From: "MANOJ KHURANA" manoj2570@yahoo.com
Subject: [NMusers] Plcebo Corrected PK/PD 
Date: Wed, March 2, 2005 12:07 pm 

Dear Group,
I am trying to understand the way and need of doing placebo correction
for PD response in PK/PD modeling.  I went through NONMEM archives but
couldn't find much.  One reference that I came accross was "P. Mortin et al.
Pharmacodynamic and Pharmacokinetic profile of S17092, a new orally active
prolyl endopeptidase inhibitor, in elderly healthy volunteers. A phase I
study. BJCP, 50, 350-359" and they looked at %inhibition of prolyl endopeptidase
for escalating dose cohorts having 9 subjects per cohort (3 placebos + 6 treatment)
and placebo correction was done by subtracting the average response at each time
point.  I need to know if this averaging method is the ideal and acceptable way
of doing it when we know that there is  intersubject variability in the response
in the placebo group other than the assay variability.  Is there a way to handle
this in the model itself in NONMEM.  My next question is vvery general. Can someone
explain me under what circumstances do we need a placebo correction and how it is
done for different PD responses.  What should be our criteria to decide if we need
a placebo correction while looking at the placebo response data.  I would appreciate
if someone could provide me references and guidance for that.

Thanks in advance for your time
Manoj Khurana
_______________________________________________________

From: "Bachman, William (MYD)" bachmanw@iconus.com
Subject: RE: [NMusers] Plcebo Corrected PK/PD 
Date: Wed, March 2, 2005 1:05 pm

Manoj,
 
In a nutshell, averaging and baseline subtraction are probably
the worst way to model placebo effect.  Particularly when the
response in absence of drug is not a "flat line".  I'll let
the statisticians give all the details but basically, averaging
of anything removes information from the data and baseline
subtraction has it's own issues.  That being said, its often
done (and in best case scenarios it may be a good first
approximation but not ideal.)
 
There are any number of examples out there of better approaches
to accounting for placebo effect - the first two that come to
mind are (1) Nick Holford's tacrine model and (2) Bill Jusko's
cortisol papers that take into account circadian rhythm.
 
Bill 
_______________________________________________________

From: "Gobburu, Jogarao V" GOBBURUJ@cder.fda.gov
Subject: RE: [NMusers] Plcebo Corrected PK/PD 
Date: Wed, March 2, 2005 1:47 pm

Dear Manoj,
 
The answer really depends on the experimental design and the
modeling objective. In principle, I agree with Bill's statements.
Unfortunately, we deal with designs that might not allow anything
better than subtracting mean placebo responses. If you are dealing
with data from a fixed-dose, parallel trial, then the only way to
adjust for placebo effect is by subtracting the mean of the groups 
(placebo vs. trt).  You could use population mean or typical placebo
responses for determining the drug effect.  Even if the placebo response
has a rhythm to it, if you determine the drug effect by time (instead of
modeling the placebo response) your analysis would be equivalent to
modeling all data simultaneously (provided all you are interested
in is the drug effect size+?variance).  What you will lose is the
ability to simulate the placebo effects for future trials. There
are some other advantages such as handling missing data and unbalanced
observations. You seem to be concerned about the variability in the
placebo group. If you do not have a cross-over design, you simply
cannot estimate the true variability in the drug effect. You are
stuck to using population means or typical values for placebo effect.
No modeling can help you with it. 
 
As a general practice you should always account for placebo effects. You
should have a reason to exclude placebo data (e.g.: no signal over the
trial duration). Most primary endpoints have placebo effects, while
several biomarkers might not have a placebo effect. Placebo data
allows you to estimate the intercept of the exposure-response model
and duration of effect (I am sure you already know this). 
 
Hope this is helpful.
 
Regards,
 Joga 
_______________________________________________________

From: "Nick Holford" n.holford@auckland.ac.nz
Subject: RE: [NMusers] Plcebo Corrected PK/PD
Date: Fri, March 4, 2005 11:10 am 

Joga Gobburu wrote:

I am not sure why you conclude that the variability in drug effect cannot be
estimated from a parallel group design. I would include disease progression as part
(perhaps the major part) of the response seen in the placebo treatment group. By
making assumptions about the time course of the disease/placebo response and how it
interacts with the drug effect then one can estimate the population parameters and
between subject variability for all components (disease, placebo, drug). Making
assumptions is a necessary part of this kind of modelling. The models can be quite
helpful for interpretation of clinical trials and have been used successfully for
clinical trial simulation and prospective design of trials.

