Beta and SE errors in a model

questions concerning analysis/theory using program PRESENCE

Beta and SE errors in a model

Postby paula_cruz » Fri Aug 09, 2019 11:02 am

I'm using single-species models in PRESENCE. I'm using 5 covariates to model psi and one variable to model p. I ran the whole combination of variables (64 models), and it seems to be everything OK except for one model, which is suspiciously the best-ranked model (it doesn't seem to make sense, comparing with all other models within deltaAIC<2). This model has betas with very big values and SE with errors. Something like this has never happened to me (I've been using these models since 2012).
Does anyone know what could be happening? I can share the project.
Thank you very much.
Paula Cruz
Posts: 1
Joined: Wed Jan 27, 2016 8:08 am

Re: Beta and SE errors in a model

Postby jhines » Fri Aug 09, 2019 11:16 am

Large beta's (with large SE's) aren't necessarily a problem as it might just mean that a "real" parameter (eg., psi or p) has approached an upper limit (near 1.0) for sites associated with that covariate. For example, if all sites at elevation=0 are occupied (with detections) and occupancy decreases as elevation increases, the intercept will be a large number and the elev covariate will be negative. (The SE is large because there isn't a single beta value which yields 1.0 for the real parameter... any value above around 8 or so will give a value near 1.0 from the logit transformation.) There is nothing wrong with that.

On the other hand, sparse data and/or over-parameterized models can cause unreasonable beta estimates, so it's a good idea to know the cause of those estimates. I'd be happy to look them over if you'd like to send me the project.

Jim (
Posts: 460
Joined: Fri May 16, 2003 9:24 am
Location: Laurel, MD, USA

Re: Beta and SE errors in a model

Postby jhines » Sun Aug 11, 2019 12:27 am

Hi Paula,

I've looked over your output files and I think the top model is just over-parameterized. The model has 4 continuous covariates for occupancy with 91 sites. Depending on the variation in values of the covariates and the observed detections, 91 might be enough sites, but it doesn't appear to be in this case. I think what is happening is that for some combinations of covariates, there is not a gradual increase/decrease in occupancy as the covariates change. Instead, once one or more of the covariates reach critical values, occupancy becomes 1.0. Perhaps, this did not happen with your other analyses.

I suggest removing that model from the model-set as it doesn't really provide much useful information.


Posts: 460
Joined: Fri May 16, 2003 9:24 am
Location: Laurel, MD, USA

Re: Beta and SE errors in a model

Postby jlaake » Sun Aug 11, 2019 8:48 am

I don't normally post to this forum but this one caught my eye. So here is my 2 cents. It sounds like you have complete separation based on your variables. If you think of splitting the continuous covariates into low and high values, that means you have 2^4=16 combinations with 91 data points. Quite easy to achieve complete separation in which values are either all unoccupied or occupied in a combination. This isn't exactly how your model may be setup but may help you understand what Jim means by over-paramerterized. To help you understand this further you may want to look at your data by creating those 16 combinations of levels ( full interaction model) and looking at your data as if occupancy is known and see if that helps your understanding. The values of the beta should help guide you to where you have complete separation.
Posts: 1097
Joined: Fri May 12, 2006 12:50 pm
Location: Escondido, CA

Return to analysis help

Who is online

Users browsing this forum: No registered users and 1 guest