I would like to fit a Tweedie GLM to the data. However, even though PROC GENMOD from SAS and glm() from R gives the same coefficient estimates, they give quite different standard error. What it the reason behind this? If we run the following R code, mydata <- read.csv("https://stats.idre.ucla.edu/stat/data/binary.csv")
mydata$rank <- as.factor(mydata$rank)
mydata$rank <- relevel(mydata$rank, ref = "1")
mytweedie <- glm(admit ~ gre + gpa + rank, data = mydata,
+ family = statmod::tweedie(var.power = 1.5, link.power = 0))
summary(mytweedie) R shows that Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -3.4957360 0.7906423 -4.421 1.27e-05 ***
gre 0.0016542 0.0007611 2.173 0.030345 *
gpa 0.5274319 0.2307913 2.285 0.022825 *
rank2 -0.3265264 0.2167527 -1.506 0.132754
rank3 -0.7947102 0.2364447 -3.361 0.000852 ***
rank4 -0.9818197 0.2867970 -3.423 0.000683 *** However, if we run the same model using the same data on SAS with PROC GENMOD, PROC GENMOD data=logistic_test DESCENDING;
CLASS rank(REF = "1");
MODEL admit = gre gpa rank / DIST=tweedie(p=1.5);
RUN; The result is Parameter . DF Estimate Standard Error
Intercept 1 -3.4958 1.1387
gre 1 0.0017 0.0011
gpa 1 0.5274 0.3176
......
... View more
