How can I use functions returning vectors (like fivenum) with ddply or aggregate?

Question

I would like to split my data frame using a couple of columns and call let\'s say fivenum on each group.

aggregate(Petal.Width ~ Species, iris,          


        

Answer 1

Here is a solution using data.table (while not specifically requested, it is an obvious compliment or replacement for aggregate or ddply. As well as being slightly long to code, repeatedly calling quantile will be inefficient, as for each call you will be sorting the data

library(data.table)
Tukeys_five <- c("Min","Q1","Med","Q3","Max") 

IRIS <- data.table(iris)
# this will create the wide data.table
lengthBySpecies <- IRIS[,as.list(fivenum(Sepal.Length)), by = Species]

# and you can rename the columns from V1, ..., V5 to something nicer

setnames(lengthBySpecies, paste0('V',1:5), Tukeys_five)


lengthBySpecies



      Species Min  Q1 Med  Q3 Max
1:     setosa 4.3 4.8 5.0 5.2 5.8
2: versicolor 4.9 5.6 5.9 6.3 7.0
3:  virginica 4.9 6.2 6.5 6.9 7.9

Or, using a single call to quantile using the appropriate prob argument.

IRIS[,as.list(quantile(Sepal.Length, prob = seq(0,1, by = 0.25))), by = Species]


       Species  0%   25% 50% 75% 100%
1:     setosa 4.3 4.800 5.0 5.2  5.8
2: versicolor 4.9 5.600 5.9 6.3  7.0
3:  virginica 4.9 6.225 6.5 6.9  7.9  

Note that the names of the created columns are not syntactically valid, although you could go through a similar renaming using setnames

---

EDIT

Interestingly, quantile will set the names of the resulting vector if you set names = TRUE, and this will copy (slow down the number crunching and consume memory - it even warns you in the help, fancy that!)

Thus, you should probably use

 IRIS[,as.list(quantile(Sepal.Length, prob = seq(0,1, by = 0.25), names = FALSE)), by = Species]

Or, if you wanted to return the named list, without R copying internally

IRIS[,{quant <- as.list(quantile(Sepal.Length, prob = seq(0,1, by = 0.25), names = FALSE))
       setattr(quant, 'names', Tukeys_five)
       quant}, by = Species]

Answer 2

You can use do.call to call data.frame on each of the matrix elements recursively to get a data.frame with vector elements:

dim(do.call("data.frame",dfr))
[1] 3 7

str(do.call("data.frame",dfr))
'data.frame':   3 obs. of  7 variables:
 $ Species            : Factor w/ 3 levels "setosa","versicolor",..: 1 2 3
 $ Petal.Width.Min.   : num  0.1 1 1.4
 $ Petal.Width.1st.Qu.: num  0.2 1.2 1.8
 $ Petal.Width.Median : num  0.2 1.3 2
 $ Petal.Width.Mean   : num  0.28 1.36 2
 $ Petal.Width.3rd.Qu.: num  0.3 1.5 2.3
 $ Petal.Width.Max.   : num  0.6 1.8 2.5

Answer 3

This is my solution:

ddply(iris, .(Species), summarize, value=t(fivenum(Petal.Width)))

Answer 4

As far as I know, there isn't an exact way to do what you're asking, because the function you're using (fivenum) doesn't return data in a way that can be easily bound to columns from within the 'ddply' function. This is easy to clean up, though, in a programmatic way.

Step 1: Perform the fivenum function on each 'Species' value using the 'ddply' function.

data <- ddply(iris, .(Species), summarize, value=fivenum(Petal.Width))

#       Species value
# 1      setosa   0.1
# 2      setosa   0.2
# 3      setosa   0.2
# 4      setosa   0.3
# 5      setosa   0.6
# 6  versicolor   1.0
# 7  versicolor   1.2
# 8  versicolor   1.3
# 9  versicolor   1.5
# 10 versicolor   1.8
# 11  virginica   1.4
# 12  virginica   1.8
# 13  virginica   2.0
# 14  virginica   2.3
# 15  virginica   2.5

Now, the 'fivenum' function returns a list, so we end up with 5 line entries for each species. That's the part where the 'fivenum' function is fighting us.

Step 2: Add a label column. We know what Tukey's five numbers are, so we just call them out in the order that the 'fivenum' function returns them. The list will repeat until it hits the end of the data.

Tukeys_five <- c("Min","Q1","Med","Q3","Max") 
data$label <- Tukeys_five

#       Species value label
# 1      setosa   0.1   Min
# 2      setosa   0.2    Q1
# 3      setosa   0.2   Med
# 4      setosa   0.3    Q3
# 5      setosa   0.6   Max
# 6  versicolor   1.0   Min
# 7  versicolor   1.2    Q1
# 8  versicolor   1.3   Med
# 9  versicolor   1.5    Q3
# 10 versicolor   1.8   Max
# 11  virginica   1.4   Min
# 12  virginica   1.8    Q1
# 13  virginica   2.0   Med
# 14  virginica   2.3    Q3
# 15  virginica   2.5   Max

Step 3: With the labels in place, we can quickly cast this data into a new shape using the 'dcast' function from the 'reshape2' package.

library(reshape2)
dcast(data, Species ~ label)[,c("Species",Tukeys_five)]

#      Species Min  Q1 Med  Q3 Max
# 1     setosa 0.1 0.2 0.2 0.3 0.6
# 2 versicolor 1.0 1.2 1.3 1.5 1.8
# 3  virginica 1.4 1.8 2.0 2.3 2.5

All that junk at the end are just specifying the column order, since the 'dcast' function automatically puts things in alphabetical order.

Hope this helps.

Update: I decided to return, because I realized there is one other option available to you. You can always bind a matrix as part of a data frame definition, so you could resolve your 'aggregate' function like so:

data <- aggregate(Petal.Width ~ Species, iris, function(x) summary(fivenum(x))) 
result <- data.frame(Species=data[,1],data[,2])

#      Species Min. X1st.Qu. Median Mean X3rd.Qu. Max.
# 1     setosa  0.1      0.2    0.2 0.28      0.3  0.6
# 2 versicolor  1.0      1.2    1.3 1.36      1.5  1.8
# 3  virginica  1.4      1.8    2.0 2.00      2.3  2.5