Problems with Multi-state models

[ssaps33]

Hi,

I am fairly new to MARK and am running a multistate model. I have unequal time intervals (and I read the post on problems with this). However, I am confused on how to fix this - I wasn't sure from the posts on this problem.

I originally had 5 sampling occassions with 2 months between each occassion. I also have three groups. However, due to a cyclone I couldn't access a site on the 4th occassion, so to make things equal among all sites I took out the 4th sampling occassion. Therefore, I set my time intervals to 1 1 2 (with the 4th encounter missing). However, after reading the post on unequal time intervals the estimates of Psi, and possibly S are wrong? From my understanding, to fix this, instead of leaving out the sampling occassion I put a . ? So my encounter history, for example, would be NP0.N (N=negative for disease, P=positive)? Is this correct?

I have already run a few models (without accounting for unequal time intervals) and have come across another problem - parameter identifiability. I have tried using estimates from simpler models as starting values in more complex models, simulated annealing. I have recently found posts on Appendix F: cloning and will try that next...but I didn't want to go ahead if the unequal time intervals are messing up my estimates.

I appreciate it any advice on this problem! Thank you for your time.

Kind regards,
Sarah

[aduarte]

Hi Sarah,

I may be wrong, but I think you can set p=0 for that group on that specific occassion before running each model. In order to do this for models in which phi does not vary across time you would have to build your model in the design matrix so that you can still set p=0 for that specific occassion.

Adam

[ssaps33]

Hi Adam,

Thank you for your reply. I just wanted to confirm what my encounter histories would look like. If I make p=0 for that "missing" encounter, I must have to include it in the analysis? Therefore, would I put a 0 in the 4th sampling occassion or would I still take it out? ie. 5 vs. 4 sampling occassions?

Thanks again!

Kind regards,
Sarah

[bacollier]

Hi Adam,

Thank you for your reply. I just wanted to confirm what my encounter histories would look like. If I make p=0 for that "missing" encounter, I must have to include it in the analysis? Therefore, would I put a 0 in the 4th sampling occassion or would I still take it out? ie. 5 vs. 4 sampling occassions?

Thanks again!

Kind regards,
Sarah

Sarah,
You should use the '.' notation for the unsearched locations as detailed in the post your referenced, they are not 'possibly wrong', they are wrong as the underlying model structure has changed if you delete a sample occasion arbitrarily.

Do you have information on the other 2 sites for occasion 4? If so, I am not sure that removing that information is necessary for whatever you are doing. Why not keep and capitalize on the availability of that information, rather than not using it.

I don't think you need to be thinking about data cloning or simulated annealing until you get basic model structures running and set appropriate constraints on your parameters given your data. Have you been able to fit constrained models with the dot notation, that's the starting point.

Bret

[ssaps33]

Hi Bret,

At the moment I am playing around with my encounter histories (excluding the 4th occassion as mentioned above and excluding individuals from the site missing info and thus having equal time intervals (i.e. 5)).

What I actually have is individuals from 6 sites which I have divided into 3 site types. Only individuals from 1 site are missing data from the 4th encounter...all others have data. So, is it possible just to input dots for the data missing and include all other information for the other sites?

lets say ind a from site 1 NN0.P
ind a from site 2 N0PN0

I'm assuming I can include the dots for the individuals from that one site and then run models that are contrained? I am not 100% sure on how I can constrain certain individuals.

Thank you for your time!

[bacollier]

Hi Bret,

At the moment I am playing around with my encounter histories (excluding the 4th occassion as mentioned above and excluding individuals from the site missing info and thus having equal time intervals (i.e. 5)).

I think the point of the recent post on this topic was to not exclude the 4th occasion.

What I actually have is individuals from 6 sites which I have divided into 3 site types. Only individuals from 1 site are missing data from the 4th encounter...all others have data. So, is it possible just to input dots for the data missing and include all other information for the other sites?

lets say ind a from site 1 NN0.P
ind a from site 2 N0PN0

I'm assuming I can include the dots for the individuals from that one site and then run models that are contrained? I am not 100% sure on how I can constrain certain individuals.

Thank you for your time!

Yes, it is likely the best option to use the dot notation for the individuals from 1 site and include all the additional information.

[ssaps33]

Thank you! I will try this out and see what happens.

[aduarte]

Hi Bret,

Thanks for your input. I definitely do not want to lead someone down the wrong path. I had a couple questions though (just so I know this for the next time).

If you put a (.) in the encounter for individuals you previously captured at a site (and could not survey that site) are you not assuming that individual did not disperse to one of the other sites? Also, since you will not have new captures at that site for the 4th occasion does that mess with your estimates (say if I was estimating abundance)? If I constructed the models in the design matrix and had 4 groups (each site) and could not survey 1 site for a specific occasion could I set p=o for that group on that occasion or would that lead to biased estimates?

I would really appreciate your input on this.

Thanks
Adam

[bacollier]

Hi Bret,

If you put a (.) in the encounter for individuals you previously captured at a site (and could not survey that site) are you not assuming that individual did not disperse to one of the other sites? Also, since you will not have new captures at that site for the 4th occasion does that mess with your estimates (say if I was estimating abundance)? If I constructed the models in the design matrix and had 4 groups (each site) and could not survey 1 site for a specific occasion could I set p=o for that group on that occasion or would that lead to biased estimates?
Adam

1) Your not assuming anything regarding the dispersal process for the unsampled location if you use a dot.

2) Yes, it could mess with your estimates, but how will be model/situation dependent I think

3) setting p=0 (probably) would not matter, as you have no data for that occasion anyway, so the parameter would be inestimable. setting p=0 is tricky as you can impact your model set in weird ways (as I have learned the hard way) so I tend not to do it, but there are valid reasons to do it, I just have not run into many myself.