E-Surge vs M-Surge

[Guillaume Souchay]

Hi,

I did a selection model with M-Surge. Then, I wanted to imput some individual covariates, so I used E-Surge.
I used the same model from M-Surge to E-Surge in order to have a basis for selection model under E-Surge.
Now I have compared the estimation of survival and there is some differences:
F TO T A G E E-Surge e.s M-Surge e.s
S 1 1 1 1 1 1 0.876 0.055 0.850 0.051
S 1 1 2 2 1 1 1.000 0.000 0.606 0.092
S 1 1 3 3 1 1 0.917 0.030 0.933 0.031
S 1 1 7 7 1 1 1.000 0.000 1.000 0.000
S 2 2 1 1 1 1 1.000 0.000 1.000 0.000
S 2 2 2 2 1 1 0.893 0.090 0.472 0.075
S 2 2 3 3 1 1 0.968 0.035 0.984 0.029
S 2 2 7 7 1 1 0.988 0.014 0.993 0.015
S 3 3 2 2 1 1 0.851 0.077 0.776 0.071
S 3 3 3 3 1 1 0.986 0.022 0.977 0.022
S 3 3 7 7 1 1 0.984 0.013 0.996 0.012
S 4 4 2 2 1 1 0.527 0.109 0.860 0.108
S 4 4 3 3 1 1 0.833 0.060 0.982 0.038
S 4 4 7 7 1 1 1.000 0.000 1.000 0.000


Where is the source of such a difference in some estimation?

Thanks,

Guillaume

[CHOQUET]

Hello,

Many causes could be at the origin of such differences.

1) local minima

2) bad specification of one of the two models, perhaps not really equivalent

With so few elements, it is not possible to say more.
In particular, the information about the deviance is crucial.
At least, deviances should be the same (by removing the initial parameter in E-SURGE).

Rémi

[Guillaume Souchay]

Hi,

For the model with M-Surge, the deviance was : 13 880
With E-SUrge, the deviance has decreased: dev(PI)+dev(PHI,B) = 13443.806

This could be the result of a bad parameterization of the model with E-Surge?
With M-Surge, I could fixe 2 parameters for survival but I don't succeed to do the same with E-Surge, could it be the reason?

[CHOQUET]

In your case, the values of deviance couldn't be compared because
the inital parameters are not present in the likelihood in M-SURGE.
In E-SURGE, the initial parameters are modelised by default.

To compare the deviances, you must select in E-SURGE the option
" Models > If any factorisation > Transition and Encounter ".
By doing this, the initial parameters are not further modelised and
the deviance must be the same.

Rémi

[Guillaume Souchay]

Hi, I select the option "Transition and Encounter"
the deviance with the model for E-Surge is now:dev(PHI,B) = 13438.841
There is still a difference with the model with M-Surge.

Here is what I've done for these two models:

With E-Surge,
* Matrix
Initial state %%%%%%
* - - -
Transition %%%%%%
survival
s - - - *
- s - - *
- - s - *
- - - s *
- - - - *
transition
y * y y -
y y * y -
y y y * -
* y y y -
- - - - *
Event %%%%%%
* b - - -
* - b - -
* - - b -
* - - - b
* - - - -

* the model formula used is:
Model formula (or file) :
For Initial State: IS - Step 1 - (0): to
For Transition: S - Step 1 - (14): f(1).a(1)+f.a(2,3:6,7)+others
For Transition: T - Step 2 - (49): f(1).to.ete.a(1)+f.to.a(2:7)+others
For Event: E - Step 1 - (589): firste+nexte.to.t
I could fix to 1 "others" for survival, to 0 "others" for transition T and I fixed to 1 the parameter 121111 to set that individuals have to be seen at state 1 at first occasion.

With M-Surge
*the only matrix I changed was the transition pattern matrix; I used :
1011
1101
1110
0111 insteed of the default matrix.

*Model formula (or file) :
ForSurvival (14):f(1).a(1)+f.a(2,3:6,7)+others
ForTransition (49):f(1).to.ete.a(1)+f.to.a(2:7)+others
ForCapture (96):to.t
I fixed the parameter "ohters" to 1 for survival and to 0 for transition.

Could the difference exist because of the transition T matrix could be not well defined?

[CHOQUET]

The sentence for parameters capture in E-SURGE is incorrect
because you have at least 4 age classes for the capture.
Taking this into account(4 age classes for capture lead to 3 age classes for nexte), try the sentence:

firste+nexte(1:3).to.t

Rémi

[Guillaume Souchay]

Hello,
I don't understand why age classes are taken into count in the capture sentence if I don't put some age effect "a".
How do you calculated the number of age classes?

[Guillaume Souchay]

For the models, I fixed A=7.
How do you calculate the number of 4 age classes in this case?

Thanks,

Guillaume

[Guillaume Souchay]

I've just tried with the sentence: firste+nexte(1:3).to.t but now, the deviance is very much higher: 18000 instead of 13800.
In models I'm using, as A is set to 7, shall I enter firste+nexte.a(1:7).to.t?

[Guillaume Souchay]

I tried to understand by myself, and I saw that when you enter firste+nexte.to.t in Gemaco, in the IVFV windows, for the event parameter, the a is taken into account.
So, now I'm trying a model with firste+nexte(1:7).to.t and the number of parameters for event is 97 which is more similar to model with M-Surge than other paramterization. I will tell you if it was the good solution.

I apologize for the number of reply.

Thanks,

Guillaume