known fate analysis

[jc128809]

I have radio tracking data on vole mortalities for 4 different years. Unfortunately, each year's study went for a different time period (from 25 - 40 days). Any suggestions how to handle this? So far i have right censored voles from the end of their study to the end of the longest study - what kind of bias might this produce? Also, i keep getting survival estimates of 1 at period throughout the 40 days - is this due to small sample sizes of predated voles (~100 collared voles, ~40 deaths)?

Thanks

[abreton]

It sounds like you want to estimate survival probabilities between sampling events within a 25-40 day period; and that the number of sampling events within this (brief) period varies from one year to the next. If this is the case, then you're attempting to estimate survival over periods lasting only a few days - which explains you're observation that some estimates of survival are "1.0" - I suspect the SE of these estimates are large relative to most of the other estimates? If this is the case, then the "1.0" estimates are not reliable, they're bogus. What's happening (I suspect) is that the solution routine for the likelihood function that is your 'model' is struggling to locate estimates of survial that are near a boundary (0.0, 1.0), in this case near 1.0. One option is to look carefully at your data and determine if survival is in fact 1.0 for any of these short intervals and 'fix' these to 1.0 using the option of the "Setup Numerical Estimation Run" window in MARK (see button labeled "Fix Parameters"). Deciding which parameters need to be fixed will certainly test your understanding of the relationship between the PIM and design matrix (and the linear model that your specifying when you build your PIM or DM models). If you get lost here, see the "linear models" chapter in the MARK manual. Alternatively, you could consider dropping some of your sampling events thereby increasing the duration of your sampling intervals and increasing the chances that more voles will be lost and survival will move away from the problematic 1.0 boundary. I should add, that small sample sizes will compound the boundary problem - less data means less information which often leads (even when parameters are not near a boundary) convergence failure (i.e. some or all of hte paameters in the model cannot be estimated given the data).

I'm not sure how you're handling the intervals that occur between the end of one 'study' and the start of the next; and I assume that each study (25-40 day period) occurs within a summer. It also sounds like animals contribute to only one 'study', i.e., they are initially marked, subsequently monitored, and ultimately removed all in the same summer. If this is the case, then you're likely not considering the between summer intervals and that's okay. By right censoring, I assume you're parameterizing your PIM structure so that animals from study i survive and are detected with probabilities 0.0 in all subsequent 'studies' (summers)? If you've gone this 'right censoring' route, then you'll also have to remove encounters (1's in the EHistories) associated with animals from summers in which they were not marked. If e.g. animal 54 was marked and followed in summer i and also detected in summer i+1, replace the detection "1" in the i+1 summer to "0" (not seen). Otherwise, estimates of survival will be increasingly and negatively biased from the start to end of each summer (study) - Paul Doherty (Colorado State U.) modeled this phenomenon but those results have not been published.

Finally, how to deal with an unequal number of sampling events among summers. I faced the same problem with a slimy sculpin dataset taken from five streams in New Brunswick, Canada. Following this post, I'll send you a pdf of the manuscript - and anyone else that might be interested, send me a request at abreton@sfos.uaf.edu. In brief (see page 282 in the ms, paragraph starting "Streams that remained..."), we reduced the number of sampling events to 17; these included only those sampling events where all streams were visited on 'about' the same day. We then took the median interval length between sampling events and set these as 'Time Intervals' when data were imported into MARK.

[abreton]

I wasn't able to send you a 'private message' - send me your email address (abreton@sfos.uaf.edu) and I'll send the ms.

[egc]

I wasn't able to send you a 'private message' - send me your email address (abreton@sfos.uaf.edu) and I'll send the ms.

Yes - the ability to PM (private message = send email to...) from within the phidot.org forum is turned off - to prevent spammers (and the like) from using PM to email everyone.

However, I tweaked the code to allow you to reply directly to the email that gets generated whenever someone posts. You only see this email *if* you've set the option on your account to send you an email for every posting to the forum. The default is to send you an email only if some posts a reply to something you've posted.

To set this option, login to your account, open up your profile, and look for the appropriate 'flag' to flip.