This vignette provides comparisons with other packages that provide similar functionality. This is a work in progress – for a more detailed / complete / coherent comparison with other packages which provide wide-to-tall data reshaping, see my paper.
Sometimes you want to melt a “wide” data table which has several distinct pieces of information encoded in each column name. One example is the familiar iris data, which have flower part and measurement dimension encoded in each of four column names:
library(data.table)
data.table(iris)
#> Sepal.Length Sepal.Width Petal.Length Petal.Width Species
#> <num> <num> <num> <num> <fctr>
#> 1: 5.1 3.5 1.4 0.2 setosa
#> 2: 4.9 3.0 1.4 0.2 setosa
#> 3: 4.7 3.2 1.3 0.2 setosa
#> 4: 4.6 3.1 1.5 0.2 setosa
#> 5: 5.0 3.6 1.4 0.2 setosa
#> ---
#> 146: 6.7 3.0 5.2 2.3 virginica
#> 147: 6.3 2.5 5.0 1.9 virginica
#> 148: 6.5 3.0 5.2 2.0 virginica
#> 149: 6.2 3.4 5.4 2.3 virginica
#> 150: 5.9 3.0 5.1 1.8 virginica
The goal in this section will be to convert these data into a format with a column for each flower part (Sepal
and Petal
) so we can easily make a facetted scatterplot to visually examine whether or not sepals or larger than petals. The easiest way to perform this conversion is with packages which provide a function for melting into multiple output columns:
<- list(
iris.parts nc=nc::capture_melt_multiple(
iris,column=".*?",
"[.]",
dim=".*"),
tidyr=if(requireNamespace("tidyr"))tidyr::pivot_longer(
iris, cols=1:4,
names_to=c(".value", "dim"),
names_sep="[.]"),
stats=stats::reshape(
iris,direction="long",
timevar="dim",
varying=1:4,
sep="."),
"data.table::melt"=melt(
data.table(iris),
measure.vars=patterns(
Sepal="^Sepal",
Petal="^Petal")
data.table(
)[variable=factor(1:2), dim=c("Length", "Width")
on=.(variable)],
), if(requireNamespace("cdata"))cdata::rowrecs_to_blocks(
iris,controlTable=data.frame(
dim=c("Length", "Width"),
Petal=c("Petal.Length", "Petal.Width"),
Sepal=c("Sepal.Length", "Sepal.Width"),
stringsAsFactors=FALSE),
columnsToCopy="Species"))
#> Loading required namespace: cdata
$nc
iris.parts#> Species dim Petal Sepal
#> <fctr> <char> <num> <num>
#> 1: setosa Length 1.4 5.1
#> 2: setosa Length 1.4 4.9
#> 3: setosa Length 1.3 4.7
#> 4: setosa Length 1.5 4.6
#> 5: setosa Length 1.4 5.0
#> ---
#> 296: virginica Width 2.3 3.0
#> 297: virginica Width 1.9 2.5
#> 298: virginica Width 2.0 3.0
#> 299: virginica Width 2.3 3.4
#> 300: virginica Width 1.8 3.0
It is clear from the code above that each package is capable of the conversions. However the syntax and level of explicitness varies:
nc::capture_melt_multiple
requires a regular expression: (most implicit, least repetition)
column
group are used for the output column names.tidyr::pivot_longer
and stats::reshape
require specification of the input columns to melt along with a separator.
stats::reshape
assumes the output columns names occur in the part of the input column name before the separator.tidyr::pivot_longer
assumes the output columns occur in the part which corresponds to the .value
element of the names_to
argument.data.table::melt
requires a join to recover the dim
output column.cdata::rowrecs_to_blocks
requires explicit specification of a control table. (most explicit, most repetition)Any of the results can be visualized via:
if(require(ggplot2)){
ggplot()+
theme_bw()+
theme(panel.spacing=grid::unit(0, "lines"))+
facet_grid(dim ~ Species)+
coord_equal()+
geom_abline(slope=1, intercept=0, color="grey")+
geom_point(aes(
Petal, Sepal),data=iris.parts$nc)
}
It is clear from the plot above that sepals are larger than petals, for every measured flower.
