You are here: CTSPedia>CTSpedia Web>ResearchTopics>ChiVersusFisher (05 Aug 2011, MaryBanach)EditAttach

To me the definitive paper on the subject is Lydersen et al *Stat in Med* 28:1159; 2009, which states that Fisher's exact test should never be used unless the mid-P correction is applied. This paper points out the advantages of newer unconditional tests.

Also see:

http://www.childrensmercy.org/stats/weblog2007/DifferencesBetweenTests.asp

http://www.childrensmercy.org/stats/ask/fishers.asp

We seem to have a direct contradiction below between, “If your sample size is small (or you have expected cell sizes <5), you should use Fisher's Exact test,” from Charity Moore on May 27 versus, “ordinary chi-square tests are generally ‘more accurate’ than ‘exact’ tests” and “the old adage that chi-square tests are not accurate if an expected cell frequency is less than 5 is not correct,” from Frank Harrell on May 22.

This may reflect a schism in the wider field of statistics, although my understanding is that Fisher’s exact test (and other exact methods) are known to be too conservative. Some regard conservatism as a good property, but really it is a form of bias and accuracy is a better goal. So I believe that avoiding Fisher’s exact test is the best recommendation, even if this means using the ordinary chi-square test.

I agree with Frank and Peter that "exact" tests are often less appropriate than asymptotic tests, in particular when the data are not "exact", i.e., rounded or discrete surrogates for continuous concepts (e.g, parity for fertility). In the latter case, "exact" tests may not even be conservative, because the results are "conditioned on the discretization", while the investigator may be rather interested in the phenomenon before the discretization took place. Hence, "exact" tests should be avoided, in particular, when the ties are not exact, because "exact" tests may give the exact answer to the wrong question, rather than an approximate answer to the right question. Wittkowski (1998, *Biometrics* **54**:789–791), Randles (2001, The American Statistician 55:96-101).

Edit | Attach | Print version | History: r2 < r1 | Backlinks | Raw View | Raw edit | More topic actions

Topic revision: r2 - 05 Aug 2011 - 10:25:59 - MaryBanach

Copyright &© by the contributing authors. All material on this collaboration platform is the property of the contributing authors.

Ideas, requests, problems regarding CTSPedia? Send feedback

Ideas, requests, problems regarding CTSPedia? Send feedback