RE: Different EBE estimation between original and enriched dataset with MDV=1
Dear All,
Thanks for your detailed response and tricks. I am trying to address each
of them after several trial and errors with your suggestions:
1) I have only time-invariant covariates. Buth thanks to Robert and Bill
for mentioning it. I will remember!
2) I did not use the EVID=2 for my dummy times. Now I am using them, but
it does not help.
3) Starting from non optimized parameters rather than $MSFI as suggested
by Joachim does not help. But I like your explanation. Nevertheless I
can't live with "the differences [...] within the range you would also
find if you did a bootstrap" since those differences change the profiles I
observe.
4) The nice trick suggested by Heiner (After the last time point of an ID
you may add a line with EVID=3 (reset event) with the TIME (TIMERESET>the
last datapoint of the ID of interest) may work, but would probably be too
complex to implement for my special dataset since I have a long history of
not evenly spaced dosing. But thanks, Heine, I will also remember this
one.
5) Increasing the TOL is the only thing that improves the prediction.
Thanks Leonid you are right when you write "the problem is in the
precision of the integration routine". But with the data I have, I cannot
increase it beyond 8. By the way, in my model I am estimating the initial
condition at baseline in one of my compartment using a random effect. When
the slope after the baseline is large, I got almost no bias. But when it
is a moderate slope, the bias prediction with dummy points appears and is
increasing when the slope is decreasing. This probably confirms the issue
of the precision with integration routine.
6) The only solution which I mention in in my 1st Email and that was also
suggested by Jean Lavigne : one separate run for the estimation of the
EBEs and one from the simulation on dummy time points.
7.2) Thanks Robert. I am glad to learn that in 7.3 there will be an option
to automatically "fill in extra records with small time increments,
to provide smooth plots". I imagine that using this utility program will
not change the precision of the integration routine since it will be build
in. I will just have to wait a little bit for getting access to it.
Kind regards,
Pascal
PS
As someone who used to live by the Lake Leman would have said, NONMEM,
sometimes, "It's a kind og magic!" :-)
Quoted reply history
From: Herbert Struemper <[email protected]>
To: "[email protected]" <[email protected]>
Date: 26/11/2012 16:13
Subject: RE: [NMusers] Different EBE estimation between original
and enriched dataset with MDV=1
Sent by: [email protected]
Pascal,
I had the same issue a while ago with time-invariant covariates. Back then
with NM6.2, adding an EVID column to the data set and setting EVID=2 for
additional records preserved the ETAs of the original estimation (while
only setting MDV=1 for additional records did not).
Herbert
Herbert Struemper, Ph.D.
Clinical Pharmacology, Modeling & Simulation
GlaxoSmithKline, RTP, 17.2230.2B
Tel.: 919.483.7762 (GSK-Internal: 7/8-703.7762)
-----Original Message-----
From: [email protected] [mailto:[email protected]]
On Behalf Of Bauer, Robert
Sent: Sunday, November 25, 2012 9:11 PM
To: Leonid Gibiansky; [email protected]
Cc: [email protected]
Subject: RE: [NMusers] Different EBE estimation between original and
enriched dataset with MDV=1
Pascal:
There is one more consideration. If your model depends on the use of
covariate data, then during the numerical integration from time t1 to t2,
where t1 and t2 are times of two contiguous records, which have values of
the covariate c1 and c2, respectively, NONMEM uses the covariate at time
t2 (call it c2)during the interval from t>t1 to t<=t2. During your
original estimation, your data records were, perhaps, as an example:
Time covariate MDV
1.0 1.0 0
1.5 2.0 0
With the filled in data set, perhaps you filled in the covariates as
follows:
Time covariate MDV
1.0 1.0 0
1.25 1.0 1
1.5 2.0 0
Or perhaps you made an interpolation for the covariate at the inserted
time of 1.25, to be 1.5. But NONMEM made the following equivalent
interpretation during your original estimation:
Time covariate MDV
1.0 1.0 0
1.25 2.0 1
1.5 2.0 0
That is, when the time record 1.25 was not there, it supplied the
numerical integrater with the covariate value of 2.0 for all times from
>1.0 to <=1.5, as stated earlier.
Even though MDV=1 on the inserted records, NONEMM simply does not include
the DV of that record in the objective function evaluation, but will still
use the other information for simulation, by simulation I mean, for the
numerical integration during estimation.
In short, your model has changed regarding the covariate pattern based on
the expanded data set.
