NONMEM Users Network Archive

Hosted by Cognigen

Re: More Levels of Random Effects

From: Nick Holford <n.holford>
Date: Tue, 21 Oct 2008 07:36:23 +1300

Mike,

I don't agree with you that occasion is necessarily linked through time.
As already mentioned by Leonid, in the example we are discussing it is
assumed that you are unable to find a predictable relationship between
occasions in order to predict a future occasion. This would mean that
there is no relationship of occasion to time or any other factor.

It makes no difference how you use the occasion value (say 1, 2 and 3)
to the sequence of ETAs that you associate with the occasion e.g. you
could use ETA(1) with occasion 2, ETA(2) with occasion 3 and ETA(3) with
occasion 1. There is no intrinsic ordering to the occasion value if (as
we assume) there is no relationship of IOV to time. Thus the occasions
are non-ordered categories like race.

I agree that for race you may have some other clues to help you with
expected proportions when you randomly select a category for future
predictions. But if occasion is really random and you have no idea of
the future proportions then a uniform distribution of occasions seems
reasonable.

I dont understand your point about number of simulations. For whatever
purpose you are planning to use the original model the need to randomly
predict unknown occasions should not affect the number of simulations.
It just becomes part of the prediction model which presumably already
has stochastic elements in it anyway.

Best wishes

Nick

Michael.J.Fossler
>
> Hi Nick;
>
> In your example, I would also do the latter, since I would have a
> rough idea of the proportion of each of those groups living in the UK.
> But, I am not sure your example is analagous to IOV, principally
> because, in your example of Scots, Irish and Welsh, time is not
> involved. If you have 4 occasions, and you fit 4 variability terms as
> separate (no BLOCK SAME) don't they now become ordered categories?
> If there is a definite pattern to the variability changes as a
> function of time, wouldn't that change be better modeled explicitly
> as a time-dependent change, rather than in an implicit way?
>
> I think your suggestion of running simulations based on the 1-4th
> occasion when extrapolating is reasonable; however, the number of
> simulations that might have to be performed may, in certain cases
> quickly add up.
>
>
>
> Mike Fossler
> GSK
>
>
>
> *"Nick Holford" <n.holford
> Sent by: owner-nmusers
>
> 17-Oct-2008 15:38
>
>
> To
> "nmusers" <nmusers
> cc
>
> Subject
> Re: [NMusers] More Levels of Random Effects
>
>
>
>
>
>
>
>
>
> Mike,
>
> So how do you deal with other non-ordered categorical variables? Suppose
> you do your studies in Scotland, Ireland and Wales then need to predict
> what will happen in England? Assuming you found 'significant'
> differences in between subject variability in clearance between the
> Scots, Irish and Welsh and wanted to predict a population in England do
> you think it would better to take the average of the Scots, Irish and
> Welsh (equivalent to using SAME) or do you think it would be better to
> randomly choose from the 3 groups knowing that representatives of these
> 3 groups might be found living in England?
>
> I would think the latter approach would be more realistic. I would
> consider doing something similar for between occasion variability (aka
> IOV) if I find 'significant' differences across 3 occasions and need to
> predict a study which has 4 occasions. Rather than assume the 4th
> occasion is the average of the other 3 I would consider randomly
> assigning the 4th occasion data item to 1, 2 or 3.
>
> Nick
>
> Michael.J.Fossler
> >
> >
> > I suppose it really comes down to what you are going to do with the
> > model. Many times I have checked the SAME assumption when modeling
> > inter-occasional variability, and found that sometimes, removing it
> > does indeed improve the fit significantly. In almost every case I've
> > retained it (despite the better fit) for the exact reasons Leonid
> > cites: it makes your model completely data-dependent. I suppose if the
> > model was meant as a description or summary of the data, then it would
> > not matter, but I make all of my models work for a living...
> >
>
> --
> Nick Holford, Dept Pharmacology & Clinical Pharmacology
> University of Auckland, 85 Park Rd, Private Bag 92019, Auckland, New
> Zealand
> n.holford
> http://www.fmhs.auckland.ac.nz/sms/pharmacology/holford
>
>
>

--
Nick Holford, Dept Pharmacology & Clinical Pharmacology
University of Auckland, 85 Park Rd, Private Bag 92019, Auckland, New Zealand
n.holford
http://www.fmhs.auckland.ac.nz/sms/pharmacology/holford
Received on Mon Oct 20 2008 - 14:36:23 EDT

The NONMEM Users Network is maintained by ICON plc. Requests to subscribe to the network should be sent to: nmusers-request@iconplc.com.

Once subscribed, you may contribute to the discussion by emailing: nmusers@globomaxnm.com.