Occupancy and avian point counts: intensive sampling

[Kiel]

For estimating occupancy I understand that generally speaking more visits = 'better'. Of course logistics play a major role in how many and how often sites can be surveyed. I am wondering if anyone has attempted, or at least considered, employing a methodology similar to Alldredge et al.'s time-of-detection method in the context using point counts for occupancy estimation?

Let's say we are conducting a point count where we will divide a n-minute survey into 1-minute intervals, denoting the presence/absence of our species of interest during each interval. That is, if we've conducted a 10-minute point count, and we would like to limit the number of times we visit a site; we visit a site once but the resulting encounter history is an 10-character array of 1' and 0's. Does anyone have insights on how this might work, or not, for occupancy estimation?

Kiel

[ganghis]

Hi Kiel,

With regard to bird point counts, one of the main issues has to due with availability bias. That is, if birds are present in a site but do not sing during your 10 minute interval, they are 'unavailable' for detection. My guess is that singing behavior is largely Markovian so I wouldn't be surprised to get a '1111111111' history and a '0000000000' history at the same site even if it is occupied both times. In this case there would be some serious confounding between occupancy and availability.

-Paul

[darryl]

In my mind Paul is referring to 2 slightly different issues here (Paul, feel free to correct me if I'm thinking along different lines to you). If the singing behaviour is Markovian that means the probability of singing at time t depends on whether singing was occurring at time t-1 (or perhaps earlier). This is a form of auto-correlation, or a lack of independence between time periods. One way of accounting for that would be to use the equivalent of a Mb closed-population mark-recapture model, where you allow the probability of re-detection to be different to the probability of first detection. Now if there's a statistical god up there (or perhaps most would think such a being resides DOWN there), then at least the time of first detection would be somewhat random, so rather than just having all 1's or all 0's, you might get histories like 0001111111 or 0000000011, from which you could get estimates of psi's and p's, the key would be how you interpret them.

DarrylThe species would be unavailable to your sampling if it has zero probability of being detected at least once during your repeated surveys. So if the species (not individual) might remain silent for long periods of time, then it would be difficult to differentiate unavailable from absence. You can also get this is the species is temporarily absent from the point count location (eg it is present at the location over some longer time period, but just happen to be elsewhere when you do your point count). Really you have to consider how likely some of this stuff is going to be for the species you're working with, then design your study to get around them if they're going to be a problem.

Darryl

[ganghis]

In some ways what one needs is a model similar to the one Pledger proposed at EURING for stopover duration. When there are 'bouts' of singing you might end up with alot of observations like 00011100000. This differs from a behavioral model in that p_8 differs from from p_5 (for example). In this case a non-Markovian model is needed, which might employ a parametric model for the time an average singing bout lasts, as well as the time between singing bouts. This would be relatively data hungry and/or require auxiliary information.

As Darryl says, the best approach might be to take advantage of the biology of the study species - in particular space them far enough apart as to get conditionally independent observations (conditional on the site & it's characteristics). One ideally wants an observation to be an iid Bernoulli draw with the same detection probability at each draw.

Paul

[darryl]

In some ways what one needs is a model similar to the one Pledger proposed at EURING for stopover duration. When there are 'bouts' of singing you might end up with alot of observations like 00011100000. This differs from a behavioral model in that p_8 differs from from p_5 (for example). In this case a non-Markovian model is needed, which might employ a parametric model for the time an average singing bout lasts, as well as the time between singing bouts. This would be relatively data hungry and/or require auxiliary information.

Other option would be a model where p is different if the species detected was detected in the previous occasion. So rather than a permanent behavioural effect, it's a temporary one.

[Doherty]

From a field logistics point of view....

if a "behavior" response is evident then most of the information on p is going to come from that initial detection -right?

In that case someone might find using a "removal" method - i.e. don't bother recording individuals after the first detection - easier to do in the field and meanwhile not losing too much in the estimation of p?

[darryl]

From a field logistics point of view....

if a "behavior" response is evident then most of the information on p is going to come from that initial detection -right?

In that case someone might find using a "removal" method - i.e. don't bother recording individuals after the first detection - easier to do in the field and meanwhile not losing too much in the estimation of p?

Good point Dr Doherty, most of the info for estimating occupancy comes from 1st detection, although the removal approach may not save you much time in the field when you're working with multiple species. However, there's no reason why you can't just ignore the remainder of your data after first detection when it come to analysis. On the flip side though, with data form a (pseudo) removal design you can loose some flexibility for reliably modelling the detection process with covariates, which may be important to account for heterogeneity between point count locations. As usual, it's a horses for courses approach.