Searching’s not always so easy. Especially in pdfs (I need to move the manual to HTML, but it’s huge). You won’t find any mention of the R-specific notation
NA in the manual. At least I don’t recall putting it in there.
What’s missing is also Ben’s really nice approach. You can decompose the full matrix with missing data and observed data conceptually as
X = X_miss + X_obs;
where the observed data matrix
X_obs is sparse and has zeroes where data is missing;
X_miss has parameters where data is missing and zeroes where its observed.
What you can do rather than actually adding them and doing
X * beta, is to instead use
X_miss * beta + X_obs * beta
using the sparse matrix multiplication function
csr_matrix_times_vector for the multiplies and plain old addition for the additon.