Precision of parameter estimates (CV% in NLME model)

Phoenix Certara Forums Modelling and Simulation
1

Why does NLME sometimes not give a CV% (precision) for parameter estimates? For example, when I run different scenrarios, some have CV% and CI’s for the parameter estimates and others don’t (even though the AIC value may be lower for the scenario for which no CV% is given).

2

this is common in any regression software where you need to invert a matrix. The hessian might not be obtainable and then you cannot compute standard error. More often than not the model is overparametrized/unstable.

While a good indicator you should not just rely on CV% or the standard error computations to decide whether a model is good

3

Here is what I do suggest to try.

Copy the model to the workflow, click on parameters/fixed and accept all fixed and random.

Then click on run options and select QRPEM and ask for 1 iteration and standard errors.

the Hessian inversion as Samer indicated is usually the issue and is more stable when using QRPEM (no link with the QRPEM method al all but just different code).

Now QRPEM requires a mu kind of modeling that for simple models and log-linear covariate relationships is defaulted but if your code is more complicate, then you need to have all the stparm statements with only fixed effects and random effects.

If you have problems with that, best is to send just the coding part. I do not need the data set. If it is built in models and the covariate relationship selected are the “power” kind, then no problem with QRPEM.

Best Regards

Serge

4

Thank-you both for your replies. Serge, I tried using QRPEM, but it still did not give me any CV%. The model is not very complex. How can I send you the code?

5

Ronette, you could just copy and paste the code in to the text here, or perhaps if you can attach the whole project as that might make tweaking settings etc. easier.

Simon.

6

Dear Serge and Simon,

Here is the code. What I find puzzling is that, when I run scenarios, I get CV%s for the typical parameter values for some scenarios and not for others. And the ones I don’t get CV%s for are the ones with the lowest AIC values. I’m attaching two of the results files that I think make my problem clearer. I’m hesitant to post the whole project file on this forum as the data have not yet been published, but could e-mail it to you individually if that would help.

Thanks!

test(){

cfMicro(A1, Cl / V, first = (Aa = Ka))
dosepoint(Aa)
C = A1 / V
error(CEps = 0.409528)
observe(CObs = C * (1 + CEps))
stparm(Ka = tvKa * (1+dKadFed1*(Fed==1)) * exp(nKa))
stparm(V = tvV * (Weightg/mean(Weightg))^dVdWeightg * exp(nV))
stparm(Cl = tvCl * (Weightg/mean(Weightg))^dCldWeightg * exp(nCl))
fcovariate(Fed())
fcovariate(Weightg)
fixef(tvKa = c(, 0.464714, ))
fixef(tvV = c(, 1309.68, ))
fixef(tvCl = c(, 413.961, ))
fixef(dKadFed1(enable=c(0)) = c(, -0.533343, ))
fixef(dVdWeightg(enable=c(1)) = c(, 1.6833, ))
fixef(dCldWeightg(enable=c(2)) = c(, 1.95815, ))
ranef(diag(nV, nKa, nCl) = c(3.9289587E-06, 0.099044731, 0.016953372))
}
Overall.xls (10.5 KB)Theta.xls (12 KB)