|
|
Selected random variables transformations used in classical linear regression
Tejkal, Martin ; Michálek, Jaroslav (oponent) ; Hübnerová, Zuzana (vedoucí práce)
Classical linear regression model and the respective tests are based on an assumption of normally distributed response variables and on an assumption of variance equality. If the normality assumption is not fulfilled, then the response variables are usually transformed. In the first part of this work variance stabilising transformations are discussed. Great deal of attention is given to random variables of Poisson and negative binomial distribution, for which generalised variance stabilising transformations with addition constants in their arguments are studied. Optimal values of the constants for the generalised transformations are determined. The second part aims to provide a comparison of the transformations introduced in the first part and some other commonly used transformations. The comparison is done within the ANOVA framework by testing the hypothesis of equality of expectations among p random samples via F test. The properties of the distribution of the F test under the assumptions of equal and unequal variances are studied. Finally a comparison of the power functions of the F test applied to p random samples from Poisson distribution transformed via square root, logarithmic and Yeo-Johnson transformation, and to p random sample of negative binomial distribution transformed via argument of hyperbolic sine, logarithmic and the Yeo-Johnson transformation is carried out theoretically and by simulations.
|
|
|
Transformace stabilizujicí rozptyl
Kuželová, Noemi ; Omelka, Marek (vedoucí práce) ; Komárek, Arnošt (oponent)
Abstrakt. Mnohdy zkoumáme data, jejichž výběrový průměr konverguje k nor- málnímu rozdělení, jehož rozptyl však obecně závisí na neznámém parametru. K tomu, abychom se této závislosti zbavili, lze někdy využít metodu tak zvané transformace stabilizující rozptyl. Tato práce nejprve metodu detailně vysvětlí a najde obecný postup, jak vhodné transformace hledat. Poté se zaměří na data pocházející z Poissonova a binomického rozdělení s neznámými parametry. Pro tato data najde transformace, jež stabilizují (asymptotický) rozptyl, a porovná je s ještě "vylepšenými" transformacemi z článku Anscombe (1948). Právě tvaru těchto transformací je věnována většina práce. Nakonec na simulaci pro výběr z Poissonova rozdělení ukážeme, že je opravdu vhodné tuto metodu využívat a srovnáme odvozenou transformaci s její Anscombeovou verzí.
|
|
|
Selected random variables transformations used in classical linear regression
Tejkal, Martin ; Michálek, Jaroslav (oponent) ; Hübnerová, Zuzana (vedoucí práce)
Classical linear regression model and the respective tests are based on an assumption of normally distributed response variables and on an assumption of variance equality. If the normality assumption is not fulfilled, then the response variables are usually transformed. In the first part of this work variance stabilising transformations are discussed. Great deal of attention is given to random variables of Poisson and negative binomial distribution, for which generalised variance stabilising transformations with addition constants in their arguments are studied. Optimal values of the constants for the generalised transformations are determined. The second part aims to provide a comparison of the transformations introduced in the first part and some other commonly used transformations. The comparison is done within the ANOVA framework by testing the hypothesis of equality of expectations among p random samples via F test. The properties of the distribution of the F test under the assumptions of equal and unequal variances are studied. Finally a comparison of the power functions of the F test applied to p random samples from Poisson distribution transformed via square root, logarithmic and Yeo-Johnson transformation, and to p random sample of negative binomial distribution transformed via argument of hyperbolic sine, logarithmic and the Yeo-Johnson transformation is carried out theoretically and by simulations.
|
host ::
RSS.