,
14 tweets,
2 min read
Read on Twitter
Scientists? Please, please, please don't take any statistics theory or computation tutorial from applied scientists at face value.
I received my PhD in experimental physics where I taught myself statistics and thought I was hot shit.
Then I spent six years with really good statisticians learning just how incredibly wrong I was.
I had the overall picture okay, but I had all of the details wrong. But those details are _critical_ to a robust analysis.
And every intrafield tutorial I see also gets those details wrong, albeit often in different ways. Some quick lessons to keep in mind:
Lesson 1: Comp stats math is not easy, but its important and you cannot be flippant about it without compromising your results.
Corollary 1: In MCMC it doesn't matter whether your chains converge (they don't) but rather _how_ they converge. No one talks about that!
Lesson 2: Writing a comp stats algorithm and verifying that it that works for your problem is really, really hard.
Corollary 2: You _really_ shouldn't write your own MCMC sampler for your analysis. For fun or learning? Definitely!
Corollary 3: Be very wary of packages written exclusively by scientists without any stats collaborators. #subtweet
I tried to simplify MCMC as much as I could with arxiv.org/abs/1701.02434 and it still took 60 pages.
Lesson 3: Don't be shy asking for help! The good statisticians love science and love making it happen.
Corollary 4: discourse.mc-stan.org.
<deep breath> Thanks for indulging my rant. 😬
