Candidate Set

questions concerning analysis/theory using program PRESENCE

Candidate Set

Postby phapeman » Tue Aug 15, 2017 5:21 pm


I have a dataset with 34 sites surveyed 3-5 times for two weeks. I have run a number of single season models with covariates. I am trying to narrow down the entire dataset to a set of candidate models. I haven't quite decided on the criteria for choosing the models to include in the candidate set but it will be something like models with a delta AIC value <2 or all the models whose weights sum to 0.95.

My question is what to do with models that could be included in the candidate set but have standard errors that are either excessively large or listed as -1.#IND00? In one case I have a model that clearly is the best model and the standard errors look fine but the untransformed value for psi leads me to a derived occupancy estimate that is almost zero despite the naive estimate of 0.33. Should I delete models from the dataset that have those types of issues or is there some other way I should treat them? Thank you!
Posts: 2
Joined: Wed Mar 19, 2014 5:03 pm

Re: Candidate Set

Postby darryl » Tue Aug 15, 2017 5:30 pm

You shouldn't be getting an occupancy estimate from the standard single season that is lower than the naive estimate. Or are you talking about a model that has covariates in it? It that case the individual estimate for some sites might be lower. Regardless, if you're getting errors you should be trying to resolve them before moving on because it could be that the models hasn't been fit properly for some reason (eg problem with the data or user error).

Note, your 'candidate model set' is typically the set of models that you decide to fit to the data in the first place (that represent the combinations of biological hypotheses of interest), not just the set of models you use for your final inferences.

Posts: 438
Joined: Thu Jun 12, 2003 3:04 pm
Location: Dunedin, New Zealand

Re: Candidate Set

Postby jhines » Tue Aug 15, 2017 6:04 pm

My guess is that you have models with more than a few covariates on phi and/or p. With only 34 sites and 3-5 surveys, I'd limit the covariates per model to 1 or 2. You can add more, but what usually happens is that you reduce the data for estimating each possible parameter. Sometimes all sites for a particular value of a covariate might be occupied (or unoccupied) which causes the occupancy beta parameter to be large ( or large-negative) with a high standard error, or an inestimable standard error ( -1#IND00). Other times, a group of sites with a particular combination of covariates might have no detections, which makes estimation of phi and p impossible for those sites. I recommend starting simple (no covariates, one covariate on phi, one covariate on p,...) and work your way through your list of plausible models.
Posts: 427
Joined: Fri May 16, 2003 9:24 am
Location: Laurel, MD, USA

Re: Candidate Set

Postby phapeman » Wed Aug 16, 2017 11:15 am

Thank you for the replies!

I have kept the models simple with only 1-2 covariates for each model. The case of the derived estimate of psi being much lower than the naive estimate is from a model with two covariates and happens to be the single best model. It sounds like models with specific combinations of covariate data can lead to issues as you said but I am not sure if you would know that until after the data was collected.

I would still like to hear your opinion of what to do with models that have issues with standard errors, or in my case with occupancy, when they are in the candidate set and you are trying to make final interpretations. Should those models with excessively large or inestimable standard errors be removed from the candidate set or is there some other way I should treat them to include them in the final interpretations? Appreciate the help!
Posts: 2
Joined: Wed Mar 19, 2014 5:03 pm

Return to analysis help

Who is online

Users browsing this forum: No registered users and 1 guest