Stepwise Error

[will_weeden]

Hey guys,

I am attempting to run a stepwise function that is connected to my logistic regression model as well as my input data being fed through a create samples filter. I'm working with 30 different predictor variables and after about 9 minutes of runtime, I came across this error message: Error: Stepwise (131): Tool #5: Error in step(Log_Reg, direction = "both", k = 2). Has anyone else run into this error message? And if so how was it resolved? I have attached a picture of the workflow with the error message below. 

Will

[JoeS]

Hi @will_weeden

It's a tough one to diagnose as that's just the R call function that has failed being returned in the error, so understand your confusion.

In your workflow you have sent in different data to the stepwise tool compared the data sent to the logisitic regression model.

If you change it so you use the same data for both, I think that should stop the error.

The stepwise tool needs to receive the same data as the main model, as then its able to work through choosing the optimal predictors.

Thanks

Joe

[will_weeden]

Okay I actually got it to work yesterday but I can't remember what I changed to make it work lol. I think it had something to do with the data types in a few of my fields being set to "string" instead of "Int32" or "float" or vice versa. 

I'm currently running a boosted model, logistic regression, spline model, and a forest and decision tree model, all receiving data from the create samples tool that is set at 80% estimated/20% verification. Do I need to make sure that the same data is being processed by each model like you were mentioning with the stepwise model? And if so how do I ensure that the same data is being processed by each respective model? 

Will

[JoeS · Accepted solution]

Whilst I can't say with all seeing wisdom (AKA I am not a data scientist). 

I think you would be best to use the same data for all the models. As this means that they are all starting from the same place. 

If you use different data for each model, then it doesn't make the comparisons of the model as easy/fair.

To do so, you can just create as many connections as needed from the 80% part of the create samples tool.

That's not to say that there couldn't be reason to change the data around though.

[will_weeden]

Okay cool I will do that thank you for your responses!

[JoeS]

No problem.

I am not sure if you are aware but we have a great set of courses available around learning predictive, you can access them for free on the links here

[sgrabish1]

I had this issue and using the same data solved the issue