Auswertung von Stichproben normalverteilter,quantitativer Merkmalsgrößen

Estimates for the standard deviation of samples of metrical, normally distributed statistical variables can be determined by a number of different approaches leading to different results. In order to evaluate these approaches, a large number of samples were generated randomly with the help of a computer from a normally distributed population and evaluated in relation to the mean x and the standard deviation s. The hereby generated Characteristic statistical values x and s can then be compared to the Values of the underlying normal distribution. As this study shows, only the position probabilities Pì (normal ranks) derived by Henning and Wartmann deliver standard deviations ture to the expected values. Rosows commonly applied equation P_i = (3i – 1)/(3n + 1) is equally correct. All other methods examined result in systematic statistical deviations (bias) from the expected value. This includes the estimate s = $ sqrt v $, taken from the variance v, which is used almost exclusively. The bias, however, is known due to this study and therefore can be corrected. Should outliers be eliminated, none of the appreaches investigated deliver standard deviations true to the expected values. The treatment of outliers is complex, however, and discussed in the paper. The confidence fpr the estimate of the standard deviation is independent of the method chose, i.e. none of the examined methods has an advantage over the others.