RE: [NMusers] BAYES Estimation: OBJECTIVE FUNCTION IS INFINITE

From: Klaas Prins <klaas.prins_at_qpharmetra.com>
Date: Fri, 12 Jun 2015 10:10:34 +0000

Hi Xinting,

I frequently use METH=BAYES and have had the same observation at various occasions. First let me say that BAYES a wonderful estimation engine that has helped me to model PK, PKPD and odd-type data in cases where FOCE failed, even where SAEM and IMP(MAP) failed (recently shown at PAGE). In all situations of infinite OFV I managed to get it to run by making sure the search would not lead to extremely large or extremely small ( positive ) numbers. For example if you are looking for a power estimate and you have some idea about it going to be in the order of 1, then ensure that your safety valves in the code prevent it from going close to zero or larger than e.g. 3 (arbitrary number). While the estimate at extreme values may not be problematic per se, one could think that in combination with the other parameters (I believe BAYES considers all parameters ‘completely free to choose’ within the initial estimate bounds during estimation) completely weird predictions could lead to the OFV being blown up. If I read about your problem being high-dimensional, there may be many relationships in your model at large to make the OFV go poof. One argument against directing the parameters in a certain direction is that it may introduce bias, so it comes at a certain cost. Furthermore I can recommend to look into the bayes.ext file you’re writing to disk and you may see at what parameter combinations OFV goes bazurk; that may inform you further on extreme values.

Apparently BAYES is more flexible in the search which provides combinations of parameters that will send the OFV into the wild. Perhaps Bob Bauer can add some insights here – I know too little about what happens under the hood.

By the way, why are you considering BAYES when you get SAEM and IMP to run just fine? Why not use SAEM results and run IMP with EONLY=1 afterwards?

Hope this helps,
Klaas Prins

________________________________________
Klaas Prins, PhD
Vice President EU
[QPharmetra_noTag_FA little.png]
+31 618 802 292 (M)
+31 248 457 290 (O)
klaas.prins_at_qpharmetra.com<mailto:kevin.dykstra_at_qpharmetra.com>

This e-mail communication is confidential and is intended only for the individual(s) or entity named above and others who have been specifically authorized to receive it. If you are not the intended recipient, please do not read, copy, use or disclose the contents of this communication to others. Please notify the sender that you have received this e-mail in error by replying to the e-mail or by telephoning +31 618 802 292. Please then delete the e-mail and any copies of it. Thank you.



From: owner-nmusers_at_globomaxnm.com [mailto:owner-nmusers_at_globomaxnm.com] On Behalf Of Xinting Wang
Sent: Friday, June 12, 2015 4:06 AM
To: nmusers_at_globomaxnm.com
Subject: [NMusers] BAYES Estimation: OBJECTIVE FUNCTION IS INFINITE

Dear All,
I am using a cascade of the EM estimation methods implemented in Nonmem 7.2 to solve a high-dimension PK/PD problem. The estimation chain is coded as below:

$EST MET=SAEM INTER NBURN=2000 NITER=1000 PRINT=5 NOABORT NOPRIOR=0 ISAMPLE=10 SIGL=6 CTYPE=3 IACCEPT=0.4 CONSTRAIN=1 SEED=145612 FILE=Saem.ext

$EST MET=IMP INTER NITER=2000 ISAMPLE=1000 PRINT=5 SEED=581987 SIGL=6, CTYPE=3 MAPITER=0 FILE=Imp.ext

$EST MET=BAYES INTER NBURN=3000 NITER=25000 PRINT=10 NOABORT NOPRIOR=0 FNLETA=0 ISAMPLE_M1=2 ISAMPLE_M2=2 ISAMPLE_M3=2 IACCEPT=0.4 SEED=231457 FILE=Bayes.ext

​The first two estimation methods run successfuly with a convergence, but when it comes to BAYES, an error message always occurs:

​ OBJECTIVE FUNCTION IS INFINITE. PROBLEM ENDED
1THERE ARE ERROR MESSAGES IN FILE PRDERR

​I put safety brakes in the code to prevent ​0**power or square root of negative values, but the problem still occured. Setting lower and upper bound of the initial estimation, but it did not help either.

I am wondering if anyone here encountered this issue previously? It seems quite strange to me that SAEM and IMP could succeed, but BAYES fails. Could you please let me know of the possible reasons behind this? Thanks very much.
Best Regards

--
Xinting

image001.png
(image/png attachment: image001.png)

Received on Fri Jun 12 2015 - 06:10:34 EDT

This archive was generated by hypermail 2.3.0 : Fri Sep 27 2019 - 16:44:38 EDT