A number of statisticians, including Grizzle (1967) and Conover (1974), have argued against using a continuity correction on the grounds that it results in an unwarranted loss of power of the test.

However Fleiss (1981) and others have argued in favour of the correction on the basis that it brings probabilities obtained from the t-test (and the chi squared test) into closer agreement with [conventional] exact probabilities derived from Fisher's exact test. Fleiss argues that, since the researcher always uses the marginal frequencies to estimate expected proportions, this is equivalent to considering marginal frequencies as fixed - hence the appropriateness of Fisher's exact test based on the hypergeometric probability distribution.

The conventional wisdom at present is that the continuity correction should be used - although a strong case can be made to regard *P*-values calculated without the correction approximate to exact mid-*P*-values.