Nick

Holford NHG, Peace KE. Results and validation of a population pharmacodynamic model
for cognitive effects in Alzheimer patients treated with tacrine. Proceedings of the
National Academy of Sciences of the United States of America 1992;89(23):11471-11475

Lockwood P. Application of clinical trial simulation in Alzheimer’s disease. In:
Danhof M, Karlsson MO, Powell RJ, editors. 4th International Symposium on
Measurement and Kinetics of In Vivo Drug Effects; 2002 April 26; Noordwijkerhout,
The Netherlands; 2002.
--
Nick Holford, Dept Pharmacology & Clinical Pharmacology
University of Auckland, 85 Park Rd, Private Bag 92019, Auckland, New Zealand
email:n.holford@auckland.ac.nz tel:+64(9)373-7599x86730 fax:373-7556
http://www.health.auckland.ac.nz/pharmacology/staff/nholford/
_______________________________________________________

From: "Gobburu, Jogarao V" GOBBURUJ@cder.fda.gov
Subject: RE: [NMusers] Plcebo Corrected PK/PD
Date: Fri, March 4, 2005 11:38 am 

Dear Nick,

  Thanks for your note. I wrote 'cannot estimate the true variability in the
drug effect", "true" being the qualifier.  Yes, models with appropriate
assumptions can be used and are being used to estimate the variability in
the drug effect. However, unless we really know how each patient behaves on
both placebo and drug (cross-over), we do not have the right data to
estimate the bsv and wsv of the drug effects. Our reference for placebo
effect, otherwise, would be a typical patient and not that particular
patient (in my parallel design example). This ignores the bsv of placebo
effect while determining the drug effect in each patient. I thought it is
important to note that. Sure, those models will still be useful.

Regards,
 Joga
_______________________________________________________

From: "Nick Holford" n.holford@auckland.ac.nz
Subject:  Re: [NMusers] Plcebo Corrected PK/PD 
Date: Fri, March 4, 2005 1:54 pm

Joga,

I think it is possible to estimate both the disease progress AND the drug effect in
an individual in one arm of a parallel group design. Suppose the model is:

S(t) = S0 + alpha*t + beta*ce(t)

S is disease status ('response'), S0 is the baseline status, alpha is the rate of
progression, beta is the slope of a linear PD model, Ce(t) is drug conc at time t.

The drug effect (beta*ce(t)) will be superimposed on the disease progress line and
with a suitable design and reasonable parameters then the observed status could be
used to estimate the parameters of this model with random effects (i.e. BSV) on each
parameter.

Of course with the over sparse designs that are commonly used in clinical trials or
if the time course of disease progression is too closely correlated with the time
course of drug effect then there may be an a posteriori identifiability problem but
in theory the parameters are a priori identifiable.

Nick
--
Nick Holford, Dept Pharmacology & Clinical Pharmacology
University of Auckland, 85 Park Rd, Private Bag 92019, Auckland, New Zealand
email:n.holford@auckland.ac.nz tel:+64(9)373-7599x86730 fax:373-7556
http://www.health.auckland.ac.nz/pharmacology/staff/nholford/
_______________________________________________________

From: "Mats Karlsson" Mats.Karlsson@farmbio.uu.se
Subject: Re: [NMusers] Plcebo Corrected PK/PD 
Date: Fri, March 4, 2005 3:13 pm

Nick,


Would the model not be identifiable only if you assume that the covariance of
the random effects for alpha and beta is zero? 

[At least if Ce is constant. If it is variable, and variable enough, it starts
to resemble a cross-over experiment, at least when you have frequent
measurement.]

