- On 13 Apr 2006 at 11:12:10, aude.dunand.aaa.pierre-fabre.com sent the message

Back to the Top

The following message was posted to: PharmPK

Dear all,

I'm performing a compartmental analysis for intravenous data of 12

subjects and I was surprised to find large discrepancies between the

results of Kinetica and Winnonlin. The estimation CV are largely greater

with Kinetica and estimate parameters are completely different,

especially

for 1/Y^2pred weight.

Please tell me your point of view.

Aude Dunand

[How complex is the model, i.e. are you reaching proper convergence

to the global minimum in all cases? Are the weights defined the same

(from memory WinNonlin was a little different from what I expected

and don't know how Kinetica does it)? However, if you only have

plasma/blood, it shouldn't effect the parameter estimates - db] - On 13 Apr 2006 at 14:32:22, "Shawn D. Spencer" (shawn.spencer.-at-.famu.edu) sent the message

Back to the Top

Aude,

The only expected difference between two software systems is your

choice of minimization algorithm (e.g., Simplex, Levenberg). If you

have chosen the same minimization/optimization algorithm, you will

get the same estimates, provided all your other input variables/

parameters are the same. Once you have verified the same iterative

procedure, one would double check they have assumed the same error

distribution when weighting/reweighting, same limits, initial

estimates, etc. Are their any other significant differences in the

output other than CV? That may give some indication of any

discrepancy in your input.

Hope that helps

Shawn D. Spencer, Ph.D., R.Ph.

Assistant Professor of Biopharmaceutics and Pharmacokinetics

College of Pharmacy and Pharmaceutical Sciences

Florida A&M University

Tallahassee, FL 32307

shawn.spencer.aaa.famu.edu

[Any reasonable optimization routine should be capable of finding the

global minimum, but may also get last from time to time. Have you

tried refitting with different inital estimates? - db] - On 13 Apr 2006 at 14:55:10, DDubins.-at-.allied-research.com sent the message

Back to the Top

The following message was posted to: PharmPK

Hi Aude,

Parameter estimation will be largely dependent on the model you

define, any

assumptions built into it, the quality of the data itself, starting

values,

and the numerical method used by the program. I wouldn't at all be

surprised to find out if Winnonlin uses a different method than

Kinetica.

The question comes down to what you would like to believe. When

plotted, do

both models fit the data equally well? Why have you chosen to use a 1/

Y^2

weighting? Is it giving you a better fit? Do the two programs provide

the

same results in the absence of weighting? A friend of mine once told

me, "a

man with two thermometers doesn't know what the temperature is." The

same

can be said with fitting programs. Parameter estimation is a very

sensitive

and fickle procedure; a program may very well spit out a correlation

matrix

and associated errors with parameters, but in the end fitting is as

much an

art as a science, and the results just as subjective. My advice, for

what

little it may be worth, is to choose one program and stick with it.

-Dave - On 14 Apr 2006 at 10:00:08, "Ma Guangli" (guanglima.-a-.gmail.com) sent the message

Back to the Top

Nonlinear regression is the core of compartment model

computation to estimate parameters. There are many

nonlinear regression algorithms. They give very different

results. Even to a algorithm, its implementation by

different programmer or under different programming

enviroment make the PK results a little different.

As Shawn D. Spencer said, the parameters for the models

are key solution to minimize the difference. The different

softare gives the different default option. You can adjust

any possible paramters to get the results you want.

Believe your experience, and then to judge.

Ma Guangli

--

Ph.d. Candidate, Zhejiang University, China

A data analyst for drug development - On 14 Apr 2006 at 10:17:20, aude.dunand.at.pierre-fabre.com sent the message

Back to the Top

The following message was posted to: PharmPK

Winnonlin and Kinetica use indeed different minimization algorithms :

Winnonlin uses Gauss-Newton algorithm whereas Kinetica uses

Levenberg-Marquardt. I agree the discrepancies come from it.

Marquardt algorithm is known to be more robust than Gauss, but I find

using

it very large CVs and large number of iterations. The question is :

can I

trust these results ?

I tried to put same initial parameters for the 2 softwares and I find

anyway the same differences.

I also notice that the discrepancies are larger when I use greater

weights,

(i.e. 1/Ypred and 1/Y^2pred), but I can't explain myself why.

I'm looking for your help,

Aude Dunand

[Have you tried different initial estimates? I'd be curious to look

at 2-3 sets with boomer if you want to share the data, model,

parameter estimates and weighting suggestions. Using boomer with the

simplex method a few times results in multiple runs with different

initial estimates that can sometimes help. Maybe I can break the

tie ;-) - db (david.at.boomer.org) - weights are 1/Ypred? - are you

using iteratively reweighted least squares? Why not ordinary weighted

least squares, i.e. 1/Yobs] - On 14 Apr 2006 at 10:33:22, "Serge Guzy" (GUZY.at.xoma.com) sent the message

Back to the Top

The following message was posted to: PharmPK

If you are looking for help, best thing would be to send the data.

If you have many patients, best strategy would be to perform a

population analysis (Winnonlin cannot do it but other software programs

can do it, NONMEM, KINETICA, PDX-MC-PEM).

