[R-sig-ME] constructed level 2 predictors/random effects

[Paul Johnson]

I see people wanting to average survey responses to manufacture
contextual variables. They just take the average of individual level
scores and treat it as if it were context.  Or in studies of
education, they average class Socio Economic or religious variables
and use the means for context.   Have you ever seen R packages that
try to facilitate this kind of work?

I am thinking about problems like this.

1. Can we account for the standard error of the mean at the group
level?  Will the come back to "not with lme4, but the old lme had
varIdent for weights?"

Of course, if there is 1 or 2 people within a cluster, and 50 in
another, we'd have an especially big reason to try to fix this.

2. It seems to me they should calculate a "leave one out" estimate for
each row, excluding that case's impact on the group-level average.

I'm thinking about the education studies that want to both create mean
SES as a predictor, and then look at individual variations against
that predictor.  If there are 100 people within each classroom, I
don't guess it matters.  But sometimes they have 2 or 5 people within
each group.

3. They are using raw averages, not pooled estimates for these
constructed level 2 variables.  It looks to me like we ought to fit a
multi level model on those variables, using the group as a random
effect. Then take the BLUPs as estimates of the context.  Otherwise,
these means at the group level are just as inefficent as the
one-regression-per group approach.

Even if you don't know of R work on this, I'd appreciate any pointers
to literature or such.

-- 
Paul E. Johnson
Professor, Political Science       Director
1541 Lilac Lane, Room 504      Center for Research Methods
University of Kansas                 University of Kansas
http://pj.freefaculty.org               http://crmda.ku.edu