Logistic regression

[KavitaShah]

Hello,

We are trying to analyze an open source survey data to find a correlation between categorical dependent and independent variables. We have performed chi square tests to test the null hypotheses and also would like to perform logistic regression to find a correlation between these variables. However, I understand that logistic regression is predictor model, more specifically, binary classifier model which means that it tries to assign exactly one of the 2 classes to every observation and it produces probability with which each observation belongs to each class. It can be more accurately applied to variables which have exactly 2 answers like yes or no; survived or dead; male or female and such observations.

So now, according to our data, this does not hold true as we have many possible answers to the dependent variable like fair, poor, good, very good and excellent. However, we are considering only 2 answers namely poor and fair for our analysis.

Can we use logistic regression for correlation of such a data-set? We have also used Spearman Rank Correlation Coefficient to determine the correlation

Also, is there any way to correlate the 'c value' from the logistic regression and the Spearman Rank Correlation Coefficient?

Any guidance on this is highly appreciated.

Thank you in advance.

-KS.

[Reeza]

There's multinomial logistic regression as well or ordinal logistic regression which are more suited to your question.

If you look at the documentation for PROC LOGISTIC you'll find the example section has some examples. Your best bet is to walk through those examples. 

You may also want to consider PLS regression methods.

[KavitaShah]

Thank you for the suggestion.

[StatDave · Accepted solution]

In addition to the binomial (2-level) response, logistic regression models can be applied to multinomial (ordinal or nominal) responses that have several levels. SAS offers several procedures that can fit all of these models. See this note that details these and other types of logistic models and the procedures that can be used. For example, the LOGISTIC and SURVEYLOGISTIC procedures can be used to fit binomial and multinomial response models. If you have survey data, then you will probably want to look at PROC SURVEYLOGISTIC. These modeling procedures do not deal with "correlation" in the simple two variable sense. But the tests of the predictors in the model provide you with tests of the association of each predictor with the response after adjusting for the effects of the other predictors.

[KavitaShah]

Thank you for your inputs. i am working through the suggestions you have given.

[Ksharp]

PROC ADAPTIVEREG and PROC GAMPL also can do logistic regression.

[KavitaShah]

Sure, thank you.