R Undo Dummy Variables - r

Question

I have a data set where a bunch of categorical variables were converted to dummy variables (all classes used, NOT n-1) and some were not. I'm trying to recode them in a single column.
For instance
Q1.1 Q1.2 Q1.3 Q1.NA Q2 Q3.1 Q3.2
1 0 0 0 3 0 1
0 1 0 0 4 1 0
0 0 1 0 2 0 1
Is there a simple way to convert this to:
Q1 Q2 Q3
1 3 2
2 4 1
3 2 2
Right now I'm just using strsplit() (as all the dummied variable names contain '.') with a couple loops but feel like there should be a better way. Any suggestions?

Answer 1

I wrote a function a while back that did this sort of thing.
MultChoiceCondense<-function(vars,indata){
tempvar<-matrix(NaN,ncol=1,nrow=length(indata[,1]))
dat<-indata[,vars]
for (i in 1:length(vars)){
for (j in 1:length(indata[,1])){
if (dat[j,i]==1) tempvar[j]=i
}
}
return(tempvar)
}
If your data is called Dat, then:
Dat$Q1<-MultChoiceCondense(c("Q1.1","Q1.2","Q1.3"),Dat)

Answer 2

Here's an approach that uses melt from "reshape2" and cSplit from my "splitstackshape" package along with some "data.table" fun. I've loaded dplyr so that we can pipe all the things.
library(splitstackshape)
library(reshape2)
library(dplyr)
mydf %>%
as.data.table(keep.rownames = TRUE) %>% # Convert to data.table. Keep rownames
melt(id.vars = "rn", variable.name = "V") %>% # Melt the dataset by rownames
.[value > 0] %>% # Subset for all non-zero values
cSplit("V", ".") %>% # Split the "V" column (names) by "."
.[is.na(V_2), V_2 := value] %>% # Replace NA values with actual values
dcast.data.table(rn ~ V_1, value.var = "V_2") # Go wide.
# rn Q1 Q2 Q3
# 1: 1 1 3 2
# 2: 2 2 4 1
# 3: 3 3 2 2
Here's a possible base R approach:
## Which columns are binary?
Bins <- sapply(mydf, function(x) {
all(x %in% c(0, 1))
})
## Two vectors -- part after the dot and before
X <- gsub(".*\\.(.*)$", "\\1", names(mydf)[Bins])
Y <- unique(gsub("(.*)\\..*$", "\\1", names(mydf)[Bins]))
## Use `apply` to subset the X value based on the
## logical version of the binary variable
cbind(mydf[!Bins],
`colnames<-`(t(apply(mydf[Bins], 1, function(z) {
X[as.logical(z)]
})), Y))
# Q2 Q1 Q3
# 1 3 1 2
# 2 4 2 1
# 3 2 3 2
At the end, you can just reorder the columns as required. You may also need to convert them to numeric since in this case, Q1 and Q3 will be factors.

Answer 3

another base R approach
dat <- read.table(header = TRUE, text = "Q1.1 Q1.2 Q1.3 Q1.NA Q2 Q3.1 Q3.2
1 0 0 0 3 0 1
0 1 0 0 4 1 0
0 0 1 0 2 0 1")
## this will take all the unique questions; Q1, Q2, Q3; test if
## they are dummies; and return the column if so or find which
## dummy column is a 1 otherwise
res <- lapply(unique(gsub('\\..*', '', names(dat))), function(x) {
tmp <- dat[, grep(x, names(dat)), drop = FALSE]
if (ncol(tmp) == 1) unlist(tmp, use.names = FALSE) else max.col(tmp)
})
# [[1]]
# [1] 1 2 3
#
# [[2]]
# [1] 3 4 2
#
# [[3]]
# [1] 2 1 2
do.call('cbind', res)
# [,1] [,2] [,3]
# [1,] 1 3 2
# [2,] 2 4 1
# [3,] 3 2 2

Answer 4

I'm assuming your data looks like this, where the categorical columns are encoded using a dot at the end. You may also have a case where all of the values in a row are zero, which indicates a base level (such as how dummyVars in caret works with fullRank=FALSE). If so, here is a vectorized solution.
library(dplyr)
dummyVars.undo = function(df, col_prefix) {
if (!endsWith(col_prefix, '.')) {
# If col_prefix doesn't end with a period, include one, but save the
# "pretty name" as the one without a period
pretty_col_prefix = col_prefix
col_prefix = paste0(col_prefix, '.')
} else {
# Otherwise, strip the period for the pretty column name
pretty_col_prefix = substr(col_prefix, 1, nchar(col_prefix)-1)
}
# Get all columns with that encoding prefix
cols = names(df)[names(df) %>% startsWith(col_prefix)]
# Find the rows where all values are zero. If this isn't the case
# with your data there's no worry, it won't hurt anything.
base_level.idx = rowSums(df[cols]) == 0
# Set the column value to a base value of zero
df[base_level.idx, pretty_col_prefix] = 0
# Go through the remaining columns and find where the maximum value (1) occurs
df[!base_level.idx, pretty_col_prefix] = cols[apply(df[!base_level.idx, cols], 1, which.max)] %>%
strsplit('\\.') %>%
sapply(tail, 1)
# Drop the encoded columns
df[cols] = NULL
return(df)
}
Usage:
# Collapse Q1
df = dummyVars.undo(df, 'Q1')
# Collapse Q3
df = dummyVars.undo(df, 'Q3')
This uses dplyr, but only for the pipe operator %>%. You could certainly remove that if you'd prefer to do base R instead.