They should give something like a random forests a shot on the raw multinomial/one-hot encoded responses: https://www.reddit.com/r/gwern/comments/9vy62y/your_personality_data_can_do_more_items_provide/e9g0nyz/ … I bet there's some more predictive power to squeeze out, and would demonstrate the point much more dramatically.
More on 'give up theory' in science: Really kickass paper on predictive analytics of personality by Mõttus, @timothycbates, @DMCpersonality, Mroczek @WilliamRevelle.
https://psyarxiv.com/4q9gv/ pic.twitter.com/1EbzsGj0Nn
-
-
-
RF type stuff didn't produce any more validity for the cognitive tests, was slightly worse than ridge regression, and >10x slower to fit. But for personality, one would have a stronger prior for interaction effects, whereas not for cognitive data.
-
Hm, maybe not enough data then. Just for the NEO-PI-R here, unfolding Likert items without facets/Big Fives would give 1200 variables to search over.
-
Throw some xgboost at it and I bet you get some nice results.
-
If RF doesn't pick up anything beyond a dumb linear regression, I'd expect XGboost to likewise add only a little. They're not that different, after all, and if there's not enough data, there's not enough data.
End of conversation
New conversation -
-
-
If you are going to give up theory, don't use questionnaires at all. You can get 12-15% EV on a trait like IQ just looking at FB statuses. Could be considered 40k questions of form: "how often do you use X word?"
-
I would like to know EV for outcomes as well as traits, but that hasn't been researched as much
End of conversation
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.