Transformational and taxic approaches to character coding [was Re: Philosophies for Character Ordering]

[jonathan.r.wagner@mail.utexas.edu]

Okay, the following is a very biased, potentially misleading, probably worthless
rant based on a paper I will probably never write. Accept or reject at your own
risk.

Naturally, any percieved personal offense is truly unintentional.


Mickey wrote:
"Wilson, 1999- "Ordering these implies a developmental model in which
vertebrae and phalanges are added or lost incrementally.  Embryological data
from living organisms, however, do not support this transformational model."

"TRANSFORMATIONAL" CHARACTER ANALYSIS

Wilson appears to be taking an explicitly transformational approach to character
coding, in which he uses development as a model of evolutionary change. While
it is true (at least, for vertebrates) that every heritable morphological
change entails a change in development, I am not sure that this means anything.
Naturally, since development is change in an individual organism through time,
phylogenetic change in morphology (which remember, happens ACROSS or between
the lives of organisms, not within them) MUST change development. Is this
really saying anything new?

Some transformationalists (perhaps many) seem to be looking for the Holy Grail
of morphology, the rules for how morphology changes. Naturally, they look to
the transformation they CAN observe, development. Now, if we could get such
rules, maybe they would help, but then again maybe it wouldn't. Incorporating
such rules into analysis does not seem to me to be straightforward. Maybe it is
just because we DON'T have those rules, but I can't really picture how it could
be at all objective.

I must confess I am not a developmental biologist, but in reading about the
discipline, I am always impressed with how flexible and plastic it is, at least
at the scale I am interested in (piddly differences among vertebrates). Given
this perspective, it is unclear what role development should play in character
coding. For my money, it makes more sense to make fewer assumptions about what
CAN'T happen, and allow for more that CAN happen, given an observed similarity
a priori of analysis.

"TAXIC" CHARACTER ANALYSIS

The other major approach to character analysis is the taxic approach. In brief:
characters are observations made among the terminals, and transformations (in
the sense of transformational homologies, not inferred character state
transformations) need not be limited by niggling about homology or worrying
development a priori of analysis. From a taxic perspective, character ordering
is not a model of developmental (or evolutionary) change, but a convenient way
of summarizing the information contained in several, non-mutually exclusive
characters. Why code ilium: 0) no processes, 1) ilium with postacetabular
process AND ilium: 0) no process of postacetabular process only, 1) ilium with
preacetabular and postacetabular process, when you can code them as one
three-state ordered character and get the same effect (assuming you don't have
taxa with a PreAc and no PostAc!). It isn't a model of how you think it
evolved... it is data reduction, and it has the same effect in a parsimony
analysis. Similarly, ordering numbers of phalanges does not represent an
explicit model of development or evolution, but it recognizes the potential for
"two or more digits," "three or more digits" or "four or moredigits" to be just
as synapomorphic (or symplesiomorphic) as "one digit" and "five digits."

That's not to say that anyone (I hope) would necessarily approve of ordering
characters willy-nilly. For my money, characters should be unordered if their
states are mutually exclusive (i.e., they contain no overlapping information).
Characters should be combined into a multistate if some of the states partially
overlap, but don't have to be (e.g., "greater than five" overlaps partially
with "greater than seven"). Remember, a simple, ordered multistate can be
easily broken down into binary characters.

Where do those states come from? Observation! How do you decide which to use?
Total evidence! What about convergence/development/you ad hoc hypothesis about
the group? Go with the data!

The taxic perspective is not just about character ordering, but I'll try to stay
on track...

WHY ONE MIGHT PREFER TAXIC/ ESCHEW TRANSFORMATIONAL

Note that, as implied by Mickey's discussion, character ordering is IMPLICITLY
present in some studies that purport to use only unordered characters. Consider
this character: number of cervical vertebrae: 0) seven to ten, 1) eleven to
thirteen, 2) fourteen or more. When unordered, the presence of eleven to
thirteen vertebrae could be the only character diagnosing a clade. But, if you
believe that we should not impose ordering on characters, how do you justify
grouping eleven, twelve, and thirteen vertebrae as the same state to begin
with? What do those particular morphological conditions have in common besides
being within a range of numbers? You effectively ordered within the
morphologies at each state (granted, with a transformation cost of 0), but not
between.

Any time a cut-off value is given (great-than/ less than), or a relative measure
(long, short), you are grouping "different characters" (under a
transformational paradigm) under one state. This is ordering, it is "hidden"
because you only coded two states!. This is often done semi-deliberately, by
taking an "obviously" multistate character (as if characters existed in nature
for us to scoop up), and breaking it into two or more discrete characters so
they will retain their information value in the analysis.

As alluded to (by Jaime, I believe), unordered characters have the unsettling
effect of allowing intermediate (non-exclusive) states to be apomorphic for
taxa. Why would I ever consider having a femur of "medium length" to diagnose a
clade, and separate it from a clade containing of long-legged or short-legged
critters? This is analogous to how we got paraphyletic taxa in the old days! In
the vertebrae example, that character could diagnose a clade to the exclusion
of neighboring nodes diagnosed by the other two states, thus having a clade
diagnosed by 11-13 cervicals in an "unordered" analysis! Does it make sense to
you that having three digits should unite all coelurosaurs to the exclusion of
both ceratosaurs and tyrannosaurs?