Serge Guzy

President POP-PHARM; Inc. - On 14 Apr 2006 at 16:41:08, "David S. Farrrier" (DFarrier.-at-.SummitPK.com) sent the message

Back to the Top

Dear Aude & PharmPKers

PK Solutions provides an excellent and easy tool for producing

initial estimates of pharmacokinetic parameters prior to applying

iterative curve fitting routines such as used in Kinetica and

Winnonlin. PK Solutions runs in Excel and performs an easy and

flexible curve stripping of blood level data to derive these initial

estimates. It also uses these in a compartment-free approach to

calculate some 75 additional PK parameters, including projection of

steady state parameters from single dose inputs. You are invited to

take a look at PK Solutions at www.SummitPK.com

Regards,

Dr. David S. Farrier

/\ /\

SummitPK.com /\ / \ /\ / \

/ / / /\ / \

=David S. Farrier, Ph.D. Phone: 970-249-1389

Summit Research Services Fax: 970-249-1360

68911 Open Field Dr. Email: DFarrier.-at-.SummitPK.com

Montrose, CO 81401 Web: http://www.SummitPK.com - On 15 Apr 2006 at 10:01:19, "Ma Guangli" (guanglima.aaa.gmail.com) sent the message

Back to the Top

Dear Aude Dunand,

Marquardt and Gauss-Newton are both Gauss-like algorithms. So the

initial values are vital to the results. If you put same initial

parameters for the two algorithms, the two algorithms try to down the

minimum. The difference is where they stop. So, there is another key

value (tolerance? or epsilon? I forget how to call it). All of these

make the results different.

The large number of iterations maybe is because of initial values.

The start points make downhill process difficult. I wrote a grid

search program 3years ago to search the initial values of Marquardt

for global minimized in Delphi (a programming language). I do not

know how to implement it in Winnonlin or Kinetics. Anyway, I do not

have these two software packages.

The weighting routine is the key to make small values more important

to the final results. So, it is not strange to make the larger

discrepancies. As I know, IRWS is better than the others theoretically.

Simplex is another class algorithm. They are quiet different from the

Gauss-like. But the initial values and the (tolerance or epsilon) are

keys as Gauss-like.

What I can suggest is:

1) As Serge Guzy said, population analysis is a solution to such

a problem.

2) Forget these, use the mean value of the whole data set to get

the parameters.

3) If your data set is very large, cut them into small groups.

Then do some work on each group.

4) Judge by your experience.

Ma Guangli

[Ph.d. Candidate, - On 18 Apr 2006 at 16:20:20, "J.H.Proost" (J.H.Proost.at.rug.nl) sent the message

Back to the Top

The following message was posted to: PharmPK

Dear Dave,

You wrote in reply to Aude:

> Parameter estimation is a very sensitive

> and fickle procedure; a program may very well spit out

> a correlation matrix

> and associated errors with parameters, but in the end

> fitting is as much an

> art as a science, and the results just as subjective.

The choice for a model and weighting scheme are indeed

subjective, and therefore the results are subjective. But

if these choices are made in an appropriate way, the

results should be independent of optimization algorithm

and computer implementation.

> My advice, for what

> little it may be worth, is to choose one program and

>stick with it.

I don't understand this advice. In case of thermometers

this may be a good advice, but not for fitting programs.

If one finds differences, as Aude reported, one can learn

a lot about the particular properties of a program. Things

that remain invisible using a single program.

Best regards,

Hans Proost

Johannes H. Proost

Dept. of Pharmacokinetics and Drug Delivery

University Centre for Pharmacy

Antonius Deusinglaan 1

9713 AV Groningen, The Netherlands

tel. 31-50 363 3292

fax 31-50 363 3247

Email: j.h.proost.at.rug.nl - On 18 Apr 2006 at 16:29:24, "J.H.Proost" (J.H.Proost.aaa.rug.nl) sent the message

Back to the Top

The following message was posted to: PharmPK

Dear Ma Guangli,

You wrote in reply to Aude:

> Nonlinear regression is the core of compartment model

> computation to estimate parameters. There are many

> nonlinear regression algorithms. They give very

> different results.

I don't understand this statement. If the appropriate

model and weighting scheme are chosen, each program should

provide similar results. Any difference larger than

expected from rounding errors is highly suspect, and

should be investigated closely. See my message to Dave.

> Even to a algorithm, its implementation by

> different programmer or under different programming

> enviroment make the PK results a little different.

OK, this true, but for properly working algorithms the

differences should be very small.

> You can adjust

> any possible paramters to get the results you want.

This sounds very bad! I trust you mean that one should

choose the settings of the program in such a way that

reliable results are obtained, e.g. as judged from

residuals plots.

Best regards,

Hans Proost

Johannes H. Proost

Dept. of Pharmacokinetics and Drug Delivery

University Centre for Pharmacy

Antonius Deusinglaan 1

9713 AV Groningen, The Netherlands

tel. 31-50 363 3292

fax 31-50 363 3247

Email: j.h.proost.-a-.rug.nl

Want to post a follow-up message on this topic? If this link does not work with your browser send a follow-up message to PharmPK@boomer.org with "Winnonlin vs Kinetica for PK fitting" as the subject

PharmPK Discussion List Archive Index page

Copyright 1995-2010 David W. A. Bourne (david@boomer.org)