Sir,

We highly appreciate Dr AD Tsakok's interest in our recently published paper.1 In his letter,2 he suggests a different approach to the statistical problem which we solved using either Student's two-sample t-test or analysis of variance (ANOVA). Dr Tsakok argues that the shortcoming of our statistical approach lies in the assumption of equal variances between groups (Behrens-Fisher problem3, 4). In his opinion, this assumption renders the applied tests unsuitable for the purposes to which they were put. Dr Tsakok advocates a statistical test that he himself has developed to compare quantitative data between multiple groups,5 and which is already available as commercial software.

Dr Tsakok endeavoured to prove his point by applying his test to data that appears in several publications.1, 6, 7 According to his findings, we failed to detect significant differences in the total number of breaks and in the best-corrected visual acuity between scleral buckling and vitrectomy 6 months after surgery (Table 31).

We used the widely applied statistical package SPSS 11.5 (Chicago, IL, USA). The data were tested for normality of distribution using the Shapiro-Wilks test, and the equality of variance was confirmed using Levene's test. SPSS computes two-test statistics for the two-sample t-test: one for cases in which the variances in both groups are equal, and the other for cases in which they differ. If the variances differed significantly, we implemented the latter test in conjunction with the relevant significance values. Furthermore, due to the retrospective nature of our study, we stressed that the findings might not be generally applicable.1 According to currently widely accepted standards,8, 9 we are still convinced that the statistical methodology employed in our study was appropriate.

We agree with Dr Tsakok respecting the importance of the Behrens-Fisher problem. According to our literature search, the Tsakok test has as yet neither generally been recognized within the scientific community nor widely applied for the solution of comparable statistical problems. It may well prove to be superior to the statistical tests currently applied to clinical data, but it must first be validated by independent statisticians.