Calculate the mean by group

Question

I have a large data frame that looks similar to this:

df <- data.frame(dive = factor(sample(c("dive1","dive2"), 10, replace=TRUE)),
                 speed = runif(10)
                 )
> df
    dive      speed
1  dive1 0.80668490
2  dive1 0.53349584
3  dive2 0.07571784
4  dive2 0.39518628
5  dive1 0.84557955
6  dive1 0.69121443
7  dive1 0.38124950
8  dive2 0.22536126
9  dive1 0.04704750
10 dive2 0.93561651

My goal is to obtain the average of values in one column when another column is equal to a certain value and repeat this for all values. i.e. in the example above I would like to return an average for the column speed for every unique value of the column dive. So when dive==dive1, the average for speed is this and so on for each value of dive.

Answer

There are many ways to do this in R. Specifically, by, aggregate, split, and plyr, cast, tapply, data.table, dplyr, and so forth.

Broadly speaking, these problems are of the form split-apply-combine. Hadley Wickham has written a beautiful article that will give you deeper insight into the whole category of problems, and it is well worth reading. His plyr package implements the strategy for general data structures, and dplyr is a newer implementation performance tuned for data frames. They allow for solving problems of the same form but of even greater complexity than this one. They are well worth learning as a general tool for solving data manipulation problems.

Performance is an issue on very large datasets, and for that it is hard to beat solutions based on data.table. If you only deal with medium-sized datasets or smaller, however, taking the time to learn data.table is likely not worth the effort. dplyr can also be fast, so it is a good choice if you want to speed things up, but don't quite need the scalability of data.table.

Many of the other solutions below do not require any additional packages. Some of them are even fairly fast on medium-large datasets. Their primary disadvantage is either one of metaphor or of flexibility. By metaphor I mean that it is a tool designed for something else being coerced to solve this particular type of problem in a 'clever' way. By flexibility I mean they lack the ability to solve as wide a range of similar problems or to easily produce tidy output.

Examples

base functions

tapply:

tapply(df$speed, df$dive, mean)
#     dive1     dive2 
# 0.5419921 0.5103974

aggregate :

aggregate takes in data.frames, outputs data.frames, and uses a formula interface.

aggregate( speed ~ dive, df, mean )
#    dive     speed
# 1 dive1 0.5790946
# 2 dive2 0.4864489

by :

In its most user-friendly form, it takes in vectors and applies a function to them. However, its output is not in a very manipulable form.:

res.by <- by(df$speed, df$dive, mean)
res.by
# df$dive: dive1
# [1] 0.5790946
# ---------------------------------------
# df$dive: dive2
# [1] 0.4864489

To get around this, for simple uses of by the as.data.frame method in the taRifx library works:

library(taRifx)
as.data.frame(res.by)
#    IDX1     value
# 1 dive1 0.6736807
# 2 dive2 0.4051447

split :

As the name suggests, it performs only the "split" part of the split-apply- combine strategy. To make the rest work, I'll write a small function that uses sapply for apply-combine. sapply automatically simplifies the result as much as possible. In our case, that means a vector rather than a data.frame, since we've got only 1 dimension of results.

splitmean <- function(df) {
  s <- split( df, df$dive)
  sapply( s, function(x) mean(x$speed) )
}
splitmean(df)
#     dive1     dive2 
# 0.5790946 0.4864489 

External packages

data.table :

library(data.table)
setDT(df)[ , .(mean_speed = mean(speed)), by = dive]
#    dive mean_speed
# 1: dive1  0.5419921
# 2: dive2  0.5103974

dplyr :

library(dplyr)
group_by(df, dive) %>% summarize(m = mean(speed))

plyr (the pre-cursor of dplyr)

Here's what the official page has to say about plyr:

It’s already possible to do this with base R functions (like split and the apply family of functions), but plyr makes it all a bit easier with:

• totally consistent names, arguments and outputs
• convenient parallelisation through the foreach package
• input from and output to data.frames, matrices and lists
• progress bars to keep track of long running operations
• built-in error recovery, and informative error messages
• labels that are maintained across all transformations

In other words, if you learn one tool for split-apply-combine manipulation it should be plyr.

library(plyr)
res.plyr <- ddply( df, .(dive), function(x) mean(x$speed) )
res.plyr
#    dive        V1
# 1 dive1 0.5790946
# 2 dive2 0.4864489

reshape2 :

The reshape2 library is not designed with split-apply-combine as its primary focus. Instead, it uses a two-part melt/cast strategy to perform a wide variety of data reshaping tasks. However, since it allows an aggregation function it can be used for this problem. It would not be my first choice for split-apply-combine operations, but its reshaping capabilities are powerful and thus you should learn this package as well.

library(reshape2)
dcast( melt(df), variable ~ dive, mean)
# Using dive as id variables
#   variable     dive1     dive2
# 1    speed 0.5790946 0.4864489

Benchmarks

10 rows, 2 groups

library(microbenchmark)
m1 <- microbenchmark(
  by( df$speed, df$dive, mean),
  aggregate( speed ~ dive, df, mean ),
  splitmean(df),
  ddply( df, .(dive), function(x) mean(x$speed) ),
  dcast( melt(df), variable ~ dive, mean),
  dt[, mean(speed), by = dive],
  summarize( group_by(df, dive), m = mean(speed) ),
  summarize( group_by(dt, dive), m = mean(speed) )
)