What if we wanted to compare dimensions rather than parts?
<- list(
iris.dims nc=nc::capture_melt_multiple(
iris,part=".*?",
"[.]",
column=".*"),
stats=stats::reshape(
structure(iris, names=sub("(.*?)[.](.*)", "\\2.\\1", names(iris))),
direction="long",
timevar="part",
varying=1:4,
sep="."))
$nc
iris.dims#> Species part Length Width
#> <fctr> <char> <num> <num>
#> 1: setosa Petal 1.4 0.2
#> 2: setosa Petal 1.4 0.2
#> 3: setosa Petal 1.3 0.2
#> 4: setosa Petal 1.5 0.2
#> 5: setosa Petal 1.4 0.2
#> ---
#> 296: virginica Sepal 6.7 3.0
#> 297: virginica Sepal 6.3 2.5
#> 298: virginica Sepal 6.5 3.0
#> 299: virginica Sepal 6.2 3.4
#> 300: virginica Sepal 5.9 3.0
The code above shows that the syntax is mostly the same for this example. The biggest difference is for stats::reshape
which assumes that each input column name is composed of (1) the output column name, (2) a delimiter, and (3) some additional information to be stored in the output column given by timevar
. Therefore we need to pre-process column names using sub
for it to work.
if(require(ggplot2)){
ggplot()+
theme_bw()+
theme(panel.spacing=grid::unit(0, "lines"))+
facet_grid(part ~ Species)+
coord_equal()+
geom_abline(slope=1, intercept=0, color="grey")+
geom_point(aes(
Length, Width),data=iris.dims$nc)
}
It is clear from the plot above that Length
is larger than Width
for every measured flower part.
Consider the following wide data set:
<- data.table::data.table(
TC age.treatment=c(1, 5),
sex.control=c("M", "M"),
sex.treatment=c("F", "F"),
age.control=c(10, 50))
It is clear from the column names how the data should be grouped when they are converted to tall format. However the columns do not appear in regular order (age is before sex for treatment, but age is after sex for control), which causes a problem for stats and data.table:
<- list(
input.list "nc"=nc::capture_melt_multiple(
TC,column=".*?",
"[.]",
group=".*"),
"cdata"=if(requireNamespace("cdata"))cdata::rowrecs_to_blocks(
TC,controlTable=data.frame(
group=c("treatment", "control"),
age=c("age.treatment", "age.control"),
sex=c("sex.treatment", "sex.control"),
stringsAsFactors=FALSE)),
"data.table"=data.table::melt(TC, measure.vars=patterns(
age="age",
sex="sex")),
"stats"=stats::reshape(
TC,varying=1:4,
direction="long"),
"tidyr"=if(requireNamespace("tidyr"))tidyr::pivot_longer(
1:4,
TC, names_to=c(".value", "group"),
names_sep="[.]"))
<- list()
output.list for(pkg in names(input.list)){
<- input.list[[pkg]]
df.or.null if(is.data.frame(df.or.null)){
<- data.table::data.table(df.or.null)[order(age)]
output.list[[pkg]]
}
}
output.list#> $nc
#> group age sex
#> <char> <num> <char>
#> 1: treatment 1 F
#> 2: treatment 5 F
#> 3: control 10 M
#> 4: control 50 M
#>
#> $cdata
#> group age sex
#> <char> <num> <char>
#> 1: treatment 1 F
#> 2: treatment 5 F
#> 3: control 10 M
#> 4: control 50 M
#>
#> $data.table
#> variable age sex
#> <fctr> <num> <char>
#> 1: 1 1 M
#> 2: 1 5 M
#> 3: 2 10 F
#> 4: 2 50 F
#>
#> $stats
#> time age sex id
#> <char> <num> <char> <int>
#> 1: treatment 1 M 1
#> 2: treatment 5 M 2
#> 3: control 10 F 1
#> 4: control 50 F 2
#>
#> $tidyr
#> group age sex
#> <char> <num> <char>
#> 1: treatment 1 F
#> 2: treatment 5 F
#> 3: control 10 M
#> 4: control 50 M
sapply(output.list, function(DT)identical(DT$sex, c("F", "F", "M", "M")))
#> nc cdata data.table stats tidyr
#> TRUE TRUE FALSE FALSE TRUE
In conclusion, when the input column names to melt do not appear in the same order across groups or output columns, then the correct tall data can be computed using one of nc::capture_melt_multiple
, tidyr::pivot_longer
, cdata::rowrecs_to_blocks
.
