[visionlist] analyzing medians

William McIlhagga w.h.mcilhagga at bradford.ac.uk
Thu Aug 5 10:27:42 GMT 2010

Hi Guillaume, good points about the ANOVA. Your suggestion to do an
ANOVA with bootstrapping is a good one.

>There is no loss of data with trimmed means.

There can be a loss of data when you trim. The philosophy behind
trimming is that the data is a mixture of "good" data near the centre
and "bad" or "outlier" data at the extremes; bad in the sense that it
is generated by a process different from that which generated the good
data, and one which you aren't interested in. If that assumption is
wrong, trimming is wrong. (Come to think of it, that probably means my
suggestion to trim & do an ANOVA could be a very stupid one - if the
long tail of the RT distribution is meaningful, trimming would be
precisely the wrong thing to do)

> It is really a bad idea to make a decision about which statistical test to use based on p
> values.

I don't think Todd was proposing this, because it would be very
naughty indeed. I think he was just annoyed that the p values of the
two analyses he tried were on either side of the sacred 0.05, and that
presented him with the problem of which one should be offered up to
the gods of statistics. If 0.05 weren't so sacred, he wouldn't
actually have a problem, he could go with either one, since both
analyses really have arrived at the same conclusion if their p values
aren't very different.

Dr. William McIlhagga
Bradford School of Optometry & Vision Science,
Bradford University
Great Horton Road
Bradford BD7 1DP

Room G23, Richmond tel. (44) (1274) 235957

More information about the visionlist mailing list