No one at holmestice ever ran statistical tests. Instead they'd evaluate stylometric elements on a pass/no-pass basis, using them to eliminate potential authors for a work from the pool. (A pool which had already been narrowed by non-stylometric deductions that had to do with the structure of the exchange, mod biases in matching, number of known pinch-hits, etc.) Preferred stylometric elements tended to be punctuation, sentence structure, style/philosophy of author notes, Brit-picking errors, even characteristic typos, looking for things that were distinctive and hard (or psychologically painful!) to fake.

However! There have been people who have attempted to use stylometry to attribute some of the Sherlock Holmes stories to Conan Doyle's wives! c.f. this post and my critique of the technique in the comments. I have no idea of those critiques are applicable to the Dynamiter analysis; I've only given the website the quickest of scans.