Gather multiple sets of columns

Question

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,

Answer 1

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 tidyr_0.8.3.9000, 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

Answer 2

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.

Answer 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.

Answer 4

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.

Answer 5

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