Estimating Quasi-Poisson Regression with GLIMMIX in SAS

This article is originally published at https://statcompute.wordpress.com

When modeling the frequency measure in the operational risk with regressions, most modelers often prefer Poisson or Negative Binomial regressions as best practices in the industry. However, as an alternative approach, Quasi-Poisson regression provides a more flexible model estimation routine with at least two benefits. First of all, Quasi-Poisson regression is able to address both over-dispersion and under-dispersion by assuming that the variance is a function of the mean such that VAR(Y|X) = Theta * MEAN(Y|X), where Theta > 1 for the over-dispersion and Theta < 1 for the under-dispersion. Secondly, estimated coefficients with Quasi-Poisson regression are identical to the ones with Standard Poisson regression, which is considered the prevailing practice in the industry.

While Quasi-Poisson regression can be easily estimated with glm() in R language, its estimation in SAS is not very straight-forward. Luckily, with GLIMMIX procedure, we can estimate Quasi-Poisson regression by directly specifying the functional relationship between the variance and the mean and making no distributional assumption in the MODEL statement, as demonstrated below.

proc glimmix data = credit_count;
  model MAJORDRG = AGE ACADMOS MINORDRG OWNRENT / link = log solution;
  _variance_ = _mu_;
  random _residual_;
run;
  
/*
              Model Information
 
Data Set                     WORK.CREDIT_COUNT
Response Variable            MAJORDRG        
Response Distribution        Unknown         
Link Function                Log             
Variance Function            _mu_             
Variance Matrix              Diagonal        
Estimation Technique         Quasi-Likelihood
Degrees of Freedom Method    Residual        
 
              Fit Statistics
 
-2 Log Quasi-Likelihood           19125.57
Quasi-AIC  (smaller is better)    19135.57
Quasi-AICC (smaller is better)    19135.58
Quasi-BIC  (smaller is better)    19173.10
Quasi-CAIC (smaller is better)    19178.10
Quasi-HQIC (smaller is better)    19148.09
Pearson Chi-Square                51932.87
Pearson Chi-Square / DF               3.86
 
                       Parameter Estimates
                         Standard
Effect       Estimate       Error       DF    t Value    Pr > |t|
 
Intercept     -1.3793     0.08613    13439     -16.01      <.0001
AGE           0.01039    0.002682    13439       3.88      0.0001
ACADMOS      0.001532    0.000385    13439       3.98      <.0001
MINORDRG       0.4611     0.01348    13439      34.22      <.0001
OWNRENT       -0.1994     0.05568    13439      -3.58      0.0003
Residual       3.8643           .        .        .         .   
*/

For the comparison purpose, we also estimated a Quasi-Poisson regression in R, showing completely identical statistical results.

summary(glm(MAJORDRG ~ AGE + ACADMOS + MINORDRG + OWNRENT, data = credit_count, family = quasipoisson(link = "log")))
  
#               Estimate Std. Error t value Pr(>|t|)   
# (Intercept) -1.3793249  0.0861324 -16.014  < 2e-16 ***
# AGE          0.0103949  0.0026823   3.875 0.000107 ***
# ACADMOS      0.0015322  0.0003847   3.983 6.84e-05 ***
# MINORDRG     0.4611297  0.0134770  34.216  < 2e-16 ***
# OWNRENT     -0.1993933  0.0556757  -3.581 0.000343 ***
# ---
# Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
# 
# (Dispersion parameter for quasipoisson family taken to be 3.864409)
# 
#     Null deviance: 24954  on 13443  degrees of freedom
# Residual deviance: 22048  on 13439  degrees of freedom
# AIC: NA

Thanks for visiting r-craft.org
This article is originally published at https://statcompute.wordpress.com
Please visit source website for post related comments.