Model convergence issues with models including continuous covariates (distance)

[Melissa Breyer]

Hello,

I ran models including continuous site covariates "patch size" and "distance to nearest occupied patch" with single-season, single-species data in unmarked R.  I surveyed 14 sites for 4 independent surveys each, and detected the species a total of 5 survey days.  

I encountered model convergence problems with models including "distance to nearest occupied patch". I encountered the same problems when running models in Program PRESENCE.  Site covariate values were scaled in both cases.  What may be the reason for this lack of model convergence? I have attached the input files used.  

Thank you,

Melissa

[Jeffrey Royle]

hi Melissa,

 you have almost no information in your data set (detections at 2 sites, non-detections at 12 sites). This is such an extreme case that I think the model parameters are simply not identifiable. It's possible that you could estimate an intercept only model but beyond that there really isn't much there.

 Can you go back into the field and survey a few more sites? 

 I would also suggest a simulation study to investigate sample size requirements.

regards

andy

--
You received this message because you are subscribed to the Google Groups "unmarked" group.
To unsubscribe from this group and stop receiving emails from it, send an email to unmarked+u...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Allyson Lenora Degrassi]

Hello,

I hope you are all having a great day.

I apologize, but I need some help with my models.
I am attaching a sample code and 2 csvs with the data I am trying to model.

The dataset includes 392 rows (sites).
When I run the model with all the rows, the model breaks with the error:
"Error: Hessian is singular. Try providing starting values or using
fewer covariates."

I did some tests and I noticed that the model will work if I only use
343 out of the 392 rows.
I cannot explain why this is happening. Do you have any ideas or suggestions?

Thank you for your time.

Best,
Allyson

[Jeffrey Royle]

hi Allyson,

if you provide some starting values that are slightly better then it seems to work (See below).  In many cases it helps to provide good starting values.

regards

andy

 > (Occ.Habitat <- occu(~1 ~ Habitat, Occ.umf,start=c(-1,-2,0,0,0)))

Call:
occu(formula = ~1 ~ Habitat, data = Occ.umf, starts = c(-1, -2,
    0, 0, 0))

Occupancy:
                Estimate    SE      z P(>|z|)
(Intercept)       0.6752 0.350  1.927  0.0540
HabitatHardwood   1.3514 0.803  1.684  0.0922
HabitatHemlock   -0.0652 0.455 -0.143  0.8860
HabitatLogged     0.1349 0.472  0.286  0.7750

Detection:
 Estimate     SE     z  P(>|z|)
    -2.08 0.0896 -23.2 9.1e-119

AIC: 2174.072

AIC: 2176.047

[Allyson Lenora Degrassi]

Hello Andy,

Thank you very much. I will do that!
Have a great day.

Thank you,
Allyson

--
Allyson L. Degrassi, PhD Candidate
National Science Foundation Graduate Research Fellow 2013
Department of Biology, University of Vermont
Marsh Life Science 211
120A Marsh Life Science 109 Carrigan Dr.
Burlington, VT 05405
---
website: http://www.uvm.edu/~adegrass/

[Carla Mere]

Hello!

I am encountering the same error as Melissa, after I run a single-season model using unmarked. I surveyed 26 sites, 15 surveys and I got 53 detections. Is there any suggestion for what it may be the error?  I have been using the same procedure for other species, but there are some species that have this convergence problem.

nPars    AIC delta  AICwt cumltvWt

p(.)psi(dist + elev)       4 193.27  0.00 0.3763     0.38

p(temp)psi(dist+ elev)     5 193.69  0.43 0.3042     0.68

p(temp)psi(dist)           4 193.98  0.72 0.2631     0.94

p(.)psi(dist)              3 198.78  5.52 0.0238     0.97

p(temp)psi(elev)           4 199.02  5.75 0.0212     0.99

p(temp)psi(.)              3 200.78  7.52 0.0088     1.00

p(.)psi(elev)              3 203.28 10.02 0.0025     1.00

summary (model)

Warning message:

In .local(object, ...) :

  Model did not converge. Try providing starting values or increasing maxit control argment.

Thank you very much.