrombotic case study used for all hands-on
sessions: Clinical development of ME-2
Example 1: Analysis of Phase 1 study in healthy volunteers in which
PK and inhibition of factor Xa activity are measured.
Nonlinear regression of time-averaged biomarker as a
function of time-averaged plasma drug concentration.
Example 2: Ditto except we now apply a direct action PK/PD model to
the time-matched biomarker and plasma drug
concentration measurements.
Example 3: Analysis of PK data from Phase I and a Phase IIa PoC
trial where the drug is used to prevent post-op VTEs.
Example 4: Analysis of neutrophil count data from Phase I using a
semi-mechanistic PK/PD model.
Hands-on session 1
Hands-on session 1
PK-PD modeling of time-averaged biomarker and PK data
Phase 1 single dose study in healthy volunteers
Parallel dose-escalation design
8 subjects per dose arm
Single doses of ME-2
Placebo, 1.25, 5, 10, 15, 20, 30, 40, 60 and 80 mg
PK: plasma concentrations of parent drug
Biomarker: ex vivo inhibition of factor Xa activity in plasma
PK and biomarker measured at 0, 0.083, 0.167, 0.25, 0.5, 0.75, 1,
1.5, 2, 3, 4, 6, 8, 12, 18 and 24 hours after dose.
Hands-on exercise:
Model relationship between time-averaged factor Xa inhibition and
time-averaged ME-2 plasma concentrations
Hands-on session 1
Hands-on session 1
EDA: PK and biomarker data
time (h)
M
E
2

p
l
a
s
m
a

c
o
n
c
e
n
t
r
a
t
i
o
n

(
n
g
/
m
L
)
0
5
1
5
2
5
0 5 10 20
1.25 mg
0
4
0
8
0
1
2
0
5 mg
0
1
0
0
2
5
0
0 5 10 20
10 mg
0
1
0
0
3
0
0
15 mg
0
2
0
0
5
0
0
20 mg
0
2
0
0
6
0
0
30 mg
0
4
0
0
8
0
0
40 mg
0 5 10 20
0
5
0
0
60 mg
0
5
0
0
1
5
0
0
80 mg
time (h)
f
a
c
t
o
r

X
a

i
n
h
i
b
i
t
i
o
n

(
%
)
2
0
0
2
0
05 15 25
0 mg
2
0
0
2
0
4
0
1.25 mg
2
0
2
0
6
0
05 15 25
5 mg
2
0
2
0
6
0
10 mg
0
4
0
8
0
15 mg
0
4
0
8
0
20 mg
0
5
0
1
0
0
30 mg
0
5
0
1
0
0
40 mg
0
5
0
1
0
0
60 mg
05 15 25
0
5
0
1
0
0
80 mg
Hands-on session 1
Hands-on session 1
EDA: Relationship between biomarker and PK data
ME2 plasma concentration (ng/mL)
f
a
c
t
o
r

X
a

i
n
h
i
b
i
t
i
o
n

(
%
)
0
50
100
0.4 0.2
0 mg
05 15 25
1.25 mg
0 4080
5 mg
0100250
10 mg
0100 300
15 mg
0 200 500
20 mg
0 400
30 mg
0 400 1000
0
50
100
40 mg
0
50
100
0500
60 mg
0 1000
80 mg
average ME2 plasma concentration (ng/mL)
a
v
e
r
a
g
e

f
a
c
t
o
r

X
a

i
n
h
i
b
i
t
i
o
n

(
%
)
0
20
40
60
80
0 100 200 300
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
qq
qq
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
qqq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
Hands-on session 1
Hands-on session 1
Proposed model
Sigmoid Emax model relating time-averaged % inhibition of factor
Xa activity to time-averaged ME-2 plasma concentration in the i
th
subject:
E
24,i
N
_
E
24,i
,
2
_
E
24,i
=
E
max
c
24,i
EC
50
+ c
24,i
Some possible weakly informative prior distributions:
E
max
U (0, 100)
log(EC
50
) N
_
0, 10
6
_
U (0, 10)
U (0, 1000)
Hands-on session 1
Hands-on session 1
Files
Example folder: me2HandsOn1
Model: me2HandsOn1.odc
Also contains data and initial estimates
Topics in Bayesian model development using WinBUGS I Using WinBUGS scripts
WinBUGS scripting language
Permits batch-mode operation of WinBUGS 1.4
Can be launched from any program that can run DOS commands,
e.g., R
Can write a single R script that creates the WinBUGS data sets
and initial estimates, runs WinBUGS, and analyzes the
WinBUGS-generated samples
In fact you can create an R script that simulates multiple clinical
trials and analyzes each one using WinBUGS
Topics in Bayesian model development using WinBUGS I Using WinBUGS scripts
WinBUGS script for linear regression example
display ( log )
check(C:/MIBayesCourse/linear/linear1.odc)
data( C:/MIBayesCourse/linear/data.txt)
compile(3)
inits (1, C:/MIBayesCourse/linear/inits1. txt )
inits (2, C:/MIBayesCourselinear//inits2. txt )
inits (3, C:/MIBayesCourse/linear/inits3. txt )
gen. inits ()
set(a)
set(b)
set(sigma)
set(ypred)
set(deviance)
update(1000)
coda(, C:/MIBayesCourse/linear/coda)
stats(a)
stats(b)
stats(sigma)
density(a)
density(b)
density(sigma)
history(a)
history(b)
history(sigma)
autoC(a)
autoC(b)
autoC(sigma)
save(C:/MIBayesCourse/linear/log.odc)
quit ()
Topics in Bayesian model development using WinBUGS I R tools for running WinBUGS and analyzing MCMC simulations
R tools for WinBUGS
R2WinBUGS
R package of functions for running WinBUGS or OpenBUGS from R
CODA
R package of functions for analyzing MCMC results
bugs.tools.R
A few handy-dandy functions we have provided, mostly for
manipulating and plotting MCMC results
Topics in Bayesian model development using WinBUGS I R tools for running WinBUGS and analyzing MCMC simulations
R tools demo
Run linear regression example using R tools for:
Preparation of les containing data and initial estimates
Running WinBUGS
Analyzing and plotting MCMC results
Saving it all in a specied directory
This is done using the R script linear1b.R
Topics in Bayesian model development using WinBUGS I Assessing convergence & adequacy of sample sizes
Assessing convergence & adequacy of sample sizes
Early samples may be unrepresentative of the target distribution
MCMC samples within a chain are autocorrelated
Inferences based on MCMC samples are less precise than those
from the same number of independent samples
Autocorrelation also inuences the rate of convergence
Assessing convergence & adequacy of sample sizes
Use a burn-in phase, i.e., discard early iterations
Monitor convergence via multiple chains with different starting
points
Look for chains to converge to a common distribution
You want chain history plots to look more like straight horizontal
fuzzy caterpillars than wiggly snakes
Monitor Gelman-Rubin diagnostics and/or Gelman-Rubin-Brooks
plots
Essentially ratios of total variance to within chain variance.
Should approach 1 for all parameters of interest on convergence
How many samples?
Number of samples depends on the inference(s) of interest and
the desired precision
For independent samples:
A posterior mean of a parameter is estimated with an error of
n
where s
is the standard deviation of the simulated values of

and n is the number of samples.
A probability p is estimated with an error of
_
p(1p)
n
.
In general more samples are required for estimating tail quantiles
and probabilities than for central tendencies and probabilities near
0.5
How many samples?
Suppose the posterior distribution of is N ( = 10, = 2).
What if we use simulation to estimate various features of that distribution?
The following are bootstrap estimates of error due to simulation:
How many MCMC samples?
Given a set of MCMC samples it is possible to adjust for
autocorrelation to estimate:
The equivalent number of independent samples
The standard error in the estimated posterior mean
The R functions effectiveSize and summary.mcmc from the CODA
package provide such estimates
Guidance based on independent samples may then be applied
For more rigorous and comprehensive treatment see Robert and
Casella 2004 and 2010 [8, 7].
Topics in Bayesian model development using WinBUGS I Programming hierarchical models
Programming hierarchical models (e.g. population
models)
Consider the usual case where you want to model 2 levels of
variability, e.g., inter-individual and residual
The bulk of the data set consists of a set of equal length vectors
Data value(s) for each observation
Individual identier: Use a consecutive sequence of integers
Covariates (dose, time, etc.)
May include constants such as numbers of observations and
patients, parameters of prior distributions, etc.
WinBUGS demo
Hierarchical model for dose-response
Data: sample mean response from 5 clinical trials
Illustration of meta-analysis where inter-trial differences are modeled as random
dose
r
e
s
p
o
n
s
e
2
4
6
8
10
0 20 40 60 80
Mean
Study Dose n Response
1 0 40 2.02
1 10 40 8.17
1 20 40 9.87
1 40 40 10.51
2 0 40 2.19
2 40 40 10.72
2 80 40 10.99
3 0 200 1.97
3 40 200 11.44
4 0 50 1.87
4 20 50 8.86
4 40 50 10.03
5 0 20 1.99
5 10 20 8
5 20 20 9.47
5 30 20 10.33
WinBUGS demo
Hierarchical model for dose-response
Emax model with inter-trial variation in E
0
& E
max
:
log
_
E
ij
_
N
_
log
_
E
ij
_
,

2
n
ij
_
E
ij
= E
0,j
+
E
max,j
D
ij
ED
50
+ D
ij
log
_
E
0,j
_
N
_
E
0
,
2
E
0
_
log
_
E
max,j
_
N
_
E
max
,
2
E
max
_
Prior distributions
E
0
N
_
0, 10
4
_

