John E. Overall
Abstract:Fisher's exact probability test is severely conservative when interpreted with reference to conventional alpha levels due to the discontinuity of the sampling distribution for 2 × 2 tables. An adjustment of the cell frequencies is proposed that results in a correction for continuity with appropriate alpha protection and increased power.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is that the Fisher's exact test is too conservative when dealing with 2x2 contingency tables. Specifically, due to the discontinuity of the sampling distribution, when compared with the traditional significance levels (such as α = 0.05), the actual significance level of the Fisher's exact test is often much lower than expected, resulting in a decrease in the power of the test. The author proposes a method of adjusting cell frequencies to achieve continuity correction, thereby increasing the power of the test while maintaining appropriate α - level protection.
### Background and Problem Description
The Fisher's exact test is a hypothesis - testing method for 2x2 contingency tables, used to determine whether the response probabilities between two treatment groups are equal. However, when using the conventional significance levels (such as α = 0.05), the Fisher's exact test is usually too conservative, that is, the actual significance level is much lower than the nominal α - level. This conservatism leads to a decline in the power of the test, especially more obvious in the case of small samples.
### Solution
The author proposes a method of adjusting the cell frequencies in 2x2 contingency tables to achieve continuity correction. The specific adjustment methods are as follows:
- If \( A\times D>B\times C \), add 1 to each of cells \( A \) and \( D \).
- If \( B\times C>A\times D \), add 1 to each of cells \( B \) and \( C \).
Through this adjustment, the results of the Fisher's exact test can be made closer to the nominal α - level, thereby increasing the power of the test.
### Experimental Verification
To verify the effectiveness of the proposed continuity correction method, the author conducted Monte Carlo simulation experiments. The experimental results show that the actual significance level of the uncorrected Fisher's exact test at each α - level is much lower than expected, while the actual significance level of the Fisher's test after continuity correction is closer to the nominal α - level and has a higher power of the test.
### Conclusion
The author concludes that the Fisher's test with continuity correction is more effective than the uncorrected Fisher's test and the Yates - corrected χ² test. It can improve the power of the test while maintaining appropriate α - level protection. In addition, this method is particularly important in small - sample studies because it can provide more accurate unconditional probability estimates and is helpful for making more reasonable decisions in practical applications.