By the way, there is a utility program called finedeata, that actually
facilitates data record filling, with options on how to fill in
covariates, in nonmem7.3 beta. I will send the e-mail to this shortly.
If you are not using covariates in the manner I described above, then
please ignore my lengthy explanation.
Robert J. Bauer, Ph.D.
Vice President, Pharmacometrics, R&D
ICON Development Solutions
7740 Milestone Parkway
Suite 150
Hanover, MD 21076
Tel: (215) 616-6428
Mob: (925) 286-0769
Email: [email protected]
Web: www.iconplc.com
-----Original Message-----
From: [email protected] [mailto:[email protected]]
On Behalf Of Leonid Gibiansky
Sent: Friday, November 23, 2012 12:15 PM
To: [email protected]
Cc: [email protected]
Subject: Re: [NMusers] Different EBE estimation between original and
enriched dataset with MDV=1
Hi Pascal,
I think the problem is in the precision of the integration routine. With
extra points, you change the ODE integration process and the results. I
would use TOL=10 or higher in the original estimation. I have seen cases
when changing TOL from 6 to 0 or 10 changed the outcome quite
significantly.
Leonid
--------------------------------------
Leonid Gibiansky, Ph.D.
President, QuantPharm LLC
web: www.quantpharm.com
e-mail: LGibiansky at quantpharm.com
tel: (301) 767 5566
On 11/23/2012 11:08 AM, [email protected] wrote:
> Dear NM-User community,
>
> I have a model with 2 differential equations and I use ADVAN6 TOL=5.
> In $DES, I am using T the continuous time variable. The run converges,
> $COV is OK, and the model gives a reasonable fit. In order to compute
> some statistics which cannot be obtained analytically, I need to
> compute individual predictions based on individual POSTHOC parameters
> and an extended grid of time for interpolating the observed times.
>
> So I have
> 1) added to my original dataset extra points regularly spaced with
> MDV=1. To give you an idea, my average observation time is 25, with a
> range going from 5 to 160. So my grid was set so that I have a dummy
> observation every 1 unit of time.
> 2) rerun my model using $MSFI to initialize the pop parameters, with
> MAXEVAL=0 and POSTHOC options so that individual empirical Bayes
> estimates (EBE) parameters for each patient would be first
> re-estimated, then the prediction would be computed.
>
> Then I
> 3) checked that my new predictions computed from the extended dataset
> match the predictions of the original dataset at observed time points.
> I had the surprise to see that for some individuals those predictions
> match, for some others they slightly diverge, and for few others they
> are dramatically different. I checked the EBEs and they were clearly
> different between the original dataset and the one with the dummy
points.
> 4) I decided to redo the grid with only one dummy point every 1/4 of
> time unit. The result was less dramatic, but still for most of my
> individuals the EBEs predictions were diverging from the original ones
> computed without the dummy times.
>
> Of course the solution for me is to estimate the EBEs from the
> original dataset, export them in a table and reread them to initialize
> the parameter of my individuals using only dummy time points and no
> observations.
>
> This problem reminds me something that was discussed previously on
> nm-user, but I could not recover the source in the archive.
>
> Anyway is this something known and predictable that when adding dummy
> points with MDV=1 to your original dataset you sometimes get very
> different EBEs ? Are there cases/models/ADVAN where the problem is
> likely to happen? Is their a way to fix it it in NONMEM other than the
> trick I used?
>
> Thanks for your replies!
>
> Kind regards,
>
> Pascal Girard, PhD
> [email protected]
> Head of Modeling & Simulation - Oncology Global Exploratory Medicine
> Merck Serono S.A. * Geneva
> Tel: +41.22.414.3549
> Cell: +41.79.508.7898
>
> This message and any attachment are confidential and may be privileged
> or otherwise protected from disclosure. If you are not the intended
> recipient, you must not copy this message or attachment or disclose
> the contents to any other person. If you have received this
> transmission in error, please notify the sender immediately and delete
> the message and any attachment from your system. Merck KGaA,
> Darmstadt, Germany and any of its subsidiaries do not accept liability
> for any omissions or errors in this message which may arise as a
> result of E-Mail-transmission or for damages resulting from any
> unauthorized changes of the content of this message and any attachment
> thereto. Merck KGaA, Darmstadt, Germany and any of its subsidiaries do
> not guarantee that this message is free of viruses and does not accept
> liability for any damages caused by any virus transmitted therewith.
>
> Click _ http://www.merckgroup.com/disclaimer_to access the German,
> French, Spanish and Portuguese versions of this disclaimer.