Another data set where it is useful to do column name pattern matching followed by melting is the World Health Organization data:
if(requireNamespace("tidyr")){
data(who, package="tidyr")
else{
}<- data.frame(id=1, new_sp_m5564=2, newrel_f65=3)
who
}names(who)
#> [1] "country" "iso2" "iso3" "year" "new_sp_m014"
#> [6] "new_sp_m1524" "new_sp_m2534" "new_sp_m3544" "new_sp_m4554" "new_sp_m5564"
#> [11] "new_sp_m65" "new_sp_f014" "new_sp_f1524" "new_sp_f2534" "new_sp_f3544"
#> [16] "new_sp_f4554" "new_sp_f5564" "new_sp_f65" "new_sn_m014" "new_sn_m1524"
#> [21] "new_sn_m2534" "new_sn_m3544" "new_sn_m4554" "new_sn_m5564" "new_sn_m65"
#> [26] "new_sn_f014" "new_sn_f1524" "new_sn_f2534" "new_sn_f3544" "new_sn_f4554"
#> [31] "new_sn_f5564" "new_sn_f65" "new_ep_m014" "new_ep_m1524" "new_ep_m2534"
#> [36] "new_ep_m3544" "new_ep_m4554" "new_ep_m5564" "new_ep_m65" "new_ep_f014"
#> [41] "new_ep_f1524" "new_ep_f2534" "new_ep_f3544" "new_ep_f4554" "new_ep_f5564"
#> [46] "new_ep_f65" "newrel_m014" "newrel_m1524" "newrel_m2534" "newrel_m3544"
#> [51] "newrel_m4554" "newrel_m5564" "newrel_m65" "newrel_f014" "newrel_f1524"
#> [56] "newrel_f2534" "newrel_f3544" "newrel_f4554" "newrel_f5564" "newrel_f65"
Each column which starts with new
has three distinct pieces of information encoded in its name: diagnosis type (e.g. sp or rel), gender (m or f), and age range (e.g. 5564 or 1524). We would like to use a regex to match these column names, then using the matching columns as measure.vars in a melt, then join the two results. The most convenient way to do that is via:
<- list(
who.chr.list nc=nc::capture_melt_single(
who,"new_?",
diagnosis=".*",
"_",
gender=".",
ages=".*"),
tidyr=if(requireNamespace("tidyr"))tidyr::pivot_longer(
who,:newrel_f65,
new_sp_m014names_to=c("diagnosis", "gender", "ages"),
names_pattern="new_?(.*)_(.)(.*)"))
Note the result includes additional column value
which contains the melted data. There is also a column for each capture group in the specified pattern. The following example shows how to rename the value
column, remove missing values, and use numeric type conversion functions:
<- "new_?(.*)_(.)((0|[0-9]{2})([0-9]{0,2}))"
who.pattern <- function(y)ifelse(y=="", Inf, as.numeric(y))
as.numeric.Inf <- list(
who.typed.list nc=nc::capture_melt_single(
who,"new_?",
diagnosis=".*",
"_",
gender=".",
ages=list(
ymin.num="0|[0-9]{2}", as.numeric,
ymax.num="[0-9]{0,2}", as.numeric.Inf),
value.name="count",
na.rm=TRUE),
tidyr=if(requireNamespace("tidyr"))try(tidyr::pivot_longer(
who,cols=grep(who.pattern, names(who)),
names_transform=list(
ymin.num=as.numeric,
ymax.num=as.numeric.Inf),
names_to=c("diagnosis", "gender", "ages", "ymin.num", "ymax.num"),
names_pattern=who.pattern,
values_drop_na=TRUE,
values_to="count")))
str(who.typed.list)
#> List of 2
#> $ nc :Classes 'data.table' and 'data.frame': 76046 obs. of 10 variables:
#> ..$ country : chr [1:76046] "Afghanistan" "Afghanistan" "Afghanistan" "Afghanistan" ...
#> ..$ iso2 : chr [1:76046] "AF" "AF" "AF" "AF" ...
#> ..$ iso3 : chr [1:76046] "AFG" "AFG" "AFG" "AFG" ...
#> ..$ year : num [1:76046] 1997 1998 1999 2000 2001 ...
#> ..$ diagnosis: chr [1:76046] "sp" "sp" "sp" "sp" ...
#> ..$ gender : chr [1:76046] "m" "m" "m" "m" ...
#> ..$ ages : chr [1:76046] "014" "014" "014" "014" ...
#> ..$ ymin.num : num [1:76046] 0 0 0 0 0 0 0 0 0 0 ...
#> ..$ ymax.num : num [1:76046] 14 14 14 14 14 14 14 14 14 14 ...
#> ..$ count : num [1:76046] 0 30 8 52 129 90 127 139 151 193 ...
#> ..- attr(*, ".internal.selfref")=<externalptr>
#> $ tidyr: tibble [76,046 × 10] (S3: tbl_df/tbl/data.frame)
#> ..$ country : chr [1:76046] "Afghanistan" "Afghanistan" "Afghanistan" "Afghanistan" ...
#> ..$ iso2 : chr [1:76046] "AF" "AF" "AF" "AF" ...
#> ..$ iso3 : chr [1:76046] "AFG" "AFG" "AFG" "AFG" ...
#> ..$ year : num [1:76046] 1997 1997 1997 1997 1997 ...
#> ..$ diagnosis: chr [1:76046] "sp" "sp" "sp" "sp" ...
#> ..$ gender : chr [1:76046] "m" "m" "m" "m" ...
#> ..$ ages : chr [1:76046] "014" "1524" "2534" "3544" ...
#> ..$ ymin.num : num [1:76046] 0 15 25 35 45 55 65 0 15 25 ...
#> ..$ ymax.num : num [1:76046] 14 24 34 44 54 ...
#> ..$ count : num [1:76046] 0 10 6 3 5 2 0 5 38 36 ...
The result above shows that nc::capture_melt_single
(1) makes it easier to define complex patterns (2) supports type conversion without a post-processing step, and (3) reduces repetition in user code. There are several sources of repetition in tidyr
code:
ymin.num
appears only once for nc
but twice for tidyr
.ymax.chr
appears only once for nc
but three times for tidyr
.who
appears only once for nc
but twice for tidyr
.nc
but twice for tidyr
.Other packages for doing this include:
if(requireNamespace("tidyr")){
<- tidyr::gather(
gather.result
who,"variable",
"count",
grep(who.pattern, names(who)),
na.rm=TRUE)
<- tidyr::extract(
extract.result
gather.result,"variable",
c("diagnosis", "gender", "ages", "ymin.int", "ymax.int"),
who.pattern,convert=TRUE)
<- base::transform(
transform.result
extract.result,ymin.num=as.numeric(ymin.int),
ymax.num=ifelse(is.na(ymax.int), Inf, as.numeric(ymax.int)))
str(transform.result)
}#> 'data.frame': 76046 obs. of 12 variables:
#> $ country : chr "Afghanistan" "Afghanistan" "Afghanistan" "Afghanistan" ...
#> $ iso2 : chr "AF" "AF" "AF" "AF" ...
#> $ iso3 : chr "AFG" "AFG" "AFG" "AFG" ...
#> $ year : num 1997 1998 1999 2000 2001 ...
#> $ diagnosis: chr "sp" "sp" "sp" "sp" ...
#> $ gender : chr "m" "m" "m" "m" ...
#> $ ages : int 14 14 14 14 14 14 14 14 14 14 ...
#> $ ymin.int : int 0 0 0 0 0 0 0 0 0 0 ...
#> $ ymax.int : int 14 14 14 14 14 14 14 14 14 14 ...
#> $ count : num 0 30 8 52 129 90 127 139 151 193 ...
#> $ ymin.num : num 0 0 0 0 0 0 0 0 0 0 ...
#> $ ymax.num : num 14 14 14 14 14 14 14 14 14 14 ...
Note that tidyr::gather
requires two post-processing steps, which cause the same two types of repetition as tidyr::pivot_longer
:
base::transform
is used for converting age range variables to numeric, since default types are int with convert=TRUE
.tidyr::extract
is used to convert the melted column into several output columns; this results in repetition in the code because the regex is also used to define the columns to melt/gather.The reshape2
package suffers from the same two issues:
<- if(requireNamespace("reshape2")){
reshape2.result :::melt.data.frame(
reshape2
who,measure.vars=grep(who.pattern, names(who)),
na.rm=TRUE,
value.name="count")
}#> Loading required namespace: reshape2
Interestingly, data.table::patterns
can be used to avoid repeating the data set name, who
. However it supports neither type conversion nor regex capture groups.
<- data.table::melt.data.table(
dt.result data.table(who),
measure.vars=patterns(who.pattern),
na.rm=TRUE,
value.name="count")
Neither cdata nor stats provide an na.rm option:
<- data.frame(who)
who.df <- grepl(who.pattern, names(who))
is.varying names(who.df)[is.varying] <- paste0("count.", names(who)[is.varying])
<- stats::reshape(
stats.result
who.df,direction="long",
timevar="variable",
varying=is.varying)
if(requireNamespace("cdata")){
<- cdata::rowrecs_to_blocks(
cdata.result
who, ::build_unpivot_control(
cdata"variable",
"count",
grep(who.pattern, names(who), value=TRUE)),
columnsToCopy=grep(who.pattern, names(who), value=TRUE, invert=TRUE))
}
## Example 1: melting a wider iris data back to original.
library(data.table)
<- data.table(
iris.dt i=1:nrow(iris),
1:4],
iris[,Species=paste(iris$Species))
print(iris.dt)