E
max
N
_
0, 10
4
_
log(ED
50
) N
_
0, 10
4
_
U
_
0, 10
4
_

E
0
U
_
0, 10
4
_

E
max
U
_
0, 10
4
_
WinBUGS demo
WinBUGS model
Model implemented with dose-response1.txt & dose-response1.R
Modeling inter-trial variation
for ( i in 1:nstudy){
# inter trial variation
log. emax[i] dnorm(mu.emax,tau.emax)
log. e0[ i ] dnorm(mu.e0,tau.e0)
emax[i] < exp(log.emax[i])
e0[ i ] < exp(log.e0[i])
}
WinBUGS demo
WinBUGS model
Modeling the observations (treatment arm means)
for ( i in 1:narm){
# model for observed data (likelihood function)
# note use of study to index trial specic parameters
resp.median[i] < e0[study[i]] + emax[study[i]]dose[i ]/( ed50+dose[i])
mu.resp[i] < log(resp.median[i])
tau[ i ] < tau.resp n[i ]
# WinBUGS feature for data transformation
log. resp[ i ] < log(resp[i ])
log. resp[ i ] dnorm(mu.resp[i],tau[i])
}
WinBUGS demo
WinBUGS model
Prior distributions
# prior distributions
mu.emax dnorm(0,0.0001)
mu.e0 dnorm(0,0.0001)
log. ed50 dnorm(0,0.0001)
ed50 < exp(log.ed50)
med.emax < exp(mu.emax)
med.e0 < exp(mu.e0)
omega.emax dunif(0,10000)
omega.e0 dunif(0,10000)
sigma dunif(0,10000)
tau.emax < 1/(omega.emaxomega.emax)
tau.e0 < 1/(omega.e0omega.e0)
tau.resp < 1/(sigmasigma)
cv.emax < 1/sqrt(tau.emax)
cv.e0 < 1/sqrt(tau.e0)
cv.resp < 1/sqrt(tau.resp)
WinBUGS demo
MCMC history plots for model parameters
sample
v
a
l
u
e
1
0
0
4
0
0 200 400 600 800 1000 1200
deviance
5
1
0
1
5
ed50
2
3
4
5
6
med.e0
5
1
0
1
5
med.emax
0
.
0
1
.
0
omega.e0
0
.
0
1
.
0
omega.emax
sample
v
a
l
u
e
0
.
2
0
.
6
0 200 400 600 800 1000 1200
sigma
WinBUGS demo
Posterior marginal densities of model parameters
value
f
r
e
q
u
e
n
c
y
0
.
0
0
0
.
0
1
0
.
0
2
0
.
0
3
0
.
0
4
100 80 60 40 20
deviance
0
.
0
0
.
1
0
.
2
0
.
3
0
.
4
0
.
5
5 10 15
ed50
0
1
2
3
4
2 3 4 5 6
med.e0
0
.
0
0
.
2
0
.
4
0
.
6
0
.
8
5 10 15
med.emax
0
5
1
0
0.0 0.5 1.0
omega.e0
0
5
1
0
0.0 0.5 1.0 1.5
omega.emax
0
2
4
6
8
0.0 0.2 0.4 0.6 0.8
sigma
WinBUGS demo
Summary statistics for posterior marginal distributions of model
parameters
Naive Time-series Effective
Mean SD SE SE 2.5% 25% 50% 75% 97.5% N
deviance -76 11.2 0.187 0.607 -92.8 -84.2 -77.8 -69.5 -49.7 345
med.emax 9.98 0.601 0.01 0.0235 8.84 9.66 9.97 10.3 11.1 664
med.e0 2 0.112 0.00186 0.00188 1.82 1.96 2 2.04 2.18 3730
ed50 6.26 1.01 0.0168 0.0719 4.28 5.72 6.25 6.75 8.31 158
omega.emax 0.0961 0.0777 0.0013 0.00166 0.0349 0.0581 0.0778 0.11 0.255 2840
omega.e0 0.0832 0.065 0.00108 0.00142 0.0174 0.0484 0.068 0.0978 0.238 2680
sigma 0.166 0.0782 0.0013 0.00393 0.0793 0.115 0.145 0.194 0.368 430
WinBUGS demo
Model predictions (median + 90% prediction intervals) compared
to observed data
dose
r
e
s
p
o
n
s
e
2
4
6
8
10
12
0 20 40 60 80
1 2
0 20 40 60 80
3
4
0 20 40 60 80
2
4
6
8
10
12
5
dose
r
e
s
p
o
n
s
e
2
4
6
8
10
12
0 20 40 60 80
1 2
0 20 40 60 80
3
4
0 20 40 60 80
2
4
6
8
10
12
5 Posterior
predictions
for new data
in the
same study
Posterior
predictions
for a new
study
Hands-on session 2
Hands-on session 2
Population PK-PD modeling of time-matched biomarker and PK
data
Now we analyze the time course data from the same Phase I study we
analyzed in hands-on session 1:
Phase 1 single dose study in healthy volunteers
Single doses of ME-2
Placebo, 1.25, 5, 10, 15, 20, 30, 40, 60 and 80 mg
Biomarker: ex vivo inhibition of factor Xa activity in plasma
PK and biomarker measured at 0, 0.083, 0.167, 0.25, 0.5, 0.75, 1,
1.5, 2, 3, 4, 6, 8, 12, 18 and 24 hours after dose.
Hands-on exercise:
Apply a direct action PK/PD model to the time-matched factor Xa
inhibition and ME-2 plasma concentrations.
Hands-on session 2
Hands-on session 2
EDA: PK and biomarker data
time (h)
M
E
2

p
l
a
s
m
a

c
o
n
c
e
n
t
r
a
t
i
o
n

(
n
g
/
m
L
)
0
5
1
5
2
5
0 5 10 20
1.25 mg
0
4
0
8
0
1
2
0
5 mg
0
1
0
0
2
5
0
0 5 10 20
10 mg
0
1
0
0
3
0
0
15 mg
0
2
0
0
5
0
0
20 mg
0
2
0
0
6
0
0
30 mg
0
4
0
0
8
0
0
40 mg
0 5 10 20
0
5
0
0
60 mg
0
5
0
0
1
5
0
0
80 mg
time (h)
f
a
c
t
o
r

X
a

i
n
h
i
b
i
t
i
o
n

(
%
)
2
0
0
2
0
05 15 25
0 mg
2
0
0
2
0
4
0
1.25 mg
2
0
2
0
6
0
05 15 25
5 mg
2
0
2
0
6
0
10 mg
0
4
0
8
0
15 mg
0
4
0
8
0
20 mg
0
5
0
1
0
0
30 mg
0
5
0
1
0
0
40 mg
0
5
0
1
0
0
60 mg
05 15 25
0
5
0
1
0
0
80 mg
Hands-on session 2
Hands-on session 2
EDA: Relationship between biomarker and PK data
f
a
c
t
o
r

X
a

i
n
h
i
b
i
t
i
o
n

(
%
)
0
50
100
0.4 0.2
0 mg
05 15 25
1.25 mg
0 4080
5 mg
0100250
10 mg
0100 300
15 mg
0 200 500
20 mg
0 400
30 mg
0 400 1000
0
50
100
40 mg
0
50
100
0500
60 mg
0 1000
80 mg
f
a
c
t
o
r

X
a

i
n
h
i
b
i
t
i
o
n

(
%
)
0
50
100
0 500 1000 1500
Hands-on session 2
Hands-on session 2
Proposed model
Sigmoid Emax model relating % inhibition of factor Xa activity to
ME-2 plasma concentration on the i
th
occasion in the j
th
subject:
E
ij
N
_
E
ij
,
2
_
E
ij
=
E
max
c
ij
EC
50,j
+ c
ij
log
_
EC
50,j
_
N
_
log
_
EC
50
_
,
2
EC
50
_
E
max
U (0, 100) log
_
EC
50
_
N
_
0, 10
6
_
U (0, 10)
EC
50
U
_
0, 10
5
_
U (0, 1000)
Hands-on session 2
Hands-on session 2
Files
Data: fxa.data.csv
Model: me2HandsOn2.txt
R script: me2HandsOn2.R
Alternative model (Emax = 100)
Model: me2HandsOn2b.txt
R script: me2HandsOn2b.R
Topics in Bayesian model development using WinBUGS II Bayesian model evaluation and comparison
Bayesian model evaluation & comparison: The ideal
Mixture of all possible true models including all known
substantive information
Extends the notion of Bayesian modeling of uncertainty to include
uncertainty in model structure
Under such a grand mixture model the tasks of model evaluation,
model comparison and sensitivity analysis are all handled within a
single analysis
Each model in the mixture is associated with a posterior probability
that it is the true (or best) model. This provides a metric for
comparing models.
Sensitivity to model choice is obtained by comparing model-specic
posterior predictions
Like most ideals it cannot be achieved in full, only approached via
practical compromises
Bayesian model evaluation & comparison
Philosophical consequences of the ideal
Model selection is often avoided in favor of using a fuller model
that contains potential competing models as special cases
Particularly avoided is the classic approach of performing a
sequence of hypothesis tests
The fuller model approach acknowledges and quanties the
uncertainty in the relative correctness of competing models.
Types of fuller models
Discrete mixture models, e.g.,
p (y|) = Pr (M
1
) p (y|, M
1
) + Pr (M
2
) p (y|, M
2
) +
+ Pr (M
n
) p (y|, M
n
)
Continuous mixture models, e.g., models containing all covariates
of potential interest
Typical practical Bayesian model development
Propose initial model structure based on available prior
information.
Realistically exploratory analysis of the new data also inuences
this process,
Assess whether model is consistent with the data and prior
information
Posterior predictive checking
Are model inferences, predictions and values of parameters or
other derived quantities consistent with other knowledge?
Assess sensitivity to potentially inuential assumptions, e.g.,
choice of prior distributions
If deciencies are discovered that could adversely affect important
inferences then explore model revisions and reassess as before.
Typical practical Bayesian model development (cont.)
Use the model resulting from this process for pre-planned
inferences.
Optionally there may be additional hypothesis-generating
activities, e.g.,
Data mining efforts to explore the inuence of covariates not
considered in the original model
Further exploration of alternative model structures not considered
in the previous sensitivity analyses
Posterior predictive checking (PPC)
Graphical checks
Comparing data & predictions
Comparing data summaries and model predictions/inferences
Residuals
Formal numerical checks
Posterior predictive p-values based on a test statistic
Graphical checks: Comparing data & predictions
Posterior medians & 90% prediction intervals compared to observed
data
40 mg
individual predictions
time (h)
i
n
h
i
b
i
t
i
o
n

o
f

f
a
c
t
o
r

X
a

a
c
t
i
v
i
t
y

(
%
)
0
50
100
0 5 10 15 20 25
49 50
0 5 10 15 20 25
51
52 53
0
50
100
54
0
50
100
55
0 5 10 15 20 25
56
population predictions
i
n
h
i
b
i
t
i
o
n