Mats
_______________________________________________________

From:  "Bhattaram, Atul" BhattaramA@cder.fda.gov
Subject: RE: [NMusers] Plcebo Corrected PK/PD
Date: Fri, March 4, 2005 4:05 pm

Hello Nick

I dont think you can estimate the "true variability" unless one does a
cross-over design. If you dont have placebo data in the same subject for
symptomatic drugs (anti-hypertensives) how would you estimate the "true
variability"?.

Venkatesh Atul Bhattaram
Pharmacometrics, FDA
_______________________________________________________

From: "Bachman, William (MYD)" bachmanw@iconus.com
Subject: RE: [NMusers] Plcebo Corrected PK/PD
Date: Fri, March 4, 2005 4:23 pm

This is done all the time (measure the placebo effect in some subjects and
the drug effect in others (- at least where ethical).  Look at all those
analgesic studies that have been done.  It's clear you don't have the
variability in the "same" subjects, but, in a well designed study with
sufficient power the estimate of variability in placebo effect is
sufficiently accurate for the purpose.  In many cases, there is no other way
to do it.  Obviously, the study design is dependent on the indication,
severity of disease state and practicality/ethicality (you can't pull teeth
from a subject without medication one week and with medication the next).
_______________________________________________________

From:  "Nick Holford" n.holford@auckland.ac.nz
Subject: Re: [NMusers] Plcebo Corrected PK/PD 
Date: Fri, March 4, 2005 5:47 pm

Atul,

It all depends on the model you have in mind. If the time course of the placebo
effect is essentially the same as the time course of the drug effect then I agree
that the two components of the overall response cannot be distinguished (as I
indicated in my previous posting). But if the placebo response (or disease progress)
is different from the time course of drug effect then the components should be
identifiable.

Nick
--
Nick Holford, Dept Pharmacology & Clinical Pharmacology
University of Auckland, 85 Park Rd, Private Bag 92019, Auckland, New Zealand
email:n.holford@auckland.ac.nz tel:+64(9)373-7599x86730 fax:373-7556
http://www.health.auckland.ac.nz/pharmacology/staff/nholford/
_______________________________________________________

From: "Nick Holford" n.holford@auckland.ac.nz
Subject: Re: [NMusers] Plcebo Corrected PK/PD 
Date: Fri, March 4, 2005 5:56 pm

Mats,

As I indicated earlier the ability to distinguish the disease progress (aka placebo)
from the drug effect will depend on modelling assumptions. Its not clear to me why
one must assume zero covariance of alpha and beta so perhaps you could expand more
on why this is a necessary condition?

It requires a very special design to ensure that Ce does not change with time! So I
would always expect changing Ce. If you wish this is some kind of natural cross over
design as effect increases and decreases.  The problem is usually being able to
design the study to define the time course of effect well enough. 

But my point is that given a suitable model and reasonable parameters which allow
separation of the time course of disease progress and drug effect then it is
possible to estimate both components independently. 

Nick
--
Nick Holford, Dept Pharmacology & Clinical Pharmacology
University of Auckland, 85 Park Rd, Private Bag 92019, Auckland, New Zealand
email:n.holford@auckland.ac.nz tel:+64(9)373-7599x86730 fax:373-7556
http://www.health.auckland.ac.nz/pharmacology/staff/nholford/
_______________________________________________________

From: "Gobburu, Jogarao V" GOBBURUJ@cder.fda.gov
Subject: RE: [NMusers] Plcebo Corrected PK/PD 
Date: Fri, March 4, 2005 6:44 pm

Dear Nick,

  1. Let us first acknowledge what we agree on: For non-progressive
responses one cannot estimate the true variability in drug effect reliably.
You might be able to estimate the mean effects though. I would also like to
make it clear that I am not advocating not-estimating the drug effect
variance during modeling or other analysis. I simply want us to note the
limitations.

  2. Now, where you seem to be having a difference in opinion (along with
others?) is when the disease is progressive. In my opinion, the disease
progression sub-model is not going to change the consequences of
interpreting  parallel trial results. May be I am missing something. In any
case, I would like to seek your opinion on a specific problem. Let us say,
the disease is progressing linearly. All patients start with a zero
intercept at randomization (randomized to either placebo or trt). This
patient who received the drug had responses of 0, 0, 0 at 1, 3 and 6 months.
Basically the drug froze the disease from progressing. Let us even make the
analysis simpler by assuming no drop-outs, no missing observations, one dose
level only, no effect delay, etc.. So, it does not matter if you used one or
two-stage analysis. What reference control effect would you use to determine
the drug (ONLY) effect for this patient?


