Hi all,
I am running several fairly complicated presence/absence (binary) models, each of which includes ~700 data points and between 8 and 13 predictor variables (a mix of continuous and factor variables). I'm using logistic regression, and first fit these without spatial effects using glm(). Since we're concerned about residual spatial autocorrelation, I also added spatial effects (with an exponential correlation structure) in spGLM. After a few attempts and many (500,000) iterations, these appear to be converging quite nicely. However, the sigma^2 values are much bigger than we expected (35, 50, 100). As a result (I suspect), my parameter coefficients are also much more extreme than they were in the non-spatial models. For example, without the spatial term my coefficients ranged from about -1.5 to 1.5, and now they range from -5 to 7. Since this is on the logistic scale, these result in nearly perfect 0 or 1 predicted probabilities. This feels like something has gone wrong, but I'm having trouble placing my finger on exactly what. If not, what is the interpretation? (As a side note, the phi values are within the range we expected). Any insights would be greatly appreciated! Thanks, Sama Sama Winder MS Statistics University of Alaska, Fairbanks _______________________________________________ R-sig-Geo mailing list [hidden email] https://stat.ethz.ch/mailman/listinfo/r-sig-geo |
Hi Sama,
This post probably better belongs to https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models. Personally, I have no experience with spGLM. In our group we always use mgcv::glmmPQL. We check the effect of the included spatial correlation structure by fitting semivariograms of the model residuals of the glm (without correlation structure) and the glmm (with correlation structure). It is normal that the coefficients change but I can not comment on whether your magnitude of the coefficient change is suspicious or not. However, if the predictive accuracy changes than I would assume that something has gone wrong because pred. acc. should not be affected by the inclusion of an spatial correlation structure (afaik). I might take a look into ‘spGLM’ in the next days. In any case, this post is more a comment than an answer. Maybe someone else with more experience can help here. Cheers, Pat On 1. Aug 2017, 22:07 +0200, Sama Winder <[hidden email]>, wrote: > Hi all, > > I am running several fairly complicated presence/absence (binary) > models, each of which includes ~700 data points and between 8 and 13 > predictor variables (a mix of continuous and factor variables). > > I'm using logistic regression, and first fit these without spatial > effects using glm(). Since we're concerned about residual spatial > autocorrelation, I also added spatial effects (with an exponential > correlation structure) in spGLM. After a few attempts and many > (500,000) iterations, these appear to be converging quite nicely. > > However, the sigma^2 values are much bigger than we expected (35, 50, > 100). As a result (I suspect), my parameter coefficients are also much > more extreme than they were in the non-spatial models. For example, > without the spatial term my coefficients ranged from about -1.5 to > 1.5, and now they range from -5 to 7. Since this is on the logistic > scale, these result in nearly perfect 0 or 1 predicted probabilities. > > This feels like something has gone wrong, but I'm having trouble > placing my finger on exactly what. If not, what is the interpretation? > (As a side note, the phi values are within the range we expected). > > Any insights would be greatly appreciated! > > Thanks, > Sama > > Sama Winder > MS Statistics > University of Alaska, Fairbanks > > _______________________________________________ > R-sig-Geo mailing list > [hidden email] > https://stat.ethz.ch/mailman/listinfo/r-sig-geo [[alternative HTML version deleted]] _______________________________________________ R-sig-Geo mailing list [hidden email] https://stat.ethz.ch/mailman/listinfo/r-sig-geo |
In reply to this post by Sama Winder
Correction: MASS::glmmPQL, not mgcv::
On 1. Aug 2017, 22:07 +0200, Sama Winder <[hidden email]>, wrote: > Hi all, > > I am running several fairly complicated presence/absence (binary) > models, each of which includes ~700 data points and between 8 and 13 > predictor variables (a mix of continuous and factor variables). > > I'm using logistic regression, and first fit these without spatial > effects using glm(). Since we're concerned about residual spatial > autocorrelation, I also added spatial effects (with an exponential > correlation structure) in spGLM. After a few attempts and many > (500,000) iterations, these appear to be converging quite nicely. > > However, the sigma^2 values are much bigger than we expected (35, 50, > 100). As a result (I suspect), my parameter coefficients are also much > more extreme than they were in the non-spatial models. For example, > without the spatial term my coefficients ranged from about -1.5 to > 1.5, and now they range from -5 to 7. Since this is on the logistic > scale, these result in nearly perfect 0 or 1 predicted probabilities. > > This feels like something has gone wrong, but I'm having trouble > placing my finger on exactly what. If not, what is the interpretation? > (As a side note, the phi values are within the range we expected). > > Any insights would be greatly appreciated! > > Thanks, > Sama > > Sama Winder > MS Statistics > University of Alaska, Fairbanks > > _______________________________________________ > R-sig-Geo mailing list > [hidden email] > https://stat.ethz.ch/mailman/listinfo/r-sig-geo [[alternative HTML version deleted]] _______________________________________________ R-sig-Geo mailing list [hidden email] https://stat.ethz.ch/mailman/listinfo/r-sig-geo |
Thanks Pat.
I will check out glmmPQL to see if I get similar results as I do in spBayes::spGLM, since that could certainly be instructive. Could you tell me more about how you fit the semivariograms? Specifically, which residuals do you use, and then which semivariogram function? I have explored this a bit but ran into a few threads suggesting that semivariograms were more appropriate for normal data and linear trends and never came to a solution I was happy with. And, if I don't hear back from anyone else perhaps I will try the r-sig-mixed-models group. Thanks! Sama On Tue, Aug 1, 2017 at 2:18 PM, Patrick Schratz <[hidden email]> wrote: > Correction: MASS::glmmPQL, not mgcv:: > > > On 1. Aug 2017, 22:07 +0200, Sama Winder <[hidden email]>, wrote: > > Hi all, > > I am running several fairly complicated presence/absence (binary) > models, each of which includes ~700 data points and between 8 and 13 > predictor variables (a mix of continuous and factor variables). > > I'm using logistic regression, and first fit these without spatial > effects using glm(). Since we're concerned about residual spatial > autocorrelation, I also added spatial effects (with an exponential > correlation structure) in spGLM. After a few attempts and many > (500,000) iterations, these appear to be converging quite nicely. > > However, the sigma^2 values are much bigger than we expected (35, 50, > 100). As a result (I suspect), my parameter coefficients are also much > more extreme than they were in the non-spatial models. For example, > without the spatial term my coefficients ranged from about -1.5 to > 1.5, and now they range from -5 to 7. Since this is on the logistic > scale, these result in nearly perfect 0 or 1 predicted probabilities. > > This feels like something has gone wrong, but I'm having trouble > placing my finger on exactly what. If not, what is the interpretation? > (As a side note, the phi values are within the range we expected). > > Any insights would be greatly appreciated! > > Thanks, > Sama > > Sama Winder > MS Statistics > University of Alaska, Fairbanks > > _______________________________________________ > R-sig-Geo mailing list > [hidden email] > https://stat.ethz.ch/mailman/listinfo/r-sig-geo _______________________________________________ R-sig-Geo mailing list [hidden email] https://stat.ethz.ch/mailman/listinfo/r-sig-geo |
Free forum by Nabble | Edit this page |