"Jones Philip R" <[log in to unmask]> wrote:
>> Can anyone point me towards a non-parametric equivalent of a standard 2-way
>> Analysis of Variance. The data has 2 factors (each with 4 options) and each
>> of the 16 (4x4) combinations of the 2 factors has 4 observations.
>>
>> If not, thoughts on the applicability of using the parametric test on
>> ordinal data would be appreciated
>>
>> Unfortunately, Siegel and Castellan (Non-parametric Statistics for the
>> Behavioural Sciences) only offers the Friedman test - which is only
>> applicable for matched samples - and in effect only provides a 1-way
>> comparison
Such an anova is pretty uninformative anyhow. Finding that a factor with
four levels has a significant association with your outcome variable doesn't
tell you much of use.
Do the factors have a natural 'baseline' or 'reference' category against
which you could contrast the other categories using regression dummies?
And is your outcome variable incapable of being transformed to something
normal-ish? You can always run robust regression to confirm your findings.
If it's genuinely ordinal, consider ordered logit regression.
I am always reluctant to use anova on the grounds that I can't really
interpret the results of any test that has more than one df in practical
terms.
Ronan M Conroy ([log in to unmask])
Lecturer in Biostatistics
Royal College of Surgeons
Dublin 2, Ireland
+353 1 402 2431 (fax 2764)
--------------------
|