Randomization at the critical value (in the Neyman-Pearson Lemma style) is
the usual solution in case of discretely distributed test statistics. For
two-sided tests, however, the allocation of alpha between he left- and
right-hand sides should be given some care in order to ensure unbiasedness.
M. Hallin.
On 1/09/08 10:54, "Bland, M." <[log in to unmask]> wrote:
> In the past, I have been asked by publishers to review three different
> proposals for introductory statistics textbooks which contained this
> idea. "We should not do t tests for samples with fewer than six
> observations but should always use rank methods." This is the reverse
> of my own view, as the Wilcoxon, Sign, Spearman, Kendall, and
> Mann-Whitney methods cannot produce two-tailed P values less than 0.05
> for any data, all orderings arising in more than 5% of possible
> samples. I did not discover where this misleading idea came from,
> because none of these authors gave a reference and as it was
> confidential I could not ask them. Does anybody know of a published
> work which contains this idea?
>
> Martin
_____________________________________________________
Marc Hallin
Université libre de Bruxelles CP 210
Institut de Recherche en Statistique
ECARES
B-1050 Bruxelles
Belgium
_____________________________________________________
Tel +32 2 650 58 86
Fax +32 2 650 58 99
E-mail [log in to unmask]
_____________________________________________________
|