forward or backward stepwise logistic regression

KNIME Analytics Platform
1

Hi  - i  am a new KNIME user wondering if there is a  forward or backward stepwise logistic regression module in KNIME or if there is an alternative method to accomplish forward or backward stepwise logistic regression within KNIME?

 

2

Hi,

You can do this by placing our regression nodes between a Backward Feature Elimination Start and a Backward Feature Elimination End node.

Best,

Roland

3

Hi Roland!

Can you provide a workflow?

I do a lot of regressions but would be nice to test this process in my analysis.

Thank you!

4

Hi Fabio,

I have created an example for you, hope that helps!

Cheers,

Roland

5

Hi Roland!

Really cool! Can I ask you some questions about this process?

I'm planning to use the same workflow for Linear Regression with all numeric variables (around 12 independent variables). In my case:

  • Backward Elimination consider the predictor with the highest P Value to be removed. Is it possible to know wich was the predictor and what was the result of the model in every step?
  • Is it necessary to set up one of the Backward Feature nodes before start?
  • There are 2  Backward Feature Elimination Start node. Is it possible to give a short explanation?

Thank you for all your support!

6

Hi Fabio!

- You can see which predictor was removed in each step by looking at the "Removed feature" column of the Backward Feature Elimination End node

- The Backward Feature nodes do not require any configuration. You only have to specify the correct target/prediction columns in the Backward Feature Elimination End

- The 2:2 node differs from the 1-port node in that it has actually two ports on both sides. This makes it easier to feed training and test data into the flow.

Hope that helps!

Cheers,

Roland

7

Now with 4.1.2 things have changed.
My issue is that I can’t really use this, as nothing from linear regression learner goes as p-values to the loop end, so I am not able to perform stepwise regression filtering each step the variable with the highest p-value.
It’s a very common and used feature in other stat software, and it would be great to have it included as an option in the learner node.

8

Hi there @deicide_bg,

have you figured it out with help from @Martyna in this topic: Linear Regression Learner Output coefficients or still have questions?

Br,
Ivan

9

Yes, but this topic about something else. :slight_smile:

10

Hi @deicide_bg -

The workflow linked above is a bit outdated. Have you tried using the Feature Selection Loop start and end nodes? I know the loop start node allows four different selection strategies.

I guess I’m not sure what has changed in 4.1.2, that you are not able to do something now that you could before? Can you expand on that a little bit?

EDIT: Some discussion from a prior thread here that may be useful: Backward Elimination

11

Essentially, a high p-value would kick a variable out.
I’m not able to perform this with the current “elimination” metanode loops. The End loop node just won’t understand p-values, or I don’t know how to do it. Anyway, such straight forward functionality should be out-of-the-box with KNIME. I mean, it has DL, but not a simple stepwise regression.Duh!