Schedule

In many practical applications, a description of the conditional independence structure does not directly address the substantive questions raised by researchers. For example, a researcher may be concerned with trying to discover whether one variable (e.g. cellular phone usage) has a causal influence on another (e.g. risk of cancer), particularly in contexts where it is not possible to carry out randomized controlled experiments.

Such an analysis faces considerable problems: first, there may be many different causal models that are compatible with a given conditional independence structure 'correlation is not causation'; second, in most situations there may be many causally relevant quantities that have not been measured (often called 'confounding variables'). The first problem poses difficulties for any approach which begins by considering a particular causal model: compatibility of the hypothesized model with data in no way precludes the existence of causally different, but statistically equivalent models, from which one might draw radically different causal conclusions. The second problem presents difficulties for traditional methods, such as regression, which typically assume that there are no unmeasured confounding variables.

Directed graphs are a natural language for exploring such questions since directed graphs can be used to represent a causal structure (X -> Y iff 'X is a cause of Y') and conditional independence structure, as a graphical Markov model.

The purpose of this seminar is to investigate different assumptions relating causal relations and conditional independence relations and further, to examine the way in which, given these assumptions, causal structure is underdetermined by conditional independence structure.

INVITED SPEAKERS