Re: st: Centering variables in Multilevel Modeling

[Stas Kolenikov <skolenik@gmail.com>]

Centering wrt to the grand mean does not change anything (except for
the intercept of the model), while centering wrt to the level 2 means
affects the estimates of the slopes and  interpretation of the
coefficients, and is said to remove variability of the predictor
between level 2 units from the intercept(s). In my view, these are
rather different models; both rely on generated regressors, and hence
produce understated standard errors (especially the one with level 2
centering, as it needs to estimate more parameters to generate these
regressors).

With 30 units, especially unbalanced, I would not trust the estimates
of variability at level 2. In all likelihood, you have complex survey
data that have stratification, clustering and unequal probabilities of
selection that you need to account for.

On Thu, Mar 24, 2011 at 11:11 AM, Owen Corrigan <ocorrigan@gmail.com> wrote:
> In my multilevel regression model I have 30 level-2 units (countries) and
> about 20,000 cases (people), where the cases are very unevenly distributed
> across level-2 units. I have predictors at both levels. I want to center my
> level-2 continuous predictors, but am faced with a dilemma:
>
> 1. Do I center level-2 variables at the mean *solely* for those 30 units?
> OR
> 2. Do I center the level-2 variables across the entire dataset of 20,000
> cases, even though some level-2 units have many more cases than others, and
> thus the 'mean' will be weighted here towards those larger level-2 units?
>
> There are obviously large differences in the means arrived at depending on
> approach taken.
> All advice appreciated!


-- 
Stas Kolenikov, also found at http://stas.kolenikov.name
Small print: I use this email account for mailing lists only.
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/