topic Re: Can proc logistic deal with continuous response regression? in Statistical Procedures

Can proc logistic deal with continuous response regression?

kimmygzc — Mon, 25 Oct 2021 07:49:39 GMT

I read this

"logistic regression describes the relationship between a categorical response variable and a set of predictor variables. "

here:Proc Logistic and Logistic Regression Models (ucla.edu)

But theoretically, a logistic regression is continuous function, why not continuous input for the model? I tried do this, but all my values of the respond variable are transited into a kind of categories, which is upset...

So, Can proc logistic deal with continuous response regression?

Re: Can proc logistic deal with continuous response regression?

PaigeMiller — Mon, 25 Oct 2021 10:38:53 GMT

@kimmygzc wrote:

I read this

"logistic regression describes the relationship between a categorical response variable and a set of predictor variables. "

here:Proc Logistic and Logistic Regression Models (ucla.edu)

But theoretically, a logistic regression is continuous function, why not continuous input for the model? I tried do this, but all my values of the respond variable are transited into a kind of categories, which is upset...

So, Can proc logistic deal with continuous response regression?

PROC REG performs regression on continuous responses. Although PROC LOGISTIC creates continuous predictions, it cannot have continuous response variable values. Another difference is that PROC REG will make predictions in the range negative infinity to positive infinity, while PROC LOGISTIC will make predictions in the range of 0 to 1.

Re: Can proc logistic deal with continuous response regression?

kimmygzc — Mon, 25 Oct 2021 10:45:53 GMT

Thanks for your reply, Sir.

I tried proc reg but only in linear regression, and read about pro nlin. Both have complex parameters to configure.

I'm looking for a method to do logistic regression with y range [0, +inf), where y is a continuous response variable but not a probability.

Do you have anything recommend?

Re: Can proc logistic deal with continuous response regression?

PaigeMiller — Mon, 25 Oct 2021 11:04:51 GMT

@kimmygzc wrote:

I tried proc reg but only in linear regression, and read about pro nlin. Both have complex parameters to configure.

As far as I know, there are no complex parameters to configure in PROC REG. Can you be specific what "complex parameter" you are talking about?

I'm looking for a method to do logistic regression with y range [0, +inf), where y is a continuous response variable but not a probability.

What about this is "logistic"?

There are other SAS PROCs that will fit models of the range [0, +inf), such as PROC GLIMMIX, but to get this to work, you have to specify the distribution of the continuous Y variable, and then you also have to pick a distribution that spans the range [0, +inf). For PROC GLIMMIX, you can see the list of possible distributions here: https://documentation.sas.com/doc/en/pgmsascdc/9.4_3.4/statug/statug_glimmix_syntax17.htm#statug.glimmix.gmxmoddist

Re: Can proc logistic deal with continuous response regression?

SteveDenham — Mon, 25 Oct 2021 13:07:18 GMT

PROC NLIN is pretty much suited for this. Do a google search on "four parameter logistic model" and you should find several examples, some with code.

SteveDenham

Re: Can proc logistic deal with continuous response regression?

Rick_SAS — Tue, 26 Oct 2021 09:57:40 GMT

> Can proc logistic deal with continuous response regression?

PROC LOGISTIC can fit a response variable that is a proportion in the interval [0,1]. The most familiar proportion model is the binomial model, which uses the event/trial syntax. However, as Steve says, there are other models for fitting proportions. See this note:

Usage Note 22871: Types of logistic (or logit) models that can be fit using SAS

The note discusses many logistic-type models, including the 4- and 5-parameter models mentioned by Steve.

Are you fitting survival/failure data? There are various models in survival analysis that have "logistic" in their name. See the documentation for the LIFEREG procedure.

If you provide example data and the model you are trying to fit, it will be easier to help you.