ABSTRACT
If we analyze the data with the aid of classical statistical procedures, based on parametric models, we usually tacitly assume that the regression is linear, the observations are independent and homoscedastic, and assume the normal distribution of errors. However, when today we can simulate data from any probability distribution and from various models with our high–speed computers and follow the graphics, which was not possible before, we observe that these assumptions are often violated. Then we are mainly interested in the two following questions:
When should we still use the classical statistical procedures, and when are they still optimal?
Are there other statistical procedures that are not so closely connected with special models and conditions?