o
f

f
a
c
t
o
r

X
a

a
c
t
i
v
i
t
y

(
%
)
0
50
100
0 500 1000 1500
Graphical checks: Comparing data summaries and
model predictions/inferences
Posterior densities of predicted mean peak effect and area under the
effect curve compared to observed means.
mean peak inhibition of factor Xa activity (%)
D
e
n
s
i
t
y
0.00
0.05
0.10
0.15
0.20
15 20 25
Pr(pred >= obs)
=0.376
0 mg
2025303540
Pr(pred >= obs)
=0.182
1.25 mg
40 50 60
Pr(pred >= obs)
=0.764
5 mg
65 75 85
Pr(pred >= obs)
=0.895
10 mg
70 80 90
Pr(pred >= obs)
=0.0928
15 mg
80 90 100
Pr(pred >= obs)
=0.243
20 mg
85 95 105
Pr(pred >= obs)
=0.666
30 mg
85 95 105
0.00
0.05
0.10
0.15
0.20
Pr(pred >= obs)
=0.318
40 mg
0.00
0.05
0.10
0.15
0.20
90 100 110
Pr(pred >= obs)
=0.614
60 mg
95 105 115
Pr(pred >= obs)
=0.213
80 mg
mean inhibition of factor Xa activity AUC
D
e
n
s
i
t
y
0.000
0.002
0.004
0.006
0.008
0.010
0.012
100 0 100
Pr(pred >= obs)
=0.461
0 mg
50 100 250
Pr(pred >= obs)
=0.529
1.25 mg
200 400
Pr(pred >= obs)
=0.57
5 mg
300 500
Pr(pred >= obs)
=0.733
10 mg
400 600 800
Pr(pred >= obs)
=0.557
15 mg
700 1000
Pr(pred >= obs)
=0.396
20 mg
900 1200
Pr(pred >= obs)
=0.986
30 mg
900 1200
0.000
0.002
0.004
0.006
0.008
0.010
0.012
Pr(pred >= obs)
=0.125
40 mg
0.000
0.002
0.004
0.006
0.008
0.010
0.012
1200 1500
Pr(pred >= obs)
=0.05
60 mg
1300 1600
Pr(pred >= obs)
=0.545
80 mg
Graphical checks: Residuals
Uncertainty in residuals and predictions can be depicted with lines
representing prediction intervals
predicted inhibition of factor Xa activity (%)
r
e
s
i
d
u
a
l
s
60
40
20
0
20
40
60
20 0 1020
q
q
q
q
q
q
q
q
qq q
q
q
q
q
q
q
q
qq
q q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q q
q
q
q
q
qq
q
q
q
q
q
qq
q
q
q q
q
q
q
q
q
q q
q q
q
q
q
q
qqq
q
q
qq q
q
q
q
q
q
q
q
q
q
q q
q
q
qq
q q q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
0 mg
20 0 20 40
q
q
qq q
q q
q
q
q
q
q
q
q q
q
q
q
q
qq
q q
q
q q
q
q
q
q
q
q
q
q
q
q q
q
q
q q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q q q
q
qq
q
q
q q
q
q q
q
q q q q
q q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q q
q
q q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
1.25 mg
20 20 60
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q q
q q
q
q
q
q
q
q
q q
q
q
q
q
q q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q q q
q
q
q q
q
q q q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
qq
q
q q q
q
q
q
q
q q q
q
q
q
q
q
q
q
q
q q
q
q q
q
5 mg
20 20 60
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq q
q q q
q q q
q
q
q
q
q
q q
q
q
qq
q
q
q
q
q
q q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
qq
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q q q
q
q
q q
q
q q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
10 mg
20 20 60
q
q q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q q
qq
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q q q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q q
q
q
q
q
q
q q q
q q
q
q
q
q
15 mg
2020 60100
q
q q
q
qq
q
q
q
q
q
q
q q
q q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q q
q
q
q
q
q q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q q q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q q q q
q
q
q
q q
q
q
q
q
q
q
20 mg
2020 60100
q
q
q
qqq
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q q
q
q
qq
q
q
q
q q
q
q
q
q q
q
q
q q
q
q
q
q
q
q q
q
q q q
q
q
q
q
q
q q
q
q
q
q
q
q
qq
q
q q
30 mg
0 50 100
60
40
20
0
20
40
60
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q q
q
q
q
q
q
q
q
q
q q
q q q
q
q
q
qq
q
q
q
q q q
q
q
q
q
q
q
q q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
qq
q qq
q
q q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
40 mg
60
40
20
0
20
40
60
0 50 100
q
q q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q q
q q
q
q
q
q
q
q
q
q
q
q
q
q q
q q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
60 mg
0 50 100
q
q
q
qq q
q
q
q
q
q
q q
q
q
q q
q
q
q
q
q
q
q q
q
q
q
q
q
q q
q
q
q
q
q
qq
q
q
q
q
q q q
q
q
q q
q
q
q
q
q
q q
q
q
q
q
q
q q
q q
qqq
q
q q
q qq
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
qq
q q q q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
qq
q
q
q
q q
q
q
q
q
q
q
q q
80 mg
time (h)
r
e
s
i
d
u
a
l
s
60
40
20
0
20
40
60
0 510 20
q
q
q
q
q
q
q
q
qqq
q
q
q
q
q
q
q
qq
qq
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q q
q
q
q
q
qq
q
q
q
q
q
qq
q
q
q q
q
q
q
q
q
qq
qq
q
q
q
q
q q q
q
q
qqq
q
q
q
q
q
q
q
q
q
q q
q
q
qq
qqq
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
0 mg
0 510 20
q
q
qqq
qq
q
q
q
q
q
q
q q
q
q
q
q
qq
qq
q
qq
q
q
q
q
q
q
q
q
q
qq
q
q
qq
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q q q
q
qq
q
q
qq
q
qq
q
q q q q
q q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
qq
q
qq
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
1.25 mg
0 510 20
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q q
qq
q
q
q
q
q
q
qq
q
q
q
q
q q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
qqq
q
q
q q
q
q qq
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
qq
q
qqq
q
q
q
q
q qq
q
q
q
q
q
q
q
q
q q
q
q q
q
5 mg
0 510 20
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qqq
qqq
qqq
q
q
q
q
q
qq
q
q
qq
q
q
q
q
q
q q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
qq
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
qq q
q
q
q q
q
qq
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
10 mg
0 510 20
q
qq
q
q
q
q
q
q
q
q
q
q
q
q q
q
qq
qq
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
qqq
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
qq
q
q
q
q
q
qqq
q q
q
q
q
q
15 mg
0 510 20
q
qq
q
qq
q
q
q
q
q
q
q q
q q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
qq
q
q
q
q
q q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
qq
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q q q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
qqqq
q
q
q
qq
q
q
q
q
q
q
20 mg
0 510 20
q
q
q
qqq
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
qqq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q q
q
q
qq
q
q
q
qq
q
q
q
q q
q
q
qq
q
q
q
q
q
qq
q
q q q
q
q
q
q
q
qq
q
q
q
q
q
q
q q
q
q q
30 mg
0 510 20
60
40
20
0
20
40
60
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q qq
q
q
q
q
q
q
q
q
q q
q q q
q
q
q
qq
q
q
q
qqq
q
q
q
q
q
q
qq
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
qq
qqq
q
qq
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
40 mg
60
40
20
0
20
40
60
0 510 20
q
qq
qq
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q q
q q
q
q
q
q
q
q
q
q
q
q
q
q q
q q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
60 mg
0 510 20
q
q
q
qqq
q
q
q
q
q
q q
q
q
q q
q
q
q
q
q
q
qq
q
q
q
q
q
q q
q
q
q
q
q
qq
q
q
q
q
q q q
q
q
qq
q
q
q
q
q
qq
q
q
q
q
q
q q
qq
qqq
q
qq
qqq
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q q
q q q q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
qq
q
q
q
qq
q
q
q
q
q
q
q q
80 mg
Sensitivity analysis
Generally desirable to assess sensitivity of key inferences to
model assumptions
True for any modeling strategy, not just Bayesian approaches
Model assumptions to be evaluated
Model structure, i.e., functional forms of the likelihood and prior
distribution
Quantitative assumptions, i.e., values of the parameters of the prior
distribution
Lets focus on the prior distribution since that is unique to Bayesian
modeling
Sensitivity analysis: Prior distribution
When your intention is for the prior distribution to be
non-informative
Compare effect of different prior distributions on important
inferences or predictions
Different degrees of weak informativeness, e.g., normal distributions
with various large variances
Different distribution types, e.g., t-distributions with various degrees of
freedom instead of a normal distribution
Sensitivity analysis: Prior distribution
When your intention is for the prior distribution to be informative
Desirable to assess the relative inuence of prior knowledge and
new data on important inferences.
Sensitivity to the prior distribution is not necessarily a bad thing,
but it is desirable to know the extent to which the prior affects your
inferences.
Should explore a range of priors consistent with the range of
knowledge and beliefs held by other stakeholders in the analysis.
For example:
When evaluating the probable benet of some treatment, consider
priors that reect more or less skepticism than the one used for the
primary analysis.
If the primary analysis concludes the treatment is more benecial
than a comparator:
Consider identifying the least skeptical prior that would make the
conclusion unconvincing.
Assess whether that skeptical prior is plausible.
Bayesian model comparison
Competing models may be compared with respect to the various
modeling evaluation methods just described, i.e.,
Posterior predictive checks
Consistency with other substantive knowledge
Specic metrics for comparing models include:
Expected deviance
DIC
Expected deviance as a measure of prediction error
Deviance
-2 log-likelihood: D(y, ) = 2 log(p (y|))
Like the NONMEM objective function except that it is in the form of
a univariate posterior distribution instead of a single value
Two possible single value metrics for model comparison
Deviance at a point estimate of , e.g., the posterior mean of :
D
(y) = D
_
y,

(y)
_
Expected deviance:
D
avg
(y) = E (D(y, ) |y)
Expected deviance is generally regarded as the better summary
of model error
MCMC samples of deviance can be monitored in WinBUGS
Deviance Information Criterion (DIC)
If the objective is to select the model with the best predictive
power then the expected predictive deviance would be a good
criterion of model t:
D
pred
avg
(y) = E
_
D
_
y
pred
,

(y)
_
|y
_
where y
pred
are predicted observations from the same design as
the original data
The expected predictive deviance may be approximated by the
deviance information criterion (DIC):
D
pred
avg
(y) DIC = 2D
avg
(y) D
(y)
= D
avg
(y) + p
D
where
p
D
= D
avg
(y) D
(y) eective number of parameters

Thus DIC may also be viewed as expected deviance adjusted for
model complexity
Deviance Information Criterion (DIC)
All else being equal, a model with a lower DIC is preferred
DIC may be estimated from posterior samples i according to the
following sequence:
D
avg
(y) =
1
N
N
i =1
D(y,
i
)
p
D
=
1
2
1
N 1
N
i =1
_
D(y,
i
)

