Multi state recapture model

[kmk]

I have a dataset of daily telemetry locations that is probably most appropriate for a multi state live-dead model. We have regular daily locations on one set of birds with less than 2% of approximately 6,000 of these being 'zeros' in the encounter history for not found. However, I cannot get MARK to run even the simplest model on these data = probably because of all these zeros (?) and my computer has about 6 Gigs of RAM so it's not running very slow. I am thinking of transferring over to a recaptures only multi state model in the hopes that it will have fewer parameters to run through and thus, might actually run. My question is 1) is this appropriate in any way? and 2) if so, does it make sense to include the last location for birds that died in the encounter history so that the apparent 'deaths' appear in the appropriate stratum?

Thanks.

[bacollier]

kmk,
For the simplest model, I am assuming you are using a dot (.) model, where survival is constant over the entire study period? Does MARK converge, crash your computer, give a weird error, blue screen of death, etc., its hard for anyone to diagnose your issue without any of that information.

What do you mean 'with <2% being zeros'? So you mean out of 6000 individuals, <2% have a 00 pair somewhere in the capture history? Having censored (right, left, interval) in a capture history should not be a problem and would not influence MARK during the computations unless you have something coded wrong (e.g., a 01 instead of a 10 in a known fate encounter history).

If you have telemetry data, then a multi-state model will not reduce the number of parameters, rather it should increase them as you will then have to estimate transitions, recapture probabilities, etc.. But, why you would use a multi-state approach on telemetry data where p is supposedly =1 and transitions should be known (e.g., if P=1, and you are tracking the animals, then the transition probability is equal to the proportion that move from A to B, in the simplest sense) is not easy to understand. With telemetry data, the number of parameters is maxed out (typically, although their are exceptions with covariates I suppose) at the number of encounter occasions by group, so I don't think you would have less.

If you want help we will need some more details on your issues,

Bret

[kmk]

Bret,
Thanks for your reply. Yes - for the dot model MARK gives me a "DYNAMIC MEMORY LOW" error when I run it.

By saying less than 2% were zeros I meant that less than 2% of these 6000 locations were coded as "no data".

Our ultimate goal in using the multistate live-dead models was to test for differences in transition and survival within four different habitat strata (prairie, restoration, agriculture, and fescue) between 2010 resident, 2011 resident, and 2011 translocated birds. We would like to use daily time steps as we think that larger time steps would result in a significant loss of transitions between habitats. However, some of these birds were only located on a weekly basis and I am assuming that the multistrata live dead model requires a constant search effort across all sampled individuals? Maybe this is why the computer keeps crashing, because there are too many "nodata" cells?

Thanks again.

[bacollier]

Bret,
Thanks for your reply. Yes - for the dot model MARK gives me a "DYNAMIC MEMORY LOW" error when I run it.

This sounds like you are running our of RAM.

By saying less than 2% were zeros I meant that less than 2% of these 6000 locations were coded as "no data".

You have 6000 locations, how many individuals (rows) do you have, how many days are you trying to predict to? Do you have 20 telemetered over 300 days and you have a daily location for each? What is the breakdown on number tagged and number of days?

Our ultimate goal in using the multistate live-dead models was to test for differences in transition and survival within four different habitat strata (prairie, restoration, agriculture, and fescue) between 2010 resident, 2011 resident, and 2011 translocated birds.

Ok, here is where you kind of lose me as to your approach needing a multistate method. A typical multistate model does not rely on telemetry data (e.g., the detection probability with telemetry data is 1, so unless there is censoring e.g., a individual moves out of the area tracked or a transmitters shuts off) then you know where everyone is at for each encounter occasion that you tracked ( you said daily I think). If you know where everyone is and whether they are alive or dead, then the percentage of times they are found in A on day n and found in B on day n+1 is the transition probability, because there is no error in identifying who is where, when. If you know where they are, then the survival estimate for each strata is the number in that strata that survived over the total in that strata. Of course you could probably trick a multistate model into estimating stuff for you with some clever restrictions on the parameters, but that will require a lot more thought and study on your end.

We would like to use daily time steps as we think that larger time steps would result in a significant loss of transitions between habitats.

If you are using radiotelemetry and you want to estimate those transitions of individuals moving between habitats on a daily basis, then if you track every day the estimate is the percentage I described above and does not need estimates (in my opinion). With telemetry you don't have (or at least should not) have unobservable states, or p<1. You can look over the multistate chapter of the manual for more details on this.

However, some of these birds were only located on a weekly basis and I am assuming that the multistrata live dead model requires a constant search effort across all sampled individuals?

If you tracked weekly, then you cannot estimate daily survival, unless you assume that for those individuals tracked weekly that S was constant daily. You might be able to do it in the nest survival (ragged telemetry approach).

Maybe this is why the computer keeps crashing, because there are too many "nodata" cells?

They are not 'nodata' cells, they are sample occasions where you did not locate the individual, which typically does not occur in a telemetry study. I have datasets with >60% being '0' records, so that is likely not your problem.

Based on what you have said (assuming I have understood it all correctly), my opinion is that you focus on using the telemetry data (see the known fate chapter in the book) to estimate daily survival (assuming you tracked daily for most of the birds, if you tracked weekly you would probably have to adjust the time intervals in MARK to address this). Also, assuming you have daily locations from radio-telemetry data, you likely do not want to use a multistate approach (or at least I cannot see any logical benefit to it) as the percentage of locations in each state is equal to the transition probability when p=1 and S is known.

Bret