... should exclude anything mentioned in value.var

[MichaelChirico]

Consider:

set.seed(120)
DT <- data.table(V1 = sample(10))
DT[ , paste0("V", 2:20) := replicate(19, sample(10), simplify = FALSE)]
DT[ , paste0("X", 1:50) := replicate(50, rnorm(10), simplify = FALSE)]

DT_m <- melt(DT, measure.vars = patterns("^X"), value.name = "X")

I want to reshape DT_m back to being "wide". I want V1:20 to be as originally in DT, but to aggregate all of X1:50 by summing into a single column.

It seems the way to do this is:

dcast(DT_m, V1 + V2 + V3 + [...] + V20 ~ ., fun.aggregate = sum)
#    V1 V2          .
#1   1  5   5.452721
#2   2  9   2.855705
#3   3  2  -1.775939
#4   4  6   7.033915
#5   5  1 -10.456389
#6   6  8   9.050576
#7   7  4   4.917982
#8   8  7   1.901975
#9   9 10   3.969899
#10 10  3  -1.367950

But obviously it's not optimal to need to type all the variables on the LHS. I thought that was the point of ..., but this doesn't work:

dcast(DT_m, ... ~ ., value.var = "X", fun.aggregate = sum)

... has included X, so the returned table is just DT_m (with X renamed to .).

[franknarf1]

I want to reshape DT_m back to being "wide". I want V1:20 to be as originally in DT, but to aggregate all of X1:50 by summing into a single column.

You can do

DT_m[, variable := NULL]
dcast(DT_m, ... ~ ., value.var = "X", fun = sum)

However it won't be "as originally", since dcast sorts by the LHS of ~. Not sure if I'm missing something here, but this result does match the direct one DT[, Reduce(+, .SD), keyby=V1:V20, .SDcols=X1:X50]

[MichaelChirico]

Indeed you're right. However my particular application here is part of an intro to R so I'm trying to demonstrate reshaping, and Reduce-based stuff should probably be reserved for well beyond when people have gotten their feet wet.

Anyway, I don't think it's unheard of to find data in a .csv in DT_m form to start with, so that we'd need this instead of the Reduce approach.

Good suggestion on just deleting variable... may be the only way to go since I know this FR probably runs up against compatibility with reshape2::dcast.

[franknarf1]

I agree. I wasn't suggesting the Reduce way as an alternative, just using it to verify that my result was correct (since I was seeing different numbers from those you showed in the OP).