Hi all, I am having some issues when estimating two residual error parameters
with the SAEM algorithm (NM7.1.2, PsN 3.2.4).
Below is my example, where each EPS is turned on/off by an indicator variable,
not sure what the behavior is like for error models such as Y=IPRD*EXP(EPS(1))
+ EPS(2), but I suspect it works fine.
I have two subject populations, TYPE = 2 and TYPE = 3. Good reason to suspect
TYPE = 3 is more variable, so I implemented the following in $ERROR of my PD
model
Y=(IPRD+ERR(1))*(3-TYPE)+(IPRD+ERR(2))*(TYPE-2)
$SIGMA
30 ; EPS type2
30 ; EPS type3
This is the estimation sequence
$ESTIMATION METHOD=ITS
PRINT=10 NOABORT NITER=100 NSIG=3 SIGL=6 CTYPE=2
$ESTIMATION METHOD=SAEM
ISAMPLE=2 NBURN=500 NITER=2000
PRINT=10 NSIG=3 SIGL=6 SEED=123345 CTYPE=3
$ESTIMATION METHOD=IMPMAP
EONLY=1 NITER=10 ISAMPLE=1000
PRINT=1 NSIG=3 SIGL=6 SEED=123345 CTYPE=3 CITER=10 CALPHA=0.05
Results from ITS are shown below, these estimates make sense.
ITS:
EPS1 EPS2
EPS1
+ 2.17E+01
EPS2
+ 0.00E+00 2.51E+01
After SAEM (which converges before end of burn in), the estimate of SIGMA2 has
blown up (at least I think it has, the "E" is missing, which I think is a
result of truncation since the exponent has 3 digits). All other THETA and
OMEGA estimates are reasonable and consistent with prior models.
SAEM:
EPS1 EPS2
EPS1
+ 2.23E+01
EPS2
+ 0.00E+00 2.95+304
I do not understand why this, would have thought the study has a reasonable
number of TYPE2:TYPE3 to estimate the two values (70:37), ITS think so.... I
would like to confirm if it is a data issue or a NONMEM issue.
Consequently, the IMPMAP (EONLY=1) estimation goes off the rails with very
large OBJ (which is usually in the range of 3000-4000), largely driven by
several clusters of subjects with very high individual contributions to IMPMAP
OBJ.
#METH: Objective Function Evaluation by Importance/MAP Sampling
EM/BAYES SETUP
THETAS THAT ARE MU MODELED:
1 2 4 5 6 7 8 9 10 11 12 13 14
THETAS THAT ARE SIGMA-LIKE:
MONITORING OF SEARCH:
iteration 0 OBJ= 178895.840910174
iteration 1 OBJ= 178911.252369277
iteration 2 OBJ= 178904.447527235
iteration 3 OBJ= 178901.025144545
iteration 4 OBJ= 178911.754485815
iteration 5 OBJ= 178902.859882922
iteration 6 OBJ= 178904.436625787
iteration 7 OBJ= 178909.875233486
iteration 8 OBJ= 178905.268884384
iteration 9 OBJ= 178908.849070488
Elapsed estimation time in seconds: 5830.79
iteration 10 OBJ= 178909.810873438
Any thoughts?
Cheers
Brendan