6
$\begingroup$

I have to run an analysis using different steps, to include first the control variables, second the independent variables, and finally the interaction terms (moderation). However, I don't know if I should use hierarchical regression or stepwise.

Could anyone explain me what are the differences between both methods? When do we use each one of them?

Thanks a lot!!

$\endgroup$
3
  • $\begingroup$ To add just a detail to the answer and comment you already have: From your description, it seems possible that you don't need either of them. It seems you simply want a regular multiple regression, possibly comparing several nested models (Why did you decide you needed several steps? Are you planning to run tests or interpret the results after each step? Select variables to remove/keep in the model?) $\endgroup$
    – Gala
    Commented Jul 21, 2013 at 12:28
  • $\begingroup$ Thanks for your comments! @GaëlLaurans, I thought of using hierarchical regression to assess the contribution of the control variables first, then the IVs, and finally also the interaction terms in the model. I have seen that in some tutorials about hierarchical regression analysis... My idea is interpreting the results after each step. Would that be correct? $\endgroup$ Commented Jul 21, 2013 at 16:53
  • $\begingroup$ What you should do will also depends on the number of variables, observations and what exactly you want to know but basically it sounds reasonable. As others have said, just avoid stepwise regression. $\endgroup$
    – Gala
    Commented Jul 21, 2013 at 20:42

1 Answer 1

12
$\begingroup$

In hierarchical regression you decide which terms to enter at what stage, basing your decision on substantive knowledge and statistical expertise.

In stepwise, you let the computer decide which terms to enter at what stage, telling it to base its decision on some criterion such as increase in $R^2$, AIC, BIC and so on.

When to use which? Use hierarchical regression when you have knowledge of the field in which you are building a model. As for stepwise... well, I am tempted to say "don't use it". If you must use an automated procedure, you should use one that penalizes models for complexity, such as LASSO or LAR. The problems of stepwise have been discussed here many times, searching for stepwise should find lots of posts.

$\endgroup$
1
  • 5
    $\begingroup$ To add to Peter's excellent advice, failure to use subject matter expertise will result in such problems as finding that an "important" interaction terms is explained away by a missed main effect. $\endgroup$ Commented Jul 21, 2013 at 12:00

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.