Surprising number of estimates

questions concerning analysis/theory using programs M-SURGE, E-SURGE and U-CARE

Surprising number of estimates

Postby AlexAv » Fri Mar 08, 2013 6:39 am

Hi,

I am fitting a multievent model. Hereafter one of the matrix of constraint I am using:
* sa - - sb - - - - - - -
- * - - - - - - - - - -
- - * - - - - - - - - -
- - sa+ * - sb - - - - - -
- - - - * - - - - - - -
- - - - - * - - - - - -
- - - - - - * - - - - -
- - - - - - - * - - - -
- - - - - - sb+ - * sa - -
- - - - - - - - - * - -
- - - - - - - - - - * -
- - - - - - - sb+ - - sa+ *
- - - * - sb - - - - - -
- - - * - sb - - - - - -
- - - - - - - - * sa - -
- - - - - - - - * sa - -
- - - - - - - - - - sa+ *
- - - - - - - - - - sa+ *
- - - - - - - sb+ - - - *
- - - - - - - sb+ - - - *

All the states are mutually exclusive. I want all 'sa' parameters to be identical (and the same for sa+, sb and sb+ respectively).
so for example, for 'sa' I am using the gemaco sentence 'from(1 9 15 16).to(2 10)' or simply 'to(2 10)'.
Using either the first or the second sentence, I have indeed four parameters, one for each s, as I expected. But surprisingly, I get 3 different estimates per parameters, for example, for sa, I get one estimate
for transitions from(1).to(2), one for transition from(9).to(10) and one for transition from(15 16).to 10. I don't understand why.

Is this result mean that I can't constrain the parameters to be equal using this matrix?

thanks in advance

Alex
AlexAv
 
Posts: 12
Joined: Tue Sep 18, 2012 8:19 am
Location: Sweden and France

Re: Surprising number of estimates

Postby simone77 » Fri Mar 08, 2013 7:07 pm

Hi Alex,

I believe this sentence should do what you want:
f(1).to(2)&f(9).to(10)&f(15).to(10)&f(16).to(10)
simone77
 
Posts: 197
Joined: Mon Aug 10, 2009 2:52 pm

Re: Surprising number of estimates

Postby AlexAv » Mon Mar 11, 2013 5:38 am

Hi Simone

Thanks for your reply
I thought about that too and I am actually trying that sentence. But I still do not understand why these sentences give the same number of parameters...

Alex
AlexAv
 
Posts: 12
Joined: Tue Sep 18, 2012 8:19 am
Location: Sweden and France

Re: Surprising number of estimates

Postby AlexAv » Mon Mar 11, 2013 8:44 am

Hi again

I ve run the model with the sentence proposed; the pb still remains, the number of parameter is correct but I got three different estimates per parameter...

Alex
AlexAv
 
Posts: 12
Joined: Tue Sep 18, 2012 8:19 am
Location: Sweden and France

Re: Surprising number of estimates

Postby simone77 » Mon Mar 11, 2013 12:05 pm

Hi Alex,

I believe nothing wrong is going on with the GEMACO syntax, the problem must be elsewhere.
I have noticed that for from(1) and from(9) you can have three transitions (respectively to 1, 2 and 5, and to 7, 9, and 10) whereas from(15) and from (16) you can have just two transitions (to 9 and 10).
I don't know if there may be some relationship with the unexpected result. To me it is difficult to think something also because you have given just a few details, for example on the GEPAT structure (which parameter type, and step it refers?) the number of states, events. Or perhaps the answer is out of my possibilities and I wish you someone else may give you an answer (by the way I am curious to know what is going wrong).

Good luck!
simone77
 
Posts: 197
Joined: Mon Aug 10, 2009 2:52 pm

Re: Surprising number of estimates

Postby AlexAv » Mon Mar 11, 2013 1:36 pm

Hi!

Indeed yes it's very strange. This matrix is a State-transition matrix. All the transition described in that matrix are possible. My results look like I have three different intercepts (beta 0) per parameters. And as you noticed, the different estimates are always derived from the three different transitions. I tested different sentences and I always got three estimates, whereas the number of (mathematical) parameter is actually one. This is the same for Sa, Sa+, Sb and Sb+

...

if anyone has already met something similar?

thanks again

Alex
AlexAv
 
Posts: 12
Joined: Tue Sep 18, 2012 8:19 am
Location: Sweden and France

Re: Surprising number of estimates

Postby CHOQUET » Tue Mar 12, 2013 1:10 pm

Hello Alex,

This problem is linked to the multinomial logit.

When several estimates are involved in each rows then
constraints between rows can be destroyed by
the multinomial logit (if all the constraints are
not strictly identical between two rows).

This is not the case with the identity link.

Sincerely,

Rémi
CHOQUET
 
Posts: 211
Joined: Thu Nov 24, 2005 4:58 am
Location: CEFE, Montpellier, FRANCE.

Re: Surprising number of estimates

Postby AlexAv » Tue Mar 12, 2013 2:17 pm

Hi Rémi

Great! many thanks for your reply. So I understand that the deviance of a given model is different depending on the link... :?

Thanks again and good evening

Alex
AlexAv
 
Posts: 12
Joined: Tue Sep 18, 2012 8:19 am
Location: Sweden and France

Re: Surprising number of estimates

Postby AlexAv » Wed Mar 13, 2013 12:02 pm

Hi again..

I have another question regarding the use of the identity link function: the algorithm used to achieve convergence is said to be slower in the E-Surge manual: I ve tried to run the model with the identity link function, but I could not see the deviance scrolling one hour after having push the run button (the model has a maximum transition matrice size of 41*41 and the data set is more than 5000 capture histories of 152 occasions).
Do I need to specify the algorithm to be used before? as I understand it is not either the quasi newton or the EM or the hybrid algorithm. The last one is called PEM but I could not find any explanation in the Manual about the PEM algorithm... Should I use this one? or is the algorithm changing by default when choosing the identity link function?

Thanks in advance,

Sincerly

Alex
AlexAv
 
Posts: 12
Joined: Tue Sep 18, 2012 8:19 am
Location: Sweden and France


Return to analysis help

Who is online

Users browsing this forum: No registered users and 11 guests