2 ms·
Interesting thought, can you talk more about the methods and their flaws? Also what’s the article you’re referring to?
by edgyquant 4y ago
Interesting thought, can you talk more about the methods and their flaws? Also what’s the article you’re referring to?
- RcouF1uZ4gsC 4y agoThe discussion is https://news.ycombinator.com/item?id=32438368#32442794 https://news.ycombinator.com/item?id=32438368#32442794 Here is my comment: This is actually very poor statistics. Take a look at the actual study: https://www.jhep-reports.eu/article/S2589-5559(22)00122-7/fulltext https://www.jhep-reports.eu/article/S2589-5559(22)00122-7/fu... And look at table 3. They are testing 6 different types of PFAS, and only 1 is statistically significant. They need to be using the Bonferroni Correction because they are checking multiple hypothesis. To do that, you divide the required p-value (0.05) by the number of tests (6). If you do the Bonferroni Correction, none of the PFAS is statistically significant. Relevant XKCD: https://xkcd.com/882/ https://xkcd.com/882/ Thus the real conclusion is that the study did not find any statistically significant link between 'forever chemicals' and liver cancer.
- mbreese 4y agoIn the text they mention an FDR correction. In which case you typically don’t use 0.05 as the threshold. Bonferroni isn’t the only way to do that correction, but it a valid method (depending on the pvalue distribution). But with only 6 tests, it’s not really that critical. I usually only think of multiple testing correction in the context of thousands of tests. Given your comment I was expecting something like 5 patients per group. But they had 50 cases and 50 controls. In this context, an uncorrected pvalue of 0.02 is pretty solid.