Strange State Assignment estimates and EM(20) not working

[simone77]

Hi,

I have done lot of trials and thought lot about that but I can't find a reasonable answer.

I have 16 seasons with data coming from resightings in winter. I am doing a simple analysis of survival and using E-SURGE also because I have sex uncertainty (only a part of marked individuals were genetically sexed when chicks).

I have done GOF in U-CARE on unistate encounter histories on both the whole data set and the data set with just males and females (two groups). These suggested to use as global model the Phi(g*a2-t/t) p(g*t) (TEST3.SM + TEST2.CT + TEST2.CL was no significant and the c-hat=1.2).

So I have three states (female, male, dead) and 4 events (not seen, seen as female, seen as male, seen as unknown) and two ages. There are no transitions among states.

First two years there are only unsexed individuals. I have checked the m-array in U-CARE of both males, females and all birds together and there are no evident strange things, the only thing is a few recaptures in t(4,10) because it was a dry year.

In GEPAT I have:
IS=*-
S=diagonal matrix (3X3)
C=diagonal matrix (3X3)
SA=
* - - -

- p - *

- - p *

In GEMACO:
IS=t(1 2 3,4_16)
S=a(1).t(1,2) + a(1).f(1,2).t(3_15) + a(2).t(2) + a(2).f(1,2).t(3_15)
C=firste + nexte.t(2) + nexte.f.t(3_16)
SA=f.t

In the IVFV I have fixed first capture to one and first four SA to zero (first to years there are no sexed individuals).



Two results are concerning me:

1) Only Quasi-Newton works (this is not the first time it occurs to me), EM(20)+Quasi Newton doesn't, here below the error I get.



2) estimates of SA for males are all "1" with SE=0 (intuitively the p a resighted individual is assessed should be more or less the same so that it doesn't seem to be a boundary estimate trouble).

Any idea on what might be going on?

Thanks in advance for any help

Simone

[CHOQUET]

In reply to:

First two years there are only unsexed individuals. I have checked the m-array in U-CARE of both males, females and all birds together and there are no evident strange things, the only thing is a few recaptures in t(4,10) because it was a dry year.

In GEPAT I have:
IS=*-

Normally, IS=a *

S=diagonal matrix (3X3)

Normally, S is not really diagonal but death must be possible

C=diagonal matrix (3X3)

SA=
* - - -

- p - *

- - p *


In GEMACO:
IS=t(1 2 3,4_16)


should be changed to t(1 2 3,4_16).to

S=a(1).t(1,2) + a(1).f(1,2).t(3_15) + a(2).t(2) + a(2).f(1,2).t(3_15)
C=firste + nexte.t(2) + nexte.f.t(3_16)
SA=f.t

In the IVFV I have fixed first capture to one and first four SA to zero (first to years there are no sexed individuals).

In that case, the error has to be modelised and set equal to the first year where there is sex individuals.

Two results are concerning me:

1) Only Quasi-Newton works (this is not the first time it occurs to me), EM(20)+Quasi Newton doesn't, here below the error I get.

Which version of E-SURGE did you use ?

2) estimates of SA for males are all "1" with SE=0 (intuitively the p a resighted individual is assessed should be more or less the same so that it doesn't seem to be a boundary estimate trouble).

Any idea on what might be going on?

see my remarks about the modelisation above


Rémi

[simone77]

Hi Remi,

Thanks for the help.

The IS GEPAT was just a typing error (it was IS= p *).

The S GEPAT was defined by hitting the "diagonal matrix button" but it allow death, in fact it is:
p - *
- p *
- - *

Now I have done two trials:
1) I have just changed the GEMACO sentence for IS as you suggested to do
2) I have changed the GEMACO sentence for IS and changed the GEMACO of SA from "f.t" to "f.t(1 2 3,4_16)" because this is what I have understood when you said "In that case, the error has to be modelised and set equal to the first year where there is sex individuals" (but I am not sure I understood the right thing).

I have tried again but I have still the same problem, I also have noticed that sometimes are SA of females to be estimated with 1 and zero SE and sometimes are the males.

My version of E-SURGE is 1.8.5.

Simone

[simone77]

I want to add something to the previous answer.

This week I have been doing many trials to try to understand what is going on.
It seems to me that perhaps the more relevant one to track the problem, has been the elimination of first two years (when there were no sexed individuals). So now I have 14 sessions instead of 16, it has been at the expense of losing only a few encounter histories.

In GEMACO I have (the GEPAT structure it is still the same as before - see my previous answer):
IS: t
S: g(1).a(1).f.t + [g(1).a(2).f.t][g(2).f.t] (Now I have two groups, first with individuals seen first time as juveniles and second of individuals seen first time as adults, new GOF results suggest transient effect in the first and not in the second group - anyway I would not expect the modeling of S to affect SA estimates)
C: firste + nexte.f.t
SA: f.t

Now, as expected in an over-parameterized umbrella model the results are quite poor in terms of confidence intervals of many parameters, but there are no huge problems with estimability (except for Survival of both genders in 8th year when only a few individuals were captured). The SA estimates are still "1" for males up to 8th session (any relationship with S??) and after they are estimated (and not very big CIs).

Anyway, this morning, after a long night running with Multiple Random= 32, the estimates for males make sense and ALL the estimates for females are "1". This confirms what I told in the first post that it seemed that sometimes the non sensible estimates were the males and sometimes the females ones. It might be that the "good" estimates of males after 8th occasion I got in the previous running were due somehow to local minima.



