[R] bootstrapping respecting subject level information

[Andrew Robinson]

I'd like to preface this answer by suggesting that if you have multiple
measurements within subjects then you should possibly be thinking about
using mixed-effects models.  Here you have a balanced design and seem
to be thinking about a constrained bootstrap, but I don't know whether
the resulting distribution will have the right properties - others on
the list may.

That said, I think that this needs the 'strata' argument of the boot function.
See the help.  Something like

data.boot <- boot(data, boot.huber, 1999, strata = Subject)

(not tested)

Cheers

Andrew

On Thu, Jul 4, 2013 at 12:19 PM, Sol Lago <solcita.lago at gmail.com> wrote:
> Hi there,
>
> This is the first time I use this forum, and I want to say from the start I am not a skilled programmer. So please let me know if the question or code were unclear!
>
> I am trying to bootstrap an interaction (that is my test statistic) using the package "boot". My problem is that for every resample, I would like the randomization to be done within subjects, so that observations from different subjects are not mixed. Here is the code to generate a dataframe similar to mine:
>
> Subject = rep(c("S1","S2","S3","S4"),4)
> Num     = rep(c("singular","plural"),8)
> Gram    = rep(c("gram","gram","ungram","ungram"),4)
> RT      = c(657,775,678,895,887,235,645,916,930,768,890,1016,590,978,450,920)
> data    = data.frame(Subject,Num,Gram,RT)
>
> This is the code I used to get the empirical interaction value:
>
> summary(lm(RT ~ Num*Gram, data=data))
>
> As you can see, the interaction between my two factors is -348. I want to get a bootstrap confidence interval for this statistic, which I can generate using the "boot" package:
>
> #Function to create the statistic to be boostrapped
> boot.huber <- function(data, indices) {
> data <- data[indices, ] #select obs. in bootstrap sample
> mod <- lm(RT ~ Num*Gram, data=data)
> coefficients(mod)       #return coefficient vector
> }
>
> #Generate bootstrap estimate
> data.boot <- boot(data, boot.huber, 1999)
>
> #Get confidence interval
> boot.ci(data.boot, index=4, type=c("norm", "perc", "bca"),conf=0.95) #4 gets the CI for the interaction
>
> My problem is that I think the resamples should be generated without mixing the individual subjects observations: that is, to generate the new resamples, the observations from subject 1 (S1) should be shuffled within subject 1, not mixing them with the observations from subjects 2, etc... I don't know how "boot" is doing the resampling (I read the documentation but don't understand how the function is doing it)
>
> Does anyone know how I could make sure that the resampling procedure used by "boot" respects the subject level information?
>
> Thanks a lot for your help/advice!
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.



-- 
Andrew Robinson
Deputy Director, CEBRA
Senior Lecturer in Applied Statistics                      Tel: +61-3-8344-6410
Department of Mathematics and Statistics            Fax: +61-3-8344 4599
University of Melbourne, VIC 3010 Australia
Email: a.robinson at ms.unimelb.edu.au    Website: http://www.ms.unimelb.edu.au

FAwR: http://www.ms.unimelb.edu.au/~andrewpr/FAwR/
SPuR: http://www.ms.unimelb.edu.au/spuRs/