The statistics in the Pras & Guastavino paper seem a bit suspect. They do a lot of significance tests with no multiple comparisons compensation, and find few st

The statistics in the Pras & Guastavino paper seem a bit suspect. They do a lot of significance tests with no multiple comparisons compensation, and find few statistically significant results. The problem is that with enough testing, you're very likely to get statistically significant results just by chance.