Discovery problems: asymptotic reliability if assumptions are correct; produce all graphs compatible with evidence and procedure (same principle as confidence regions, though they don't put it that way).
``Sufficient'' discovery problems: Algorithms which work when we know all causal variables, except perhaps for uncorrelated (independent?) noise.
``Insufficient'': Algorithms which can distinguish the existnece of latent, unmeasured causes, including common causes that affect several variables.
The work-horse of quantitative social science, and even much of the natural sciences, is multivariable linear (``multilinear'') regression. This is actually a very bad tool for detecting causal relationships, and DAG methods are distinctly superior. Discuss their critique of regression.
Critique of the dogma that causation can only be inferred from experiments; suitable observations can be used. But this is not to say that experiment is not important, and they have very nice discussions of the design of experiments, and especially of manipulation.
Excursuses on the causes of smoking, the design of observational studies, etc. A good selection of open problems, at least in the 1st ed...
Proofs relegated to separate chapter. (Still?)
What's new in this edition? One hopes an improvement over the dreadful layout of Springer...
David Hume famously argued that it was impossible to demonstrate a ``causal connexion'' between two classes of events; all that one could hope to show was a ``constant conjunction''. The first to claim this, apparently, and with very nearly the same arguments as Hume, was the great Muslim theologian al-Ghazali, who spoke of ``habits'' of the world instead of constant conjunction. None of the modern work on causal inference, of which SGS is a leading light, purports to refute the philosophers. Instead, and perhaps more interestingly, they accept that all we can really work with are statistical patterns of conjunction, and set out to show how those habits can be recognized amid the background chatter of the world, and put to use. On the perhaps deeper point which does seem to belong to Hume, that all induction, and with it all empirical knowledge, is more or less unreliable, depending as it does on the part of the world we have not examined being like what we have, remains untouched. SGS do not refute Hume, or for that matter al-Ghazali, by providing a way of proving the existence of a causal connection. What they --- and the many others who have shared and are sharing in this work --- is in a way more interesting than a refutation would have been. What these methods do is, in effect, accept that we never really deal with anything more than a statistical pattern of association (a ``habit,'' in al-Ghazali's word), and then go on to show how such patterns can be discerned amid the background noise of the world, and put to use. Of course (and here the point really is due to Hume), even that isn't really guaranteed; we have techniques which are consistent (reliability of discovery in SGS's sense), and whose results we can continue to use --- if things don't change.