Statistical hypothesis testing--how exact are exact p-values?
OBJECTIVES AND BACKGROUND: When testing a hypothesis statistically, a principle is generally accepted that exact p values shall be stated in the treatise. Researchers have the choice of many statistical computer programmes with implemented hypothesis tests. Are exact p values calculated in the same statistical tests by diverse statistical programmes identical? METHODS: The respective zero hypothesis were tested in 5 artificially created data sets by the parametric unpaired t-test, non-parametric Mann-Whitney test, two-tailed F-test. The calculations were carried out by the following programmes: Statistix, version 7.1 (source www.statistix.com), Analyse-it, version 1.62 (source www.analyse-it.com), MedCalc, version 6.14 (source www.medcalc.be). The p values in the same tests were mutually compared. RESULTS: All three programmes calculated identical exact p values for the t-test. In the remaining two tests in case of 26 out of 44 calculations (59.1 per cent; 95 per cent confidence interval 43-73 per cent) different p values were calculated. The greatest difference was 18.35 per cent. In two cases the values oscillated about 0.05 and this fact caused essentially different interpretation of results. CONCLUSIONS: Using the significance test in the biomedical research has been subject to criticism for a longer period of time. The testing of the zero hypothesis on the arbitrary significance level of 0.05 should be substituted by other methods. Our discoveries should undermine the ungrounded belief of the users of statistical tests--physicians in ununderminable accuracy of mathematical procedures. The use of confidence intervals deems much more suitable although there are objections against them as well. (Tab. 4, Fig. 1, Ref. 19.).