The number of statistical tests performed has increased by several orders of magnitude through the use of computers and ready-made statistical packages. The vast majority of these tests neither physically can be, nor ever are, published. Only tests which show a strong statistical correlation between some variables seem to be sorted out for publication. This practice, which can be found in applied sciences such as the social and medical sciences, is investigated and shown in many cases to lead to a considerable increase in the percentage of published false results. A new function to be performed by statistical packages is then proposed which will reduce this percentage. © 1979 IFMBE.