One interesting aspect of transformationalism is the use of ad hoc arguments,
often based on development or presumed character correlation, to avoid coding
similar morphology as the same state. See some of the arguments about snake and
turtle evolution; little differences are emphasized, and the taxa are given
different (unordered) states based on them. This a priori obviates the
possibility of homology; it strips that character of the power to unite the two
taxa in question. If one takes such an approach to character coding, one will
inevitably end up missing information that is informative about really deep
nodes in the tree. Such information is carried in similarities which have
changed the most, due to the long divergence times; the detailed morphology
often differs greatly among taxa. In this case too, developmental differences
are often employed, amounting to the assumption that development cannot evolve.
However, to take what I consider a transformational point of view: which part
of the phenotype seems more likely to be selected for, the developmental
process, or the product?

Ultimately, transformationalism emphasizes differences: DIFFERENT morphologies 
get DIFFERENT character states (and we must ask our analysis to show us the
direction of their transformation from one to the other. This also implies that
characters somehow have their own reality beyond their status as observations.
A taxic approach, quantifying and coding similarities among organisms without
recourse to ad hoc developmental models, emphasizes SIMILARITIES (which, after
all, are what we are after, right?). The _reductio ad absurdam_ of unordered
characters is a data matrix containing only one character ("morphology"), with
a different state for each taxon ("morphology of _Velocisaurus unicus_," etc.).
I am sure you can imagine how that tree looks!

OBJECTIONS TO THE TAXIC APPROACH

I am perhaps not the best person to write in support of the transformational
ideology, because I (quite obviously) do not subscribe to it.

Some might argue that the transformational approach is the one explicitly
employed in molecular sequence studies and other more "biological" analyses,
and therefore better. I don't see it this way... I feel that modern,
probabalistic methods account for additional information beyond a strict
transformationalist interpretation of the data (e.g., "just A, C, T, and G"),
but they take advantage of the biological comparability of these data to
incorporate additional information as global parameters rather than characters
or parameters of the characters already there. The overall effect is as much
taxic as transformational, because additional information about similarity
among character states in groups is used.

Taxic approaches may be associated with pattern cladistics by some people
(amusing, considering transformationalism's emphasis on developmental data).
This may be because both approaches emphasize minimizing assumptions about
evolutionary processes, and because the taxic approach is apparently advocated
by some pattern cladists. However, while pattern cladistics proceeds without an
a priori evolutionary explanation, that is not necessarly part of a taxic
approach. I very deliberately set out to estimate/ reconstruct phylogeny,
assuming that evolution did in fact occur (I feel that's a pretty parsimonious
assumption).

A complaint I have heard is that analysis should be an explicit model of
evolution. I must confess that I am not an expert in the philosophy of science,
but I feel that this something of a misrepresentation of what a model is in a
statistical sense. My feeling is that the taxic approach is a model of
evolution, in the sense that you can model dune migration as stacked bricks, or
you can model the ecomony using complex statistics, or you can model human
behaviour using a deck of playing cards and a wire hanger. The mechanics of the
model need not directly model the evolutionary process in great detail; the
measure of a good model is not whether it catches all the details of the
system, but whether its predictions match the observed pattern (in the case of
phylogeny, this works a little differently, but the idea is the same). The
taxic philosophy, to my mind, makes fewer assumptions about the nature of the
process to be modelled, which fits both with the generally gross character of
the model and our lack of understanding of the processes.

THE argument I see the most is against ordered characters: "you shouldn't force
evolution to go a certain way." Now, if we interpret the model literally, yes,
ordered characters force passage through both states. However, since states are
just (hopefully informed) lumpings of different morphologies, evolution would
have had to pass through a (unknowable) number of "states" to get from A to B
anyway. All ordering does is give an intermediate state, based on similarities
observed in the terminals, some power to group taxa. Alternately, if we look at
the model as simply a way to match the pattern of evolution (which it does by
reconstructing transformations, but that's another argument), then the actual
path each character follows is subordinate to the resulting tree anyway.

A LAST WORD

Probably no one actually spends all their professional time at either extreme. I
think about transformations all the time; they often limit where I look for
characters, and they probably subconsciously influence how I code characters.
As noted above, taxic ideas sneak into transformationalists' work too. I am
pretty sure there cannot be a "right way" to code characters, and I doubt we'll
ever see a formulaic approach to coding morphology. Transfomationalist datasets
"feel" better, because they appear to be making fewer assumptions (are they?).
Taxic thinking can ratchet your character search off of ideological "islands"
(or ruts) by suggesting characters you wouldn't have noticed. I've found that
looking for characters that violate your favorite hypothesis, which I
personally would call taxic thinking, can have a similar effect. Many people
have suggested to me that you should run your data both ways (usually meaning
"ordered and unordered") to see if it makes a difference. I would caution you,
however, to beware of hidden ordering!

There. I need to home!

Wagner