NB: Bill, there are cross-over studies for your 'teeth-pulling example' in
the literature.

Thanks.
joga
_______________________________________________________

From: "Mats Karlsson" Mats.Karlsson@farmbio.uu.se
Subject: Re: [NMusers] Plcebo Corrected PK/PD
Date: Fri, March 4, 2005 11:51 pm

Nick,

I read your model carelessly. The point I made does not apply to this
(intercept/symptomatic) model, but to one where the drug effect slows disease
progression, i.e 

S(t) = S0 + (alpha + beta*ce(t))*t

With variations in Ce being unimportant, it is probably clear that with knowing
only the variance of alpha from the placebo group, many combinations of
variance in beta and covariance(alpha,beta) may exlain the variability in the
trt group. If you want to say that the resolution comes from variations in Ce,
I won't argue, but it may not always be the case in practise.

Mats
_______________________________________________________

From: "Nick Holford" n.holford@auckland.ac.nz
Subject: Re: [NMusers] Plcebo Corrected PK/PD
Date: Sat, March 5, 2005 4:43 pm

Joga,

>"Gobburu, Jogarao V" wrote:
> 
> Dear Nick,
> 
>   1. Let us first acknowledge what we agree on: For non-progressive
> responses one cannot estimate the true variability in drug effect reliably.
> You might be able to estimate the mean effects though. I would also like to
> make it clear that I am not advocating not-estimating the drug effect
> variance during modeling or other analysis. I simply want us to note the
> limitations.
-----
I am afraid I dont think I agree with you on this. Non-progressive responses are
just a special case of progression where alpha=0. I think it is a simple matter to
estimate the variability in drug effect in this case -- under the model of no
disease progression then any systematic change in response with time must be
attributed to drug effect. The parallel design with a non-progressing placbeo group
would let you estimate the residual error and thus be able to distinguish true drug
effect from noise.
----- 
>   2. Now, where you seem to be having a difference in opinion (along with
> others?) is when the disease is progressive. In my opinion, the disease
> progression sub-model is not going to change the consequences of
> interpreting  parallel trial results. May be I am missing something. In any
> case, I would like to seek your opinion on a specific problem. Let us say,
> the disease is progressing linearly. All patients start with a zero
> intercept at randomization (randomized to either placebo or trt). This
> patient who received the drug had responses of 0, 0, 0 at 1, 3 and 6 months.
> Basically the drug froze the disease from progressing. Let us even make the
> analysis simpler by assuming no drop-outs, no missing observations, one dose
> level only, no effect delay, etc.. So, it does not matter if you used one or
> two-stage analysis. What reference control effect would you use to determine
> the drug (ONLY) effect for this patient?
-----
The situation when the drug affects the rate of progression is harder to identify
(this is the case that Mats also thought was problematic). With sufficient variation
in Ce within a subject then I think one can identify the protective drug effect on
rate of progression. As Mats has pointed out the variation in Ce acts as a form of
natural cross over within the treatment period although the basic design does not
demand a formal crossover of treatment. In the special case where there is no
variation in Ce within a subject then I agree one cannot separate drug effect from
disease progression.
----
Nick
--
Nick Holford, Dept Pharmacology & Clinical Pharmacology
University of Auckland, 85 Park Rd, Private Bag 92019, Auckland, New Zealand
email:n.holford@auckland.ac.nz tel:+64(9)373-7599x86730 fax:373-7556
http://www.health.auckland.ac.nz/pharmacology/staff/nholford/
_______________________________________________________

From: "Bachman, William (MYD)" bachmanw@iconus.com
Subject: RE: [NMusers] Placebo Corrected PK/PD 
Date: Mon, March 7, 2005 8:07 am 

>NB: Bill, there are cross-over studies for your 'teeth-pulling example' in
>the literature.

>Thanks.
>joga

I doubt they would pass a current IRB.
_______________________________________________________