relative rate test

Joe Felsenstein joe at evolution.genetics.washington.edu
Fri Jul 19 12:00:40 EST 1996


In article <4sjta2$poo at nntp3.u.washington.edu>,
Spencer Muse <muse at kurtz.bio.psu.edu> wrote:

>In article <4sgfu9$mcv at nntp3.u.washington.edu>,
>joe at evolution.genetics.washington.edu (Joe Felsenstein) writes:
>|> Actually likelihood ratio tests can also be done with three species,
>|> it is just that one of the key programs (DNAML) happens not to be 
>|> able to cope with three species, for purely silly reasons.  We're 
>|> working on fixing that
>|> in the next major version.  
>
>[useful citations to many statistical relative rate tests done by a many
>people.]
>
>|> No one has ever
>|> specified what one does with the RRT with more than three species, 
>|> nor even how to do the three species case statistically.
>
>I'm not sure what you mean here, Joe. Clearly valid statistical versions
>of the relative rate test exist. Could you clarify?

I stand corrected by your useful citations.  Clearly there *are* lots of
statistically formulated relative rate tests for three taxa.  I note that
often people who tell me they're doing "the relative rate test" are not
doing anything statistical, nor were the original formulations of the RRT.

>|> If you want to know which branch or clade in the tree is responsible
>|> for the rate heterogeneity, one could fit trees that were clocklike
>|> except for having that branch (or all of that clade, alternatively)
>|> evolving at R times the rate of the rest of the tree.  This is not 
>|> easy to do with present-day programs, alas.  Aside from making that
>|> possible, there are very interesting questions about how to test 
>|> which branch (clade) it was that was going R times as fast.
>|> 
>|> But however that is to be done in a likelihood framework, the RRT has
>|> more problems as it cannot tell you how to combine all the 
>|> three-taxon tests.
>
>Agreed. BUT, that does not imply that doing many or all pairwise
>comparisons is a useless thing to do. And, in fact, many of the tests
>_are_ independent (this can be argued along the lines from Felsenstein's
>1985 (?) article on independent contrasts). Patterns arising from
>exhaustive pairwise tests are often amazingly strong; so strong that it
>would take incredibly high levels of correlation to explain them.

But when they *aren't* that strong you are at a loss what to do.

>But I
>will reiterate Joe's comment that it isn't at all clear how the
>collection of results should be properly interpreted. Rejection of a
>global clock test should certainly be a prerequisite for looking further
>at individual tests, reminiscent of so-called "F-protected" pairwise
>comparisons in traditional ANOVA, where differences among particular
>treatments aren't checked for unless a global F test indicates real
>treatment differences. 

Certainly we need these multiple-tests procedures, analogous to
contrasts among effects in analysis of variance.

>On a cynical final note, I do find it a bit intriguing that so much
>concern is given to multiple correlated tests in one paper, but the
>problem is almost completely ignored when a result from one paper is
>followed up (in a correlated way) in other papers. (This is meant to be a
>general comment, not poined at the rr test problem in particular.)

Here I am less clear where the correlations come from.  If you look at a
different locus in the second paper, how is that result correlated with
the previous one?

-- 
Joe Felsenstein         joe at genetics.washington.edu     (IP No. 128.95.12.41)
 Dept. of Genetics, Univ. of Washington, Box 357360, Seattle, WA 98195-7360 USA



More information about the Mol-evol mailing list