Re: [NMusers] SIMULATION QUESTIONS

Hi Live,
In terms of GOF, look on the eta distributions (histograms and Q-Q
plots): they should be centered and reasonably normal. Look on eta
correlation (ETAs vs ETAs plots): they should be in agreement with
correlations in your OMEGA matrix. Look on DV vs PRED: there should be
no obvious bias. CWRES vs PRED plots are helpful. If all these plots
looks good, simulation should give reasonable values. To start, do
predictive check simulation: predict Cmax, AUC, Cmin distributions for
100 studies (as a rough test), and compare those with distributions of
the observed data set (there are many papers/posters that describe how
to do it; let me know if you need help with this). If predictive check
turn out to be good, I would look for some typo in your current
simulation code. If predictive check will indicate some problems, you
will see where the problems are and will need to refine/correct the
model. In terms of terminology, if you search for Erlang distribution on
the web, you will see some probability distribution description. Most
likely, it has a relation to what you do, but the code below is more
known as transit compartment model (with 6 transit compartments) rather
than Erlang distribution model (not a problem, but easier to understand).

For splitting: you start by dividing the data set into 10 parts, but
then you take 9 parts, combine together (i.e., exclude just 10% of the
data from the full data set), then fit the model to those 90% and
predict into the excluded 10%, is this what you have done? I cannot
believe that you may end up with significant over-prediction if
diagnostic plots are good and predictive check show good agreement of
simulated values with observed ones. Your parameter estimates (for
those 10 runs) should be pretty close to the final model; anything
larger than 20-25% would indicate instability of the model. How many
data points you have ? 5 random effects could be too many unless you
have a very dense data.

Leonid


livest_at_student.matnat.uio.no wrote:
> Hi Leonid,
>
> The Erlang distribution is specified in the model. Here is more of the
> control file:
> $MODEL COMP=(DEPOT,DEFDOSE) ;
> COMP=(DELA1)
> COMP=(DELA2)
> COMP=(DELA3)
> COMP=(DELA4)
> COMP=(DELA5)
> COMP=(DELA6)
> COMP=(CENTRAL,DEFOBS)
> COMP=(PERIPH)
>
> $PK
> K12=THETA(1)*EXP(ETA(1)) ;lagtime
> K23=K12
> K34=K12
> K45=K12
> K56=K12
> K67=K12
> K78=K12
>
> CL=THETA(2)*EXP(ETA(2)) ; clearance 1
> V8=THETA(3)*EXP(ETA(3)) ; volum 1
> V9=THETA(4)*EXP(ETA(4)) ; volum 2
> Q=THETA(5)*EXP(ETA(5)) ; clearance 2
>
> K80=CL/V8
> K89=Q/V8
> K98=Q/V9
> S8=V8
>
> The reason for using ADVAN 5 is because of the Erlang distribution, which
> was found to best describe the lagged absorption. And since the optimum
> number of lag-compartment was found to be 6, this gives CMT 8 (the
> observation compartment).
>
> The GOF plots look actually good, as far as I can see. However, you might
> have right about that the simulation indicates deficiency of the model
> rather then insufficient information given to NONMEM.
>
> Regarding what I’m trying to do:
> I’m doing a data-splitting analysis, where I’ve divided the dataset into
> 10 subsets. Each subset was analysed by NONMEM using the final model. This
> was done in order to compare the parameter estimates with that of the
> final model. All the values for the subsets were in the range of ± 2 SD.
> Also I want to predict the concentrations in the remaining 10 % from each
> of the 10 subsets using the estimates from each subset. This is what I
> asked help for.
>
> Thanks for your reply!
>
>
> Live
>
>
>
>
Received on Fri Sep 21 2007 - 10:56:30 EDT