Data preparation
A very large part of research work can consist of data gathering,
cleaning, combining and formatting. You may spend much more time on data
preparation than on analysis, modeling or visualization. R provides some
tools to make this process easier; but doing this well also requires
that you carefully consider your options. See Wickham
(2014) for a
discussion. Two very important functions in this context are reshape
and merge. Sorting of data can also be helpful.
reshape
reshape allows you to rearrange data from a wide to a long form
and vice versa. This can be a very important step to simplify data
analysis. In the wide form, each variable is a column and each case
(e.g. individual) is a singe row. This is the common ‘spreadsheet’
approach. In the long form there is column that indicates the variable
name and a column that has its value. Other columns identify the cases,
and these can be repeated many times. The long form can be much easier
for use in data analysis than the wide form. In many cases you may want
to go from one form to the other and back.
The function to go between wide and long form is called reshape.
Unfortunately reshape is rather difficult to use. There is only one
function to go from wide to long and vice versa, creating confusion
about what arguments to use when. It is also poorly documented, and its
error messages are bewildering. (There is a package called reshape
that you can use instead. But in this Introduction I want to stick with
base R idiom.)
There is a good example of how to use reshape here. And I will re-use that and expand on it.
wide to long
Get some example data on student’s performance in different subjects. Note how you can read a text on a web server (http address).
dd <- read.csv('https://stats.idre.ucla.edu/stat/data/hsb2.csv')
d <- dd[1:3, c(1:2, 8:9)]
d
## id female write math
## 1 70 0 52 41
## 2 121 1 59 53
## 3 86 0 33 54
This is a “wide” form dataset. To go to a long form, you need to provide
an argument varying that indicates the variables that are combined
into one new variable. v.names is the name of this new variable.
wvars <- c("write", "math")
x <- reshape(d, varying=wvars, v.names="score", direction = "long")
x
## id female time score
## 70.1 70 0 1 52
## 121.1 121 1 1 59
## 86.1 86 0 1 33
## 70.2 70 0 2 41
## 121.2 121 1 2 53
## 86.2 86 0 2 54
x has two new variables “time”, and “score”. As you can see, “score”
has the values for “write” and “math” in the d. The “time” variable
identifies which is which. “write” is identified with a 1, and “math” is
identified with a 2. To have clear labels rather than such numbers, you
can use the times argument; and add the timevar argument to
rename “time” to something more meaningful. Note that the argument names
are based on the idea that the data refer to different measurements over
time. But this is not always the case.
x <- reshape(d, varying=wvars, v.names="score", times=wvars, timevar = "subject", direction = "long")
x
## id female subject score
## 70.write 70 0 write 52
## 121.write 121 1 write 59
## 86.write 86 0 write 33
## 70.math 70 0 math 41
## 121.math 121 1 math 53
## 86.math 86 0 math 54
The row names also identify how the records were created. To get rid of the row names do
rownames(x) <- NULL
x
## id female subject score
## 1 70 0 write 52
## 2 121 1 write 59
## 3 86 0 write 33
## 4 70 0 math 41
## 5 121 1 math 53
## 6 86 0 math 54
Variables “id” and “female” were unchanging, but they were duplicated because two variables (“write” and “math”) were combined into one (“subject”)
long to wide
To go from long to wide we need to use “idvar” and “timevar”. “idvar” identifies the variables that identify a single case (e.g. a single person, or other observational unit). In contrast, “timevar” indicates the variable that has the identifiers that become variables in the wide format. The remaining variable should have the values that match these new variables.
w <- reshape(x, idvar=c("id", "female"), timevar = "subject", direction = "wide")
w
## id female score.write score.math
## 1 70 0 52 41
## 2 121 1 59 53
## 3 86 0 33 54
Note that w is identical to d, except for the last two column
names that now have “score.” prepended to them. We can change that like
this:
cn <- colnames(w)
cn <- gsub("score.", "", cn)
colnames(w) <- cn
w
## id female write math
## 1 70 0 52 41
## 2 121 1 59 53
## 3 86 0 33 54
merge
A common situation is to have multiple data.frames with data for the
same cases (e.g., individuals, fields, countries, …). Such data.frames
may need to be joined such that they can be analyzed. For example here
we have a and b:
a <- dd[, 1:3]
# random sample of 100 records)
set.seed(1)
b <- dd[sample(nrow(dd), 100), c(1, 7:10)]
a has 200 records. It has a unique identifier for each student and
information about their sex (female or not) and race (4 groups).
dim(a)
## [1] 200 3
head(a)
## id female race
## 1 70 0 4
## 2 121 1 4
## 3 86 0 4
## 4 141 0 4
## 5 172 0 4
## 6 113 0 4
table(a$female)
##
## 0 1
## 91 109
table(a$race)
##
## 1 2 3 4
## 24 11 20 145
b has the same unique identifier as a (but only for 100
students) and it has the grades for four subjects.
