NONMEM Users Network Archive

Hosted by Cognigen

Re: PRED for BLQ-like observations

From: Pavel Belo <nonmem>
Date: Fri, 20 Nov 2015 17:14:09 -0500 (EST)

Hello Nick,

I run NONMEM jobs with 2 files every time.  One file is for quick=
parameter estimates and the other one is for graphics.  The =
records can increase the run time considerably particularly there ~100=
more missing than non-missing DV (sparse data case).  My best guess is=
that 2-fold increase in the number of point will not increase the run
less than 2-fold, but it will increase it.

 On Fri, Nov 20, 2015 at 03:38 PM, Nick Holford wrote:
 > Pavel,
> Did you test the run time with double the records?
> I would expect that the MDV=1 records would be largely ignored in the=
> estimation step and not contribute much to run time.
> Nick
> On 21-Nov-15 08:59, Pavel Belo wrote:
>> Thank you Bill,
>> In my case it exactly doubles the number of records... The records
>> are daily measures and the code is running slow enough. I'll split
>> the code into estimation part and one that that is redundant, but
>> uses a larger file and creates an output. It will be something like
>> I guess the best future way is modify something in NONMEM so there is
>> an option to provide only PRED in the PRED column (version 7.4?).
>> Thanks!
>> Pavel
>> On Fri, Nov 20, 2015 at 01:06 PM, Denney, William S. wrote:
>> Hi Pavel,
>> The easiest way that I know is to generate your data file with
>> one
>> set of rows for estimation with M3 and another row just above or
>> below with MDV=1. NONMEM will then provide PRED and IPRED in the
>> rows with MDV=1.
>> Thanks,
>> Bill
>> *From:*owner-nmusers
>> [mailto:owner-nmusers
>> *Sent:* Friday, November 20, 2015 11:47 AM
>> *To:* nmusers
>> *Subject:* [NMusers] PRED for BLQ-like observations
>> Hello The NONMEM Users,
>> When we use M3-like approach, the outputs has PRED for
>> non-missing
>> observations and something else for BLQ (is that PRED=CUMD?). As
>> in the diagnostic figures PRED for BLQs looks like noise, I
>> remove
>> them. It is not always perfect, but OK in for most frequent
>> cases.
>> When we use count data such as a scale with few possible values
>> (for example, 0, 1, 2, 3, 4, 5), it makes more sense to use PHI
>> function (home-made likelihood) for all observations rather than
>> to treat the count as a continuous variable an apply M3-like
>> approach to 1 and 5 while only (as we know, they are like LLOQ
>> and
>> ULOQ). In this case, all PRED values look like noise. A hard
>> way
>> to replace the noise with PRED value is to simulate PRED for each
>> point and merge them with the DV and IPRED data. Is there an easy
>> way?
>> (The model runs well and better than when the count is treated as
>> a continuous variable.)
>> Thanks!
>> Pavel
> --
> Nick Holford, Professor Clinical Pharmacology
> Dept Pharmacology & Clinical Pharmacology, Bldg 503 Room 302A
> University of Auckland,85 Park Rd,Private Bag 92019,Auckland,New
> Zealand
> office:+64(9)923-6730 mobile:NZ+64(21)46 23 53
> email: n.holford
> Holford SD, Allegaert K, Anderson BJ, Kukanich B, Sousa AB, Steinman
> A, Pypendop, B., Mehvar, R., Giorgi, M., Holford,N.H.G.
> Parent-metabolite pharmacokinetic models - tests of assumptions and
> predictions. Journal of Pharmacology & Clinical Toxicology.
> 2014;2(2):1023-34.
> Holford N. Clinical pharmacology = disease progression + drug action.=
> Br J Clin Pharmacol. 2015;79(1):18-27.
Received on Fri Nov 20 2015 - 17:14:09 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: