Fastest way to reshape variable values as columns

Question

I have a dataset with about 3 million rows and the following structure:

PatientID| Year | PrimaryConditionGroup
---------------------------------------
1        | Y1   | TRAUMA
1        | Y1   | PREGNANCY
2        | Y2   | SEIZURE
3        | Y1   | TRAUMA

Being fairly new to R, I have some trouble finding the right way to reshape the data into the structure outlined below:

PatientID| Year | TRAUMA | PREGNANCY | SEIZURE
----------------------------------------------
1        | Y1   | 1      | 1         | 0
2        | Y2   | 0      | 0         | 1
3        | Y1   | 1      | 0         | 1

My question is: What is the fastest/most elegant way to create a data.frame, where the values of PrimaryConditionGroup become columns, grouped by PatientID and Year (counting the number of occurences)?

Answer 1

There are probably more succinct ways of doing this, but for sheer speed, it's hard to beat a data.table-based solution:

df <- read.table(text="PatientID Year  PrimaryConditionGroup
1         Y1    TRAUMA
1         Y1    PREGNANCY
2         Y2    SEIZURE
3         Y1    TRAUMA", header=T)

library(data.table)
dt <- data.table(df, key=c("PatientID", "Year"))

dt[ , list(TRAUMA =    sum(PrimaryConditionGroup=="TRAUMA"),
           PREGNANCY = sum(PrimaryConditionGroup=="PREGNANCY"),
           SEIZURE =   sum(PrimaryConditionGroup=="SEIZURE")),
   by = list(PatientID, Year)]

#      PatientID Year TRAUMA PREGNANCY SEIZURE
# [1,]         1   Y1      1         1       0
# [2,]         2   Y2      0         0       1
# [3,]         3   Y1      1         0       0

EDIT: aggregate() provides a 'base R' solution that might or might not be more idiomatic. (The sole complication is that aggregate returns a matrix, rather than a data.frame; the second line below fixes that up.)

out <- aggregate(PrimaryConditionGroup ~ PatientID + Year, data=df, FUN=table)
out <- cbind(out[1:2], data.frame(out[3][[1]]))

2nd EDIT Finally, a succinct solution using the reshape package gets you to the same place.

library(reshape)
mdf <- melt(df, id=c("PatientID", "Year"))
cast(PatientID + Year ~ value, data=j, fun.aggregate=length)

Answer 2

There are fast melt and dcast data.table specific methods implemented in C, in versions >=1.9.0. Here's a comparison with other excellent answers from @Josh's post on 3-million row data (just excluding base:::aggregate as it was taking quite sometime).

For more info on NEWS entry, go here.

I'll assume you've 1000 patients and 5 years in total. You can adjust the variables patients and year accordingly.

require(data.table) ## >= 1.9.0
require(reshape2)

set.seed(1L)
patients = 1000L
year = 5L
n = 3e6L
condn = c("TRAUMA", "PREGNANCY", "SEIZURE")

# dummy data
DT <- data.table(PatientID = sample(patients, n, TRUE),
                 Year = sample(year, n, TRUE), 
                 PrimaryConditionGroup = sample(condn, n, TRUE))

DT_dcast <- function(DT) {
    dcast.data.table(DT, PatientID ~ Year, fun.aggregate=length)
}

reshape2_dcast <- function(DT) {
    reshape2:::dcast(DT, PatientID ~ Year, fun.aggregate=length)
}

DT_raw <- function(DT) {
    DT[ , list(TRAUMA = sum(PrimaryConditionGroup=="TRAUMA"),
            PREGNANCY = sum(PrimaryConditionGroup=="PREGNANCY"),
              SEIZURE = sum(PrimaryConditionGroup=="SEIZURE")),
    by = list(PatientID, Year)]
}

# system.time(.) timed 3 times
#         Method Time_rep1 Time_rep2 Time_rep3
#       dcast_DT     0.393     0.399     0.396
#    reshape2_DT     3.784     3.457     3.605
#         DT_raw     0.647     0.680     0.657

dcast.data.table is about 1.6x faster than normal aggregation using data.table and 8.8x faster than reshape2:::dcast.