Groups | Search | Server Info | Login | Register

Groups > comp.soft-sys.sas > #866

Proc Glimmix-Overdispered Poisson regression model with nested random effects

Show all headers | View raw

Hi All,

I’m using proc glimmix to investigate the effects of environmental
factors on the number of wading birds observed foraging in 1392
different 2km x 2 km grid cells during the breeding season.  This is a
long-term data set, where data are collected once a month (in every
cell) from Jan. through May (5 survey months).  I am using data from
2002-2009, thus I have 8 years of data equaling 47,071 observations.
My dependent variable is a count variable and therefore I am using a
poisson distribution. I am not interested in the difference between
years or survey months, just want to account for the yearly and within
year variation. Thus, I am treating these as random effects.  Survey
month is nested within year. Additionally, my data is overdispersed
with spatial autocorrelation. I attempted to use "random _residual_  /
type = sp(exp)(lat lon)" random statement to account for
overdispersion and autocorrelation, but I receive "Interger overflow
on computing amount of memory required" and "SAS stopped processing
because of insufficient memory" errors. I do not believe SAS can
handle this random statement with a data set as large as mine. So I am
trying to do an RSMOOTHING to handle the spatial aspect. However, I
get a "QUANEW Optimization cannot be complete" error. I am not sure if
I am using the correct code to run this. It is listed below:

proc glimmix data = final method=laplace; /**1**/
title 'Great Egret: Global ';
class  basinid rev year surveymonth;
model ln_greg = basinid wd wd2 rev rec2 sqrec2 dsd cat sawmarsh
fwmarsh avep wd*rec2 / s cl dist=poisson  ;
random lat lon /type=rsmooth; /*accounting spatial autocorrelation*/
random int / subject = year;
random int / subject = surveymonth(year); /*survey month nested within
year*/
output out=predA pred=pred;
ods output fitstatistics =a  parameterestimates=fixeda; /*outputs
parameter estimates, AIC and Logliklihood*/
run;

I am using the method=LAPLACE in order to calculate a maximum
likelihood rather than a pseudo-maximum likelihood in order to
calculate AIC values for a model selection approach. I have been
reading however, that proc glimmix is not a good framework for model
selection. Is this correct? Does using method=laplace correct this by
not using pseudo-maximum likelihood?  I have also run the model above
with a negative binomial distribution to account for overdispersion,
but the get "Floating Point Zero Divide" and "Termination due to
Floating Point Exception" errors.

Does anyone have any suggestions on modeling this overdispersed count
data?

Cheers,
Michelle

Back to comp.soft-sys.sas | Previous | Next | Find similar

Thread

Proc Glimmix-Overdispered Poisson regression model with nested random effects miimes1 <michelle.imes@gmail.com> - 2011-06-03 12:46 -0700

csiph-web