FOLLOW-UP OF FISHER F TEST WITH SIGNIFICANTLY LOW VALUES IN SMALL SAMPLES

Fisher’s F-test, a common statistical choice for testing differences between groups, has been found to return unusually small, significant values. This paper presents a follow-up of two examples from medical surveys which registered significantly low F-values when testing for differences between evaluations (marks) given by physicians in different specialties. A new wave of the two surveys has been conducted, but did not register low F-ratios. The datasets of the two waves are compared for specific traits which may have caused the low F-ratios in the original wave. Results confirmed some of the previous findings (e.g. normality should always be tested because non-normal data may “hide” behind normal F-ratios), but showed similar conditions of sample size, non-normality of data and influence of a second variable (locality). The main conclusion of the follow-up is that neither of these causes, alone or acting together, guarantees that significantly low F-ratios will occur. This makes generalisation, or preventing low F-ratios, a difficult, if not impossible task.