Centering predictors

noetsi

Fortran must die
#1
The only place I have encountered this argument is in multilevel approaches, but obviously it could include logistic and linear regression generally.

" As with other forms of regression models, the analysis of multilevel data should incorporate the centering of predictor variables to yield the correct interpretation of intercept values and predictors of interest (Enders & Tofighi, 2007; Hofmann & Gavin, 1998).Predictor variables can be group-mean or grand-mean centered, depending on the primary focus of the research question. For a well-written discussion on the pros and cons of the two approaches to centering predictor variables, we refer readers to Enders and Tofighi (2007). In the examples presented in our paper, we utilized both binary and standardized predictors, therefore, our variables already had a “meaningful zero” value for interpretation of the intercept and other predictors of interest."

I have two questions.

Is it best practice to center predictors if the models are multilevel?
If your predictors are all dummy variables or standardized variables is it the case you do not center these? I assume if some are standardized or dummies then you do not center these, you center only interval variables.

I have rarely run into this advice in linear or logistic regression.
 

hlsmith

Not a robit
#2
The econ people love centering and it is definitely in the social sciences. If you center, you should also make sure the reader knows what a stdev was valuewise since it may get lost and be different between samples. I try to remember to standardize, but it isnt a huge thing in medicine. It is required in Lasso, so that the model can compare between variable effects, kind of like standardized results in linear reg to compare vars. I think it is alsi recommended when you have an interaction.

Correct, if vars are binary or already standardized you are fine. Not sure about interval vars.

I think in MLM it may help with complexity in computations.