Based on Paul Allison's booklet "Fixed Effect Regression Models"
(2009), the FE model can be estimated by person-mean centering all of your
variables (but not the outcome), and then including a random intercept for each
person. The centering gives you the FE model estimates, and the random
intercept adjusts the standard errors for clustering by individuals. Note that
your data must be in person-period (or long) format to do this.
In case you are unfamiliar with person-mean centering, that simply means taking
the mean of each person's values for a given variable for all of the periods
in your data, and then calculating a deviation from that mean at each time
period. For example, a person's average income over four years might be
$50,000, but in each year their actual income would be slightly higher or lower
than this (these would be the person-mean deviations). In symbolic form, your
code might look something like this:
library(lme4)
variable_pmcentered = variable - person_mean
mod = lmer(outcome ~ variable_pmcentered + person_mean + other predictors +
(1|personID))
The advantage of this method (which Allison calls a "hybrid" method)
over traditional FE models is that you get the benefits of a FE model
(subtracting out time-invariant omitted variables) along with the benefits of
random effect models (e.g., estimating coefficients for time-invariant
variables, estimating interactions with time, letting intercepts and slopes
varying randomly, etc.) See Allison's booklet for more details on this
method.
Allison, Paul D. 2009. Fixed Effects Regression Models. Los Angeles, C.A.: Sage.
Andrew Miles
On Feb 7, 2012, at 5:00 PM, cariboupad@gmx.fr wrote:
> Dear R-helpers,
>
> First of all, sorry for those who have (eventually) already received that
request.
> The mail has been bumped several times, so I am not sure the list has
received it... and I need help (if you have time)! ;-)
>
> I have a very simple question and I really hope that someone could help me
>
> I would like to estimate a simple fixed effect regression model with
clustered standard errors by individuals.
> For those using Stata, the counterpart would be xtreg with the
"fe" option, or areg with the "absorb" option and in both
case the clustering is achieved with "vce(cluster id)"
>
> My question is : how could I do that with R ?
> An important point is that I have too many individuals, therefore I cannot
include dummies and should use the demeaning "usual" procedure.
> I tried with the plm package with the "within" option, but R
quikcly tells me that the memory limits are attained (I have over 10go ram!)
while the dataset is only 700mo (about 50 000 individuals, highly unbalanced)
> I dont understand... plm do indeed demean the data so the computation
should be fast and light enough... and instead it saturates my memory and do not
converge...
>
> Do you have an idea ?
> Moreover, it is possible to obtain cluster robust standard errors with plm
?
>
> Are there any other solutions for fixed effects linear models (with the
demean trick in order not to create as many dummies as individuals) ?
> Many thanks in advance ! ;)
> John
>
>
>
> ______________________________________________
> R-help@r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
[[alternative HTML version deleted]]