Model Average, two-species

questions concerning analysis/theory using program PRESENCE

Model Average, two-species

Postby albarros » Thu Jun 01, 2017 8:38 am

Hello! I'm using two-species models to study species interaction. I have 73 sampling points, that have been surveyed 3 times. One of the species is very abundant but the other is declining. I have built 16 models testing different scenarios, with 2 detection covariates and two occupancy covariates. 

I've been having two issues. Firstly, two of the models don't converge (message: Numerical convergence may not have been reached. Parameter estimates converged to approximately 5.44 significant digits.). And a few others, although they don't show this message, the beta coefficient for the psiBa parameter is huge. Does anyone know why this happens and what I should do?

Secondly, in models using covariates, PRESENCE output shows occupancy and detection estimates for each site. Is there any way of obtaining a single, average estimate of each parameter, for each model? Either using PRESENCE or by-hand?

Also about PRESENCE, I have a doubt regarding the output file. When using two-species models, you have to enter presence/absence data for the two species in the same columns. From what I read, the program assumes the first 73 data for species A and the other 73 for species B. But in the output I have estimates for 146 sites. Is this normal or have I done something wrong when entering the data?

Thank you very much for any comment!
albarros
 
Posts: 6
Joined: Thu Jun 01, 2017 7:04 am

Re: Model Average, two-species

Postby jhines » Thu Jun 01, 2017 9:59 am

The warning message about convergence indicates that the likelihood value was still changing (by a small amount) at the end of the optimization process. In my experience, whenever the number of significant digits is >= 2, the parameter estimates are accurate to 3 or 4 decimal places. So, I wouldn't worry about the message in this case. You can sometimes avoid that message by re-doing the model with different starting values (ie., start with values close to where you expect them to end up).

While the real parameters can vary between zero and one, the beta parameters can vary between minus and plus infinity. The logit link function translates the beta parameters to real parameters. Large values for the beta parameters correspond to a real parameter near 1.0 (if it's an intercept), or a very strong effect of a covariate. If you don't have covariates on PsiBa and the beta estimate is huge, it just means the real parameter, PsiBa, is 1.0. You might see a very large standard error for those beta's as well. There is nothing wrong with that, as the real parameter will probably have a standard error of zero. We usually say that when a probability is 1.0, the standard error is undefined.

To get an average estimate for each site, you could just report the estimates from a model with no covariates on parameters. That would give you a weighted (by the inverse of the standard error) average estimate for each parameter. If you are interested in the average for each parameter for each model, I suggest cut/paste the site estimates into a spreadsheet, or use R and/or RPresence.

The output of estimates by site for the 2nd half of sites (74-146) should be duplicates of the 1st half of sites. You can ignore sites 74-146 in the output. I'll correct this in the next update to Presence.
jhines
 
Posts: 599
Joined: Fri May 16, 2003 9:24 am
Location: Laurel, MD, USA

Re: Model Average, two-species

Postby albarros » Fri Jun 02, 2017 7:06 am

Thank you so much for the quick response!!

In regard to the second half of estimates, in my case they're not duplicates of the 1st half. It's just the same value repeated from 74 to 146. I don't know if that's normal, or if I've done something wrong. Anyway, I understand I should disregard these values for the average estimates.

To obtain the average estimate for each parameter, then I can use a simple average function, or does it have to be weighted in some way?

I also wanted to clarify something. When I use covariates to model occupancy, does this allow me to separate habitat preferences from actual interaction (either competition or co-occurrence)?

Thank you again!
Ana
albarros
 
Posts: 6
Joined: Thu Jun 01, 2017 7:04 am

Re: Model Average, two-species

Postby jhines » Fri Jun 02, 2017 7:40 am

I think the 2nd set should be duplicates of the first, unless you didn't duplicate the covariates. Perhaps you could send me (jhines@usgs.gov) your latest backup-zipfile in the project folder and I could check.

You can use weighted or unweighted averages, as long as you specify which one you used in your publication. I think a simple unweighted average is most common.

With different intercept and effect beta's, you can separate habitat effects from species intaraction.

Jim
jhines
 
Posts: 599
Joined: Fri May 16, 2003 9:24 am
Location: Laurel, MD, USA

Re: Model Average, two-species

Postby albarros » Fri Jun 09, 2017 12:40 pm

Yes, you're right! I entered the data wrong. Thank you!

Also, I used the model average tool in Presence, but I noticed the output doesn't have the average estimates for the beta coefficients. How can I obtain these values? Is there a way using Presence or does it have to be done in a spreadsheet?

Sorry for all the questions, and thank you for the quick replies!
albarros
 
Posts: 6
Joined: Thu Jun 01, 2017 7:04 am

Re: Model Average, two-species

Postby jhines » Fri Jun 09, 2017 1:04 pm

PRESENCE doesn't compute model-averaged beta parameters. There is a discussion about it on the MARK forum:

viewtopic.php?f=1&t=996&p=2620&hilit=model+average+betas#p2620
jhines
 
Posts: 599
Joined: Fri May 16, 2003 9:24 am
Location: Laurel, MD, USA


Return to analysis help

Who is online

Users browsing this forum: No registered users and 6 guests