6 replies to this topic

[RB2017]

Hi everybody,

 

can anyone tell me why after running a Phoenix model, sometimes in some individuals, value of parameter estimates appears but no data of stderr, CV% and CI appears for these parameters in theta output? .

 

Thank you very much!

[Simon Davis]

Dear Maria-Rosa,

   This is not uncommon and does not necessarily mean your model is 'bad', however please see some suggeestions below (sometimes just re-executing with final estimates accepted is sufficient, without the need to change algorithms).

 

  Simon.

 

 

  The standard error and %CV calculations can be tricky, particularly when a model is over-parameterized. The FOCE methods do not always successfully compute a standard error. Phoenix NLME is using a lot of background calculations to compute the standard error of given parameters and from that, the %CV, Confidence Intervals, and VIFs are computed. The issue is that some of these systems are rather complex, making the calculation of the standard error difficult.

 

One option is to switch to the QRPEM method in the model fitting algorithm. The reason this works in some cases is because QRPEM doesn't rely on a numerical optimization procedure because QRPEM only approximates the likelihood using numerical integration instead of numerical derivatives and optimization.

 

Here are some suggestions to improve the chances of calculating standard errors (SE):

 

1. First, fit the model with FOCE-LB or FOCE-ELS. On the Parameters tab, Fixed Effects sub-tab, click Accept All Fixed and Random. On the Run Options tab, change the method to QRPEM and run with 1 iteration (N iter = 1). Then execute the model using QRPEM to calculate the standard errors.

 

2. If still no SE, then you can decrease the number of model parameters to be estimated by freezing some parameters you believe could cause the problem. For example, freezing the entire omega and/or some fixed effect parameters.

 

3. On the Run Options tab, you can also change the error estimation method by using forward rather than central or fisher score rather than hessian.

 

4. If none of this works, then use the Bootstrap method.

[RB2017]

Thank you very much Simon,

 

just re-executing with final estimates accepted works well and I can obtain the stderr and CV%!. The problem now is that they are enormous in some parameters, so...I will try another model..

 

Thank you very much!

 

MR

[joybaker]

Hello, I am going through the check list to see if any modification can get CV% and CI from my model. Unfortunately none of it works. I have a three compartment model with zero and first order absorption and several covariates. The model has been verified with GOF plot and VPC, but the bootstrap results are not right, so I wonder if I get a right setting.

Do you have any suggestion related with the setting?

 

Please notice that the Stderr and CV% are still ALL ZERO even without the categorical covariates stratification.

 

 

 

 

Attached Thumbnails

• 
• 

[Simon Davis]

it maybe indicative of some uncertaintiy in the model - it's pretty complicated looking at that boot theta sheet, can you post the original theta fit or better yet the project to try and help you further?

[joybaker]

it maybe indicative of some uncertaintiy in the model - it's pretty complicated looking at that boot theta sheet, can you post the original theta fit or better yet the project to try and help you further?

 

Dear Simon, thank you for your reply. The original theta fit is as the shown below. Actually I have send the file through support@certara.com titled "Bootstrap setting and abnormal CV% _FROM joy baker" a while ago. But I didn't get reply yet, thus why I came back here  again to look for some advice.

Attached Thumbnails

• 

[Simon Davis]

Joy - please forward your email directly to me as I can't see your case in support and i would like to work out what happened.  Simon.davis_AT_certara.com (replace the _AT_ for @ if you're not a robot ;0)