Orange Forum • View topic - Linking logistic regression & evaluation widgets

Linking logistic regression & evaluation widgets

A place to ask questions about methods in Orange and how they are used and other general support.

Linking logistic regression & evaluation widgets

Postby algol » Fri Jun 22, 2007 19:37

Question from a complete newcomer to Orange. Question confined to Widgets. I place DATA widget on canvass. Load Titanic datast. Drop in LOGISTIC REGRESSION widget (from Classify). Drop in NOMOGRAM widget (from classify). Connect them. Hit OPTIONS - Enable all links. Click on Nomogram widget and am very well-pleased. Excellent.

Now I am a Statistician... and am rather more interested in the EVALUATE widgets. ROC and LIFT and CALIBRATE. If I connect any one of them to the LOGISTIC REGRESSION widget, all I get is a pop-up ---- Selected widgets don't share a common signal type. Unable to connect ----

It seems that these 3 evaluation widgets require an input from --- Evaluation results --- (orngTest.ExperimentResults).

Where exactly is THAT widget? I am expecting to be told to look somewhere stunningly obvious and feel a fool, but having tried pretty hard, I'll take the risk.

Finally... how do I get to see the logistic regression model parameters (e.g. parameter values, standard errors, P-values. Odds-ratios... etc).

Next time I will ask a much shorter question!!! Promise.

Any helpful replies get 3 candles lit each. Thankyou.

Postby Patrick Shaughnessy » Fri Jun 22, 2007 23:07

You need to decide what sort of test to gather metrics about the regression, such as cross-validation, leave-one-out, or a separate test data set. This is done with the Test Learners widget, first button under the Evaluate tab.

I don't know how you visualize the model parameters for logistic regression.

Re: Linking logistic regression & evaluation widgets

Postby algol » Sat Jun 23, 2007 16:10

Hello Patrick.

Thank-you for your helpful (and considerate) advice. I dropped the --- Test learners --- widget (from EVALUATION tab) onto the canvas, figured out that it needed to be linked to both the Logistic regression widget AND a source data widget (FILE with Titanic dataset in this case).

Adding the Lift / ROC / Calibration widgets thereafter (connected to Test learners) delivered the goods. Brilliant. You have your three candles :o)

My only interest in seeing the ORANGE logistic regression (LR) parameter estimates is a cross-check. I have an LR model --- risk factors c/o prioritising referrals for endoscopy for possible bowel cancer --- developed using STATA. The latter has no nomogram facility. No surgeon is going to bother with fancy equations ... it has to be simple / direct. A nomogram is exactly what the doctor ordered!

STATA does not deliver a nomogram. S-plus does… cannot afford S-plus. R does… but I am too old for yet another language, even if R is brilliant. ORANGE does… interesting, but I am cautious. Devil and detail eh!

Mind you, cannot help liking a software package that on first use announces… 'Several widgets now use numpy module… do you want to download it now?' I will persist with ORANGE.

My thanks again. Paul.S. (algol)

Return to Questions & Support