«1» Rounding error and test of normality:
There is a common statistical test called Anderson-Darling (AD), (see the literature for details).
It is very often used to test the normality of a set of data and is routinely incorporated in common
computer programs. However, the normal distribution is a continuous distribution and thus numerical
values have (theoretically) an unlimited number of decimals while real measurements always carry a
limited, often small, number of decimals.
This paper shows, by simulation, how the test is influenced by number of decimals but also how
adding a small random term improves the result.
(NB that a set of data having a limited number of decimals is formally a discrete distribution and
thus, by definition, not a normal distribution and thus correctly rejected by the test. Although,
such an interpretation is usually too strict and perhaps useless).
(pages: 5, size: 324k, format: doc, language: eng)