[Rd] Statistical mode
One descriptive statistic that is conspicuously missing from core R is the statistical mode - the most frequent value in a discrete distribution. I would like to propose adding the attached 'statmode' (or a similar function) to the 'stats' package. Currently, it can be quite cumbersome to calculate the mode of a distribution in R, both for experts and beginners. The lack of a function to do this is felt, both when teaching introductory R courses, and when using sapply() or the like. Looking forward to your feedback, Arnistatmode <- function(x, all=FALSE, ...) { if(is.list(x)) { output <- sapply(x, statmode, all=all, ...) } else { freq <- table(x, ...) if(all) output <- names(freq)[freq==max(freq)] else output <- names(freq)[which.max(freq)] ## Coerce to original data type, using any() to handle mts, xtabs, etc. if(any(class(x) %in% c("integer","numeric","ts","complex","matrix","table"))) output <- as(output, storage.mode(x)) } return(output) } \name{statmode} \alias{statmode} \title{Statistical Mode} \description{ Compute the statistical mode, the most frequent value in a discrete distribution. } \usage{ statmode(x, all = FALSE, \dots) } \arguments{ \item{x}{an \R object, usually vector, matrix, or data frame.} \item{all}{whether all statistical modes should be returned.} \item{\dots}{further arguments passed to the \code{\link{table}} function.} } \details{The default is to return only the first statistical mode.} \value{ The most frequent value in \code{x}, possibly a vector or list, depending on the class of \code{x} and whether \code{all=TRUE}. } \seealso{ \code{\link{mean}}, \code{\link{median}}, \code{\link{table}}. \code{\link{density}} can be used to compute the statistical mode of a continuous distribution. } \examples{ ## Different location statistics fw <- faithful$waiting hist(fw) barplot(table(fw)) mean(fw) median(fw) statmode(fw) plot(density(fw)) with(density(fw), x[which.max(y)]) ## Different classes statmode(chickwts$feed) # factor statmode(volcano)# matrix statmode(discoveries)# ts statmode(mtcars) # data frame ## Multiple modes table(mtcars$carb) statmode(mtcars$carb) statmode(mtcars$carb, TRUE) statmode(mtcars, TRUE) } \keyword{univar} __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
Re: [Rd] Statistical mode
Arni, Here are two examples: R> statmode(iris) Sepal.Length Sepal.Width Petal.Length Petal.Width Species "5" "3""1.4""0.2" "setosa" R> table(iris$Species) setosa versicolor virginica 50 50 50 R> library(lattice) R> statmode(barley) yieldvariety year site "20.6" "Svansota" "1932" "Grand Rapids" My thoughts: 1. The mode is not so interesting for continuous data. I would much rather use something like density(). 2. Both the iris and barley data sets are balanced (each factor level appears equally often), and the current output from the statmode function is misleading by only showing one level. 3. I think the describe() function in the Hmisc package is much more useful and informative, even for introductory stat classes. I always use describe() after importing data into R. Kevin On Thu, May 26, 2011 at 3:26 PM, Arni Magnusson wrote: > One descriptive statistic that is conspicuously missing from core R is the > statistical mode - the most frequent value in a discrete distribution. > > I would like to propose adding the attached 'statmode' (or a similar > function) to the 'stats' package. > > Currently, it can be quite cumbersome to calculate the mode of a > distribution in R, both for experts and beginners. The lack of a function to > do this is felt, both when teaching introductory R courses, and when using > sapply() or the like. > > Looking forward to your feedback, > > Arni > __ > R-devel@r-project.org mailing list > https://stat.ethz.ch/mailman/listinfo/r-devel > > [[alternative HTML version deleted]] __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
Re: [Rd] Statistical mode
Thank you, Kevin, for the feedback. 1. The mode is not so interesting for continuous data. I would much rather use something like density(). Absolutely. The help page for statmode() says it is for discrete data, and points to density() for continuous data. 2. Both the iris and barley data sets are balanced (each factor level appears equally often), and the current output from the statmode function is misleading by only showing one level. Try statmode(iris,TRUE). It points out that petal lengths 1.4 and 1.5 are equally common in the data. I decided to make all=FALSE the default behavior, but I'd be equally happy with all=TRUE as the default. As for the barley data, statmode(barley,TRUE) is just the honest answer. The yield is continuous, so the discrete mode is not of interest, and the factors levels are all equally common as you point out. 3. I think the describe() function in the Hmisc package is much more useful and informative, even for introductory stat classes. I always use describe() after importing data into R. The describe() function is a verbose summary, usually of a data frame. The statmode() function is the discrete mode, usually of a vector. Importantly, describe(faithful$waiting) points out the mean, median and range, but not the mode. --- Allow me to include two more valid comments, from Sarah Goslee and David Winsemius, respectively: 4. The 'modeest' package does this and more, see for example mfv(). I think core R should come with a basic function to get the mode of a discrete vector. One option would be to lift mfv() into the 'stats' package, but something like statmode() could also cover factors and strings. Might as well provide all=TRUE/FALSE functionality, too, and retain integers as integers. It's common to find rudimentary basic functionality in the 'stats' package, and dedicated packages for more details; time series models and robust statistics come to mind. The 'modeest' package is impressive indeed. 5. Isn't this just table(Vec)[which.max(table(Vec))]? Yes it is, only less cumbersome. Much like sd(Vec) is less cumbersome than sqrt(var(Vec)). Moreover, I find it confusing to see the count as well, table(volcano)[which.max(table(volcano))] # 110 # 177 although this can be debated. Finally, I think the examples statmode(mtcars) statmode(mtcars, TRUE) demonstrate practical functionality beyond table(Vec)[which.max(table(Vec))]. The mean, median, and mode are often mentioned together as fundamental descriptive statistics, and I just find it odd that statmode() is not already in core R. Sure, we could get by without the sd() function in core R, but why should we? All the best, Arni __ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel