error convergence in POPAN

[fanny]

Good morning!

question about POPAN:


I can't run a model in POPAN because I obtain "error numerical convergence never reached".
I have read in the help that POPAN need the log link fonction and is used with the pent parameters. But if a use the log link fonction I obtain again the error. I don't now how to do for eliminate this problem of convergence. Somebody can tell me how to do?

Merci!

[Bill Kendall]

I suspect that when you ran POPAN you forgot to use the mlogit transformation for the pent's. As with transition probabilities in the multi-state model, these must add to 1.0, yet individually they can range between 0 and 1. Use the parm-specific option in the run screen. After you hit "OK" it will give you another screen where you can specify the transformation for each parameter. Use mlogit(1) for each of the pent parameters.

[sixtystrat]

If the pent's are set to Mlogit, should the other parameters be set to Logit?
I am getting reasonable answers in Popan when I use the loglog link. When I use the parameter specific link with Pents at Mlogit and the rest at logit, I get convergence but whacky estimates... Should I stick with the loglog link for everything? Thanks,
Joe Clark

[gwhite]

Joe:
You should be able to use the mlogit link on the pent parameters and any other link function on the remaining parameters, AS LONG AS you don't have some betas that are tied up with both the pent and other parameters. In other words, you want to be sure that the section of the design matrix for the pent parameters does not also have entries for other parameters. Although I suppose there might be models that would make biological sense if you violated this rule, I can't think of one off the top of my head.

I suggest you be sure to provide initial values as you start building more complex models so that the optimization starts out at a reasonable set of initial values. Start with dot models, and build towards the time-specific models.

Gary

[sixtystrat]

Thanks Gary.

[fanny]

Hello and thank for the help!!!

I have now some problems with the Qdeviance (in POPAN). I have 6 primary sessions and c=1,06.
When I test model PHI(t)P(t)PENT(t) or PHI(.)P(t)PENT(t) or PHI(.)P(.)PENT(t), the Qdeviance is null.
When I test model with PHI(t)P(t)PENT(.) I obtain Qdeviance=38885.92
I suppose the problem is with Pent because I use the mlogit(1). It is correct?
I would like to calculate the goodness of fit of my data. But with this Qdeviance, I can use the bootstrap. The only solution that I have is the test RELEASE?