Multi-Season Analysis - zero detection in first season

questions concerning analysis/theory using program PRESENCE

Multi-Season Analysis - zero detection in first season

Postby leyla » Thu Oct 13, 2016 4:25 pm

I am new to PRESENCE and occupancy modeling and am still trying to learn this method. I have a question regarding multi-season analysis.

I have 40 sites that has been surveyed in 2 seasons, 10 and 14 survey occasions in season 1 and 2, respectively. I've used simple multi-season analysis (model parameterization: Seasonal occupancy and colonization, detection). However, I have a case that no animal was detected in any of the survey occasions during the first season, and therefor I get unreliable estimates as large as 1 or very close to 1, and of course detection probabilities of less than the critical value (<0.15).
Could anyone tell me please which parameterization of multi-season analysis should I use for cases that there are 0 detections in one season (which leads to unreliable detection probabilities)?
Or is applying mutli-season analysis appropriate for these cases?

Your opinions are much appreciated in advance.

Thank you.
leyla
 
Posts: 7
Joined: Thu Oct 06, 2016 2:04 pm

Re: Multi-Season Analysis - zero detection in first season

Postby jhines » Thu Oct 13, 2016 4:42 pm

If you have no detections for one season, then it is impossible to differentiate the case where occupancy is zero vs the case where some sites are occupied and all detection probs are zero. So, you have 2 choices: 1) run the single-season model on the season which has detection data, and 2) assume detection probs are the same across both seasons and run a multi-season model. If the multi-season model can get the detection probs from the season with data and apply those detection probs to the season without data, then it might be able to estimate occupancy for both seasons. I suspect that with only 40 sites and low detection probs, you'll have to go with option 1.
jhines
 
Posts: 599
Joined: Fri May 16, 2003 9:24 am
Location: Laurel, MD, USA

Re: Multi-Season Analysis - zero detection in first season

Postby leyla » Sun Oct 16, 2016 8:31 pm

Thank you very much for your response.
For option 2, does it mean that p needs to be treated as constant, i.e. applying neither sampling covariates nor seasonal effect? or I should remove covariates only or seasonal effect only?
Thanks again.
leyla
 
Posts: 7
Joined: Thu Oct 06, 2016 2:04 pm

Re: Multi-Season Analysis - zero detection in first season

Postby jhines » Mon Oct 17, 2016 9:08 am

Detection doesn't have to be constant, but since there is no data for season 1, you can't estimate detection for season 1. So, you can estimate detection for season 2 with any covariates you would be able to use if it were a single-season analysis, then apply them to season 1. For example, if detection is a function of time-of-day for season 2, you can run a multi-season model with detection as a function of time-of-day for both seasons, but equal between season 1 and season 2 (p(timeofday)). You won't be able to run a model with detection as seasonal, plus time-of-day.
jhines
 
Posts: 599
Joined: Fri May 16, 2003 9:24 am
Location: Laurel, MD, USA

Re: Multi-Season Analysis - zero detection in first season

Postby leyla » Mon Oct 17, 2016 6:45 pm

Thanks for your explanation again. This is very interesting and sounds reasonable.
So, if I want to stick with option 2 and do a multi-season analysis, I firstly have to run a single-season using season2 data, and estimate p with covariates of interest (which in my study there are 6 of them) and then run a mutli-season. correct?
I'm sorry to keep asking questions, but If I use this approach, I am not quite sure how I can apply the detection estimates obtained from the single-season to season1? Would you mind please explaining in more details?
My other question is about when you say "if detection is a function of time-of-day for season 2, you can run a multi-season model with detection as a function of time-of-day for both seasons, but equal between season 1 and season 2 (p(timeofday))". How can I make them equal between seasons? Is it by deselecting Seasonal Effect for p? If yes, then what is the point for running a single-season prior to mutli-season?

Hope my questions are not too long. :)
leyla
 
Posts: 7
Joined: Thu Oct 06, 2016 2:04 pm

Re: Multi-Season Analysis - zero detection in first season

Postby leyla » Tue Oct 18, 2016 3:29 pm

I tried a multi-season analysis and did not include "seasonal effect" for detection probabilities. In this way, I finally got occupancy estimates that are not extreme (0 or 1) and looks reasonable by themselves only (5% for Season1 and 65% for Season2). However overall results are not satisfactory. SE is zero. Lower and upper confidence intervals are identical with occupancy estimates. and p barely exceeds 15% in Season1. All means that the estimates are not reliable.
Any thoughts?
leyla
 
Posts: 7
Joined: Thu Oct 06, 2016 2:04 pm

Re: Multi-Season Analysis - zero detection in first season

Postby jhines » Fri Oct 28, 2016 1:45 pm

If you have estimates of 0 or 1, then a SE of 0 (or NA) is expected. If you have estimates of 5% and 65% with SE of 0, then something is wrong; probably insufficient data for the model. If you'd like to send me a copy of the data (most recent backup zipfile in your project folder), off-list, I'll try to see what's going on.

Jim (jhines at usgs dot gov)
jhines
 
Posts: 599
Joined: Fri May 16, 2003 9:24 am
Location: Laurel, MD, USA

Re: Multi-Season Analysis - zero detection in first season

Postby leyla » Thu Dec 01, 2016 11:57 am

Jim,
Thank you very much for your explanations. It seems that "insufficient data" is the problem I am facing. So, whats the solution for such cases? Will I be able to use PRESENCE to run occupancy models for it? or there will be an alternative method?
p.s. Sorry for my very late respond. I was away for a while.
leyla
 
Posts: 7
Joined: Thu Oct 06, 2016 2:04 pm

Re: Multi-Season Analysis - zero detection in first season

Postby jhines » Thu Dec 01, 2016 2:40 pm

After some thought, I think you should only analyze season 2 as a single-season model. If you have zero detections in season 1 and set detection probabilities to whatever value you get for season 2, then the best estimate of occupancy in season 1 will have to be zero. The only ways to get zero detections is either occupancy = 0, or detection = 0 (or both). If occupancy is zero, then extinction is not defined since there are no occupied sites to go extinct. So, the extinction parameter cannot be estimated. The only way to get rid of the problems you're having with estimated variances would be to fix parameters which cannot be estimated, which would be equivalent to a single-season model anyway.
jhines
 
Posts: 599
Joined: Fri May 16, 2003 9:24 am
Location: Laurel, MD, USA

Re: Multi-Season Analysis - zero detection in first season

Postby leyla » Thu Dec 01, 2016 3:19 pm

Jim,
Thanks for your prompt response. I really really appreciate it.
I actually followed your advise and used single season model for those less frequent species. And the results were promising. However I still have an issue using a covariate in the models.
I have a site covariate (Lets say cov.A) which is continuous and standardized. When running occupancy models without cov.A, the occupancy probability estimates are meaningful and reasonable. However, for those problematic species (the one that I ran in single season due to no detection in season 1), when adding cov.A to the model, I will gain unique values for occupancy estimate, LCI and UCI, and obviously 0 for SE. It also causes the -2LL to increase and that does not make sense.
Do you think it could be related to insufficient data and PRESENCE cannot handle it? Is there any alternative method to estimate occupancy in such cases? Or is there any reference that I can read to understand this problem better?
Thanks A LOT.
leyla
 
Posts: 7
Joined: Thu Oct 06, 2016 2:04 pm

Next

Return to analysis help

Who is online

Users browsing this forum: No registered users and 13 guests