[R] Is there a non-parametric repeated-measures Anova in R ?

[Frank E Harrell Jr]

On 06/17/2010 07:12 PM, David Winsemius wrote:
>
> On Jun 16, 2010, at 1:43 PM, Tal Galili wrote:
>
>> Hello Jeremy,
>> Thank you for replying.
>>
>> I came across friedman test (I even wrote and published R code to easily
>> perform a post-hoc analysis of friedman
>> test<http://www.r-statistics.com/2010/02/post-hoc-analysis-for-friedmans-test-r-code/>
>>
>> ).
>> But what I am after is *multi-way* repeated-measures anova. Thank you for
>> your reply which allowed me to clarify my intentions.
>
> Many years ago I remember reading advice in Conover and Iman's
> "Practical Non-Parametric Statistics" that one could apply a rank
> transformation to the dependent and independent variables and then run a
> typical anova test. This is probably inferior in many ways to doing
> quantile regression (don't know if this has a repeated measures
> extension) or to the use of robust standard errors for examining
> inferential issues in regression models, but it certainly represents a
> useful consistency check when all you are worried about is influential
> points in a skew distributions. I cannot comment on how it would
> theoretically behave in a repeated-measures analysis, but I suspect that
> there are readers of this list who can comment with greater authority,
> and I invite them to do so.
>

David - the rank transform method doesn't handle interactions properly, 
among other problems.  The proportional odds model is the logical 
extension of the Wilcoxon-Kruskal-Wallis approach.  It relies only on 
the rank of Y and reduces to the regular nonparametric tests as special 
cases.

Frank
-- 
Frank E Harrell Jr   Professor and Chairman        School of Medicine
                      Department of Biostatistics   Vanderbilt University