Mean centering (good bad or ugly?)

[trinker]

As part of a regression course I have to do mean centering. I looked it up and have read 4-5 articles on it. Generally opinions are skewed in the direction that mean centering adds nothing to the interpretability (if that's a word) to the regression model. (see a paper on this by clicking here). Though there is counter literature as well.

I'm wondering your opinions on the subject? Do you use mean centering or would you?

EDIT: Here is a past thread on the subject (LINK)

[spunky]

thing is, without mean centering, i dont know of any other relatively easy ways to deal with the (sometimes SEVERE) multicollinearity that arises when you create an interaction or some other form of multiplicative variable to add to your model...

and if you think centering is ugly in regression, just wait 'till you go into multilevel/mixed-effects regression where all hell breaks loose!

[spunky]

i just skimmed through the article that you attached, but something caught my att'n almost at the very end, where they propose alternatives to mean centering:

"Because collinearity problems cannot be remedied after the data has been collected
in most cases, we recommend that researchers carefully design their research studies prior to
collecting their data. If feasible, one can address it by using a data collection scheme that isolates
the interaction effect (for example, a factorial design)"

factorial designs in the social sciences where research is 99% correlational and sometimes not even that?! are you F kiddin me?!

i dunno... sometimes i read this people saying "what we do is wrong, but our alternative to do it right is even worse" and cant help but wonder what were they thinking when they wrote that...

[trinker]

LMAO I appreciate the insight.

[Jake]

I'm going to go full disclosure and admit that I didn't read the attached article (and also that I'm drunk... so there's that). But I don't see what the issue really is? Perhaps the only purpose of mean centering is that it increases the interpretability of the parameters in your model, at least in many fairly common situations. It doesn't really change your model in a deep sense. After all, you're just rescaling one or more of the predictors, nothing more. It is especially helpful when you have interaction terms in the model and therefore the partial effect of (say) b1 only tells you about the case where b2 = 0. Obviously if b2 can never equal 0, which is the case often enough, then you will almost certainly make your life easier by mean centering b2 (and it probably wouldn't hurt to do b1 as well!). But again, nothing has *really* changed in the model, and any apparent multicollinearity issues are a complete red herring, because they are perfectly counterweighed by corresponding changes in the variance of your predictor variables when you choose to mean center vs. not mean center. The confidence intervals don't change! I can't see how anyone who really understands what is happening in regression could take a strong stance against mean centering.

[trinker]

That's the stance of Kromley & Foster-Johnson(1998) Mean Centering in Moderated Multiple Regression: Much Ado about Nothing. It was an interestinf read for sure. I'd link it but it's not publically available and thus would be wrong. Thank you for your drunken insight.

[trinker]

i dunno... sometimes i read this people saying "what we do is wrong, but our alternative to do it right is even worse" and cant help but wonder what were they thinking when they wrote that...

The answer is we're all pushing to get published at the expense of reason sometimes.

[Jake]

You mean I could have turned my post into a bona fide publication? ****!

[Dason]

Yup. You would have been a shoe-in for the ASAJDP (ASA's Journal of Drunken Posts).