model.matrix() with na.action=NULL?

Question

I have a formula and a data frame, and I want to extract the model.matrix(). However, I need the resulting matrix to include the NAs that were found in the original dataset. If I were to use model.frame() to do this, I would simply pass it na.action=NULL. However, the output I need is of the model.matrix() format. Specifically, I need only the right-hand side variables, I need the output to be a matrix (not a data frame), and I need factors to be converted to a series of dummy variables.

I'm sure I could hack something together using loops or something, but I was wondering if anyone could suggest a cleaner and more efficient workaround. Thanks a lot for your time!

And here's an example:

dat <- data.frame(matrix(rnorm(20),5,4), gl(5,2))
dat[3,5] <- NA
names(dat) <- c(letters[1:4], 'fact')
ff <- a ~ b + fact

# This omits the row with a missing observation on the factor
model.matrix(ff, dat) 

# This keeps the NA, but it gives me a data frame and does not dichotomize the factor
model.frame(ff, dat, na.action=NULL) 

Here is what I would like to obtain:

   (Intercept)          b fact2 fact3 fact4 fact5
1            1  0.7266086     0     0     0     0
2            1 -0.6088697     0     0     0     0
3            NA 0.4643360     NA    NA    NA    NA
4            1 -1.1666248     1     0     0     0
5            1 -0.7577394     0     1     0     0
6            1  0.7266086     0     1     0     0
7            1 -0.6088697     0     0     1     0
8            1  0.4643360     0     0     1     0
9            1 -1.1666248     0     0     0     1
10           1 -0.7577394     0     0     0     1

Answer 1

You can mess around a little with the model.matrix object, based on the rownames :

MM <- model.matrix(ff,dat)
MM <- MM[match(rownames(dat),rownames(MM)),]
MM[,"b"] <- dat$b
rownames(MM) <- rownames(dat)

which gives :

> MM
     (Intercept)         b fact2 fact3 fact4 fact5
1              1 0.9583010     0     0     0     0
2              1 0.3266986     0     0     0     0
3             NA 1.4992358    NA    NA    NA    NA
4              1 1.2867461     1     0     0     0
5              1 0.5024700     0     1     0     0
6              1 0.9583010     0     1     0     0
7              1 0.3266986     0     0     1     0
8              1 1.4992358     0     0     1     0
9              1 1.2867461     0     0     0     1
10             1 0.5024700     0     0     0     1

Alternatively, you can use contrasts() to do the work for you. Constructing the matrix by hand would be :

cont <- contrasts(dat$fact)[as.numeric(dat$fact),]
colnames(cont) <- paste("fact",colnames(cont),sep="")
out <- cbind(1,dat$b,cont)
out[is.na(dat$fact),1] <- NA
colnames(out)[1:2]<- c("Intercept","b")
rownames(out) <- rownames(dat)

which gives :

> out
     Intercept          b fact2 fact3 fact4 fact5
1            1  0.2534288     0     0     0     0
2            1  0.2697760     0     0     0     0
3           NA -0.8236879    NA    NA    NA    NA
4            1 -0.6053445     1     0     0     0
5            1  0.4608907     0     1     0     0
6            1  0.2534288     0     1     0     0
7            1  0.2697760     0     0     1     0
8            1 -0.8236879     0     0     1     0
9            1 -0.6053445     0     0     0     1
10           1  0.4608907     0     0     0     1

In any case, both methods can be incorporated in a function that can deal with more complex formulae. I leave the exercise to the reader (what do I loath that sentence when I meet it in a paper ;-) )

Answer 2

Joris's suggestion works, but a quicker and cleaner way to do this is via the global na.action setting. The 'Pass' option achieves our goal of preserving NA's from the original dataset.

Option 1: Pass

Resulting matrix will contain NA's in rows corresponding to the original dataset.

options(na.action='na.pass')
model.matrix(ff, dat) 

Option 2: Omit

Resulting matrix will skip rows containing NA's.

options(na.action='na.omit')
model.matrix(ff, dat) 

Option 3: Fail

An error will occur if the original data contains NA's.

options(na.action='na.fail')
model.matrix(ff, dat) 

Of course, always be careful when changing global options because they can alter behavior of other parts of your code. A cautious person might store the original setting with something like current.na.action <- options('na.action'), and then change it back after making the model.matrix.

Answer 3

Another way is to use the model.frame function with argument na.action=na.pass as your second argument to model.matrix:

> model.matrix(ff, model.frame(~ ., dat, na.action=na.pass))
   (Intercept)          b fact2 fact3 fact4 fact5
1            1 -1.3560754     0     0     0     0
2            1  2.5476965     0     0     0     0
3            1  0.4635628    NA    NA    NA    NA
4            1 -0.2871379     1     0     0     0
5            1  2.2684958     0     1     0     0
6            1 -1.3560754     0     1     0     0
7            1  2.5476965     0     0     1     0
8            1  0.4635628     0     0     1     0
9            1 -0.2871379     0     0     0     1
10           1  2.2684958     0     0     0     1

model.frame allows you to set the appropriate action for na.action which is maintained when model.matrix is called.

Answer 4

I half-stumbled across a simpler solution after looking at mattdevlin and Nathan Gould's answers:

 model.matrix.lm(ff, dat, na.action = "na.pass")

model.matrix.default may not support the na.action argument, but model.matrix.lm does!

(I found model.matrix.lm from Rstudio's auto-complete suggestions — it appears to be the only non-default method for model.matrix if you haven't loaded any libraries that add others. Then I just guessed it might support the na.action argument.)