topic Re: How to fit mixed Poisson/Negative binomial model with offset in Statistical Procedures

How to fit mixed Poisson/Negative binomial model with offset

SBS — Sun, 17 Feb 2019 15:02:25 GMT

I am very new to SAS and need some help fitting a poisson model first, and in case of overdispersion, fitting a negative binomial mixed model for the following context:

I want to isolate gender differences in the preferences for reports written by others using language that is high on several pre-determined categories (e.g., personal pronouns). I have raw counts of words in the report that fall into each of the various language categories as well as total number of words in the report.

I understand that I have to use mixed Poisson/Negative binomial regression models as in my data - (1) I have multiple reports bought by the same person; and (2) the same report may be bought by multiple people;

I also need to control for the random effects of two categorical variables: country and industry.

How do I fit a mixed model for this situation with a count dependent variable (raw count of personal pronouns in the report), offset parameter (total number of words in the report), gender of the rater (m=0; f=1); and four nestings (raterID; reportID; countryCode; industryCode)? How do I know if Poisson or negative binomial mixed model fits better based on regression results?

I understand that I need to use nlmixed. But I don't understand how to use it for my situation.I would really appreciate your kind help.

Re: How to fit mixed Poisson/Negative binomial model with offset

PGStats — Sun, 17 Feb 2019 19:58:08 GMT

Unless you are looking at non-linear models, you should be looking at proc glimmix, it does support Poisson and negative binomial distributed responses with offset.

Re: How to fit mixed Poisson/Negative binomial model with offset

SBS — Sun, 17 Feb 2019 20:58:26 GMT

I tried the following code with GLIMMIX

proc glimmix data=myData method=quad;
class GenderRater RaterID;
model RAW_COUNT_PRONOUNS =GenderRater / link=log s dist=negbinomial offset=log_TOTAL_WORDS;
random int / subject=RaterID;
run;

Within a second it gives the following error: The SAS system stopped processing this step because of insufficient memory

I just have 50,000 observations and 1500 Raters. What am I getting wrong?

Re: How to fit mixed Poisson/Negative binomial model with offset

SBS — Sun, 17 Feb 2019 21:09:29 GMT

I also tried NLMIXED and it gives no results, despite no error -

proc nlmixed data=myData;
xb = b0 + b1*GenderRater + u;
mu = exp(xb +log_TOTAL_WORDS);
m = 1/alpha;
ll = lgamma(RAW_COUNT_PRONOUNS+m)-lgamma(RAW_COUNT_PRONOUNS+1)-lgamma(m)
+RAW_COUNT_PRONOUNS*log(alpha*mu)-(RAW_COUNT_PRONOUNS+m)*log(1+alpha*mu);
model RAW_COUNT_PRONOUNS ~ general(ll);
random u ~ normal(0,s2u) subject=RaterID;

run;

Re: How to fit mixed Poisson/Negative binomial model with offset

PGStats — Sun, 17 Feb 2019 21:12:38 GMT

Maybe an explanation here:

http://support.sas.com/kb/60/666.html

Re: How to fit mixed Poisson/Negative binomial model with offset

SBS — Sun, 17 Feb 2019 21:16:20 GMT

Unfortunately, even the method=quad(fastquad qpoints=3) option gives the same memory error.

Re: How to fit mixed Poisson/Negative binomial model with offset

SBS — Mon, 18 Feb 2019 01:08:09 GMT

I was able to get NLMIXED running. The error was due to gender variable. It worked after dummy coding it. However, GLIMMIX still gives the same memory error.