I thought this newsgroup post by Dave Krantz was very good.

--------------------------------------------------------------------------
From: 		dhk@paradox.psych.columbia.edu (Dave Krantz)
Subject: 	ancova versus change scores
Date: 		1997/04/16
Newsgroups: 	sci.stat.edu
--------------------------------------------------------------------------

In response to John Reece's question about using pretest scores
as covariates or by calculating change scores:  it seems to me
that the real issue is not one of assumptions nor yet one of power,
but what MODEL approximately describes the STRUCTURE of the responses.

Change scores are a special case of using the pretest as covariate:
in a change score, the pretest is an implicit linear predictor, with
coefficient = 1.  If that is a good model, then by all means, use it.
Another alternative is that log(pre) contributes linearly to log(post)
with coefficient 1--i.e., the change is measured by log post/pre ratio
rather than by difference scores.  Many other models are possible.
Linear covariance analysis uses pretest as a linear predictor
with slope different from 1, usually with the same slope for all
groups (though this is often inappropriate).  If the slope is considerably
less than 1 (as it often is, in my experience) and/or varies between
groups, these are FACTS (approximations) that theories need to deal with.

So the question of of what is the better way to analyze the data
is not really an appropriate question--it almost never is a proper
question.  The right question is, what is the underlying structure,
and how can it be characterized approximately by a parsimonious
mathematical model?

(Burt Alperson is of course entirely correct to point out that
the change-score analysis and the mixed-model anova are identical
analyses.)*

Dave Krantz (dhk@columbia.edu)

--------------------------------------------------------------------------

* In other words, an unpaired t-test on the difference scores for the two
groups is equivalent to the F-test for the Group x Time interaction term 
in a between-within ANOVA.