If only 190 out of 19k people are political-compass-center, and only 190 out of 19k people are into feet or whatever, but 180 of those are overlapped (pol. compass-center & into feet), then that's an overwhelmingly strong correlation, despite the sparsity of positive responses.
Conversation
Replying to
Sure, that would be an incredibly strong correlation. But is that what we have here? Sounds more like if there were 190 people into feet, they were split 55-45-45-45 across the four quadrants. That'd be a weak correlation, and you can't go "but 19K sample size"; it's 190 people.
1
8
Replying to
We can check how much evidence the data provides for the hypothesis that the "true" correlation is as observed, vs the hypothesis that the "true" correlation is 0 and the observed correlation is noise. Aella's data supports the former hypothesis ~10B times more than the latter.
1
In other words: yes, sometimes the data is shaped such that it looks a little correlated, but we shouldn't read much into it b/c even the hypothesis saying the data is (normally distributed and) uncorrelated predicts a little observed correlation. Aella's data isn't like that.
1
1
(And tbc, some of her likelihood ratios on the points near the center are small, like 5-to-1ish. And I agree that shouting "sample size!" alone doesn't do much. But some of the LRs she's reporting are *insane*, and I don't think they should be dismissed.)
1
1
Replying to
I don't have access to the source data, but I don't understand how that's possible. She has O(200) people for something like necrophilia. The null hypothesis is that 50% of them are to the left of her median responder, and 50% are to the right.
1
(I know that's not precisely the same as the question she was asking, which instead checked which quadrant relative to (0,0) but it's extremely similar and easier to reason about).
1
It's possible to get an extremely robust conclusion from a sample size of those 200. That would require >70% of them to be on one side. But she said they were weak correlations. 55% of them being right of median isn't enough to reject the null, much less at p<.0000000001
1
1
If you had 10k people in each bucket, then yeah, 52% of a bucket being right-of-median would be extremely unlikely to occur by chance. But the test is only as good as your smallest bucket, and for the super rare ones, that's O(hundreds) at most.
2
Necrophile data (men):
no interest: 14445
slight interest: 573
moderate interest: 213
extreme interest: 281
im tryin to get the divisions by quadrant the data loading just takes forever
1
3
authleft
no interest: 4195
slight interest: 61
moderate interest: 29
extreme interest: 43
authright:
no interest: 618
slight interest: 31
moderate interest: 15
extreme interest: 19
libleft:
5811
225
81
95
libright:
2588
91
33
44
disclaimer: i wasnt super neat with the data cause sample was huge; also ignored centrist responses
1
1
also i weighted these 1-4; no interest was 1, mild interest was 2, etc.