Re: Is the t test good enough?

From: Bruce Weaver (weaverb_at_mcmaster.ca)
Date: 08/05/04 

Date: Thu, 05 Aug 2004 14:36:24 -0400

MB Wagner wrote:
> Hi Everyone
>
> I wonder why some referees in biomedical journals argue that the
> Student&#8217;s t test should not be used with small samples (n<30 or
> n<20)? As far as I know the t test was developed by W.S. Gosset (a
> brewer at Guinness) with the help of K. Pearson and R. Fisher
> precisely to be used with these small samples. Every statistics book I
> know state that the t test is robust enough to stand even moderate
> deviations form normality and its major deficiency would be in
> handling heterocedasticity for which there is a solution
> (Welch&#8217;s t&#8217; test). So, my question is: why should we use
> non-parametric tests to analyze continous data in small samples (say
> n<20) that do not seem to be skewed?
>
> Many thanks,
>
> MB Wagner.

I guess the reason is that some referees have a very limited understanding
of statistics, and so rely on dogma.

As Rich Ulrich (among others) has suggested many times in these newsgroups,
the rank-based non-parametric tests have distributional assumptions that
are nearly as strict as for the corresponding parametric tests. Many
people do not know this, and assume that rank-based tests can cope with
just about anything. 

-- 
Bruce Weaver
weaverb@mcmaster.ca
www.angelfire.com/wv/bwhomedir/