Multi-season "fixing" colonization

[wsutton]

This is a bit more of a process question than anything, but I am currently assisting a colleague with a large multi-species salamander dataset from the late 50's to current across a total of 10 or so transects that run traverse about 2400' ft. in elevation. Long story short, the sampling employed in this study has forced us to collapse the 50's and 90's data into a "historical data" set and then compare this data to the current time period. Luckily the current time period has repeat surveys so we can use these detections to model occupancy during the other primary periods (with some assumptions of course).

I am using a multi-season method and due to the lack of repeat surveys in the historical data I don't feel confident trying to model colonization (plus it usually doesn't give reliabe estimates). I tried the alternate parameterization with colonization removed and most models aren't converging. I have read through the listserv and found that many folks have also seen this same problem. I also tried "fixing" colonization to zero by entering zero in the "fix parameter" menu and in the design matrix. The resulting models make pretty good sense, but I am concerned that forcing the analysis to essentially remove colonization may result in erroneous models. Another alternative I thought of was running two single season models and comparing the resulting curves based on covariates. Anyone have any input on this?

Thanks,

WBS

[jhines]

Without some way to assess detection probabilities in the 'historical' data-sets, any estimates of changes in occupancy between then and the current time will be meaningless. The changes could be due to change in occupancy, or change in detection rates between the two time-periods.

So, you could fix detection probabilities to 1.0, which is probably the worst option. Unless detection probabilities are 1.0 in the current data, you will most likely get a result that occupancy is declining.

Or, you could assume detection rates for the historical data are the same as detection rates for the current data. Given the time-span (50's, 90's and today), I don't think I'd be comfortable with that assumption.

In some cases, people have been able to artificially partition historical data in order to create replicate 'surveys'. For example, if bird-count data were collected for 10 minutes at a site and you know the time each detection occurred, you could use counts in the 1st 5 minutes for survey 1, and counts in the 2nd 5 minutes for survey 2. Or, you might be able to break the transects into pieces and treat them as 'spatial' replicates. These options make assumptions that you'll need to be comfortable with as well.

Jim

[wsutton]

Jim,

In the analysis I have performed so far I have done exactly as you stated in your next to last paragraph. I have calculated detection probabilities with the current data and used these to correct occupancy in the historical data. My assumption was that if no repeat surveys are provided for earlier primary sampling periods, Presence will automatically take the dp's derived from the current data and use these to determine occupancy in the historical data. Please correct me if I am wrong.

Some species have very high detection probabilities (>0.90), which I feel more confident about. Other species are a bit lower, close to around 0.65-0.70. If we were to go this route (may be our only option), any advice for a potential cut-off for using current detection probabilities?

Thanks again

WBS

[jhines]

If you're modelling detection to be the same across the sets of years, then you're right that occupancy for the historical years is based on a detection probability estimated from the current years. In whatever the case may be, I think anything you say about change in occupancy will depend on how strong a case you can make for detection being the same in the historical data as the current data. You could use program GENPRES to try scenarios which would give you a feel for how big a difference in change in occupancy is possible to detect for your different levels of detection probability.

Jim

[wsutton]

Thanks Jim, I really appreciate it. I will give GenPres a try and see how the simulations turn out.

Thanks again,

WBS