Circularity & testing.

Tom DiBenedetto tdib at UMICH.EDU
Sun Oct 20 21:28:47 CDT 1996

On Sun, 20 Oct 1996 16:02:44 -0700 (PDT), James Lyons-Weiler wrote:

>You are obfuscating the point that the algorithm will display pattern when
>there is no pattern to be displayed.

What? Thats ridiculous. I think you mean "the algorithm will display
a pattern generated by random data as well as a pattern generated by
non-random data". Pattern is pattern; if random data generates a
pattern, then it generates a pattern. Either the data aint random, or
you have an artefact, but you cant deny it is a pattern. Once again,
the intrpretation of the pattern is up to you, and will be heavily
influenced by the decisions you made in putting together the

>> ???not necessarily. If the new taxa have sets of characters which are
>> fully congruent with the pattern emerging from the original analysis,
>> then the homoplasy ratio goes down. If their character sets have a
>> similar amount of incongruence as does the original analysis, the
>> ratios remain about the same. If they have more incongruence than the
>> "average" in the original set then the ratios go up. Ratios are
>> pretty simple things. None of this is relevant anyway,,the ratios are
>> only used to compare alternative topolgies for the same matrix, they
>> are meaningless when compared across two different matrices.
>Yes, necessarily.  This has been demonstrated by Donoghue and Sanderson..

What is this "demonstrated by D & S? This has been understood
absolutely throughout the history of the use of these indices.

>Homoplasy indices show an across-taxon, among study, very general trend:
>when more taxa are added, the congruence goes down.  The ratios are
>supposed to summarize congruence.  You have not addressed that.

Well, you obviously place great stock in meaningless generalizations.
No cladist I have ever known has ever used these indices in
across-study comparisons, and I cant imagine how any artefacutual
behavior of those indices in other circumstances makes one whit of
difference to anything. They arent even used in the algorithm to
assess the differnces between alternative topologies, the counting of
steps being a perfectly adequate value for this purpose. The indices
are pretty useless, they seem to be more of a sop to the numerically
inclinded. Would the practice of cladistic systematics be changed in
the slightest if noone bothered to calculate or report such numbers?

>The congruence which you call as test is nothing more than reciprocal
>illumination, referred to by Wiley (again, p. 139) when referring to "the
>idea that truth is approached aymptotically, that is, by testing and
>retesting in a system of reciprocal illumination".  A tad out of context,
>so I'll remind you that what Wiley was in fact discussing was how to
>address the hypothesis of homology.  It is not a specialized term
>referring only to patterns found to be common among research programs.

Well let me ask you this. Have you ever READ the Hull 1967 paper
which Wiley cites, and which you earlier cited as an example of how
we cladists have been warned for 30 years about this obvious
circularity? Let me quote you from the conclusion: "The purpose of
this paper has been to show that the "tangle of circular arguments"
referred to in evolutionary taxonomy by Sokal and Sneath, once
untangled are not circular at all -perhaps unwarrented but definitly
not circular. The criticism that the reasoning of the evolutionists
is circular stems from several misunderstandings, the most important
concerning the logic of discovery. What Hennig referred to as the
method of reciprocal illumination is nothing new in science....".
(I'll add here that Hennig's use was more in line with the "patterns
across research programs"  use).  I dont know how much more of the
time of this group we should spend on this; I'll just say that I am
massivly unconvinced that there is any circularity in the test of
congruence. It is simply a test of a set of hypotheses under an
expectation of congruence. Those hypotheses which pass the test are
further corroborated, and since we have no further tests in pure
systematics, are provisionally accepted.  Those hypotheses which fail
the test are modified with ad hoc hypotheses so that they are
congruent. Its really that simple.

>Induction is required when you generalize about proponderant patterns.

No, there may be induction in the general inference of taxic
relatinships from character realtionships, but the choosing of the
pattern by the algorithm is not a generalization, and is not

>Probabilities are formalized statements (observed or deduced) of events.
>This is missing from the ancient art of phylogenetic argumentation.

and rightly so.

>The fact remains that the preponderant assessment of homology (loose
>definition) may be erroneous, and yet result in a preponderant pattern.

Which is both OBVIOUS, and has nothing whatsoever to do with the
points you are making.

>> >>
>> >As I have responded twice now, that's not much of a test, expecially when
>> >the reason why you might end up abandoning correct character generality
>> >statement is because other character generality statements are erroneous.
>> Well, perhaps we are making progress; "not much of a test" is better
>> than "no test at all"! All that you are now demanding is some test
>> which will prevent us from ever being wrong! Hah!
>I'll make my own arguments, thank you very much.  I'm not demanding
>anything; I'm simply pointing out "here is a limitation, and here is a
>statistical solution".

Oh, statistics is going to show me the truth? It is going to somehow
bridge the inherent gap that we humans confront between our best
inferences and the absolute truth? I think not.
Tom DiBenedetto
Fish Division
Museum of Zoology
University of Michigan

More information about the Taxacom mailing list