8 replies to this topic

[raghava choudary]

Dear Simon,

 

Hope all is well,

 

I had a problem with attached project where i could nt get the standard errors,

Looking forward for your reply,

 

With best regards,

 

Raghav, [file name=123_IV_PO_260511_Dog.phxproj size=1298514]http://pharsight.com/extranet/media/kunena/attachments/legacy/files/123_IV_PO_260511_Dog.phxproj[/file]

Attached Files

•  123_IV_PO_260511_Dog.phxproj   1.24MB   0 downloads

[Simon Davis]

Hi Raghav,

It's not necessary to address all your messages to me - there are several hundred other users who may like to respond to your questions ;0)

I loaded your project and selected FOCE-LB algorithm and it executed fine with SEs calculated. My understanding of the FO method is that it is only intended for 'quick and dirty' runs to get/confirm initial estimates, since you have already performed a Naive Pooled run i think this FO run is redundant.

Hope that helps - enjoy the w/e, I am going to be taking a break over the next few weeks as I've had a manic couple of months travelling and training.

Simon.

[Shyeilla Dhuria]

I am having a similar problem where the standard errors and CV% are not computed for the population. I have tried using different engines but have not had any success. Any ideas on what to try next?

[Simon Davis]

Hi Shyeilla - without seeing your project file then it's hard to locate exactly what the cause might be - I seem to recall seeing the problem of no SEs was combined with a failure in the Hessian matrix. I would be careful that your initial estimate for Epsilon is reasonable - e.g. if you have a multiplicative error model you could enter 0.1 here to give an equivalent of 10%

whereas with an additive error the initial estimate here would be an absolute number e.g. if your concs ranged over 10 to 1000 and you have an value of only 1 then you are saying you expect only a 0.1% residual error at Cmax. Here is a reminder of the equivalent error models between PML (in bold) and WNL classic.

Phoenix WinNonlin Classic WinNonlin
Additive: C + CEps = Homoscedastic Uniform
Multiplicative: C*(1 + CEps) = Constant CV 1/Yhat*Yhat
Power: C + Cpower*Ceps (Yhat)-2*power
Power = 0.5: C + C0.5*CEps = Poisson 1/Yhat
Mixed: C + CEps + C*CEps*CMixRatio Custom: user-defined

Check to see if that helps you if not then please let us know the s/w version and a copy of your PHXPROJ illustrating the issue. If confidentiality of data is an issue we may be able to help via Support (I just enabled your account for self service by the way.)

Simon.

Attached Thumbnails

• 

[Shyeilla Dhuria]

Thanks for the explanation on selecting reasonable initial estimates for epsilon. I understand how to choose the estimates for additive and multiplicative error models, but it is not clear what to use for other error models (i.e., CmixRatio, power, etc.). If you could provide examples for these other error models that would be helpful.

 

Unfortunately, I cannot share the project file due to confidentiality, but I can describe the issues in more detail (using NLME v1.0):

 

I did not have a problem using the FOCE-LB engine with the additive error model. However, it does not appear to be an appropriate error model. I copied the file and evaluated the multiplicative error model with a reasonable initial estimate (stdev=0.5). FOCE-LB provided estimates with stderr, but the return code was -1, indicating an error with the convergence. When I tried other engines, an error message appeared. Only using the FO engine was I able to get estimates with stderr and a retcode=1. However my understanding is that the FO method is generally not recommended for final estimates. Any advice on how to use the FOCE-LB or FOCE-ELS methods to get an acceptable return code? Also, when comparing different error models, is it recommended to use the same engine? I also explored the CmixRatio error model but it is not clear to me what to enter for the initial estimate for epsilon. In this case the output contains parameter estimates but no stderr.

 

Many thanks for your advice and help!

[Simon Davis]

I think that giving an intial Multiplicative Stdev of 0.50 maybe a little high as it would correspond to a 50% CV 

this image helped me understand the mixed model better;

Mixed Stdev of 0.1 with CMixRatio of 1.5 means we have an additive component of 0.1 (µg/L for example) and a proportional part of: 0.1*1.5 = 0.15 = 15% CV

Does that help - a log-additive error may also be one to consider - try looking at how the code changes as you select different options ?

there is a course next month 8-10 November in King of Prussia, PA - that could be worth considering Population Modeling Methodology using Phoenix NLME

Simon,

Attached Thumbnails

• 
• 

[satyawanjadhav]

Just wondering if someone can help on standard errors and %CV (file attached). I tried multiple combinations and epsilon estimates but not able to get stderr and %CV for thetas.

 

Thanks

 

 

 

 

 

 

I think that giving an intial Multiplicative Stdev of 0.50 maybe a little high as it would correspond to a 50% CV

this image helped me understand the mixed model better; <script pagespeed_no_defer="">//

Attached Files

•  PK-1.phxproj   1.28MB   374 downloads

[Simon Davis]

Dear Satyawan, please look at the attached project, I took your second model and went back to using FOCE ELS, (from Naive pooled) and I used your last estimates for starting points for my next run. for interest I also added a 1-com model since you don't ahve a lot of data to support the 1st compartment in my opinon, however the fit was still slightly better (with 2com) also I added a model comparer object, is this what you were looking for? Simon

Attached Thumbnails

• 

Attached Files

•  PK_1saad.phxproj   2.16MB   381 downloads

Edited by Simon Davis, 20 April 2016 - 09:34 PM.

[satyawanjadhav]

Dear Satyawan, please look at the attached project, I took your second model and went back to using FOCE ELS, (from Naive pooled) and I used your last estimates for starting points for my next run. for interest I also added a 1-com model since you don't ahve a lot of data to support the 1st compartment in my opinon, however the fit was still slightly better (with 2com) also I added a model comparer object, is this what you were looking for? Simon

Thanks Simon. Actually I had also got it worked (by changing optimization method and some fine tuning of initial estimates) after I had posted it. But still wondering why it didn't work in my original model (full block) despite similar estimates with what you gave from naïve pooled.