Causality
I recently read most of this famous article, Demystifying Dependence. This is a pretty transformative paper, in my opinion, as a Nix user who is interested in what “dependence” means, and as a software engineer. I also greatly appreciated the didactic method of providing nine stories and then investigating what we can learn about dependencies from them in the new framework.
The surprise in the paper is that dependence can be analyzed in terms of causality. The authors show how using a new framework called “Halpern causation,” a definition of causation from a subdiscipline called “actual causality.” The treatment is really interesting. One of the big “a-ha” moments is realizing that you cannot discuss causality without talking about counterfactuals. To paraphrase the math:
- Create a model of the world in terms of input variables (“exogeneous”) and derived variables (“endogeneous”)
- Figure out the variable configuration for the event that happened
- Find a variable whose setting, if it were inverted, would lead to the opposite outcome
That variable is the “but-for” cause. It turns out not to be a perfect model of causality. They give an example: suppose we are worried about watering some grass, but we don’t want to water on days where it rains. We ask after a rain, why is the grass wet? According to the but-for definition, imagining a counterfactual world where it did not rain, we still have wet grass, because the sprinkler system would have come on. The authors then introduce a more sophisticated definition of causality called Halpern-Pearl causality which fixes this problem, using the idea of a contingency, which forces certain settings of variables. These settings are called witnesses to the contingency. The jargon alone is quite fun.
It does make me want to think about the Chornobyl accident in terms of actual causality. For instance, the roof being made of flammable materials made the disaster worse, but it was not a but-for cause of the explosion. But we could use these definitions of causality to answer questions like, was graphite-tipping the control rods an actual cause? It wasn’t if (and only if), holding everything else about the accident the same, the accident plays out in the exact same way without them.