why is LinearModel.stepwise() so much slower than stepwisefit() ?

why is LinearModel.stepwise() so much slower than stepwisefit(), the algorithm description of the two functions is the same.

Edit: to avoid confusion I do not mean the function stepwise(), I mean the method of the class LinearModel, "LinearModel.stepwise()".

Answer 1

[Tom Lane]

The stepwise method in the LinearModel class is written to make its selection using any of a variety of measures. Also, it considers not just single-column changes, but also changes that may involve multiple columns because the term being changed is a categorical variable with multiple levels. The stepwisefit function, on the other hand, is optimized for single-column terms using a specific measure.

I've noticed this performance issue also. I'm going to enter a request in the MathWorks request database asking that some effort be put into speeding up the LinearModel version.

[William]

So basically it is a more general implementation that can handle more cases but is less optimized for a specific case? Yes it would be nice if it were optimized as stepwisefit() is.

Answer 2

[Greg Heath]

Stepwise involves human interaction.

Thank you for formally accepting my answer.

Greg

[William]

No it does not. You are thinking of the function stepwise(). I am asking about the method of the LinearModel class, LinearModel.stepwise().