Causal Interpretation and Identification
of Conditional Independence Structures
Schedule - Seminar 1 on CAUSAL INTERPRETATION OF
GRAPHICAL MODELS
September 27 - October 8, 1999
Organized by A. P. Dawid and Glenn Shafer
Abstracts and Programme for the week of October 4 to 8, 1999
We hope that each day will be dominated by informal discussions and
small working groups. Speakers are urged to use their talks to get discussion
going. The talks themselves should be no more than 50 minutes, but the
discussion should be unlimited.
Monday, 10 am
Glenn Shafer on "Causal Logic"
Bayes nets emerged from an effort to make reasoning under uncertainty
as modular as the reasoning in rule-based expert systems. Although they
are very powerful, they fall far short of that modularity. In this talk
I report on my current work on completely modular but rigorous causal
probabilistic reasoning. This involves a logical language that permits
reference to instantaneous events at different levels of specificity
and constructs probabilistic judgments from judgements of the prudence
of overlapping gambles.
Monday, 2 pm
Paul Holland on "Where do counterfactuals hide in statistical models?"
Counterfactual statements involve unobservable causal effects. Statistical
models also involve unobservable quantities--parameters and error terms,
for example. Are these two kinds of unobservable quantities related
to each other? If so, how?
Tuesday, 10 am
Richard Scheines on "The Causality Lab"
At Carnegie Mellon we have been developing web-based software for teaching
causal reasoning with statistical data: the Causality Lab. Students
are confronted with a list of variables, and stored behind the scenes
is a causal structure (Bayesian Network). They can then set-up experiments
by randomly assigning values to a variable, collect data, form causal
hypotheses, make predictions about independence, test their predictions,
etc. By systematically removing experimental capability until students
can only collect purely observational data, the limits of what can be
learned become apparent.
Tuesday, 2 pm
Glenn Shafer on "Mediators"
Adjusting for a mediator (a covariate affected by the treatment) is
treacherous to do but frequently done. In this talk, I suggest that
such adjustment makes sense only if (1) the mediator is defined precisely
and (2) has a consistent effect in other contexts or is normative. I
hope participants in the discussion can help me square this formulation
with their way of thinking about this issue.