D
avg
(y)
_
2
DIC

D
avg
(y) +

p
D
WinBUGS has commands for calculating DIC
Other approaches not covered
Other ICs, e.g., Akaike IC, Bayesian IC, Schwartz IC,
Bayes factors
For model comparison
For model averaging
Other model averaging or mixture model approaches
Topics in Bayesian model development using WinBUGS II BUGSModelLibrary
BUGSModelLibrary
http://bugsmodellibrary.googlecode.com
BUGSModelLibrary is a prototype PKPD model library for use with
WinBUGS 1.4.3. The current version includes:
Specic linear compartmental models:
One compartment model with rst order absorption
Two compartment model with elimination from and rst order
absorption into central compartment
General linear compartmental model described by a matrix
exponential
General compartmental model described by a system of rst order
ODEs
BUGSModelLibrary
The models and data format are based on
NONMEM/NMTRAN/PREDPP conventions including:
Recursive calculation of model predictions
This permits piecewise constant covariate values
Bolus or constant rate inputs into any compartment
Handles single dose, multiple dose and steady-state dosing
histories
Implemented NMTRAN data items include:
TIME, EVID, CMT, AMT, RATE, ADDL, II, SS
BUGSModelLibrary
Calling conventions
All BUGSModelLibrary functions have the form:
<modelName>(time, amt, rate, ii, evid, cmt, addl, ss, theta)
where time, amt, rate, ii, evid, cmt, addl and ss are equal length time-ordered
vectors and are dened identically to the NONMEM variables of the same name.
theta may be either
A vector of parameters of length equal to the number of model parameters,
or
A matrix with a number of rows equal to the length of the time vector. The
i
th
row contains a vector of model parameters for the time interval
(time[i-1], time[i]]. This permits time-dependent parameters.
<modelName> is the name of a built-in model, such as OneCptModel or
TwoCptModel, or a user-dened name.
Within a WinBUGS model <modelName> is usually called once per individual
within a for block that loops over individuals.
BUGSModelLibrary: Built-in models
OneCptModel(time, amt, rate, ii, evid, cmt, addl, ss, theta)
TwoCptModel(time, amt, rate, ii, evid, cmt, addl, ss, theta)
OneCptKaModel(time, amt, rate, ii, evid, cmt, addl, ss, theta)
TwoCptKaModel(time, amt, rate, ii, evid, cmt, addl, ss, theta)
WinBUGS model
function argument parameters
model name names in theta
one compartment
model with rst order
absorption
_
k
a
>
CL
V
2
_
OneCptModel time, amt, rate, ii,
evid, cmt, addl, ss
CL, V
2
, k
a

CL
V
2
, F
1
, F
2
,
t
lag1
, t
lag2
two compartment
absorption (k
a
>
1
)
TwoCptModel time, amt, rate, ii,
evid, cmt, addl, ss
CL, Q, V
2
, V
3
, k
a

1
,
F
1
, F
2
, F
3
, t
lag1
, t
lag2
,
t
lag3
one compartment
absorption
OneCptKaModel time, amt, rate, ii,
evid, cmt, addl, ss
CL, V
2
, k
a
, F
1
, F
2
, t
lag1
,
t
lag2
two compartment
absorption
TwoCptKaModel time, amt, rate, ii,
evid, cmt, addl, ss
CL, Q, V
2
, V
3
, k
a
, F
1
,
F
2
, F
3
, t
lag1
, t
lag2
, t
lag3
BUGSModelLibrary
User-programmed models
Linear compartmental models
User species the non-zero elements of the rate constant matrix.
Linear ODEs are solved using matrix exponential methods.
General compartmental models
User species the ODEs.
The ODEs are solved using either a Runge-Kutta 4th/5th order
method or LSODA, the Livermore Solver for Ordinary Differential
equations with Automatic method switching for stiff and nonstiff
problems.
Both cases require user specication of a rate constant matrix or
ODEs in a template Component Pascal procedure that must be
compiled using the BlackBox Component Builder 1.5.
Topics in Bayesian model development using WinBUGS II Dealing with censored data in WinBUGS
Dealing with censored data in WinBUGS
Right, left and interval censored data are easily handled in
WinBUGS by modifying the likelihood distribution with
I(lower,upper). lower or upper may be blank, meaning no limit.
For example:
Suppose y[i] is normally distributed with some mean mu[i] and
precision tau and
A point value of y[i] is not observed but
You know it lies between 5 and 10
Then we say it is interval censored
The likelihood may described in WinBUGS by
y[i] dnorm(mu[i],tau)I(5,10)
The corresponding y[i] should be NA in the data set.
Topics in Bayesian model development using WinBUGS II Dealing with censored data in WinBUGS
BQL data in WinBUGS
BQL data should be treated as left censored data, i.e., you dont
know the value except that it is < QL.
Recommended approach in WinBUGS:
Enter BQL data as NA.
Include a data vector of the same length as the dependent variable
vector y. Lets call it QL.
QL[i] is the lower quantitation limit for y[i].
Set QL[i] = an impossibly large number if the y[i] is missing for
reasons other than BQL. WinBUGS wont accept an NA for the limits.
QL[i] has no effect when y[i] is not missing, so you can set it to
anything for those cases.
Specify the likelihood as: y[i] ddist(theta)I(,QL[i]) where
ddist is a WinBUGS distribution function, e.g., dnorm.
This is a rigorous approach to BQL data that appropriately
considers the information content of BQL data.
Topics in Bayesian model development using WinBUGS II Demo: Population PK using BUGSModelLibrary
Demo: Population PK using BUGSModelLibrary
Simulated single oral dose PK
20 subjects
1250 mg dose
Lower limit of quantitation = 0.1 mg/L
Samples at 0.25, 0.5, 0.75, 1, 1.5, 2, 3, 4, 5, 6, 8, 10, 16, 24, 32
and 48 h.
Two compartment model with rst order absorption
Taken from the censored example in the PKBugs documentation.
Hands-on session 3
Hands-on session 3
Population PK of ME-2
A phase IIa PoC trial of ME-2 for prevention of post-op VTEs has
just been completed.
One of your tasks is to do a pop PK analysis based on the
accumulated ME-2 PK data (Phase I SD, Phase I MD & Phase IIa)
Phase 1 single dose study in healthy volunteers described during
hands-on sessions1 & 2
Phase 1 multiple dose study in healthy volunteers
Placebo or ME-2 5, 10, 20, 40 or 80 mg bid (q12h) x 7 days
PK measured at 0, 0.083, 0.167, 0.25, 0.5, 0.75, 1, 1.5, 2, 3, 4, 6, 8,
12, 12.1, 12.2, 12.2, 12.5, 12.8, 13, 13.5, 14, 15, 16, 18, 20, 24, 36,
48, 60, 72, 84, 96, 108, 120, 132, 144, 156, 168, 168, 168, 168, 168,
169, 169, 170, 170, 171, 172, 174, 176, 180, 186 and 192 hours after
the rst dose.
Phase IIa trial design:
Treatments
ME-2 20 mg bid (q12h) x 7 days
Enoxaparin 30 mg bid (q12h) x 7 days
100 patients per treatment arm
Sparse ME-2 PK data (3-6 samples/patient)
LOQ = 10 ng/mL
Available patient-specic covariates: weight, age, gender
Hands-on session 3
Hands-on session 3
time (h)
M
E
2

p
l
a
s
m
a

c
o
n
c
e
n
t
r
a
t
i
o
n

(
n
g
/
m
L
)
0
5
1
5
2
5
0 5 10 20
1.25 mg
0
4
0
8
0
1
2
0
5 mg
0
1
0
0
2
5
0
0 5 10 20
10 mg
0
1
0
0
3
0
0
15 mg
0
2
0
0
5
0
0
20 mg
0
2
0
0
6
0
0
30 mg
0
4
0
0
8
0
0
40 mg
0 5 10 20
0
5
0
0
60 mg
0
5
0
0
1
5
0
0
80 mg
ME-2 PK data from
Phase I SD trial
Hands-on session 3
Hands-on session 3
time (h)
M
E
2

p
l
a
s
m
a

c
o
n
c
e
n
t
r
a
t
i
o
n

(
n
g
/
m
L
)
0
5
0
1
0
0
1
5
0
0 50 150
5 mg
0
5
0
1
5
0
2
5
0
10 mg
0
2
0
0
4
0
0
6
0
0
0 50 150
20 mg
0
4
0
0
8
0
0
1
2
0
0
40 mg
0 50 150
0
5
0
0
1
5
0
0
80 mg
ME-2 PK data from
Phase I MD trial
Hands-on session 3
Hands-on session 3
ME-2 PK data from Phase IIa trial
time (h)
M
E
!
2

p
l
a
s
m
a

c
o
n
c
e
n
t
r
a
t
i
o
n

(
n
g
/
m
L
)
0
200
400
600
800
0 50 150
! ! ! !
1
!!
!
!
2
0 50 150
!!
!
!
3
!
!! !
!
4
0 50 150
!
!
! !
!
5
!
!
!
!
6
!
!
!
!
!
!
7
! !
!
!
!
!
8
!
!
!!!
9
0
200
400
600
800
!
!
!
!
10
0
200
400
600
800
!
!
!
!
!
!
11
! !
!
12
!
! !
!
!
13
!
!
! !
!
14
!!
!
15
!
!
! !
16
! !
!!
17
!
!
!
18
!
!
!
!
!
19
0
200
400
600
800
!
!
!
!
20
0
200
400
600
800
!
!
!
!
21
0 50 150
!
!
!
!
!
22
!
!
!
!
!
!
23
0 50 150
!
!
!
24
!
!
!
!
!
25
time (h)
M
E
!
2

p
l
a
s
m
a

c
o
n
c
e
n
t
r
a
t
i
o
n

(
n
g
/
m
L
)
0
200
400
600
800
0 50 150
!
! !
!
!
!
26
!
!
!
!
27
0 50 150
!
! ! !!!
28
! ! !
!
!
29
0 50 150
! !
!
30
!
!
! !
!
31
!
!
!
32
!
!
!! !
!
33
!
!
!
!
34
0
200
400
600
800
!
!
!
!
! !
35
0
200
400
600
800
! !
!
!
36
!
!
!
!
! !
37
!
!
!
!
!
38
!
!
!!
!
!
39
!
!!
! !
!
40
! !
!
! !
41
! !!
!
!
42
!
!
!
43
!
!
!
!
!
!
44
0
200
400
600
800
!
!
!
45
0
200
400
600
800
! ! !
!
46
0 50 150
!
! ! !
47
!
!
!
!
48
0 50 150
!
!
!
!
49
! !
!
!
!
50
Hands-on session 3
Hands-on session 3
ME-2 PK data from Phase IIa trial
time (h)
M
E
!
2

