Causal probability (Q1868154)

Conventional Decision Theory takes the expected value of an act \(A\) having outcomes \(0\) in the set \(\mathbf{O}\) as \[ \mathbf{EV}(A) = \sum_{O \in \mathbf{O}}{{\mathbf{PROB}}(O/A)}. \] This doesn't always work. For example, suppose your choice is between smoking and not smoking, but that both a propensity to cancer and the desire to smoke are the result of a single genetic predisposition, so that smoking is a symptom of this predisposition, and thus of a propensity to cancer, but does not cause cancer. Pollock departs from the usual subjective interpretation of probability (for good resons) and instead proposes an analysis in terms of his own ``nomic probability''. Nomic probability reflects the existence of statistical laws. It is objective and that allows us to avoid the difficulties to which we are led by subjective probability. In addition, to arrive at causal probabilities, we must take account of the direction of time. Pollock defines \(\mathbf{C-PROB}_A(O)\) as a weighted average over possible backgrounds consistent with \(A\). All this is done for conditional probabilities, as well as absolute probabilities, so that policies involving contingent plans can be evaluated. It is only in the next to last section of the paper that Pollock addresses the problem of computing causal probabilities. When knowledge of these probabilities is not available, we can infer estimates of their values by defeasible inference based on statistical inference and the calculus of nomic probabilities. The system hangs together nicely, but depends on a prior understanding of projectibility, and a grasp of conditions of `defeat', which many readers will find a sticking point.