NONMEM Users Network Archive

Hosted by Cognigen

Re: FO vs FOCE, sequential vs simultaneous

From: Gastonguay, Marc <marcg>
Date: Tue, 9 Dec 2008 13:55:39 -0500

There's an additional, related point to consider with respect to
estimation method, in selecting a simultaneous vs sequential

In the case where simultaneous modeling under conditional estimation
is not feasible (run-time, convergence, etc), it is preferable to use
a sequential approach. In the first step, model PK (or parent) using
conditional estimation or FO/POSTHOC, and run the second sequential
step (e.g. PD or metabolite) conditioned on the individual estimates
obtained in the first step. By doing so, the second step (PD or
metabolite) model will be driven by individual conditional random
effect estimates obtained the first step. This is preferable to
running a simultaneous model under FO, where only the population
typical values would be used to drive the second stage endpoint (PD or
metabolite) model.

For more on this point, see:
Zhang L, Beal SL, Sheiner LB. J Pharmacokinet Pharmacodyn. 2003 Dec;
30(6):387-404. Simultaneous vs. sequential analysis for population PK/
PD data I: best-case performance.


Marc R. Gastonguay, Ph.D.
President & CEO, Metrum Research Group LLC []
Scientific Director, Metrum Institute []
Direct: 860-670-0744 Main: 860-735-7043
Email: marcg

On Dec 9, 2008, at 12:33 PM, Leonid Gibiansky wrote:

> Hi Alan,
> Here:
> I used all datasets that I had, and I was not able to find any
> problem where FO was superior to FOCE.
> Not-converged FOCE is better, in my opinion, than converged FO
> (although you can always check using diagnostic plots).
> If you cannot use FOCE due to time restrictions, it is better to use
> FO than just abandon modeling. Still, I would try to run the final
> model with FOCEI.
> Concerning sequential vs simultaneous: there are several points to
> consider, and this is usually relates to the PK-PD case. For PK-PD,
> the main question is the comparison of PK and PD variabilities.
> Usually, PK variability is smaller, and PK data are more reliable.
> Then, sequential modeling can be more warranted. If PK and PD
> variabilities are similar (both residual and inter-subject) you can
> use joint fit. I usually do PK first, then PK-PD, and then try to
> fit combined model at the very last stage.
> For parent-metabolite case, both sets of data are equally reliable
> (or not reliable), and variability is usually similar. Then the
> question boils down to time and convenience. Again, I usually do
> parent fist, then fix parameters and do metabolite, and then, if
> possible, do simultaneous fit. This often saves time: parent model
> is more simple, it can be done in standard ANDANs for 1-2
> compartment models that are much quicker. You can experiment freely
> with random effect, covariates, residual error, etc. Joint model
> often needs to be solved using ADVAN5, 7 or even $DES which are more
> CPU-consuming. You want to do minimum number of runs here. Thus, you
> want to start with good parent model, and study metabolite part
> only. The final joint run fits all parts together.
> Thanks
> Leonid
> --------------------------------------
> Leonid Gibiansky, Ph.D.
> President, QuantPharm LLC
> web:
> e-mail: LGibiansky at
> tel: (301) 767 5566
> Xiao, Alan wrote:
>> Dear All,
>> I know this is an old topic, too, but would like to see the
>> statistics. When you have a dataset with about 10% of dense Phase
>> II data (predose, 2, 4, 8, and 12 hrs post dose on day 1 and at
>> steady state, twice-daily dose regimen) and about 90% of very
>> sparse Phase III data (1-2 samples/patient), which method do you
>> prefer: FO or FOCE? or FO for model development but FOCE for model
>> refinement/finalization? If FOCE is not practical because of long
>> run-time or numerical difficulties in converge, do you stop here or
>> would you use FO?
>> Thanks,
>> Alan

Received on Tue Dec 09 2008 - 13:55:39 EST

The NONMEM Users Network is maintained by ICON plc. Requests to subscribe to the network should be sent to:

Once subscribed, you may contribute to the discussion by emailing: