Gather multiple sets of columns

r reshape dplyr qualtrics tidyr

115,118

Solution 1

This approach seems pretty natural to me:

df %>%
  gather(key, value, -id, -time) %>%
  extract(key, c("question", "loop_number"), "(Q.\\..)\\.(.)") %>%
  spread(question, value)

First gather all question columns, use extract() to separate into question and loop_number, then spread() question back into the columns.

#>    id       time loop_number         Q3.2        Q3.3
#> 1   1 2009-01-01           1  0.142259203 -0.35842736
#> 2   1 2009-01-01           2  0.061034802  0.79354061
#> 3   1 2009-01-01           3 -0.525686204 -0.67456611
#> 4   2 2009-01-02           1 -1.044461185 -1.19662936
#> 5   2 2009-01-02           2  0.393808163  0.42384717

Solution 2

This could be done using reshape. It is possible with dplyr though.

  colnames(df) <- gsub("\\.(.{2})$", "_\\1", colnames(df))
  colnames(df)[2] <- "Date"
  res <- reshape(df, idvar=c("id", "Date"), varying=3:8, direction="long", sep="_")
  row.names(res) <- 1:nrow(res)
  
   head(res)
  #  id       Date time       Q3.2       Q3.3
  #1  1 2009-01-01    1  1.3709584  0.4554501
  #2  2 2009-01-02    1 -0.5646982  0.7048373
  #3  3 2009-01-03    1  0.3631284  1.0351035
  #4  4 2009-01-04    1  0.6328626 -0.6089264
  #5  5 2009-01-05    1  0.4042683  0.5049551
  #6  6 2009-01-06    1 -0.1061245 -1.7170087

Or using dplyr

  library(tidyr)
  library(dplyr)
  colnames(df) <- gsub("\\.(.{2})$", "_\\1", colnames(df))

  df %>%
     gather(loop_number, "Q3", starts_with("Q3")) %>% 
     separate(loop_number,c("L1", "L2"), sep="_") %>% 
     spread(L1, Q3) %>%
     select(-L2) %>%
     head()
  #  id       time       Q3.2       Q3.3
  #1  1 2009-01-01  1.3709584  0.4554501
  #2  1 2009-01-01  1.3048697  0.2059986
  #3  1 2009-01-01 -0.3066386  0.3219253
  #4  2 2009-01-02 -0.5646982  0.7048373
  #5  2 2009-01-02  2.2866454 -0.3610573
  #6  2 2009-01-02 -1.7813084 -0.7838389

Update

With new version of tidyr, we can use pivot_longer to reshape multiple columns. (Using the changed column names from gsub above)

library(dplyr)
library(tidyr)
df %>% 
    pivot_longer(cols = starts_with("Q3"), 
          names_to = c(".value", "Q3"), names_sep = "_") %>% 
    select(-Q3)
# A tibble: 30 x 4
#      id time         Q3.2    Q3.3
#   <int> <date>      <dbl>   <dbl>
# 1     1 2009-01-01  0.974  1.47  
# 2     1 2009-01-01 -0.849 -0.513 
# 3     1 2009-01-01  0.894  0.0442
# 4     2 2009-01-02  2.04  -0.553 
# 5     2 2009-01-02  0.694  0.0972
# 6     2 2009-01-02 -1.11   1.85  
# 7     3 2009-01-03  0.413  0.733 
# 8     3 2009-01-03 -0.896 -0.271 
#9     3 2009-01-03  0.509 -0.0512
#10     4 2009-01-04  1.81   0.668 
# … with 20 more rows

NOTE: Values are different because there was no set seed in creating the input dataset

Solution 3

With the recent update to melt.data.table, we can now melt multiple columns. With that, we can do:

require(data.table) ## 1.9.5
melt(setDT(df), id=1:2, measure=patterns("^Q3.2", "^Q3.3"), 
     value.name=c("Q3.2", "Q3.3"), variable.name="loop_number")
 #    id       time loop_number         Q3.2        Q3.3
 # 1:  1 2009-01-01           1 -0.433978480  0.41227209
 # 2:  2 2009-01-02           1 -0.567995351  0.30701144
 # 3:  3 2009-01-03           1 -0.092041353 -0.96024077
 # 4:  4 2009-01-04           1  1.137433487  0.60603396
 # 5:  5 2009-01-05           1 -1.071498263 -0.01655584
 # 6:  6 2009-01-06           1 -0.048376809  0.55889996
 # 7:  7 2009-01-07           1 -0.007312176  0.69872938

You can get the development version from here.

Solution 4

It's not at all related to "tidyr" and "dplyr", but here's another option to consider: merged.stack from my "splitstackshape" package, V1.4.0 and above.

library(splitstackshape)
merged.stack(df, id.vars = c("id", "time"), 
             var.stubs = c("Q3.2.", "Q3.3."),
             sep = "var.stubs")
#     id       time .time_1       Q3.2.       Q3.3.
#  1:  1 2009-01-01      1. -0.62645381  1.35867955
#  2:  1 2009-01-01      2.  1.51178117 -0.16452360
#  3:  1 2009-01-01      3.  0.91897737  0.39810588
#  4:  2 2009-01-02      1.  0.18364332 -0.10278773
#  5:  2 2009-01-02      2.  0.38984324 -0.25336168
#  6:  2 2009-01-02      3.  0.78213630 -0.61202639
#  7:  3 2009-01-03      1. -0.83562861  0.38767161
# <<:::SNIP:::>>
# 24:  8 2009-01-08      3. -1.47075238 -1.04413463
# 25:  9 2009-01-09      1.  0.57578135  1.10002537
# 26:  9 2009-01-09      2.  0.82122120 -0.11234621
# 27:  9 2009-01-09      3. -0.47815006  0.56971963
# 28: 10 2009-01-10      1. -0.30538839  0.76317575
# 29: 10 2009-01-10      2.  0.59390132  0.88110773
# 30: 10 2009-01-10      3.  0.41794156 -0.13505460
#     id       time .time_1       Q3.2.       Q3.3.

Solution 5

In case you are like me, and cannot work out how to use "regular expression with capturing groups" for extract, the following code replicates the extract(...) line in Hadleys' answer:

df %>% 
    gather(question_number, value, starts_with("Q3.")) %>%
    mutate(loop_number = str_sub(question_number,-2,-2), question_number = str_sub(question_number,1,4)) %>%
    select(id, time, loop_number, question_number, value) %>% 
    spread(key = question_number, value = value)

The problem here is that the initial gather forms a key column that is actually a combination of two keys. I chose to use mutate in my original solution in the comments to split this column into two columns with equivalent info, a loop_number column and a question_number column. spread can then be used to transform the long form data, which are key value pairs (question_number, value) to wide form data.

View more solutions

115,118

Andrew

I'm currently working on a PhD in Public Policy at the Sanford School of Public Policy at Duke University, where I'm studying education and freedom of expression INGOs that work in Egypt and China. Life rocks.

Updated on February 19, 2021

Comments

Andrew about 3 years
I have data from an online survey where respondents go through a loop of questions 1-3 times. The survey software (Qualtrics) records this data in multiple columns—that is, Q3.2 in the survey will have columns Q3.2.1., Q3.2.2., and Q3.2.3.:
```
df <- data.frame(
  id = 1:10,
  time = as.Date('2009-01-01') + 0:9,
  Q3.2.1. = rnorm(10, 0, 1),
  Q3.2.2. = rnorm(10, 0, 1),
  Q3.2.3. = rnorm(10, 0, 1),
  Q3.3.1. = rnorm(10, 0, 1),
  Q3.3.2. = rnorm(10, 0, 1),
  Q3.3.3. = rnorm(10, 0, 1)
)

# Sample data

   id       time    Q3.2.1.     Q3.2.2.    Q3.2.3.     Q3.3.1.    Q3.3.2.     Q3.3.3.
1   1 2009-01-01 -0.2059165 -0.29177677 -0.7107192  1.52718069 -0.4484351 -1.21550600
2   2 2009-01-02 -0.1981136 -1.19813815  1.1750200 -0.40380049 -1.8376094  1.03588482
3   3 2009-01-03  0.3514795 -0.27425539  1.1171712 -1.02641801 -2.0646661 -0.35353058
...
```
I want to combine all the QN.N* columns into tidy individual QN.N columns, ultimately ending up with something like this:
```
   id       time loop_number        Q3.2        Q3.3
1   1 2009-01-01           1 -0.20591649  1.52718069
2   2 2009-01-02           1 -0.19811357 -0.40380049
3   3 2009-01-03           1  0.35147949 -1.02641801
...
11  1 2009-01-01           2 -0.29177677  -0.4484351
12  2 2009-01-02           2 -1.19813815  -1.8376094
13  3 2009-01-03           2 -0.27425539  -2.0646661
...
21  1 2009-01-01           3 -0.71071921 -1.21550600
22  2 2009-01-02           3  1.17501999  1.03588482
23  3 2009-01-03           3  1.11717121 -0.35353058
...
```
The tidyr library has the gather() function, which works great for combining one set of columns:
```
library(dplyr)
library(tidyr)
library(stringr)

df %>% gather(loop_number, Q3.2, starts_with("Q3.2")) %>% 
  mutate(loop_number = str_sub(loop_number,-2,-2)) %>%
  select(id, time, loop_number, Q3.2)


   id       time loop_number        Q3.2
1   1 2009-01-01           1 -0.20591649
2   2 2009-01-02           1 -0.19811357
3   3 2009-01-03           1  0.35147949
...
29  9 2009-01-09           3 -0.58581232
30 10 2009-01-10           3 -2.33393981
```
The resultant data frame has 30 rows, as expected (10 individuals, 3 loops each). However, gathering a second set of columns does not work correctly—it successfully makes the two combined columns Q3.2 and Q3.3, but ends up with 90 rows instead of 30 (all combinations of 10 individuals, 3 loops of Q3.2, and 3 loops of Q3.3; the combinations will increase substantially for each group of columns in the actual data):
```
df %>% gather(loop_number, Q3.2, starts_with("Q3.2")) %>% 
  gather(loop_number, Q3.3, starts_with("Q3.3")) %>%
  mutate(loop_number = str_sub(loop_number,-2,-2))


   id       time loop_number        Q3.2        Q3.3
1   1 2009-01-01           1 -0.20591649  1.52718069
2   2 2009-01-02           1 -0.19811357 -0.40380049
3   3 2009-01-03           1  0.35147949 -1.02641801
...
89  9 2009-01-09           3 -0.58581232 -0.13187024
90 10 2009-01-10           3 -2.33393981 -0.48502131
```
Is there a way to use multiple calls to gather() like this, combining small subsets of columns like this while maintaining the correct number of rows?
- Alex over 9 years
  
  what's wrong with df %>% gather(loop_number, Q3.2, starts_with("Q3."))
- Andrew over 9 years
  
  That gets me one consolidated column with 60 rows. I guess that could work if I then included some sort of call to seperate() to divide up the Q3.3 (and beyond) values into their own columns. But that still seems like a really roundabout hacky solution…
- Alex over 9 years
  
  use spread i am working on a solution now :p
- Alex over 9 years
  
  try this! df %>% gather(question_number, Q3.2, starts_with("Q3.")) %>% mutate(loop_number = str_sub(question_number,-2,-2), question_number = str_sub(question_number,1,4)) %>% select(id, time, loop_number, question_number, Q3.2) %>% spread(key = question_number, value = Q3.2)
- Andrew over 9 years
  
  Ooh, that works really well for the two variables. I'm curious if it's scalable though—in my real data I've got Q3.2-Q3.30, so it would need a bunch of individual calls to spread(). Though multiple calls appears inevitable anyway, whether it's a bunch of generate()s that work or nested spread()s…
- Andrew over 9 years
  
  @akrun figured out a scalable version below, with gather + separate + spread. Magic stuff.
- Alex over 9 years
  
  my code should work for any number of questions, since the question number column will have distinct Q3.2, Q3.3, ..., Q3.30.
- Andrew over 9 years
  
  Oh, I see. I misread it and thought that in spread(..., value=X), I would need to set X to each of the final column names. But that's definitely not the case.
- Arun over 9 years
  
  +1 great question. FWIW, it's neither natural nor that efficient to have to melt (gather) the whole data set only to cast (spread) back. I'll post an answer if I manage to succeed implementing melt to accomplish this directly as required in this scenario.
Andrew over 9 years

Whoa, this works perfectly. tidyr is ostensibly a replacement/upgrade for reshape--I wonder if @hadley knows of a way to do this same thing with dplyr or tidyr…
Andrew over 9 years

That is pure magic. The only thing I added was mutate(loop_number = as.numeric(L2)) before dropping L2, and it's perfect.
akrun over 9 years

@Andrew I personally prefer the reshape method for its compact code, though dplyr may be faster for big datasets.
hadley over 9 years

I've never been able to understand the reshape() function, see my solution for what I seems to me a pretty clean tidyr implementation.
skan about 8 years

Hello. I have many columns with names ending in 1 and 2, like age1, age2, weight1, weight2, blood1, blood2.... How would I apply your method here?
skan about 8 years

Hello. I have many columns with names ending in 1 and 2, like age1, age2, weight1, weight2, blood1, blood2.... How would I apply your method here?
skan about 8 years

Hello. I have many columns with names ending in 1 and 2, like age1, age2, weight1, weight2, blood1, blood2.... How would I apply your method here?
Arun about 8 years

skan, check the reshaping vignette. Good luck!
skan about 8 years

I did but I don't know how to properly embed regular expressions to split column names and pass it to melt. There is only one example with patterns, and it's too simple. In my case I would need to include many many column names inside pattern()
skan about 8 years

Imagine you have these columns: paste0(rep(LETTERS, each=3), 1:3) and you want to get the long table defined by a letter and a number
mob over 6 years

What does this part mean: "(Q.\\..)\\.(.)" What would I search for to decode what is happening there?
hadley over 6 years

@mob Regular expressions
Michael Bellhouse over 4 years

This is hands down the most succinct and easy to interpret.
LC-datascientist over 4 years

@mob "(Q.\\..)\\.(.)" is a regular expression with parentheses that define the groups of the regular expression to extract into "question" and "loop_number". More specifically, in this example, the items in key with the expression "Q.\\.." goes into the "question" column (i.e., "Q3.2" and "Q3.3"), then the part after next period, expressed as ".", goes into the "loop_number" column.
tjebo about 3 years

tidyr::pivot_longer: Now not only in the development version :)