Highly Skewed Covariates = Convergence Issues?

[MCJ2]

A colleague and I are running multi-season models in PRESENCE where we are interested in relating occupancy of certain animal species to various site-specific habitat covariates. We are experiencing many convergence problems and large beta estimates and standard errors for many of our study species. We've played around with different initial values for our betas, transformations of our data, etc. from remedies posted on the forum.

It seems we might be having most problems when tree canopy cover percentages and densities of particular tree species are used as covariates. We are working in a savanna-type landscape where these covariates are highly skewed (many very low values in the predominate open grassland and a few high values near / under trees) -- this skewness remains despite z and arcsin transformations of those covariates and elimination of "outliers". Has anyone experienced similar problems with such covariate data? Any suggestions? Should we throw out these covariates altogether?

Thanks for any help

[jhines]

You can have convergence issues when there is insufficient data, or variation in the covariate to build a relationship between the covariate and the parameter estimate. I've seen this happen quite a bit with multi-season analyses when occupancy is very high. In that case, there are very few unoccupied sites, so it's difficult to estimate colonization (gamma). It's tough to say what's happening with your data without seeing it. Perhaps you could email the latest backup-zipfile in your project folder and I'll take a look.

Jim (jhines at usgs.gov)

[ekalies]

Related to this post (I have the exact same problem with some skewed variables)- are there still useful results that can come of a model with a parameter that has a big SE? I have 12 species each being modeled in a multi-season framework with 8 occupancy covariates each (it is a huge dataset), and it all works beautifully except just one parameter has large SEs for 4 species. I just want to talk about relative sizes of betas and directions of signs for the different parameters. For each species with one "bad" parameter, can I still interpret the other parameters and feel confident that this isn't nonsense? I hate to run different models for these 4 species (omitting the "bad" parameter), as I like the apples-to-apples effect I have going on. Thanks!

[darryl]

What is the variable that is causing the issue? If you have a skewed variable (lots of small, and some large values) sometimes using a log or sqrt transformation of the variables can help as the large values get adjusted more than the small values. Untransformed, those few large values can have a lot of influence on the esitmated effect size. Not neccesarily good or bad, just what it is and you need to be aware of it.

Cheers
Darryl