Knowing the causes narrows the bounds without revealing the effect on a person

The main practical value of this work is not that it "finds" the causal effect for an individual person. It shows how to shrink the space of uncertainty when we have extra knowledge about the causal structure of the problem.

What is at stake are probabilities of causation: for instance, how likely it is that a particular outcome happened precisely because of the treatment, or that it would have happened without it. Such quantities cannot be observed directly for one and the same person in both scenarios at once — treated and untreated.

Classical methods give upper and lower bounds for these probabilities. The researchers consider three quantities: the probability of necessity, of sufficiency, and of necessity and sufficiency together. The new work extends the approach from binary cases to multi-level ones, where the treatment or the outcome takes more than two values.

The key move is to use information from covariates and mediators. This is no magic button turning correlation into proven cause: the extra variables merely help discard options incompatible with the causal model. As a result the bounds on the probabilities of causation come out tighter than in existing non-binary estimates.

The limitations here are substantial. The theoretical results are illustrated with toy examples, and the advantage of the new bounds is shown through simulations. The description contains no validation on real applied data, and the individual causal answers themselves remain directly unobservable. So the method is better treated as a tool for honest assessment of uncertainty than as a way to learn what exactly happened to a particular person.

In which tasks does it matter more to you to get an exact causal answer, or at least to narrow the range of possible explanations reliably? Source: cs.AI updates on arXiv.org