dim(b)
## [1] 100 5
head(b)
## id read write math science
## 68 107 47 39 47 42
## 167 156 50 59 53 61
## 129 91 50 49 56 47
## 162 74 57 50 50 51
## 43 168 52 54 57 55
## 14 104 54 63 57 55
Imagine we were interested in differences in reading by sex or race. We
would need to combine a and b. That, fortunately, is very
simple, because we have the merge function.
ab <- merge(a, b, by='id')
head(ab)
## id female race read write math science
## 1 3 0 1 63 65 48 63
## 2 6 1 1 47 41 46 40
## 3 8 1 1 39 44 52 44
## 4 9 0 1 48 49 52 44
## 5 12 0 1 37 44 45 39
## 6 13 1 1 47 46 39 47
Always check the dimensions of the result
dim(ab)
## [1] 100 7
In this case we expected 100 records (the lower number of the two; a
had 200 records, but b only 100). Sometimes you get fewer than
expected, suggesting that the identifiers do not match. In other cases
you might want to keep all records, and create missing values where
these are not available. You can do that like this:
ab <- merge(a, b, by='id', all.x=TRUE)
dim(ab)
## [1] 200 7
head(ab)
## id female race read write math science
## 1 1 1 1 NA NA NA NA
## 2 2 1 1 NA NA NA NA
## 3 3 0 1 63 65 48 63
## 4 4 1 1 NA NA NA NA
## 5 5 0 1 NA NA NA NA
## 6 6 1 1 47 41 46 40
Note that the “x” in all.x refers to the first argument, hence a
in this case. In other cases you might need to say all.y=TRUE or
all=TRUE. Consider these extreme cases (with no matching records):
merge(a[1:3,], b[1:3, ], by='id')
## [1] id female race read write math science
## <0 rows> (or 0-length row.names)
merge(a[1:3,], b[1:3, ], by='id', all.x=T)
## id female race read write math science
## 1 70 0 4 NA NA NA NA
## 2 86 0 4 NA NA NA NA
## 3 121 1 4 NA NA NA NA
merge(a[1:3,], b[1:3, ], by='id', all.y=T)
## id female race read write math science
## 1 91 NA NA 50 49 56 47
## 2 107 NA NA 47 39 47 42
## 3 156 NA NA 50 59 53 61
merge(a[1:3,], b[1:3, ], by='id', all=T)
## id female race read write math science
## 1 70 0 4 NA NA NA NA
## 2 86 0 4 NA NA NA NA
## 3 91 NA NA 50 49 56 47
## 4 107 NA NA 47 39 47 42
## 5 121 1 4 NA NA NA NA
## 6 156 NA NA 50 59 53 61
Now that we have ab we can compute what we needed:
tapply(ab$read, ab$female, mean, na.rm=TRUE)
## 0 1
## 53.43902 52.20339
tapply(ab$read, ab$race, mean, na.rm=TRUE)
## 1 2 3 4
## 49.69231 56.00000 45.00000 54.22535
sort
It is often useful to sort data to make it easier to inspect it. R has a
sort function but that is only for vectors. For matrices or data.frames
you need to use the order function.
sort is straightforward:
x <- sample(10)
x
## [1] 10 8 2 5 4 7 6 9 3 1
sort(x)
## [1] 1 2 3 4 5 6 7 8 9 10
Now consider order:
i <- order(x)
i
## [1] 10 3 9 5 4 7 6 2 8 1
x[i]
## [1] 1 2 3 4 5 6 7 8 9 10
order returns a vector that allows you to sort. The first value of
i is 7. This means that x[7] should be the lowest number in
x. The next number is x[6] followed by x[3] and x[2] and
so forth.
Consider data.frame x:
set.seed(0)
x <- a[sample(nrow(a), 10), ]
x
## id female race
## 142 90 1 4
## 68 107 0 4
## 167 156 1 4
## 129 91 1 4
## 162 74 1 4
## 43 168 0 4
## 14 104 0 4
## 187 158 1 4
## 51 27 0 2
## 85 58 0 4
Here is how you can use order to sort it by one column (“id” in this
case):
oid <- order(x$id)
y <- x[oid, ]
y
## id female race
## 51 27 0 2
## 85 58 0 4
## 162 74 1 4
## 142 90 1 4
## 129 91 1 4
## 14 104 0 4
## 68 107 0 4
## 167 156 1 4
## 187 158 1 4
## 43 168 0 4
Or by multiple columns. In this case, we want to sort first by “race”, then by “female” and then by “id”:
oid <- order(x$race, x$female, x$id)
x[oid, ]
## id female race
## 51 27 0 2
## 85 58 0 4
## 14 104 0 4
## 68 107 0 4
## 43 168 0 4
## 162 74 1 4
## 142 90 1 4
## 129 91 1 4
## 167 156 1 4
## 187 158 1 4