EM(20)+Quasi-Newton is still not working with these data.
I am thinking that perhaps the problem might be related to my data set but can't understand how.
Here below it is the m-array of the three groups (female, males, unknown) to get you an idea of my data set.



At the moment I can't understand what is going on.
Thanks in advance for any help

Simone

[CHOQUET]

The detected transient effect was surely due to the juvenile survival.
The way to manage it, is to separate juvenile survival from adult survival as for example:

[a(1).g(1)]+[a(2:14).g(1)&a(1:14).g(2)]

for two classes of age. This model has two parameters, one for the juvenile survival, the other
for the adult survival. For further details, see one of the example given in

Choquet R. (2008). Automatic generation of multistate capture recapture models. The Canadian Journal of Statistics 36(1): 43-57.

Rémi

[simone77]

Hi Rémi,

Definitively the transient effect is related to juvenile local survival and it makes full biological sense.
Anyway, the problems I am experiencing are others:
1. EM(20)+Quasi-newton not working (see above first post)
2. SA (State Assignment) estimates for one gender are nonsensical: are all "1s".

In the prior posts there are details on my data set, the GEPAT and GEMACO structure I have used. I want to go on with these analyses in E-SURGE (because it allows to deal with unknown genders and to get better estimates) but I really can't understand why those nonsensical SA estimates.

Thanks,

Simone

The detected transient effect was surely due to the juvenile survival.
The way to manage it, is to separate juvenile survival from adult survival as for example:

[a(1).g(1)]+[a(2:14).g(1)&a(1:14).g(2)]

for two classes of age. This model has two parameters, one for the juvenile survival, the other
for the adult survival. For further details, see one of the example given in

Choquet R. (2008). Automatic generation of multistate capture recapture models. The Canadian Journal of Statistics 36(1): 43-57.

Rémi

[CHOQUET]

1) the EM(20)-QN help when there is a lot of boundaries paramter but didn't removed the
estimates from the boundaries.
2) our model is overparametrised, all types of parameters are time dependent. Is is the best model ?

[simone77]

1) the EM(20)-QN help when there is a lot of boundaries paramter but didn't removed the
estimates from the boundaries.
2) your model is overparametrised, all types of parameters are time dependent. Is is the best model ?

1) Really I didn't expect the EM(20)+QN would have made the trick and solve the problem of SA estimates, I just don't understand why it doesn't work with this data set and if this fact could give me some hint on some particular problem related to my data set or whatever...

2) No, it is not the best model, is the Umbrella Model. I have tried to fit much simpler models than that (up to time dependency only on capture and SA) but the problem of nonsensical males SA estimates is still there. Also, I have tried to set SA as just time dependent (no difference between genders) but it has a much worse AIC value (very much!). Really, I can't understand what is going on with this analysis.

[simone77]

Hi Rémi,

I am sorry I believe I didn't explain it very well.
In my data set just a part of the population is sexed (without uncertainty), the rest is not sexed.
Individuals are physically captured all as chicks and some of them are bled to determine their gender. All the other "captures" of the same individual are resightings.

Here it is a way to deal with this situation that should take into account the multievent approach and, in this case, sex uncertainty.
I build a dataset whose encounter histories are all '0's and '1's and 6 groups:
1= sure female first captured as juvenile
2= sure male first captured as juvenile
3= unknown first captured as juvenile
4= sure female first captured as adult
5= sure male first captured as adult
6= unknown first captured as adult

This grouping is made taking into account something I found in GOF (U-CARE) where a transient effect was evident for individuals captured first time as juveniles (something that makes biological sense in my opinion).

In E-SURGE I set three states, two events, and 2 ages. Also I create three shortcuts to code for gender:
fem=g(1 4)
mal=g(2 5)
unk=g(3 6)

In GEPAT:

Initial State
p*

Survival
p-*
-p*
--*

Capture
*p
*p
*-

In GEMACO the global model may be:

IS {fem + mal + unk}* S{a(1).f + a(2).f.t} C{firste + nexte.f.t}**

*IVFV of first two parameters (corresponding to "fem" and "mal" are set respectively to "1" and "0"), this is a crucial point because corresponds to the very moment where the model is dealing with sex uncertainty.
**IVFV of first parameter as usual set to 1 (p of an individual to be captured first time).

I think it should work this way, estimates I am getting seem to make sense, does this approach make sense?

Meanwhile, thanks to Roger Pradel who made me think about something very important:
"The multievent approach is not just in the possibility to have additional codes for uncertain observations but also in the reinterpretation of the 1 itself as a possible case of uncertainty. Here, a 1 is uncertain because it leaves open the possibility that the individual is a male or a female, two different states in the model. In other words, uncertainty does not require the presence of a 2, it can start with 1."

One last thing: EM(20)+Quasi Newton seems to work with this type of analysis.

[CHOQUET]

Hello,
your new approach seems to be much better. However, I think that you should rewrite the model for survival.

Let consider two shortcuts:

Juveniles = [g(1:3).a(1)]
Adults = [g(1:3).a(2)+g(4:6)]

A simple model with two "real classes of age" is

Juveniles+Adults

A model with the same effects used for your general model is

Juveniles.f+Adults.f.t

Sincerely

p.s: see example 2 of Choquet R. (2008). Automatic generation of multistate capture recapture models. The Canadian Journal of Statistics 36(1): 43-57.