p
l
a
s
m
a

c
o
n
c
e
n
t
r
a
t
i
o
n

(
n
g
/
m
L
)
0
200
400
600
800
0 50 150
!
! !
!
51
!
!
!
!
!
!
52
0 50 150
!
!
!!
!
53
!!
!
!
!
!
54
0 50 150
!!
!
55
! !
!
!!
!
56
!
!!
!
57
!
!
!
!
!
58
!
!
!
59
0
200
400
600
800
!! !
60
0
200
400
600
800
!
!
!
61
!
!
! !
62
!
!
! !
63
!
!
!
!
64
! !
!
65
!
!
!
! !
66
!
!
!
!
67
!
!
!
! !
68
!
!
!
69
0
200
400
600
800
! ! ! !
70
0
200
400
600
800
!
!
!
!
!
71
0 50 150
!
!
!
! !
72
!
!
!
!
!
73
0 50 150
!
!
!
74
!
!
!
!
!!
75
time (h)
M
E
!
2

p
l
a
s
m
a

c
o
n
c
e
n
t
r
a
t
i
o
n

(
n
g
/
m
L
)
0
200
400
600
800
0 50 150
!
!
!
76
!
!
!
! ! !
77
0 50 150
!
!
!
!
!
!
78
!
!
!
!
79
0 50 150
!
!
!
!
!
!
80
!
!
!
81
! !
!
!
!
82
!!
!
83
!
!
!
!
!
84
0
200
400
600
800
! !
! !
!
!
85
0
200
400
600
800
!
!
!
86
!!
!
!
87
!
!
!
!
!
88
!
!
!
!
!
!
89
!
!
!
!
90
!
!
!
!
!
91
!
! !
92
!
!
!!
!
93
!
!
!!
94
0
200
400
600
800
!
!
!
!
!
95
0
200
400
600
800
!
! !
96
0 50 150
!
!
!
!
!
97
!
!
!
!
98
0 50 150
!! !
!!
99
!
!
!
!
! !
100
Hands-on session 3
Hands-on session 3
Proposed base model
Two compartment model with rst order absorption describing ME-2 plasma
concentration on the i
th
occasion in the j
th
subject as a function of time, dose and
body weight:
log
_
c
ij
_
N
_
log
_
c
ij
_
,
2
_
c
ij
= f
2cpt
_
t
ij
, D
j
,
j
, CL
j
, Q
j
, V
1j
, V
2j
, k
aj
_
log
_
CL
j
, Q
j
, V
1j
, V
2j
, k
aj
_
N
_
log
_
CL
_
bw
j
70
_
0.75
,
Q
_
bw
j
70
_
0.75
,
V
1
_
bw
j
70
_
,
V
2
_
bw
j
70
_
,
k
a
_
,
_
log
_
CL
_
N
_
0, 10
6
_
log
_
Q
_
N
_
0, 10
6
_
log
_
V
1
_
N
_
0, 10
6
_
log
_
V
2
_
N
_
0, 10
6
_
log
_
k
a

1
_
N
_
0, 10
6
_
U (0, 100)
1
Wishart
_
_
_
_
_
5
_
_
_
_
_
0.05 0 0 0 0
0 0.05 0 0 0
0 0 0.05 0 0
0 0 0 0.05 0
0 0 0 0 0.05
_
_
_
_
_
, 5
_
_
_
_
_
Hands-on session 3
Hands-on session 3
Files
Data
ME-2 plasma concentration data (including all covariates) in
NONMEM format: fxaNONMEMData.csv
Topics in Bayesian model development using WinBUGS III Use of informative priors
Use of informative priors in clinical pharmacology
applications
Physiologic/mechanistic knowledge in PB-PK/PD models
Integration of prior modeling results
Biomarker-to-clinical outcome models
Informative priors for parameters not identiable with new data set,
e.g., ka for a rapidly absorbed drug in popPK modeling of sparse
data
Historical controls in clinical trials
Synthesis of new and past data
Bayesian updating vs analysis of combined data
Topics in Bayesian model development using WinBUGS III Use of informative priors
Use of informative priors in clinical pharmacology
applications
Some issues to consider
Relevance of prior knowledge:
Are past studies really exchangeable with new ones or should the
past info be discounted in some way to account for differences
(known or unknown)?
Are the patients comparable?
Has the standard of care or measurement changed?
If using published info, is there a risk of signicant publication bias?
Is the prior acceptable to the key stakeholders in the inferences
resulting from the analysis?
You may need to construct a prior by consensus or do a sensitivity
analysis to explore how the inference changes for different priors
preferred by particular stakeholders.
Topics in Bayesian model development using WinBUGS III Use of the cut() function
Use of the cut() function
Analogy to sequential vs simultaneous PK/PD modeling
Reasonable to use cut() to prevent feedback of PD data to the PK
model when
The PK model is well-characterized by the PK data alone, and
There is substantial risk that PD model misspecication would
cause inappropriate effects on the posterior distributions of the PK
model parameters
The uncertainty in the PK model parameters is still considered in
the tting to the PD data, but the PD data has no effect on the PK
parameter estimates.
Topics in Bayesian model development using WinBUGS III Use of the cut() function
Use of the cut() function
Excerpt from a simultaneous PK/PD model:
for ( i in 1:nSubjects){
.
.
.
xhat[ start [ i ]: end[i ],1:3] < TwoCptKaModel(time[start[i]:end[i]], amt[start [ i ]: end[i ]],
rate[ start [ i ]: end[i ]], ii [ start [ i ]: end[i ]],
evid[ start [ i ]: end[i ]], cmt[ start [ i ]: end[i ]],
addl[ start [ i ]: end[i ]], ss[ start [ i ]: end[i ]],
theta[ i ,])
}
for ( i in 1:nObservations){
## PD observations
E[ i ] dnorm(EHat[i], tauPD)
EHat[i ] < e0 + emax cut(cHat[i]) / (ec50 + cut(cHat[i]))
## PK observations
logCObs[i] dnorm(logCHat[i], tau)
cHat < xhat[i, 2] / theta[ subject[ i ], 3]
}
Topics in Bayesian model development using WinBUGS III Programming ODE-based models
BUGSModelLibrary
User-programmed models
Linear compartmental models
User species the non-zero elements of the rate constant matrix.
Linear ODEs are solved using matrix exponential methods.
General compartmental models
User species the ODEs.
The ODEs are solved using either a Runge-Kutta or an adaptive
multistep (LSODA) method
Both cases require user specication of a rate constant matrix or
ODEs in a template Component Pascal procedure that must be
compiled using the BlackBox Component Builder 1.5.
Linear compartment model
model described by a system of rst order linear differential
equations with (piecewise) constant coefcients:
x
(t ) = Kx (t )
where K is a matrix.
For example K for a two compartment PK model with rst order
absorption is:
K =
_
_
k
a
0 0
k
a
(k
10
+ k
12
) k
21
0 k
12
k
21
_
_
When applicable this method is usually faster than the
Runge-Kutta or LSODA methods.
Linear compartment model
Component Pascal code for a two compartment PK model with rst
order absorption
PROCEDURE UserKMatrix(IN theta: ARRAY OF REAL;
nCmt: INTEGER):
POINTER TO ARRAY OF ARRAY OF REAL;
VAR
kMatrix: POINTER TO ARRAY OF ARRAY OF REAL;
i , j : INTEGER;
k10, k12, k21, ka: REAL;
BEGIN
NEW(kMatrix,nCmt,nCmt);
( Initialize to all zeros )
FOR i := 0 TO nCmt1 DO;
FOR j := 0 TO nCmt1 DO;
kMatrix[ i , j ] := 0;
END; END;
k10 := theta[0];
k12 := theta[1];
k21 := theta[2];
ka := theta[3];
( Assign nonzero rate constants )
kMatrix[0,0] := -ka;
kMatrix[1,0] := ka;
kMatrix[1,1] := -(k10+k12);
kMatrix[1,2] := k21;
kMatrix[2,1] := k12;
kMatrix[2,2] := -k21;
RETURN kMatrix;
END UserKMatrix;
PROCEDURE (m: MatExpModel) InitModel;
BEGIN
m.nParameter := 10;
m.F1Index := 4;
m.tlag1Index := 7;
m.nCmt := 3;
END InitModel;
Only the red portions need to be programmed by the user. The
remainder is from a template provided with BUGSModelLibrary.
Demo: Two compartment PK model with effect
compartment and rst order absorption
See the BUGSModelLibrary User Manual, pp 1113
General compartment model
model described by a system of rst order ordinary differential
equations (ODEs), i.e., differential equations of the form:
x
(t ) = f (t , x (t ))
where x and f are vector-valued functions.
For example a two compartment PK model with Michaelis-Menten
elimination:
x
1
(t ) =
_
V
max
K
m
+ x
1
(t )
+ k
12
_
x
1
(t ) + k
21
x
2
(t )
x
2
(t ) = k
12
x
1
(t ) k
21
x
2
(t )
Two ODE solving methods are available:
A Runge-Kutta 4th/5th order methodusually faster for non-stiff
problems.
LSODA, the Livermore Solver for Ordinary Differential equations
with Automatic method switching for stiff and nonstiff problems
General compartment model
Component Pascal code for a two compartment PK model with
Michaelis-Menten elimination
PROCEDURE UserDerivatives(IN theta, x: ARRAY OF REAL;
numEq: INTEGER; t: REAL; OUT dxdt: ARRAY OF REAL) ;
VAR
Vmax, Km, k12, k21: REAL;
BEGIN
Vmax := theta[0];
Km := theta[1];
k12 := theta[2];
k21 := theta[3];
( Differential equations for the model excluding piecewise )
( constant input rates provided in the data set )
dxdt[0] := -(Vmax / (Km + x[0]) + k12) * x[0] + k21 * x[1];
dxdt[1] := k12 * x[0] - k21 * x[1];
END UserDerivatives;
Demo: Two compartment PK model + indirect effect
model with drug effect on kout (inhibitory Emax)
See the BUGSModelLibrary User Manual, pp 1319
Hands-on session 4
Hands-on session 4
Population PK/PD modeling of ME-2 induced neutropenia
Neutropenia was observed in some subjects receiving higher
ME-2 doses in Phase I.
Your task is to model the relationship between neutrophil counts
and drug exposure.
Phase 1 multiple dose study in healthy volunteers
Placebo or ME-2 5, 10, 20, 40 or 80 mg bid (q12h) x 7 days
PK measured at 0, 0.083, 0.167, 0.25, 0.5, 0.75, 1, 1.5, 2, 3, 4, 6, 8,
12, 12.1, 12.2, 12.2, 12.5, 12.8, 13, 13.5, 14, 15, 16, 18, 20, 24, 36,
48, 60, 72, 84, 96, 108, 120, 132, 144, 156, 168, 168, 168, 168, 168,
169, 169, 170, 170, 171, 172, 174, 176, 180, 186 and 192 hours after
the rst dose.
LOQ = 10 ng/mL
PD: Absolute neutrophil count (ANC) measured daily for 12 days
Available patient-specic covariates: weight, age, gender
Hands-on session 4
Hands-on session 4
ME-2 PK/PD data from Phase I MD trial
time (h)
M
E
2