#> i Sepal.Length Sepal.Width Petal.Length Petal.Width Species
#> <int> <num> <num> <num> <num> <char>
#> 1: 1 5.1 3.5 1.4 0.2 setosa
#> 2: 2 4.9 3.0 1.4 0.2 setosa
#> 3: 3 4.7 3.2 1.3 0.2 setosa
#> 4: 4 4.6 3.1 1.5 0.2 setosa
#> 5: 5 5.0 3.6 1.4 0.2 setosa
#> ---
#> 146: 146 6.7 3.0 5.2 2.3 virginica
#> 147: 147 6.3 2.5 5.0 1.9 virginica
#> 148: 148 6.5 3.0 5.2 2.0 virginica
#> 149: 149 6.2 3.4 5.4 2.3 virginica
#> 150: 150 5.9 3.0 5.1 1.8 virginica
## what if we had two observations on each row?
set.seed(1)
<- iris.dt[sample(.N)]
iris.rand <- cbind(treatment=iris.rand[1:75], control=iris.rand[76:150])
iris.wide print(iris.wide, topn=2, nrows=10)
#> treatment.i treatment.Sepal.Length treatment.Sepal.Width
#> <int> <num> <num>
#> 1: 68 5.8 2.7
#> 2: 129 6.4 2.8
#> ---
#> 74: 91 5.5 2.6
#> 75: 64 6.1 2.9
#> treatment.Petal.Length treatment.Petal.Width treatment.Species control.i
#> <num> <num> <char> <int>
#> 1: 4.1 1.0 versicolor 60
#> 2: 5.6 2.1 virginica 113
#> ---
#> 74: 4.4 1.2 versicolor 57
#> 75: 4.7 1.4 versicolor 72
#> control.Sepal.Length control.Sepal.Width control.Petal.Length
#> <num> <num> <num>
#> 1: 5.2 2.7 3.9
#> 2: 6.8 3.0 5.5
#> ---
#> 74: 6.3 3.3 4.7
#> 75: 6.1 2.8 4.0
#> control.Petal.Width control.Species
#> <num> <char>
#> 1: 1.4 versicolor
#> 2: 2.1 virginica
#> ---
#> 74: 1.6 versicolor
#> 75: 1.3 versicolor
## This is the usual data.table syntax for getting the original iris back.
<- melt(iris.wide, value.factor=TRUE, measure.vars = patterns(
iris.melted i="i$",
Sepal.Length="Sepal.Length$",
Sepal.Width="Sepal.Width$",
Petal.Length="Petal.Length$",
Petal.Width="Petal.Width$",
Species="Species$"))
identical(iris.melted[order(i), names(iris.dt), with=FALSE], iris.dt)
#> [1] TRUE
## nc can do the same thing -- you must define an R argument named
## column, and another named argument which identifies each group.
<- nc::capture_melt_multiple(
(nc.melted
iris.wide,group="[^.]+",
"[.]",
column=".*"))
#> group Petal.Length Petal.Width Sepal.Length Sepal.Width Species
#> <char> <num> <num> <num> <num> <char>
#> 1: control 3.9 1.4 5.2 2.7 versicolor
#> 2: control 5.5 2.1 6.8 3.0 virginica
#> 3: control 5.6 1.4 6.1 2.6 virginica
#> 4: control 1.5 0.1 4.9 3.1 setosa
#> 5: control 1.4 0.2 5.1 3.5 setosa
#> ---
#> 146: treatment 1.6 0.2 4.8 3.1 setosa
#> 147: treatment 1.3 0.4 5.4 3.9 setosa
#> 148: treatment 5.4 2.1 6.9 3.1 virginica
#> 149: treatment 4.4 1.2 5.5 2.6 versicolor
#> 150: treatment 4.7 1.4 6.1 2.9 versicolor
#> i
#> <int>
#> 1: 60
#> 2: 113
#> 3: 135
#> 4: 10
#> 5: 1
#> ---
#> 146: 31
#> 147: 17
#> 148: 140
#> 149: 91
#> 150: 64
identical(nc.melted[order(i), names(iris.dt), with=FALSE], iris.dt)
#> [1] TRUE
## This is how we do it using stats::reshape.
<- data.frame(iris.wide)
iris.wide.df names(iris.wide.df) <- sub("(.*?)[.](.*)", "\\2_\\1", names(iris.wide))
<- stats::reshape(
iris.reshaped
iris.wide.df,direction="long",
timevar="group",
varying=names(iris.wide.df),
sep="_")
identical(data.table(iris.reshaped[, names(iris.dt)])[order(i)], iris.dt)
#> [1] TRUE
## get the parts columns and groups -- is there any difference
## between groups? of course not!
<- nc::capture_melt_multiple(
parts.wide
iris.wide,group=".*?",
"[.]",
column=".*?",
"[.]",
dim=".*")
if(require("ggplot2")){
ggplot()+
theme_bw()+
theme(panel.spacing=grid::unit(0, "lines"))+
facet_grid(dim ~ group)+
coord_equal()+
geom_abline(slope=1, intercept=0, color="grey")+
geom_point(aes(
Petal, Sepal),data=parts.wide)
}