This talk reviews the basis for inferring causation from
regression, proceeding by examples drawn from the medical and
social sciences. Invariance assumptions are required. Parameters
need to be invariant to interventions and so do errors or error
distributions. Exogeneity is a further issue. Thus, causal
relationships cannot be inferred from a data set by running
regressions, unless there is substantial prior knowledge about
the mechanism that generates the data. Such prior knowledge is
seldom available. There is a handout on the web---

http://www.stat.berkeley.edu/users/census/poliscih.pdf