p
l
a
s
m
a

c
o
n
c
e
n
t
r
a
t
i
o
n

(
n
g
/
m
L
)
0
5
0
1
0
0
1
5
0
0 50 150
5 mg
0
5
0
1
5
0
2
5
0
10 mg
0
2
0
0
4
0
0
6
0
0
0 50 150
20 mg
0
4
0
0
8
0
0
1
2
0
0
40 mg
0 50 150
0
5
0
0
1
5
0
0
80 mg
time (h)
n
e
u
t
r
o
p
h
i
l

c
o
u
n
t

(
1
0
9
/
m
L
)
0
5
10
15
050 150250
0 mg 5 mg
050 150250
10 mg
20 mg
050 150250
40 mg
0
5
10
15
80 mg
Hands-on session 4
Hands-on session 4
Proposed model
Friberg-Karlsson semi-mechanistic model for drug-induced
myelosuppression [19, 20, 21, 22, 23, 14]
Figure 2 of reference [19]
Hands-on session 4
Hands-on session 4
Friberg-Karlsson semi-mechanistic model for drug-induced
myelosuppression
dProl
dt
= k
prol
Prol (1 E
drug
)
_
Circ
0
Circ
_
k
tr
Prol
dTransit 1
dt
= k
tr
Prol k
tr
Transit 1
dTransit 2
dt
= k
tr
Transit 1 k
tr
Transit 2
dTransit 3
dt
= k
tr
Transit 2 k
tr
Transit 3
dCirc
dt
= k
tr
Transit 3 k
circ
Circ
E
drug
=
c
k
prol
= k
circ
= k
tr
MTT =
n + 1
k
tr
c plasma drug concentration

