Yikes. While this sidesteps some of the issues in the other approaches, it also doubles your memory usage and halves your potential SIMD performance. Maybe this is better than the downsides of the alternatives, but as someone working on optimizing large and slow R code with C++ extensions speaking to someone who is trying to replace R with something better, this seems like a choice that may haunt you in the future. If you can, I'd keep your options open for a "Structure of Arrays" approach or "buddy bitmap" that keeps your data contiguous and compact.
We'll keep re-evaluating this over time, but I'm currently not sure it's that serious a problem. For many numerically intensive computations, you'll want to move over to a pure array of 64-bit floating point values before the long-running computation begins, at which point you can adopt NaN == NA semantics if appropriate. I believe the solution I outlined in my updated comment above may already describe the buddy bitmap solution you're referring to. It certainly describes a solution that I would refer to as a struct-of-arrays.
