nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

Option for Stepwise Regression Whole Effect rule to respect heredity for categorical model effects?

Chris_Kirchberg

Hi All,

I have a customer asking if there is a way to get Stepwise to respect heredity when looking at categorical effects in the model with interaction terms and using the Whole Effects rule. Currently, when using Whole Effects, Heredity does not seem to be respected, at least not strong heredity from what I can tell.

Anybody know a way around this? Search on Google did not readily reveal anything (that one is for you Lou )

Thanks,

Chris

Accepted answers

DonMcCormack

I'm aware of two options:

1) Fit Model, Whole Effects, add in main effects after the fact.

2) Generalized Regression using forward selection and checking Enforce Effects Heredity.

All comments

DonMcCormack

I'm aware of two options:

1) Fit Model, Whole Effects, add in main effects after the fact.

2) Generalized Regression using forward selection and checking Enforce Effects Heredity.

Chris_Kirchberg

I just thought of another manual way. If a main effect is not entered into the final stepwise with Whole Effects, then manually remove all of the terms that contain that main effect that was not entered.

Or remove the interaction terms in the Fit Model Dialog if the main effects are there.

They were looking for something more automatic instead of manual.

GenReg sounds good if one had Pro and might be the solution to the question. I forgot about that option in GenReg. Thanks.

DonMcCormack

Should have mentioned, there is a difference in the way Least Squares and GenReg parameterize the interactions. In both cases, they split them into 1 df effects. LS uses the contrasts that you would have gotten if you used Partition with the respose as the Y and the categorical factor as the X (this way, the first split is the contrast that captures the most variability in the response, the second split the one that captures the second most, etc.) GR compares the last level with each of the other levels.