[R] what does cut(data, breaks=n) actually do?
Tony Plate
tplate at acm.org
Sat Dec 15 06:45:46 CET 2007
Peter Dalgaard wrote:
> melissa cline wrote:
>> Hello,
>>
>> I'm trying to bin a quantity into 2-3 bins for calculating entropy and
>> mutual information. One of the approaches I'm exploring is the cut()
>> function, which is what the mutualInfo function in binDist uses. When it's
>> called in the format cut(data, breaks=n), it somehow splits the data into n
>> distinct bins. Can anyone tell me how cut() decides where to cut?
>>
>>
> This is one case where reading the actual R code is easier that
> explaining what it does. From cut.default
>
> if (length(breaks) == 1) {
> if (is.na(breaks) | breaks < 2)
> stop("invalid number of intervals")
> nb <- as.integer(breaks + 1)
> dx <- diff(rx <- range(x, na.rm = TRUE))
> if (dx == 0)
> dx <- rx[1]
> breaks <- seq.int(rx[1] - dx/1000, rx[2] + dx/1000, length.out = nb)
> }
>
> so basically it takes the range, extends it a bit and splits in into
> <breaks> equally long segments.
>
> (For the sometimes more attractive option of splitting into groups of
> roughly equal size, there is cut2 in the Hmisc package, or use quantile())
>
It can be a bit dangerous to use quantile() to provide breaks for cut(),
because quantiles can be non-unique, which cut() doesn't like:
> x1 <- c(1,1,1,1,1,1,1,1,1,2)
> cut(x1, breaks=quantile(x1, (0:2)/2))
Error in cut.default(x1, breaks = quantile(x1, (0:2)/2)) :
'breaks' are not unique
>
However, cut2() in Hmisc handles this situation gracefully:
> library(Hmisc)
Attaching package: 'Hmisc'
The following object(s) are masked from package:base :
format.pval,
round.POSIXt,
trunc.POSIXt,
units
> cut2(x1, g=2)
[1] 1 1 1 1 1 1 1 1 1 2
Levels: 1 2
>
(Additionally, a potentially dangerous peculiarity of quantile() for
this kind of purpose is that its return values can be out of order
(i.e., diff(quantile(...))<0, at rounding error level), but this doesn't
actually upset cut() in R because cut() sorts the breaks prior to using
them.)
-- Tony Plate
More information about the R-help
mailing list