Circ absolute neutrophil count (ANC)
Parameters in red are system
parameters, i.e., drug-independent.
Hands-on session 4
Hands-on session 4
PK model
PK model: 2 compartment with rst order absorption
Same as before except except we will use only the Phase I MD
data.
Arguably a more comprehensive approach would be to use all
available PK data, i.e., also include Phase I SD data.
We could also do a staged modeling approach where we use the
previous PK modeling results as informative priors and t only the
PD data.
Hands-on session 4
Hands-on session 4
Stochastic model components
Inter-individual variation in PD parameters:
log
_
MTT
j
_
N
_
log
_
MTT
_
,
2
MTT
_
log
_
Circ
0j
_
N
_
log
_
Circ
0
_
,
2
Circ
0
_
log
_
j
_
N
_
log( ) ,
2
_
Log-normal residual variation in ANC:
log
_
ANC
ij
_
N
_
Circ
ij
,
2
ANC
_
Hands-on session 4
Hands-on session 4
Prior distributions
Informative priors for PD system parameters constructed using
values reported in [19, 20, 21, 22, 23, 14]
Prior mean = mean of published values
Prior sd = 2 sd of published values.
Normal priors for log
_
Circ
0
_
, log
_
MTT
_
and log()
Gamma priors for 1/
2
Circ
0
and 1/
2
MTT
mean standard deviation
log
_
Circ
0
_
log(5.4) 0.20
log
_
MTT
_
log(110) 0.16
log() log(0.16) 0.16
1/
2
Circ
0
11 21
1/
2
MTT
37 61
Hands-on session 4
Hands-on session 4
Prior distributions
Weakly informative priors for the rest...
U (0, 1)
U (0, 5)
Hands-on session 4
Hands-on session 4
Files
Data
ME-2 plasma concentration data (including all covariates) in
NONMEM format: me2HandsOn4NONMEMData.csv
BUGSModelLibrary module: TwoCptNeut1ModelRK45.odc
Selected published examples
Selected published examples of Bayesian applications
Lets revisit the motivating scenarios posed at the beginning of the
course and see how others have approached similar situations.
Selected published examples Physiologically-based PK models
You need to characterize the risk to human workers posed by
VOC-X in the air.
VOC-X is a volatile organic compound known to cause a range of
adverse health effects.
Because of its toxicity we cannot perform controlled studies
involving the administration of VOC-X to humans.
As a result we have no direct knowledge about TK, the
relationships between atmospheric exposure, systemic exposure,
and toxicity in humans.
We do have such information for 3 other species.
Limited information is available for closely related compounds.
How would you synthesize this information to make quantitative
inferences about the probable ranges of systemic exposure and
toxicity in humans?
How would you characterize the uncertainty in your inferences?
Bayesian population PK using physiologically-based
models
A Bayesian approach to PB-PK modeling is advantageous for
combining physiologic knowledge ranging in degree of uncertainty
with new data.
The result is a more comprehensive approach than the following
classical approaches to PB-PK modeling:
A priori models with xed parameter values that are only compared
with new data rather than adjusted for the data,
Or at the other extreme, attempting to apply conventional nonlinear
regression to t the new data by somewhat arbitrarily selecting only
a subset of the model parameters to be adjusted.
models
Uses complex models with physiologically relevant parameters,
e.g., tissue blood ows, tissue volumes, tissue/blood partition
coefcients, etc.
Prior distributions based on a combination of information sources:
Physiologic knowledge
Nonclinical data: animal & in vitro
Results are expressed in probabilistic terms to communicate the
uncertainty in predicted quantities and inferences.
Classically PB-PK results were expressed only as point estimates
without measures of uncertainty.
models
The following slides briey illustrate early Bayesian work of Bois et
al:
FY Bois, ET Jackson, K Pakari, MT Smith. Population toxicokinetics
of benzene. Environ Health Perspect 104(Suppl 6):1405-1411
(1996) (http:
//ehpnet1.niehs.nih.gov/docs/1996/Suppl%286%29/bois.html)
A Gelman, F Bois, J Jiang. Physiological pharmacokinetic analysis
using population modeling and informative prior distributions. JASA
91:1400-1412 (1996).
FY Bois, A Gelman, J Jiang, DR Maszle, L Zeise, G Alexeef.
Population toxicokinetics of tetrachloroethylene. Arch Toxicol
70:347-355 (1996).
Since then Bayesian methods have grown to common use in the
environmental toxicology community.
Population toxicokinetics of benzeneBois et al
Schematic representation of the
ve-compartment physiological
model used to simulate the
distribution and metabolism of
benzene.
Results: Posterior Distribution for Fraction of Benzene
Metabolized
Simulated relationship between benzene
exposure concentration and bone marrow
metabolism in humans (geometric mean
population prediction 1 and 3
population SDs including uncertainty and
variability).
Estimated population distribution of the
fraction of benzene metabolized by
humans. The histogram represents both
uncertainty and variability.
Physiological pharmacokinetic analysis using
population modeling and informative prior
distributionsGelman et al
Five key features, all of which work in combination:
1
a physiological model
2
a population model
3
prior information on the population physiological parameters
4
experimental data
5
Bayesian inference
Physiological pharmacokinetic analysis using
population modeling and informative prior
distributionsGelman et al
If any of these ve features are missing, the model will not work:
1. Without a physiological model, there is no good way to
obtain prior information on the parameters;
2. Without a population model, there is not generally enough
data to estimate the model independently on each
individual;
3&4. The parameters of a multicompartment physiological
model cannot be determined accurately by data or prior
information alone;
5. Bayesian inference yields a distribution of parameters
consistent with both prior information and data, if such
agreement is possible.
Selected published examples Bayesian decision analysis for treatment optimization
You need to recommend a dose or dose range for drug X that
optimizes the trade-offs between benets and risks.
Phase II has been completed.
Dose-response information is available for key efcacy and
safety-related outcomes of drug X treatment.
How would you synthesize the cumulative quantitative evidence on
safety and efcacy in order to predict outcomes for various doses?
How would you quantify the uncertainty in those predictions?
How would you use those predictions in conjunction with patient
and prescriber opinions regarding acceptable risk relative to
benet?
Bayesian decision analysis to optimize treatment
regimens
Bayesian PD modeling and decision analysis to
optimize the oxybutynin dosage regimen
Indication: Urinary incontinence (UI) Efcacy:
UI episodes/week
PopPD model for Poisson distributed observations (counts)
Toxicity:
Dry mouth (0-3 ordinal scale)
PopPD model for ordered categorical observations
Thumbnail sketch of Bayesian decision analysis
Utility
state of nature, e.g., parameters of a model predicting the
consequences c of an action a
p
|x
(|x) p
x
(x|) p
()
x observations
a action that has consequence c
c consequence of an action a given the state of nature
p
c
(c|, a)
u (c) utility of consequence c. [u (c) R]
U (, a) utility of an action a given the state of nature
= E (u (c) |, a)
Optimal decision is to take the action that maximizes the Bayesian
posterior expected utility (aka average gain):
a
optimal
= arg max
a
E (U (, a)) = arg max
a
_
U (, a) p
|x
(|x) d
Bayesian decision analysis for oxybutynin treatment
Proposed utility for oxybutynin treatment
U (D) = w
tox
U
tox
(D) +w
eff
U
eff
(D)
U
eff
(D) =
_
1 expected UI event rate following dose D < 6
0 expected UI event rate following dose D 6
U
tox
(D) =
_
_
_
_
_
_
_
0 Pr (severe dry mouth) < 0.1 & Pr (moderate dry mouth) < 0.7
0.5 Pr (severe dry mouth) 0.1 & Pr (moderate dry mouth) < 0.7
0.5 Pr (severe dry mouth) < 0.1 & Pr (moderate dry mouth) 0.7
1 Pr (severe dry mouth) 0.1 & Pr (moderate dry mouth) 0.7
Optimal oxybutynin dose is the one that maximizes the expected
value of U(D)
Expectation over both interpatient variation and model parameter
uncertainty
Results: Optimal dose 10 mg
Maximum expected
utility
Optimal dose
Selected published examples Bayesian adaptive trial design
You need to design a Phase II strategy for a new drug with an
unprecedented MOA for treating a life-threatening disorder for
which there is no effective treatment.
There is large uncertainty in whether the known pharmacological
effects of the drug will translate into therapeutic benet,
And even more uncertainty in dose-response for both efcacy and
safety-related responses.
As a result, a conventional dose nding trial using 3-4 active doses
has a high risk of failing to yield good dose-response information.
How would you design an efcient dose-nding and PoC strategy
that is likely to yield relatively conclusive inferences regarding
go/no-go and dose selection?
Dose-ranging trial design with Bayesian adaptive dose
assignment and adaptive stopping: The ASTIN trial
Acute Stroke Therapy by Inhibition of Neutrophils (ASTIN)
M Krams, KR Lees, W Hacke, AP Grieve, J-M Orgogozo, GA Ford.
Stroke 34:2543-2548 (2003)
Theoretical and computational method development by Don Berry
& Peter M uller
Sponsored by Pzer
Phase II dose-ranging/PoC trial.
Identify dose(s) for subsequent conrmatory trial.
Go/no go decision
Treatments
Neutrophil inhibitory factor (NIF) (UK-279276)
0 (placebo), 10, 16, 22, 27, 33, 38, 45, 52, 59, 67, 78, 84, 96, 108
and 120 mg
Single 15 min infusion administered within 6 hours of a stroke
Endpoint = Scandinavian Stroke Scale (SSS) at 90 days after stroke
Additional SSS measurements obtained at multiple visits prior to 90 days.
Dose-ranging trial design with Bayesian adaptive dose
assignment and adaptive stopping: The ASTIN trial
new patient
termination rule
transition to
pivotal trial
terminate
development
continue
find optimal dose for
learning about ED95
Adapted from Figure 2.1 of Berry, Mller et al (2000)
start
randomization to
placebo or optimal
dose
patient
treatment
patient data:
interim or
endpoint
longitudinal model to
predict endpoint
estimation of
dose-response
model
Bayesian dose assignment and stopping decisions
Adaptive dose assignment
Assign dose to the next patient that minimizes the expected
variance of the response to the optimal dose (+ random assignment
of placebo to 15% of patients).
Optimal dose = ED
95
.
Uses Bayesian modeling of the response as a function of dose and
time.
Adaptive stopping rule
Decision based on response to ED
95
: R (ED
95
) = SSS difference
from placebo
Decide among 3 possible actions each week:
Pr (R (ED
95
) < 1) > 90%& 500 evaluable patients: Stop trial and
abandon development
Pr (R (ED
95
) > 2) > 90%& 250 evaluable patients: Transition to
conrmatory trial
Otherwise: Continue trial
Maximum of 1300 evaluable patients
Probability that the treatment is
ineffective exceeded 90% at 40
weeks
>
Dose assignments & nal
estimated dose response curve
(mean & 95% credible intervals)
Conclusions/commentary
NIF development terminated based on ASTIN trial
746 evaluable patients led to unambiguous conclusion that NIF is
ineffective
Traditional Phase II dose-nding trial with 3 active doses +
placebo:
Would require 1080 patients for 80
Would be less informative about dose-reponse
More likely to leave open question regarding whether untested doses
might have been effective.
Additional topics & closing discussion Deciding between Bayesian or ML modeling
Why Bayesian analysis (vs maximum likelihood)?
Fully Bayesian treatment of uncertainty in parameters and
predictions
Estimated, predicted and derived quantities may be presented in
the form of samples from posterior distributions
Ability to combine new data with prior information
No series or linear approximations of the likelihood function
Improved estimation performance
Particularly for generalized hierarchical models, e.g, models for
categorical data
However this is not an issue for some of the more recently
implemented ML methods
Why not Bayesian analysis (vs maximum likelihood)?
Can be much more computationally intensive
BUGS vs NONMEM BAYES method
BUGS allows a much more exible stochastic structure
No restriction on number of levels of random effects
Many built-in distribution functions plus ability to add more
Easier to combine models for multiple types of data, e.g.:
Individual patient data + summary data
Preclinical + clinical data
NONMEM provides a wider range of estimation methods within a
single platform.
More rapid model exploration possible using estimation of posterior
modes
Comparative performance (accuracy, precision, computation time)
varies, i.e., sometimes WinBUGS does better, sometimes
NONMEM.
Additional topics & closing discussion Other considerations with Bayesian approaches
Other considerations with Bayesian approaches
Regulatory applications
Regulatory authorities are still skeptical of Bayesian analyses of
pivotal trials.
Recent statements by key FDA staff indicate greater openness.
But they are still likely to demand good frequentist operating
characteristics, and assessments of sensitivity to both choice of
prior distributions and model structure (good practices anyway).
There is increasing acceptance and sometimes encouragement
for model-based approaches, Bayesian or otherwise, to aid
decisions such as dose selection and trial design.
Additional topics & closing discussion Other considerations with Bayesian approaches
Other considerations with Bayesian approaches
Specifying prior distributions
Proper construction of prior distributions is (or at least should be)
hard work:
Constructing informative prior distributions is a somewhat
subjective process. Even when based on hard evidence there is
usually some subjectivity in the selection of the evidence to use.
Stakeholder agreement on priors (and the model) is a prerequisite
for their acceptance of inferences based on them.
Sensitivity to choice of priors should be evaluated even when using
supposedly uninformative priors
Additional topics & closing discussion Efforts to facilitate use of Bayesian methods in pharmacometrics
Efforts to address limitations & facilitate use of
Bayesian methods in pharmacometrics
Develop computational methods and open-source software tools for Bayesian
modeling and simulation relevant to pharmacometric applications
Short term efforts
Further development of BUGSModelLibrary
Port to OpenBUGS
Automation to simplify specication of user-programmed models
More built-in models
Tools for distributed computing of multiple chains
bugsParallel, a working prototype using MPICH2 and R with a modied
version of R2WinBUGS, is available from Metrum
Long-term plans:
Develop a more comprehensive platform for Bayesian M&S
Implement both MCMC and estimation of posterior modes. Efcient estimation
of posterior modes would facilitate rapid exploratory modeling
Open source with greater platform independence
Probably interfaced with R
Support for parallel computing
Suite of tools for analysis of MCMC samples (probably in R)
Provide short courses in Bayesian modeling for pharmacometric applications
Additional topics & closing discussion What didnt we cover
What didnt we cover
