Premium
What's normal anyway? Residual plots are more telling than significance tests when checking ANOVA assumptions
Author(s) -
Kozak M.,
Piepho H.P.
Publication year - 2018
Publication title -
journal of agronomy and crop science
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.095
H-Index - 74
eISSN - 1439-037X
pISSN - 0931-2250
DOI - 10.1111/jac.12220
Subject(s) - normality , analysis of variance , residual , statistics , homogeneity (statistics) , variance (accounting) , mathematics , statistical analysis , normality test , statistical hypothesis testing , mixed design analysis of variance , econometrics , algorithm , accounting , business
Abstract We consider two questions important for applying analysis of variance ( ANOVA ): Should normality be checked on the raw data or on the residuals (or is it immaterial which of the two approaches we take)? Should normality and homogeneity of variance be checked using significance tests or diagnostic plots (or both)? Based on two examples, we show that residuals should be used for model checking and that residual plots are better for checking ANOVA assumptions than statistical tests. We also discuss why one should be very cautious when using statistical tests to check the assumptions.