# [R] Resampling to find Confidence intervals

Ben Ward benjamin.ward at bathspa.org
Tue Jan 4 18:56:35 CET 2011

```Ok I'll check I understand:
So it's using sample, to resample d once, 10 values, because the rnorm
has 10 values, with replacement (I assume thats the TRUE part).
Then a for loop has this to resample the data - in the loop's case its
1000 times. Then it does a lm to get the coefficients and add them to
d1.coef. I'm guessing that the allboot bit with rbind, which is null at
the start of the loop, is the collection of d1.coef values, as I think
that without it, every cycle of the loop the d1.coef from the previous
cycle round the loop would be gone?

On 04/01/2011 16:24, Dieter Menne wrote:

Axolotl9250 wrote:

>> ...
>> resampled_ecoli = sample(ecoli, 500, replace=T)
>> coefs = (coef(lm(MIC. ~ 1 + Challenge + Cleaner + Replicate,
>> data=resampled_ecoli)))
>> sd(coefs)
>>
>> ...
>>
> Below a simplified and self-consistent version of your code, and some
> changes
>
> Dieter
>
> # resample
> d = data.frame(x=rnorm(10))
> d\$y = d\$x*3+rnorm(10,0.01)
>
> # if you do this, you only get ONE bootstrap sample
> d1 = d[sample(1:nrow(d),10,TRUE),]
> d1.coef = coef(lm(y~x,data=d1))
> d1.coef
> # No error below, because you compute the sd of (Intercept) and slope
> # but result is wrong!
> sd(d1.coef)
>
> # We have to do this over and over
> # Check ?replicate for a more R-ish approach....
> nsamples = 1000
> allboot = NULL
> for (i in 1:1000) {
>    d1 = d[sample(1:nrow(d),10,TRUE),]
>    d1.coef = coef(lm(y~x,data=d1))
>    allboot = rbind(allboot,d1.coef) # Not very efficient, preallocate!
> }
> head(allboot) # display first of nsamples lines
> apply(allboot,2,mean) # Compute mean
> apply(allboot,2,sd) # compute sd
> # After you are sure you understood the above, you might try package boot.
>
>
>
>
>

```