It is difficult to gather empirical data on human behavior and efficiency, so I wouldn't ding Rinard points for that (as a systems PL researcher from MIT, he surely knows how to measure numbers). You'll find the lack of useful empirical data for most any pedagogic technique (what we can measure is performance in simple tasks, by this doesn't scale up).

Math and philosophy are hella abstract, while learning through programming at least has the learner building something concrete that can be more gratifying. When maintaining the student's interest is much of the battle, it might not matter as programming is not better than the other approaches, it only needs to be good enough.