[R-sig-ME] Fitting multilevel model to a recipe and simulating the model.

[Paul Johnson]

On Fri, Jun 29, 2012 at 8:26 AM, Justice Moses K. Aheto
<justiceaheto at yahoo.com> wrote:
> Dear All,
> I am a new R user but currently working on my dissertation which request that I fit multilevel model to a given situation that can be located in the attachment to this mail and simulate the model 10000 times.
> In addition, I will like to compute the standard error, t-value and p-value for the fitted model.
> I am available to provide further clarification on the task should you request for.
> Could someone help me in this regards.
> I do appreciate any assistance given me.
> Many thanks in advance.
>
> Kind regards.
>
> Justice Moses K. Aheto
> (Chief Executive Officer)
> Statistics & Analytics Consultancy Services Ltd.

In order to get this done, you will have to learn quite a bit of R and
mixed effects modeling.  What is your time frame for the project?  I
see you are a Chief Executive Officer, it may behoove you to hire a
programmer rather than learning this yourself. I think you will see
you have to think this through in stages, you try and fail in
translating your simple idea into a final result.

I have some notes about building simulations of that sort here:

http://pj.freefaculty.org/guides/stat/MonteCarloExperiments

I have a tutorial on that similar exercise here:

http://winstat.quant.ku.edu/svn/hpcexample/trunk/Ex80-PrevSci2007

Read through the Version-X.R files to get the idea.  Putting that on a
cluster computing system is the end goal.

I'm serious, though. You have to work through one test sample very
carefully to be sure you understand your model, and then worry about
drawing more samples and summarizing.

I'm working on a tutorial to go through those steps for a mixed model
just exactly as you describe, but have only worked so far on the parts
for generating one data set and analyzing it.  Next I'll come to the
part about doing that over and over and harvesting the results.

You need to learn lmer well enough to decide which parameter estimates
are worth tracking in a simulation, and ignoring the rest.  Especially
where we are concerned with estimating variance components and the
associated degreees of freedom, or confidence intervals for
predictions, I find we have plenty of questions and not so many
definitive answers.

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