Reply to Tom, Brian, and Martin about Dunn et al

[T. Florian Jaeger]

Hi

Brian, the Dunn et al article is problematic both in terms of the
theoretical interpretation (as you pointed out) and in terms of the relation
between what they *claim* to test and what they actually test. That latter
problem is addressed in detail in Bill's contribution to the LT commentaries
(Croft, Bhattacharya, Kleinschmidt, Smith, and Jaeger, to appear). I think
you will find that this commentary is helpful in understanding what Dunn et
al do and do not show.

One thing that follows from that critique of Dunn et al is that we should be
cautious to accept the claimed lack of cross-family correlations for many of
the features. So, just because Dunn et al find two word order features (e.g.
adj-noun, gen-noun) correlate in one language family but not the other, it
does not mean that there is actually enough evidence to reject the
hypothesis that the correlation is equally strong across language families
(Dunn et al do not test this hypothesis even though it is arguably an
important test before one rejects the hypothesis that the correlations are
identical across language families).

Another things (in line with Martin's comment) that we point out in the
Croft et al commentary is that Dunn et al did not control for language
contact as an explanation for change. This will likely make the observed
correlations between different word order features weaker than they would be
if language contact was taken into consideration in the model (as language
contact adds a source of 'noise' that is unaccounted for by the Dunn et al
model).

Bill, please correct me if I am misrepresenting the commentary =).

I think the point that Tom, Brian and Martin made that language-specific
properties addition to those accounted for in the model  (e.g. modifier-head
number-gender agreement along with fusional case-marking) might affect the
likely diachronic development of languages is a great one. If Dunn et al
(or, for that matter, anybody) has enormous amounts of data, this shouldn't
matter as such differences between languages would get 'averaged out', but
given available data, such between-language difference, if unaccounted for,
are likely to   be a problem for their analysis (again leading to an
under-estimation of the stability of whatever biases and processes cause
gradient universals).

Florian