Categorical, count and time-to-event data
Custom WinBUGS functions and distributions
Population and trial simulations based on MCMC results
Selected publications
Selected References I
[1] A. Gelman, J. B. Carlin, H. S. Stern, and D. B. Rubin.
Bayesian data analysis.
Chapman & Hall/CRC, New York, 2004.
[2] Christian P. Robert.
The Bayesian Choice: From Decision-Theoretic Foundations to Computational
Implementation.
Springer, second edition, 2007.
[3] B. P. Carlin and T. A. Louis.
Bayesian Methods for Data Analysis.
Chapman & Hall/CRC, third edition, 2008.
[4] J.O. Berger.
Statistical Decision Theory and Bayesian Analysis.
[5] D.J. Lunn, A. Thomas, N. Best, and D. Spiegelhalter.
Winbugs a bayesian modelling framework: concepts, structure, and extensibility.
Statistics and Computing, 10:325337, 2000.
Selected References II
[6] D. J. Spiegelhalter, K. R. Abrams, and J. P. Myles.
Bayesian Approaches to Clinical Trials and Health-Care Evaluation.
Wiley, 2004.
[7] Christian P. Robert and George Casella.
Introducing Monte Carlo Methods with R.
Springer, 2010.
[8] Christian P. Robert and George Casella.
Monte Carlo Statistical Methods.
[9] C. Dansirikul, R. G. Morris, S. E. Tett, and S. B. Duffull.
A bayesian approach for population pharmacokinetic modelling of sirolimus.
Br.J Clin.Pharmacol., 62(4):420434, 2006.
[10] M. G. Dodds and P. Vicini.
Assessing convergence of markov chain monte carlo simulations in hierarchical bayesian
models for population pharmacokinetics.
Ann.Biomed.Eng, 32(9):13001313, 2004.
Selected References III
[11] A. Dokoumetzidis and L. Aarons.
Propagation of population pharmacokinetic information using a bayesian approach:
comparison with meta-analysis.
J Pharmacokinet.Pharmacodyn., 32(3-4):401418, 2005.
[12] L. E. Friberg, G. K. Isbister, L. P. Hackett, and S. B. Duffull.
The population pharmacokinetics of citalopram after deliberate self-poisoning: a bayesian
approach.
J Pharmacokinet Pharmacodyn, 32(3-4):571605, 2005.
[13] I. Gueorguieva, L. Aarons, and M. Rowland.
Diazepam pharamacokinetics from preclinical to phase i using a bayesian population
physiologically based pharmacokinetic model with informative prior distributions in winbugs.
J Pharmacokinet.Pharmacodyn., 33(5):571594, 2006.
[14] Steven J Kathman, Daphne H Williams, Jeffrey P Hodge, and Mohammed Dar.
A bayesian population pk-pd model for ispinesib/docetaxel combination-induced
myelosuppression.
Cancer Chemother Pharmacol, 63(3):469476, 2009 Feb.
[15] D. J. Lunn, N. Best, A. Thomas, J. Wakeeld, and D. Spiegelhalter.
Bayesian analysis of population pk/pd models: general concepts and software.
Journal of Pharmacokinetics and Pharmacodynamics, 29(3):271307, 2002.
Selected References IV
[16] S. Mu and T. M. Ludden.
Estimation of population pharmacokinetic parameters in the presence of non-compliance.
Journal of Pharmacokinetics and Pharmacodynamics, 30(1):5381, 2003.
[17] M. K. Smith, I. Jones, M. F. Morris, A. P. Grieve, and K. Tan.
Implementation of a bayesian adaptive design in a proof of concept study.
Pharm Stat, 5(1):3950, 2006.
[18] M. K. Smith and S. Marshall.
A bayesian design and analysis for dose-response using informative prior information.
J Biopharm Stat, 16(5):695709, 2006.
[19] L. E. Friberg, A. Henningsson, H. Maas, L. Nguyen, and M. O. Karlsson.
Model of chemotherapy-induced myelosuppression with parameter consistency across
drugs.
J Clin Oncol, 20(24):471321, 2002.
[20] L. E. Friberg and M. O. Karlsson.
Mechanistic models for myelosuppression.
Invest New Drugs, 21(2):183194, 2003.
Selected References V
[21] J. E. Latz, M. O. Karlsson, J. J. Rusthoven, A. Ghosh, and R. D. Johnson.
A semimechanistic-physiologic population pharmacokinetic/pharmacodynamic model for
neutropenia following pemetrexed therapy.
Cancer Chemotherapy and Pharmacology, 57(4):412426, 2006.
[22] I. F. Troconiz, M. J. Garrido, C. Segura, J. M. Cendros, P. Principe, C. Peraire, and
R. Obach.
Phase i dose-nding study and a pharmacokinetic/pharmacodynamic analysis of the
neutropenic response of intravenous diomotecan in patients with advanced malignant
tumours.
Cancer Chemother Pharmacol, 57(6):72735, 2006.
[23] S. J. Kathman, D. H. Williams, J. P. Hodge, and M. Dar.
A bayesian population pk-pd model of ispinesib-induced myelosuppression.
Clin Pharmacol Ther, 81(1):8894, 2007.
COMPLETE COURSEWARE LICENSE AGREEMENT
COMPLETE COURSEWARE LICENSE AGREEMENT I
RIGHTS RESERVED: THIS COMPLETE COURSEWARE LICENSE IS SEPARATE AND DISTINCT FROM THE OPEN
COURSEWARE LICENSE, WHICH IS PROVIDED FOR PUBLIC USE OF A SPECIFIED SUBSET OF METRUM INSTITUTE
TRAINING MATERIALS. SEE Courseware Info at http://metruminstitute.org.
THE WORK IS PROTECTED BY COPYRIGHT AND/OR OTHER APPLICABLE LAW. ANY USE OF THE WORK OTHER THAN
AS AUTHORIZED UNDER THIS LICENSE OR COPYRIGHT LAW IS PROHIBITED. BY EXERCISING ANY RIGHTS TO THE
WORK PROVIDED HERE, YOU ACCEPT AND AGREE TO BE BOUND BY THE TERMS OF THIS LICENSE. TO THE EXTENT
THIS LICENSE MAY BE CONSIDERED TO BE A CONTRACT, THE LICENSOR GRANTS YOU THE RIGHTS CONTAINED
HERE IN CONSIDERATION OF YOUR ACCEPTANCE OF SUCH TERMS AND CONDITIONS.
1. Denitions
Adaptation means a work based upon the Work, or upon the Work and other pre-existing works, such as a translation,
adaptation, derivative work, or any other form in which the Work may be recast, transformed, or adapted including in any form
recognizably derived from the original, except that a work that constitutes a Collection will not be considered an Adaptation for the
purpose of this License.
Collection means a collection of literary or artistic works, such as encyclopedias and anthologies, or performances,
phonograms or broadcasts, or other works or subject matter other than works listed in Section 1(f) below, which, by reason of the
selection and arrangement of their contents, constitute intellectual creations, in which the Work is included in its entirety in
unmodied form along with one or more other contributions, each constituting separate and independent works in themselves,
which together are assembled into a collective whole. A work that constitutes a Collection will not be considered an Adaptation
(as dened above) for the purposes of this License.
Distribute means to make available to anyone other than You, the original and copies of the Work or Adaptation, as
appropriate, through sale or other transfer of ownership.
Licensor means the individual, individuals, entity or entities that offer(s) the Work under the terms of this License.
Original Author means the individual, individuals, entity or entities who created the Work or if no individual or entity can be
identied, the publisher.
Work means the scientic work offered under the terms of this License including without limitation any production in the literary,
scientic and artistic domain, whatever may be the mode or form of its expression including digital form, such as a book,
COMPLETE COURSEWARE LICENSE AGREEMENT II
pamphlet and other writing; a lecture, address, sermon or other work of the same nature; a broadcast; a phonogram; a
compilation of data to the extent it is protected as a copyrightable work.
You means an individual or entity exercising rights under this License who has not previously violated the terms of this License
with respect to the Work, or who has received express permission from the Licensor to exercise rights under this License despite
a previous violation.
Publicly Perform means to perform public recitations of the Work and to communicate to the public those public recitations, by
any means or process, including by wire or wireless means or public digital performances; to make available to the public Works
in such a way that members of the public may access these Works from a place and at a place individually chosen by them; to
perform the Work to the public by any means or process and the communication to the public of the performances of the Work,
including by public digital performance; to broadcast and rebroadcast the Work by any means
including signs, sounds or images.
Reproduce means to make copies of the Work by any means including without limitation by sound or visual recordings and the
right of xation and reproducing xations of the Work, including storage of a protected performance or phonogram in digital form
or other electronic medium.
Commercial means any copy, application or Adaptation of the Work for the direct or indirect purpose of revenue generation
Non-Commercial means any copy, application or Adaptation of the Work for purposes that are not directly or indirectly related
to revenue generation.
2. Fair Dealing Rights. Nothing in this License is intended to reduce, limit, or restrict any uses free from copyright or rights arising
from limitations or exceptions that are provided for in connection with the copyright protection under copyright law or other
applicable laws.
3. License Grant. Subject to the terms and conditions of this License, Licensor hereby grants You fair use of this copyrighted
Work for noncommercial educational purposes or for commercial and noncommercial research purposes. For example, You may
adapt example code sets (the Work) and apply them to a commercial research project.
4. Restrictions. The license granted in Section 3 above is expressly made subject to and limited by the following restrictions:
Under the terms of this License, You may not Distribute, Reproduce, or Publicly Perform the Work. For example, You may not use
courseware (the Work or any adaptations) to create a new public training course, whether commercial or non-commercial.
COMPLETE COURSEWARE LICENSE AGREEMENT III
You may not implicitly or explicitly assert or imply any connection with, sponsorship or endorsement by the Original Author,
Licensor and/or Attribution Parties, as appropriate, of You or Your use of the Work, without the separate, express prior written
permission of the Original Author, Licensor and/or Attribution Parties.
For the avoidance of doubt:
Non-waivable Compulsory License Schemes. In those jurisdictions in which the right to collect royalties through any statutory
or compulsory licensing scheme cannot be waived, the Licensor reserves the exclusive right to collect such royalties for any
exercise by You of the rights granted under this License;
Waivable Compulsory License Schemes. In those jurisdictions in which the right to collect royalties through any statutory or
compulsory licensing scheme can be waived, the Licensor reserves the exclusive right to collect such royalties for any exercise by
You of the rights granted under this License if Your exercise of such rights is for a purpose or use which is otherwise than
noncommercial as permitted under Section 4(b) and otherwise waives the right to collect royalties through any statutory or
compulsory licensing scheme; and,
Voluntary License Schemes. The Licensor reserves the right to collect royalties, whether individually or, in the event that the
Licensor is a member of a collecting society that administers voluntary licensing schemes, via that society, from any exercise by
You of the rights granted under this License that is for a purpose or use which is otherwise than noncommercial as permitted
under Section 4(c).
5. Representations, Warranties and Disclaimer
UNLESS OTHERWISE MUTUALLY AGREED TO BY THE PARTIES IN WRITING, LICENSOR OFFERS THE WORK AS-IS AND
MAKES NO REPRESENTATIONS OR WARRANTIES OF ANY KIND CONCERNING THE WORK, EXPRESS, IMPLIED,
STATUTORY OR OTHERWISE, INCLUDING, WITHOUT LIMITATION, WARRANTIES OF TITLE, MERCHANTIBILITY, FITNESS
FOR A PARTICULAR PURPOSE, NONINFRINGEMENT, OR THE ABSENCE OF LATENT OR OTHER DEFECTS, ACCURACY,
OR THE PRESENCE OF ABSENCE OF ERRORS, WHETHER OR NOT DISCOVERABLE. SOME JURISDICTIONS DO NOT
ALLOW THE EXCLUSION OF IMPLIED WARRANTIES, SO SUCH EXCLUSION MAY NOT APPLY TO YOU.
6. Limitation on Liability. EXCEPT TO THE EXTENT REQUIRED BY APPLICABLE LAW, IN NO EVENT WILL LICENSOR BE
LIABLE TO YOU ON ANY LEGAL THEORY FOR ANY SPECIAL, INCIDENTAL, CONSEQUENTIAL, PUNITIVE OR
EXEMPLARY DAMAGES ARISING OUT OF THIS LICENSE OR THE USE OF THE WORK, EVEN IF LICENSOR HAS BEEN
ADVISED OF THE POSSIBILITY OF SUCH DAMAGES.
COMPLETE COURSEWARE LICENSE AGREEMENT IV
7. Termination This License and the rights granted hereunder will terminate automatically upon any breach by You of the terms of
this License. Subject to the above terms and conditions, the license granted here is perpetual (for the duration of the applicable
copyright in the Work). Notwithstanding the above, Licensor reserves the right to release the Work under different license terms
or to stop distributing the Work at any time; provided, however that any such election will not serve to withdraw this License (or
any other license that has been, or is required to be, granted under the terms of this License), and this License will continue in full
force and effect unless terminated as stated above.
8. Miscellaneous If any provision of this License is invalid or unenforceable under applicable law, it shall not affect the validity or
enforceability of the remainder of the terms of this License, and without further action by the parties to this agreement, such
provision shall be reformed to the minimum extent necessary to make such provision valid and enforceable. No term or provision
of this License shall be deemed waived and no breach consented to unless such waiver or consent shall be in writing and signed
by the party to be charged with such waiver or consent. This License constitutes the entire agreement between the parties with
respect to the Work licensed here. There are no understandings, agreements or representations with respect to the Work not
specied here. Licensor shall not be bound by any additional provisions that may appear in any communication from You. This
License may not be modied without the mutual written agreement of the Licensor and You. The rights granted under, and the
subject matter referenced, in this License were drafted utilizing the terminology of the Berne Convention for the Protection of
Literary and Artistic Works (as amended on September 28, 1979), the Rome Convention of 1961, the WIPO Copyright Treaty of
1996, the WIPO Performances and Phonograms Treaty of 1996 and the Universal Copyright Convention (as revised on July 24,
1971). These rights and subject matter take effect in the relevant jurisdiction in which the License terms are sought to be
enforced according to the corresponding provisions of the implementation of those treaty provisions in the applicable national law.
If the standard suite of rights granted under applicable copyright law includes additional rights not granted under this License,
such additional rights are deemed to be included in the License; this License is not intended to restrict the license of any rights
under applicable law.

MI250: Introduction To Bayesian PK-PD Modeling & Simulation

Diunggah oleh

Informasi Dokumen

Judul Asli

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

MI250: Introduction To Bayesian PK-PD Modeling & Simulation

Diunggah oleh

Hak Cipta:

Format Tersedia

MI250: Introduction to Bayesian PK-PD Modeling &

is the standard deviation of the simulated values of

(y) eective number of parameters

c plasma drug concentration

Anda mungkin juga menyukai