So hacked together the first tools for Bayesian calculations. I can now properly compare two hypotheses and update my priors! Yeah, baby! Almost finished the write-up of that first iteration. Also got myself two introductory books to Bayesian statistics, and going to improve based on that.
I’m deliberately re-inventing the wheel here because I hate magical black boxes. If I haven’t written the tools myself, I won’t understand why they work, and statistics without understanding isn’t any better than astrology. Also, hacking all this myself in Ruby makes it so much more usable in the end. Personal DSL ftw!1
Also, just typing post 20.p, 3, 4 (= calculate the
posterior, starting with a 20% prior and two observations, one at
3:1 odds, the other at 4:1 odds in favor of the hypothesis, which
outputs posterior: 4.77dB / 75.00%) is much nicer than
any R crap2. And thanks to a few tables I taped to my
wall, I can also do all those calculations by hand during thinking
sessions, with nothing more than a few lookups and simple addition.
(Log tables ftw!)
Math is fun. Especially non-arbitrarily derived math that actually makes sense (unlike frequentist bullshit), is intuitive and usable.
Currently working on supporting more than two hypotheses, and more details of the experimental protocol, particularly wrt blinding. Though I’m beginning to think that proper Bayesian analysis is much less subject to cognitive biases and stuff like that. Just roll it all into your priors, add more hypotheses and keep on updatin’.
-
The only disadvantage of this is that the tools are highly customized to my thinking style, but still, it’s trivial to adjust. If you know Ruby and are on a *NIX system. But if not, the fuck is wrong with you? Seriously. ↩
-
If you have to deal with gigabytes of data, and need to test models that require supercomputers to even run, go ahead, use R or Mathematica or whatever. I’m figuring out if caffeine fucks with my sleep, not decoding the genome or predicting the weather. ↩