[R-sig-ME] Timings with SAMM, lme4, nlme
HDoran at air.org
Thu Feb 1 17:24:54 CET 2007
Try adding this portion of code to your lmer model
control=list(gradient = FALSE, niterEM = 0)
This will change what you have below from
lmer(math~ gr + sx + eth + cltype + (1+yrs|id) + (1+yrs|sch), data=star)
lmer(math~ gr + sx + eth + cltype + (1+yrs|id) + (1+yrs|sch), data=star,
control=list(gradient = FALSE, niterEM = 0))
BTW, you don't need (1+yrs) you can reduce this to just (yrs|id).
This should make a huge difference.
> -----Original Message-----
> From: r-sig-mixed-models-bounces at r-project.org
> [mailto:r-sig-mixed-models-bounces at r-project.org] On Behalf
> Of Kevin Wright
> Sent: Thursday, February 01, 2007 10:25 AM
> To: r-sig-mixed-models at r-project.org
> Subject: [R-sig-ME] Timings with SAMM, lme4, nlme
> About a year and a half ago I did some comparison of model
> fitting with SAMM, lme4, and nlme. Since Doug Bates put out
> a request for some recent timings, I am repeating/extending
> my comparisons. In the interim, I have switched computers,
> so the timings of the samm function facilitate comparing the
> speed across the two computers.
> First, the results (view with fixed-width font)
> Setup T30 D620 T30 D620 D620 D620
> Model \ Function samm samm lmer lmer lmer2 lme
> ---------------- ---- ---- ---- ---- ----- -----
> yrs|id + yrs|sch f f 30.0 kill 65
> yrs|id + 1|sch 10.5 7.8 13.7 kill 23
> 1|id + yrs|sch 9.4 9.5 11.5 kill 7.3
> 1|id + 1|sch 6.5 5.7 5.6 kill 5.4 60
> yrs|sch 4.8 4.0 f kill 0.5
> 1|sch 3.4 2.2 0.4 0.7 0.3 4
> yrs|id 5.7 5.7 8.2 kill 15 150
> 1|id 4.1 3.4 2.4 kill 4.2 14
> In the table above, "Setup" refers to the following two computer
> T30: IBM Thinkpad T30, 1 GB ram, 1.8 Ghz processor, Windows 2000
> SAMM version 1.1, lme4 & nlme current on 5.24.2005
> D620: Dell Latitude D620, 2 GB ram, 1.8 Ghz duo core processor, WinXP
> SAMM version 1.1 lme4 & nlme current as of 1.30.2007
> I timed most model fits only once. I did a quick inspection
> of the results from the different modelling functions to
> persuade myself that I was fitting the same model (i.e. that
> estimates were similar).
> Here are the full models I used:
> lmer(math~ gr + sx + eth + cltype + (1+yrs|id) + (1+yrs|sch),
> samm(math ~ gr + sx + eth + cltype, random=~ us(link(~1+yrs)):id +
> us(link(~1+yrs)):sch, data=star, na.method.X="omit")
> Starting with these full models, I tried reduced models with
> simpler random effects structures. These are the different
> rows in the table above.
> (1) The performance of the common version of SAMM on the two
> computers suggests the Dell is slightly faster than the IBM.
> (2) On 24 May 2005, lmer and samm has roughly similar
> timings. On 31 Jan 2007, lmer is nearly unusable for this
> data (I killed the job after 5-10 minutes of waiting).
> (3) The current version of lmer2 is the only function that
> appears to fit all models.
> (4) The current version of lme is slower than sammm/lmer2 for
> those models I tried to fit.
> FYI. SAMM is officially available at
> http://www.vsni.co.uk/products/samm/(but seems not to be
> there now, perhaps in preparation for release of a new
> version). Unofficially it is available here: ftp://ftp.dpi.nsw.gov.au/
> I hope this information is useful. Thanks for the progress
> evident in the
> lme4 package.
> Kevin Wright
> [[alternative HTML version deleted]]
> R-sig-mixed-models at r-project.org mailing list
More information about the R-sig-mixed-models