Timeline for Three important issues
Current License: CC BY-SA 3.0
17 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Jul 25, 2019 at 10:59 | comment | added | kjetil b halvorsen♦ | @Nick Cox: Fisher may have used fixed levels out of necessity: He relied on tables which only contained a few levels. Also (out of memory, can't remember reference) he have metioned that one cannot trust levels beyond 0.01 (or even 0.01) because in practice, our theoretical distributions are approximations, and their quality deteriorates in the tails. | |
| Feb 1, 2014 at 12:57 | vote | accept | user 31466 | ||
| Dec 4, 2013 at 17:10 | comment | added | Nick Cox | Quite so. Wikipedia on statistics varies from very useful to utterly appalling, but since I lack the inclination and the energy to get involved, I should not complain. | |
| Dec 4, 2013 at 17:02 | comment | added | Flask | @NickCox I like wikipedia because it encourages the reader to "trust noone". That said the page I linked contains a large collection of citations to original work and historians (Stigler, Zabell) that I find useful. | |
| Dec 4, 2013 at 16:58 | comment | added | Nick Cox | Thanks. I tend not to get my history from Wikipedia but from the originals or the historians, so I have not looked at that source. Karl Pearson was one of many to use P before P-values, as it were. | |
| Dec 4, 2013 at 16:48 | comment | added | Flask | @NickCox The NHST wikipedia page has some sources on early use of significance tests. It appears that when used by e.g. Laplace the null hypothesis was considered to be a plausible research hypothesis and the first to use the "no relationship" null hypothesis was Karl Pearson, who did not seem to realize (or at least never mentioned it) that this "flipped" the logic of the test with important consequences. | |
| Dec 4, 2013 at 16:39 | comment | added | Nick Cox | I guess we agree on all substantive and substantial points. It's a matter of trying hard to be (a) correct and (b) comprehensible to learners at the same time! Significance tests in moderately broad senses go back to the 18th century. On Fisher and Neyman as both saying one thing and often doing the opposite, see e.g. D.R. Cox. 2006. Principles of statistical inference. Cambridge U.P. p.43. I was just pointing out that the wording "signal-noise ratio" is anachronistic, but equivalent ideas can be found, no doubt. | |
| Dec 4, 2013 at 16:34 | history | edited | Flask | CC BY-SA 3.0 |
clarified statements about null hypothesis
|
| Dec 4, 2013 at 16:13 | history | edited | Flask | CC BY-SA 3.0 |
added additional cite
|
| Dec 4, 2013 at 16:09 | comment | added | Flask | @NickCox see the middle bolded portion of the fisher quote where he speaks of "fluctuations". Also, I have seen mention of Pearson using significance testing in the Fisher sense but never seen this paper, if you have a source it would be appreciated. The applied work I have seen by was an attempt to distinguish between two plausible scientific hypotheses. | |
| Dec 4, 2013 at 16:03 | history | edited | Flask | CC BY-SA 3.0 |
added quotes and additional commentary
|
| Dec 4, 2013 at 15:02 | comment | added | Flask | @NickCox I agree but would add that unless the null hypothesis is determined by theory or conventional wisdom (e.g., telepathy does not exist at all, good batches of beer have specific gravity X) It is always foolish to perform a significance test rather than do parameter estimation. | |
| Dec 4, 2013 at 15:01 | comment | added | Nick Cox | Although the question obliges you to simplify, drastically, your historical comments have to be taken with a pinch of salt. Fisher in practice often used, or appeared to use, fixed-level testing even while in broad discussions he also denounced it as naive or worse, while Neyman and Pearson's practice was also wider than their theoretical papers implied. I doubt that you can find Fisher writing of "signal-noise ratio". | |
| Dec 4, 2013 at 14:57 | history | edited | Nick Cox | CC BY-SA 3.0 |
edited body
|
| Dec 4, 2013 at 14:56 | comment | added | Nick Cox | I'd suggest that 1) is a little too loose. It's true that the null hypothesis gives a reference level, but it is just foolish to use a reference level that is not of interest or (a different but related matter) to use a significance test when the science is better served by something else (notably, a confidence interval). | |
| Dec 4, 2013 at 14:40 | history | edited | Flask | CC BY-SA 3.0 |
typo
|
| Dec 4, 2013 at 14:32 | history | answered | Flask | CC BY-SA 3.0 |