Skip to content
Prev 5 / 20628 Next

lmer2 slower?

I am forwarding (with permission) a conversation with Bill Auty who
discovered that in one of his examples lmer2 is considerably slower
than lmer, although they get to the same estimates.  This can happen.

As I mentioned in one reply I hope that forcing a supernodal Cholesky
decomposition for models with crossed or partially crossed grouping
factors for the random effects will help but I haven't had time to
explore this yet.

I'll post some other timing results in another message.

---------- Forwarded message ----------
From: Douglas Bates <bates at stat.wisc.edu>
Date: Jan 26, 2007 3:38 PM
Subject: Re: lmer2 slower?
To: Bill Auty <bill at edmeasure.com>
On 1/26/07, Bill Auty <bill at edmeasure.com> wrote:
Ah, good.  That 's the answer that I wanted.

Briefly, the CHOLMOD sparse matrix code allows for two types of
Cholesky decompositions of sparse, symmetric matrices.  These are
called "simplicial" and "supernodal".  In the lmer code I forced the
supernodal decomposition.  In the lmer2 code I allow the sparse matrix
manipulation code to choose one or the other based on some criteria
which I can "tune".  I think that the default values that are
currently being used are too conservative about switching to the
supernodal decomposition so I may be able to gain back some of the
loss of speed.

Thanks again for checking this.

May I send a copy of this reply to the R-SIG-mixed-models list?
Message-ID: <40e66e0b0701270950q566d9894k69263fc4c82c9b45@mail.gmail.com>
In-Reply-To: <40e66e0b0701261338r3898b2a6ubd0af8662bcad96e@mail.gmail.com>