From: Leonid Gibiansky <*LGibiansky*>

Date: Mon, 03 Aug 2009 11:04:53 -0400

Hauke

Several comments:

It is better to parametrized the model in terms of CL, V1, Q, V2. If you

believe that VSS is the derived parameters (VSS=V1+V2), then VSS

parameterization introduced additional correlation that you do not need.

Usually, random effects on V2 and Q are not defined properly unless you

have a very rich data. I would start with the model that has only random

effects on CL and V1, then add random effect on V2 or Q (not both!), and

only then try to put all effects together. In my experience, random

effect on Q could be helpful for rich sampling, but it is unusual to see

a dataset where you need both V2 and Q etas. Based on your results

(absolute correlation of V1 and VSS) I would guess that random effect on

V2 is not needed at all.

One of the helpful diagnostics is to look on correlation of random

effect for the problem where this correlation is not explicitly

included. If the correlation is real, you will see it on the scatter

plot matrix of random effect.

Helpful diagnostic is the OMEGA value. If inclusion of correlations

substantially increases variance estimates, you may have an

over-parametrized model where extra correlation makes the model less stable.

Very helpful diagnostic is to compare models (with and without

correlations, or with and without extra random effect) by looking on

IPRED vs IPRED and PRED vs PRED plots of models under investigation. If

they show perfect correlation (coincide with the unit line) you may go

with the simplest model (assuming other diagnostics do not tell you

otherwise).

Thanks

Leonid

--------------------------------------

Leonid Gibiansky, Ph.D.

President, QuantPharm LLC

web: www.quantpharm.com

e-mail: LGibiansky at quantpharm.com

tel: (301) 767 5566

Hauke Rühs wrote:

*>
*

*> Dear NMusers,
*

*>
*

*> modelling a 2-compartmet model parameterized by CL, V1, VSS and Q, I got
*

*> to a problem with which I don’t know how to deal with: After choosing my
*

*> structural and statistic model (combined residual error model) I
*

*> estimated the covariance matrix by including an OMEGA-BLOCK(4), which
*

*> reduced the OFV by 15. The correlations between the parameters were all
*

*> estimated to be minor (< 0.8). But when I model with a BLOCK(2) on VSS
*

*> and V1, which I would expect to be positively correlated, the
*

*> correlation is estimated to be -0.99. Additionally, the inclusion of
*

*> BLOCK(2) does not significantly improve the OFV.
*

*> So does it, after all, still make sense to include the BLOCK(2)?
*

*> Generally, at which step of model-building would you recommend to test
*

*> for parameter correlation?
*

*>
*

*> Thanking you in advance,
*

*>
*

*> Hauke
*

*>
*

*> -----------------------------
*

*> Hauke Rühs
*

*>
*

*> Apotheker
*

*> Pharmazeutisches Institut
*

*> - Klinische Pharmazie -
*

*> An der Immenburg 4
*

*>
*

*> 53121 Bonn
*

*>
*

*> Tel: + 49-(0)228 73-5781
*

*> Fax: + 49-(0)228 73-9757
*

*>
*

*> www.klinische-pharmazie.info
*

*>
*

*>
*

*> *

Received on Mon Aug 03 2009 - 11:04:53 EDT

Date: Mon, 03 Aug 2009 11:04:53 -0400

Hauke

Several comments:

It is better to parametrized the model in terms of CL, V1, Q, V2. If you

believe that VSS is the derived parameters (VSS=V1+V2), then VSS

parameterization introduced additional correlation that you do not need.

Usually, random effects on V2 and Q are not defined properly unless you

have a very rich data. I would start with the model that has only random

effects on CL and V1, then add random effect on V2 or Q (not both!), and

only then try to put all effects together. In my experience, random

effect on Q could be helpful for rich sampling, but it is unusual to see

a dataset where you need both V2 and Q etas. Based on your results

(absolute correlation of V1 and VSS) I would guess that random effect on

V2 is not needed at all.

One of the helpful diagnostics is to look on correlation of random

effect for the problem where this correlation is not explicitly

included. If the correlation is real, you will see it on the scatter

plot matrix of random effect.

Helpful diagnostic is the OMEGA value. If inclusion of correlations

substantially increases variance estimates, you may have an

over-parametrized model where extra correlation makes the model less stable.

Very helpful diagnostic is to compare models (with and without

correlations, or with and without extra random effect) by looking on

IPRED vs IPRED and PRED vs PRED plots of models under investigation. If

they show perfect correlation (coincide with the unit line) you may go

with the simplest model (assuming other diagnostics do not tell you

otherwise).

Thanks

Leonid

--------------------------------------

Leonid Gibiansky, Ph.D.

President, QuantPharm LLC

web: www.quantpharm.com

e-mail: LGibiansky at quantpharm.com

tel: (301) 767 5566

Hauke Rühs wrote:

Received on Mon Aug 03 2009 - 11:04:53 EDT