EM logistic regression | how much each variable explains

[MariaD]

Hi folks, 

There is any way to view in logistic regression node results how much each variable included explains? I have continue and categorical variables. If not, there is any way to calculate?

Regards, 

[Ksharp]

You could check the estimated coefficent. The more big of it (absolute value of it),
The more explain in model.
Also you could check P value ,the more small ,the more important of this variable.

[PaigeMiller]

I think you mean ... check the standardized coefficient ... so all coefficients are on the same scale. Otherwise the value of the un-standardized coefficient are all on different scales and not really comparable.

[MariaD]

Thanks @PaigeMiller.The only problem it's that the regression node does not show the standardized coefficients for the categorical variables in the result.

I'm thinking in add a decision tree node, after the regression, with only the variables used on the regression model and with the P_1 as target, selecting the largest tree option. After that, I analyze the variable importance result of the tree decision node. What do you think of this approach?

[PaigeMiller]

Decision trees do fundamentally different things than a logistic regression. So, I'm not sure that a decision tree will answer your original question about logistic regression.

As far as not having standardized coefficients, you can first standardize the data, and then run the logistic regression (this gives you standardized coefficients)

[MariaD]

Thanks @PaigeMiller . I have the interval variables already standardized, but that does not make much sense for the categorical variables because the are used as 0/1 coding for the regression.  

[PaigeMiller]

You don't have to code categorical variables as 0 or 1, you could leave them as categories and use a CLASS statement (I'm sure EG has something similar) but however you handle categorical variables, standardizing does not apply.

[MariaD]

Yes @PaigeMiller , you are right. I don't recode the categorical variables, is the way the node treats them. So, in this case, I understand is to analyze the standardized coefficients for the interval variables and the normal coefficients for each category of the categorical variables. That's correct? 

[PaigeMiller]

I think that's correct.

[StatDave]

There is a Variable Selection node that you could look into. Separately from Enterprise Miner, see this note on variable importance that might be helpful particularly if using the procedures in SAS/STAT.