circularity

Larry Trask larryt at cogs.susx.ac.uk
Fri Aug 13 11:25:38 UTC 1999


On Thu, 12 Aug 1999 ECOLING at aol.com wrote:

[on my baseball `park effects' example]

> Actually, how they work most often is that the investigator has
> plausible evidence for a conclusion, and then AFTER THAT sets up a
> null hypothesis which the investigator already has some reason to
> think is false and can be defeated, and then proceeds to lay out the
> evidence to defeat it.

Oh, sure.  In practice, we are hardly ever able to work through an issue
in the optimal manner I have described.  Real life is inevitably
messier.  We make mistakes and eventually correct them; new evidence
becomes available; new tools become available; all sorts of things
happen along the way.  But how we do the work is one thing, while how we
present our results in order to persuade our colleagues is another.

Obviously the linguists who established the validity of the IE family
did not overtly begin with the null hypothesis `These languages are not
related'.  But, once a sufficient amount of work had been done, the
results were laid out in the handbooks, in the form of phonological and
morphological correspondences which are so pervasive and so systematic
that the null hypothesis is untenable.

We accept IE today, not because Bopp and Grimm thought it was a neat
idea, and not because the IEists have cunningly manipulated the issue,
but because the evidence assembled in the handbooks against
unrelatedness is overwhelming.

> Notice how potentially circular this is, in the way the question
> asked is designed precisely to bolster a previously drawn
> conclusion!

No, I can see no circularity at all.  If we want to show that two or
more things are related, then the only possible null hypothesis is that
they are not related, and the disconfirmation of this null hypothesis is
the only way of establishing the proposed relationship.  There is no
choice of null hypothesis; there is no scope for intellectual
intimidation; and there is no circularity.

The null hypothesis of unrelatedness among the IE languages is massively
disconfirmed by the evidence.  The null hypothesis of no park effects in
baseball is likewise massively disconfirmed.  But the null hypothesis of
unrelatedness among the families assigned to Nostratic is at present
*not* disconfirmed, and that is why few linguists accept Nostratic.
It doesn't matter that there is evidence in support of Nostratic: what
matters is that there is insufficient evidence to reject unrelatedness.

Let me return to my baseball example.  Now, thoughtful baseball fans
have believed for years that park effects were real, and probably
anybody who's ever watched a few games at that pitchers' graveyard
called Coors Field believes in park effects.  However, there exist
baseball fans who do not believe in the reality of park effects.
Even if there were none, it wouldn't matter, since a universally held
belief does not necessarily constitute a truth.

But, until recently, real evidence in favor of park effects has been
sketchy and informal -- not much better than anecdotal: "Batting
averages at Oakland-Alameda County Stadium have been lower than at other
parks in the last few years."  Observations like this one are
interesting, but they don't suffice to disconfirm the null hypothesis of
no park effects.  Maybe OAC Stadium has just happened to see a lot of
pitchers' duels recently, entirely by chance, or maybe the A's have had
great pitchers.

So, what that statistician did was to state the null hypothesis and then
to test it statistically, with one test after another.  The results are
clear: the null hypothesis is massively disconfirmed by the statistical
evidence to a very high level of confidence.  That is, the absence of
park effects could be consistent with the data only in something like
one case in a trillion (I forget the precise number, but it was big).

Larry Trask
COGS
University of Sussex
Brighton BN1 9QH
UK

larryt at cogs.susx.ac.uk



More information about the Indo-european mailing list