Quartz | Level 8

## PROC GENMOD link option values for binomial distribution

Hi, everyone

Recently I have been learning logistic regression. I know that PROC GENMOD could be used to do a logistic regression with the option link = logit. However, with the link = logit I can only get the odds ratio. I am more interested in the relative risk and what I found via Google is the same PROC GENMOD but with an option of link = log.

So, my question is: Does the PROC GENMOD still do a logistic regression with the option of link = log? If yes, what is the math behind it? If not, then what is the mathematical mode it uses? Thanks

1 ACCEPTED SOLUTION

Accepted Solutions
SAS Super FREQ

## Re: PROC GENMOD link option values for binomial distribution

Yes, using the log link you model the log of the event probability and then exponentiation of parameters can be used to estimate relative risks. But using the log link can result in estimation problems because the log link does not keep predicted values restricted to the valid probability range. However, you can also estimate relative risks using a logistic model followed by use of the NLMeans macro. These and other approaches are covered in this note

2 REPLIES 2
SAS Super FREQ

## Re: PROC GENMOD link option values for binomial distribution

Yes, using the log link you model the log of the event probability and then exponentiation of parameters can be used to estimate relative risks. But using the log link can result in estimation problems because the log link does not keep predicted values restricted to the valid probability range. However, you can also estimate relative risks using a logistic model followed by use of the NLMeans macro. These and other approaches are covered in this note

Quartz | Level 8

## Re: PROC GENMOD link option values for binomial distribution

And after some google later, I found the link = log is a model called "log-binomial regression" (https://stats.stackexchange.com/questions/581678/what-is-the-difference-between-a-logistic-regressio... )

Logistic regression always produces estimates of $$𝑝𝑖$$ that are between 0 and 1. This is not true for log-binomial regression;