Multivariate Logistic Regression

[#1]

I am using JMP Pro 14.1 for fitting a multivariate logistic regression. The response variable Y is nominal, and all the columns in design matrix X is continuous numeric. I have some questions about feature selection and inverse prediction.

1)  When I use stepwise logistic regression for feature selection, I cannot find the option for "stepwise regression" from Personality drop-down list. Is stepwise regression not applied to the case when Y is nominal? 

2) If stepwise logistic regression is not applicable with JMP, then is there a way to do cross-validation for logistic regression?

3) How to illustrate the results of inverse prediction for multivariate logistic regression? Every time I got one inverse prediction window for one variable (the one I put blank value), what are the default values for the other variables? Does that mean when other variables are fixed we have 95% confidence that when the variable of interest is within the predicted window we can have P(Y=1)=0.9 (or other values)?

[#2]

Hi @lzhang83,

Stepwise is not available with non-continuous responses. Since you have Pro, there is a Validation column role in the Fit Model dialog.  You'll need to create a validation column first (shown here: https://www.youtube.com/watch?v=M5_mECc4NAg).  Unfortunately, I don't have a ready answer on the inverse-prediction stuff and I have to leave.  If no one else answers in the mean time, I'll try to come back to this.

[#3]

You could also consider changing the fitting Personality setting to Generalized Regression. You can select estimation methods and options after you click Run and the platform launches. Stepwise is available for nominal responses. You several other choices for variable selection, such as LASSO.

[#4]

I tried generalized linear model with binomial as distribution and logit as link function, but I still cannot see any options for stepwise regression from the red triangle drop-down list. Could you show which one I can select from the list?

[#5]

Follow this example.

1. Open Big Class data table from Sample Data folder.
2. Select Analyze > Fit Model.
3. Complete the launch dialog as shown below:
   
4. Click Lasso (default method) and select stepwise method, such as Backward Elimination, and click Go.
   

In your case, you should probably use the Poisson distribution, as you have count data.

[#6]

Now I know what is the issue. When I put the nominal type of variable in Y field, I can not see the option “generalized regression” from Personality list. Type of Y has to be continuous numeric. In my case, Y value is either 0 or 1. I believe I should choose Binomial as Distribution.

[#7 · Accepted solution]

Are you sure it's not there?  You can see it in Mark's screenshot with a nominal response variable.  I'm guessing that you aren't seeing the scroll bar on the Personality dropdown list.  You need to scroll up a bit to see it.

[#8]

Yes, now I see that. Thank you all for your help.