Hi

I have taken the liberty of cc:ing the
R-SIG-Mixed-Models at R-project.org mailing list on my reply.  It is best
to send such queries to that list as several of those who read the
list are able to respond, often much faster than I will.

Dear Professor Douglas Bates,

My name is James Cho. I work for marketing analytics company.

I am working with an internal development team on linear mixed modeling, and
we have been testing R(lmer) and SAS (Proc Mixed) in terms of their
performance and compatibility to determine if we could incorporate R into
our company?s statistical application.

One thing I noticed is that Proc Mixed and lme() give very close results but
lmer() sometimes gives very different estimates when the model has a poor
fit (for example two of random effects are highly negatively correlated).

Could you provide an example, please?  If the model specifications are
equivalent then lme and lmer should provide corresponding values of
the REML criterion or the deviance.

Does lmer() use different algorithm than lme()?

Yes.  lmer() uses a profiled REML criterion or deviance and a
different specification of the variance-covariance matrix of the
random effects.  They also use different optimizers which may have an
effect on the results.  We are currently evaluating the use of
different optimizer functions for generalized linear mixed models, in
particular, but would be very interested in situations where the
linear mixed models did not converge appropriately.  This is why a
reproducible example, perhaps with anonymized data, would help.

Lmer() has default max iteration as 300. If the loglik does not converge
within this number of iterations, does it keep going until it does? It seems
like with max iteration specification it runs much faster than without
specification.

Before trying to increase the maximum number of iterations I would
first use a verbose setting to see what is happening in the
optimization.  The "brute force" attack of using the same model
specification and setting the maximum number of iterations to a large
value is not always a good approach.

Thank You

James









James Cho

Analytic Consultant

Ipsos-Synovate-MMA

James.cho at synovate.com

O: 212 293 6668

M: 917 270 6360