[R-sig-ME] Does lmer have it's own random number generator?

[Douglas Bates]

On Mon, Jan 3, 2011 at 4:01 AM, ONKELINX, Thierry
<Thierry.ONKELINX at inbo.be> wrote:
> Dear Mike,
>
> Are you sure? Below the results of 100 simulations. I get the same
> results each time.
>
> Best regards,
>
> Thierry
>
>> set.seed(1)
>> tmp <- replicate(100, {
> +       fit <- lmer(
> +               formula = rt ~ (1|subnum) +
> +                       cue*flank*group*block +
> +                       cue*flank*group*I(block^2) +
> +                       cue*flank*group*I(block^3)
> +               , data = ANT[ANT$error==0,]
> +       )
> +       fit at deviance
> + })
>> apply(tmp, 1, var)
>       ML      REML      ldL2     ldRX2   sigmaML sigmaREML     pwrss
> disc
>        0         0         0         0         0         0         0
> 0
>     usqr      wrss       dev      llik   NULLdev
>        0         0        NA        NA        NA
>> sessionInfo()
> R version 2.12.0 (2010-10-15)
> Platform: i386-pc-mingw32/i386 (32-bit)
>
> locale:
> [1] LC_COLLATE=Dutch_Belgium.1252  LC_CTYPE=Dutch_Belgium.1252
> [3] LC_MONETARY=Dutch_Belgium.1252 LC_NUMERIC=C
> [5] LC_TIME=Dutch_Belgium.1252
>
> attached base packages:
>  [1] grid      grDevices datasets  splines   graphics  stats     tcltk
>
>  [8] utils     methods   base
>
> other attached packages:
>  [1] ez_2.1-0           lme4_0.999375-37   Matrix_0.999375-45
> lattice_0.19-15
>  [5] stringr_0.4        ggplot2_0.8.8      proto_0.3-8
> reshape_0.8.3
>  [9] plyr_1.2.1         car_2.0-7          nnet_7.3-1         MASS_7.3-7
>
> [13] rj_0.5.0-5         svSocket_0.9-51    TinnR_1.0.3        R2HTML_2.2
>
> [17] Hmisc_3.8-3        survival_2.36-1
>
> loaded via a namespace (and not attached):
> [1] cluster_1.13.2 nlme_3.1-97    rJava_0.8-8    stats4_2.12.0
> svMisc_0.9-61
> [6] tools_2.12.0
>
>
>> -----Oorspronkelijk bericht-----
>> Van: r-sig-mixed-models-bounces at r-project.org
>> [mailto:r-sig-mixed-models-bounces at r-project.org] Namens Mike Lawrence
>> Verzonden: maandag 3 januari 2011 3:52
>> Aan: r-sig-mixed-models at r-project.org
>> Onderwerp: [R-sig-ME] Does lmer have it's own random number generator?
>>
>> Hi folks,
>>
>> Does lmer have it's own internally seeded random number
>> generator? I just came across an example where I get
>> different results across repeated identically specified lmer
>> fits even when I use set.seed():
>>
>> library(ez)
>> data(ANT)
>>
>> #repeat the below a few times, the result # on the last line
>> should vary occasionally, # sometimes TRUE, sometimes FALSE,
>> # and sometimes there will be a warning # after the fit
>>
>> set.seed(1)
>> fit = lmer(
>>       formula = rt ~ (1|subnum) +
>>               cue*flank*group*block +
>>               cue*flank*group*I(block^2) +
>>               cue*flank*group*I(block^3)
>>       , data = ANT[ANT$error==0,]
>> )
>> any(is.finite(attr(fit,'deviance')))

A call to lmer does not, as far as I know, result in any calls to
random number generators.  There is only one point in the lme4 package
where set.seed is called and that is in the simulate method.