Date: Tue, 23 Nov 1999 09:44:55 +0000
From: James <J.G.Wright@ncl.ac.uk>
Subject: IOV effects (inter-occasion variability)

Dear List,

New Question:

If you are modelling a population where some patients have a several occasions and some patients have only one, should the patients with only one occasion have an IOV random effect?

James

 

 

*****

 

 

Date: Tue, 23 Nov 1999 12:15:08 +0100
From: Mats Karlsson <Mats.Karlsson@biof.uu.se>
Subject: Re: IOV effects

Hi James,

Yes all patients should have an IOV random effect. For example, think about the case where IOV is much larger than IIV (or IIV actually is zero), then you would predict data to be much less variable than they truly are for subject with only one occasion.

Mats

 

*****

 

 

Date: Tue, 23 Nov 1999 14:30:13 +0100
From: Pascal Girard <pg@upcl.univ-lyon1.fr>
Subject: Re: IOV effects

Hi Mats and James,

I understand James question

"should the patients with only one occasion have an IOV random effect?"

as

"When I have a dataset mixing patients with several occasions, and other with only one occasion, do I have to implement an IOV model within NONMEM for every patients, OR can I implement it only for patients with several occasions and leave those with one occasion with only IIV?"

Mats says that in this case you should always have an IOV random effect. I think that before answering about the principle, we should answer, from a technical point of view, that we are left with little choice: either you implement IIV and IOV for every patient or you implement only IIV for every patient. The thing that you can have in one model is parameters affected by IIV & IOV for <<every>> patients and other parameters affected <<only>> by IIV for <<every>> patients.

It would be tricky <<and>> useless to have in one model a parameter affected by IIV & IOV for a subset of patients with several occasions, and IIV only for another subset of patients with only one occasion. This would be useless because of the way IOV is modelled: when a patient has only one occasion he carries absolutely no information about IOV. This is true whateve method you use (NONMEM, NLME Splus function, MCMC, ...).

With NONMEM, if you try to estimate IOV using a dataset with patients having only one occasion, you should get an estimate of IOV omega close to 1E-8, which is the sign that NONMEM has not enough information to estimate this random effect (we found this when estimating inter-study variability -paper in press in JPS-, and I think that what Mats found in his 1993 JPB paper). You have the same difficulties when you try to separate IIV from residual variability with a dataset where you have only one observation/patient. This can only be done by preliminary fixing residual variability to a certain value using a parametric model. This is currently performed with non-parametric methods as NPML or NPEM. If you are confident with some prior knowledge about IOV, maybe you can try to fix IOV to a certain value.

Now what happens when you estimate IOV on a dataset with some patients with several occasions and most with only one occasion? My guess is that you will have to pay some attention to the precision with which random IOV effect parameters are estimated: the less number of patients with several occasions you have in your dataset, the more imprecisely estimated those parameters will be. And you may even find IOV useless in certain cases, except ... if you trust IOV exists and set a prior distribution for it. But this takes us back to the previous discussion ...

Pascal
--
Pascal Girard
------------------------------------------------------------------
Service Pharmacologie Clinique PG@upcl.univ-lyon1.fr
Faculte RTH Laennec Tel : +33 (0)4 78 78 57 26
BP 8071, rue Guillaume Paradin Fax : +33 (0)4 78 77 69 17
69376 LYON Cedex 08 FRANCE http://www.spc.univ-lyon1.fr
------------------------------------------------------------------

 

*****

 

 

From: "Piotrovskij, Vladimir [JanBe]" <VPIOTROV@janbe.jnj.com>
Subject: RE: IOV effects
Date: Tue, 23 Nov 1999 15:52:36 +0100

It is the question of study design. If you are going to estimate IOV separately and independently from IIV you should design your study accordingly. In case you have only a small proportion of subjects tested at different occasions mixed-effects modelling will not help you irrespective to the program you use (NONMEM, nlme, a new SAS PROC (I forgot the name), etc.).

Vladimir
----------------------------------------------------------------------
Vladimir Piotrovsky, Ph.D.
Janssen Research Foundation
Clinical Pharmacokinetics
B-2340 Beerse
Belgium
Email: vpiotrov@janbe.jnj.com

 

 

*****

 

 

Date: Tue, 23 Nov 1999 14:50:16 +0000
From: James <J.G.Wright@ncl.ac.uk>
Subject: Re: IOV effects

Dear Matts & Pascal,

This was my feeling also, but I was concerned about the way maximum likelihood would use these fixed effects. Basically, it will take the IOV random effect for an individual with only one course and use it to account for some of the variability that would normally be taken up by the IIV random effect. This is fair enough but it may affect the estimates of IOV and IIV (and it will always be in the direction most convenient fo the model).

James

 

 

*****

 

 

Date: Tue, 23 Nov 1999 15:27:56 +0000
From: James <J.G.Wright@ncl.ac.uk>
Subject: RE: IOV effects

Dear List,

Please ignore the word "fixed" in my last email, it was onre of the whole-word typo.

The situation I have is one where I have 45 patients, with say 40 or so studied on two occasions, but a few with an occasion missing. I didn't want to throw these people away because I hoped they could improve estimation of IIV.

James

 

 

*****

 

 

Date: Tue, 23 Nov 1999 16:55:16 +0100
From: Mats Karlsson <Mats.Karlsson@biof.uu.se>
Subject: Re: IOV effects

Hi
I agree with all what Pascal wrote apart from

> when a patient has only one occasion he carries absolutely no information
> about IOV.

An equally true statement is that he carries no information about IIV. That truth being that he carries info only about the sum of the two and information from other individuals will have to be used to partition the total variability.

> With NONMEM, if you try to estimate IOV using a dataset with patients
> having only one occasion, you should get an estimate of IOV omega close
> to 1E-8

Actually, IIV and IOV are unidentifyable, they can take on a multitude of values, but should be highly correlated.

Mats