1 reply to this topic

[Yassine Kamal Lyauk]

Hi,

 

I am modeling with covariates but it seems I am trying to incorporate too many at a time as Phoenix tells me there is a limit of a maximum of 32 covariates (it makes sense since I am checking for covariates on all parameters of my three compartment base model).

 

Therefore I separated my covariate investigation into three separate runs, each containing different covariates. My question is is this wise since I might be ruling some co-linearity between covariates and if so should I go with the covariate model scoring the best? Do you see other way of performing this task?

 

I am new to covariate modeling so please bear with me.

 

Best regards,

 

Yassine

 

PS: I am using the stepwise covariate search option.

[serge guzy]

Dear Yassine

Here is what I suggest you to do to decrease the number of covariate relationship to which you will apply the covariate search.

Another option is to use scenarios (one of the options in "run options") as a first step instead of directly covariate search.

Steps to perform

1: run the base model (simple scenario)

2: copy the base model after it ran to the workflow and accept all fixed and random, now go to parameter_structural and click on all potential relationships of interest

3: go to run options and select scenarios as option (not more than 30 please, in the new version coming next year, no limit). If you need more than 30, just do twice that part

4: go to parameters/scenarios and click on add to define the scenarios one by one. Here it is

easy because you will select only one covariate with one model parameter as a scenario (see the attached project)

5: run all these scenarios and look at results (click on overall tab to see al the results)

6: remember only the one that lead to a change in -2LL compare to the base model of more than 4 (this is arbitrary but to choose the good one it is a good number, it corresponds to a p value of about5%)

7: copy the model to the workflow and go to parameters/structural and select only the pairs (parameter/covariate) that were remembered as statistically significant

8: now choose covariate search (check before that you do not have more than 32 covariate relationship) and run it. When you look at the results, (click on parameters/scenarios to see the final selected model(the one having the box "use" checked.

 

 

In the attached example, after the scenarios step, V with sex and V with weight are the only one to consider. When I then ran the covariate search, since the program suggests you to use 10 as stat difference, the final result is the best covariate model is without any covariate relationship. Here this example was to illustrate how you can in 2 steps avoids the use of the covariate search without too many covariate relationships

 

Best Regards;

Serge [file name=covariate_official_version.phxproj size=2331939]http://pharsight.com/extranet/media/kunena/attachments/legacy/files/covariate_official_version.phxproj[/file]

Attached Files

•  covariate_official_version.phxproj   2.22MB   498 downloads