> print(m1, signif = 3)
Unit: microseconds
                                           expr  min   lq   mean median   uq  max neval      cld
                    by(df$speed, df$dive, mean)  302  325  343.9    342  362  396   100  b      
              aggregate(speed ~ dive, df, mean)  904  966 1012.1   1020 1060 1130   100     e   
                                  splitmean(df)  191  206  249.9    220  232 1670   100 a       
  ddply(df, .(dive), function(x) mean(x$speed)) 1220 1310 1358.1   1340 1380 2740   100      f  
         dcast(melt(df), variable ~ dive, mean) 2150 2330 2440.7   2430 2490 4010   100        h
                   dt[, mean(speed), by = dive]  599  629  667.1    659  704  771   100   c     
 summarize(group_by(df, dive), m = mean(speed))  663  710  774.6    744  782 2140   100    d    
 summarize(group_by(dt, dive), m = mean(speed)) 1860 1960 2051.0   2020 2090 3430   100       g 

autoplot(m1)

As usual, data.table has a little more overhead so comes in about average for small datasets. These are microseconds, though, so the differences are trivial. Any of the approaches works fine here, and you should choose based on:

• What you're already familiar with or want to be familiar with (plyr is always worth learning for its flexibility; data.table is worth learning if you plan to analyze huge datasets; by and aggregate and split are all base R functions and thus universally available)
• What output it returns (numeric, data.frame, or data.table -- the latter of which inherits from data.frame)

10 million rows, 10 groups

But what if we have a big dataset? Let's try 10^7 rows split over ten groups.

df <- data.frame(dive=factor(sample(letters[1:10],10^7,replace=TRUE)),speed=runif(10^7))
dt <- data.table(df)
setkey(dt,dive)

m2 <- microbenchmark(
  by( df$speed, df$dive, mean),
  aggregate( speed ~ dive, df, mean ),
  splitmean(df),
  ddply( df, .(dive), function(x) mean(x$speed) ),
  dcast( melt(df), variable ~ dive, mean),
  dt[,mean(speed),by=dive],
  times=2
)

> print(m2, signif = 3)
Unit: milliseconds
                                           expr   min    lq    mean median    uq   max neval      cld
                    by(df$speed, df$dive, mean)   720   770   799.1    791   816   958   100    d    
              aggregate(speed ~ dive, df, mean) 10900 11000 11027.0  11000 11100 11300   100        h
                                  splitmean(df)   974  1040  1074.1   1060  1100  1280   100     e   
  ddply(df, .(dive), function(x) mean(x$speed))  1050  1080  1110.4   1100  1130  1260   100      f  
         dcast(melt(df), variable ~ dive, mean)  2360  2450  2492.8   2490  2520  2620   100       g 
                   dt[, mean(speed), by = dive]   119   120   126.2    120   122   212   100 a       
 summarize(group_by(df, dive), m = mean(speed))   517   521   531.0    522   532   620   100   c     
 summarize(group_by(dt, dive), m = mean(speed))   154   155   174.0    156   189   321   100  b      

autoplot(m2)

Then data.table or dplyr using operating on data.tables is clearly the way to go. Certain approaches (aggregate and dcast) are beginning to look very slow.

10 million rows, 1,000 groups

If you have more groups, the difference becomes more pronounced. With 1,000 groups and the same 10^7 rows:

df <- data.frame(dive=factor(sample(seq(1000),10^7,replace=TRUE)),speed=runif(10^7))
dt <- data.table(df)
setkey(dt,dive)

# then run the same microbenchmark as above
print(m3, signif = 3)
Unit: milliseconds
                                           expr   min    lq    mean median    uq   max neval    cld
                    by(df$speed, df$dive, mean)   776   791   816.2    810   828   925   100  b    
              aggregate(speed ~ dive, df, mean) 11200 11400 11460.2  11400 11500 12000   100      f
                                  splitmean(df)  5940  6450  7562.4   7470  8370 11200   100     e 
  ddply(df, .(dive), function(x) mean(x$speed))  1220  1250  1279.1   1280  1300  1440   100   c   
         dcast(melt(df), variable ~ dive, mean)  2110  2190  2267.8   2250  2290  2750   100    d  
                   dt[, mean(speed), by = dive]   110   111   113.5    111   113   143   100 a     
 summarize(group_by(df, dive), m = mean(speed))   625   630   637.1    633   644   701   100  b    
 summarize(group_by(dt, dive), m = mean(speed))   129   130   137.3    131   142   213   100 a     

autoplot(m3)

So data.table continues scaling well, and dplyr operating on a data.table also works well, with dplyr on data.frame close to an order of magnitude slower. The split/sapply strategy seems to scale poorly in the number of groups (meaning the split() is likely slow and the sapply is fast). by continues to be relatively efficient--at 5 seconds, it's definitely noticeable to the user but for a dataset this large still not unreasonable. Still, if you're routinely working with datasets of this size, data.table is clearly the way to go - 100% data.table for the best performance or dplyr with dplyr using data.table as a viable alternative.