www.phidot.org

I am working with 10 years of radio-monitoring data where fates of animals (i.e., live/dead) were determined monthly and entries of animals into the study are staggered. My objective is to estimate survival. Originally, my approach was to use the known-fate data type in MARK. After compiling the histories, I see numerous issues with the data that may preclude use of Pollock's staggered entry design. In particular:

1) Monitoring data was not collected during approximately 20% of the months during the study and sometimes not for several consecutive months
2) Individual animals often were monitored, lost their transmitters or were not monitored for numerous consecutive months, and then re-entered the dataset when captured and recollared with a new transmitter or monitoring for them resumed
3) Some individuals occasionally went missing for a month or 2 because monitoring personnel could not find them

Because the missing data is so extensive, the amount of censoring likely would be problematic. I considered using Burnham's joint live-dead model, because the recapture probabilities for months with missing data could be fixed to 0. However, from my understanding, that model would not appropriately handle the trailing pairs of zeros for many animals whose transmitters were known to have dropped off while they were alive and were never seen again (i.e., right-censoring In known-fate). I also considered using mult-state live-dead models in some fashion, but cannot determine off hand if that is possible.

Does anyone have any helpful suggestions of how to deal with these missing data/censoring issues or can anyone point me in the right direction to other similar applications?

Many thanks in advance!

These problems must come up in all long-term studies, so I expect there's a standard answer in the literature. My 2 cents: I would be surprised if (1) is really a problem given the ability to specify varying sampling intervals, and why not simply treat a broken encounter history as two individuals, the second starting from when transmission resumed?
Murray

Thanks for your response, Murray. I had not considered specifying different sampling intervals for the occasional months when all animals were not monitored. It is possible to break apart an encounter history if an animal goes "off air" for an extended period of time. I have seen where known-fate analyses for long-lived species (e.g., bears) "fill in" missing data with 10's on an individual basis if the animal was lost for only 1 or 2 months even though this theoretically introduces a sampling bias (sensu 16.7 of 'the book'). I certainly expected it would be problematic to do that when all animals have missing data for 1 or 2 months, which is quite frequent with my dataset. Below is a 3 year snippet from an animal demonstrating the extent of the issues with which I am dealing:

101010101000001000100000000000000010001010001010101010000010000010100010000000000000000000000000

Obviously, the underlined portions are months where the animal was not monitored for various reasons. The shorter 'missed' intervals are times when no animals were monitored (e.g., scheduled monitoring flight cancelled due to weather). The long interval in the middle was when the animal lost its transmitter (because the transmitter released by design) and was later recaptured and recollared. The ending interval is when the animal again lost its collar by design. Because the causes of missing data are so varied, my original concern was that standard staggered entry design may not work well when such extensive censoring must occur. I'll keeping digging through the literature...

Thanks again
Jared

Jared
I'm not hearing any argument against breaking encounter histories, and that seems to me the rigorous thing to do. Perhaps it's a little conservative (I suggest you test by simulation), but you seem to have enough data. It may change the apparent sample size for AICc, but from previous discussion on this forum that remains a murky and minor consideration.
Murray

Hi Jared,
This might be a little late but it still might help... I have ragged telemetry data similar to yours. I am using the nest survival model in MARK/RMark to get survival estimates. The Nest Survival model is also good for "Ragged Telemetry Data". Unless you have a particular aversion to using it, I would check it out in Chapter 17 of the MARK Book.