c++ error: dpotrf A.alpha failed

[Daniela Zaffignani]

Hi everyone,

First, thank you for all the questions and answers shared in this group; they have been incredibly helpful in resolving many of my past issues.

I'm now encountering a persistent error while trying to fit a spatial, multispecies occupancy model. The model fails consistently with the error:
c++ error: dpotrf A.alpha failed

This error occurs after the first MCMC chain finishes successfully. Subsequent chains fail immediately.

Model Details:

• Data: 256 sampling sites, 2 replicates, 130 bird species.

• Covariates: I have four covariates, two of which I'm including as quadratic terms (total of six predictors):

  1. Two bamboo species proportions (range 0–1, means ~0.11 and 0.19). These are the ones I need to be as quadratic terms.

  2. Altitude (range 73–627, mean 325).

  3. Forest area (range 8–325, mean 249).

• Scaling Issue: The problem is directly related to covariate scaling.

  • The model runs perfectly if I use scale(variable, scale = FALSE) for the bamboo variables and full scale() for the others.

  • The model consistently fails if I use the default scale(variable) or  scale(variable, scale = FALSE) for all variables, including the bamboo proportions.

Troubleshooting Attempts:

• Adjusted all priors as suggested (particularly, phi.unif = list(3/(max.maxdist/2), 1/(max.mindist)) bounds from resources like the Doser Lab website).

• Verified no duplicate coordinates exist (models with scale(bamboos, scale = FALSE) runs ok, and then only changing the scaling runs to the same error).

• Increased MCMC settings substantially (n.batch = 10,000, batch.length = 25, n.thin = 50), but the error persists.

As I understand and looked up, the issue seems to be that full standardization (dividing by the standard deviation) of the bamboo proportion variables creates a numerically unstable covariance matrix. Since these are already on a defined 0–1 scale, they may not need the same scaling as continuous variables like altitude.

I definitly run out of options, any ideas of what to do next? 

Thank you for your time and expertise.

[Jeffrey Doser]

Hi Daniela, 

Thanks for the question, and for providing all the helpful background information. The error you received happens when running the MCMC sampler for the detection regression coefficients in the model. Because it's with the detection covariates, this suggests to me that there could be something wrong with the format of your "y" data matrix, or your detection covariates that is somehow passing through the model checks. The problem could be related to scaling of the covariates, and it's interesting that it runs with scale = FALSE but not when it is by default set to TRUE. Here are some things I would suggest doing to further troubleshoot: 

1. Run the model with all the scaling as normal with scale = TRUE for only one chain and take a look at the output. If you look at traceplots of each parameter, you may very well see that one or multiple of the detection covariates look extremely odd. That would help narrow down what covariate may be causing the problem. 
2. Multi-species models in spOccupancy require each species in the data array "y" to have the same number (and location) of missing values (NAs). Oftentimes the model may not detect unevenness in missing values (e.g., if the data were formatted incorrectly) and the model can run and lead to problems like what you're finding. 
3. Check to make sure that the NA values in your detection covariates are in the same spots as the NA values in your "y" detection-nondetection data. If you confirm all missing values are the same across species as I recommend above, then you can just compare it to one species in the "y" array. It's okay if there are more missing values in the "y" data, but you should not have an NA value in a detection covariate that is not an NA value in the "y" data. 
4. The standardization could certainly be the problem, but I am skeptical that it is the sole problem leading to this behavior. If the model runs successfully without standardizing, there is no requirement that you need to standardize every covariate in the usual way when fitting a model. It will just change how you interpret things in the resulting output. If you decide to move forward with the model you successfully fit, I would closely inspect the model output and traceplots to ensure that the model is functioning properly and it is not a scenario where the model successfully fits even though the data are not formatted correctly. 

Hope that helps! If you can't figure out what's causing the problems, feel free to send me your data and code and I can take a look. 

Jeff

--
You received this message because you are subscribed to the Google Groups "spOccupancy and spAbundance users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to spocc-spabund-u...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/spocc-spabund-users/f0b1d572-8873-48c2-8f9b-44c37b4b9fadn%40googlegroups.com.

--

Jeffrey W. Doser, Ph.D.

Assistant Professor

Department of Forestry and Environmental Resources

North Carolina State University

Statistical Ecology and Forest Science Lab

Pronouns: he/him/his

[Daniela Zaffignani]

Hi Jeff,

Thank you so much for your detailed answer.

I realize I should have clarified earlier—the variables causing the issue are occurrence covariates, not detection covariates. My apologies for the confusion.

I will implement your suggestions and let you know how it goes.

Thanks again for your help, and happy new year!

D.