From: joannellyn.y.chiu@gsk.com
Subject: [NMusers] Error when combining proportional residual error,
Date: Wed, 2 Aug 2006 10:41:49 -0400

Hi,

My model had 4 proportional (eps 1,3,5,7) and 4 (eps 2,4,6,8)  additive residual errors.  I had
wanted to combine the proportional errors (eps 1,3 & eps 5,7) but wanted to leave the additive
residual errors individually.  when i did this, I received an error that states:

0PROGRAM TERMINATED BY OBJ
 ERROR IN CELS   WITH INDIVIDUAL  934   ID=0.93400000E+03
 WEIGHTED SUM OF "SQUARED" INDIVIDUAL RESIDUALS IS INFINITE
 MESSAGE ISSUED FROM ESTIMATION STEP
 AT INITIAL OBJ. FUNCTION EVALUATION

when i removed subject 934 and 935 from the dataset, it still came up with the same error.  I've
even tried increasing the initial estimates.  If anyone can help, it'll be much appreciated.

thank you,
Joannellyn

the code is:


IPRD=F

IF (STDY.EQ.1) THEN
Y=IPRD*EXP(EPS(1))+EPS(5)
ENDIF
IF (STDY.EQ.2) THEN
Y=IPRD*EXP(EPS(2))+EPS(6)
ENDIF
IF (STDY.EQ.3) THEN
Y=IPRD*EXP(EPS(3))+EPS(7)
ENDIF
IF (STDY.EQ.4) THEN
Y=IPRD*EXP(EPS(4))+EPS(8)
ENDIF


$SIGMA BLOCK(1)  1          
$SIGMA BLOCK(1) SAME  
$SIGMA BLOCK(1)  1    
$SIGMA BLOCK(1) SAME  
$SIGMA
 20 ; EPS(5)
 1000 ; EPS(6)  
 1000 ; EPS(7)
 1000 ; EPS(8) 
_______________________________________________________

From: "Serge Guzy" GUZY@xoma.com
Subject: RE: [NMusers] Error when combining proportional residual error,
Date: Wed, 2 Aug 2006 09:14:55 -0700

I think the problem is that you did not define really a combined proportional with
additive but instead an exponential error + additive.
Can you try to write instead
IF (STDY.EQ.1) THEN
Y=IPRD* (1+EPS(1)) +EPS(5)
ENDIF
IF (STDY.EQ.2) THEN
Y=IPRD* (1+EPS(2)) +EPS(6)
ENDIF
IF (STDY.EQ.3) THEN
Y=IPRD* (1+EPS(3)) +EPS(7)
ENDIF
IF (STDY.EQ.4) THEN
Y=IPRD*(1+EPS(4)) +EPS(8)
ENDIF
 
In the MCPEM program for example, an exponential error variance does not really exist. When you
assume exponential error variance, the data are log-transformed and the error model is then
assumed to follow a constant variance model. In the MCPEM it does not make any sense to combine
exponential with additive.

I guess something similar is used with NONMEM.

Serge Guzy
President POP-PHARM
_______________________________________________________

From: Leonid Gibiansky leonidg@metrumrg.com
Subject: RE: [NMusers] Error when combining proportional residual error,
Date: Wed, 02 Aug 2006 13:22:29 -0400

No, this is not the case with NONMEM. It is perfectly OK to combine exponential and
additive errors. My guess is that STDY is not equal to 1, 2, 3, or 4 for some observation,
or there exist some other model coding error that is difficult to find without seeing
the entire code.

To avoid STDY problem, try

     Y=IPRD*EXP(EPS(1))+EPS(5)
     IF (STDY.EQ.2) THEN
     Y=IPRD*EXP(EPS(2))+EPS(6)
     ENDIF
     IF (STDY.EQ.3) THEN
     Y=IPRD*EXP(EPS(3))+EPS(7)
     ENDIF
     IF (STDY.EQ.4) THEN
     Y=IPRD*EXP(EPS(4))+EPS(8)
     ENDIF
Leonid 
_______________________________________________________

From: "Serge Guzy" GUZY@xoma.com
Subject: RE: [NMusers] Error when combining proportional residual error,
Date: Wed, 02 Aug 2006

Hi Leonid.

When you use exponential error model, don't you think NONMEM transform
your Concentration data into LogC???
If not, how do you compute the Log-Likelihood?? Of course I think in
terms of MCPEM but I cannot understand how you can combine an
exponential error with additive and compute a log-likelihood.
More specifically, how do you compute the standard deviation that is
needed in the computation of the log-likelihood?

Log-likelihood= -log(sd)-((Cobs -CPREDICTED)/(SD))2/2

Serge
_______________________________________________________

From: Leonid Gibiansky leonidg@metrumrg.com
Subject: RE: [NMusers] Error when combining proportional residual error,
Date: Wed, 02 Aug 2006 14:18:48 -0400

NONMEM does not transform data into the log form. As to the likelihood step
evaluation, I think NONMEM uses Taylor expansion

exp(eta)=1+eta

effectively fitting the proportional rather than exponential error model. On the
simulation step the exact exponential model is implemented.

Leonid 
_______________________________________________________