How selective facts can convince us that A causes B — and why disproving that link is often harder
This paper studies how someone can persuade another person about a cause‑and‑effect claim by choosing which facts to reveal. The sender can disclose some variables and their true joint distribution and then propose a causal story that links them. The receiver accepts that story only if the revealed data makes the causal link clear enough. The authors call this setup “causal persuasion.”
The researchers build a formal model in which the real causal relationships come from a directed acyclic graph (DAG). A DAG is a simple map of cause‑and‑effect links that does not loop back on itself. The sender is assumed to see the true DAG and all the data. The receiver may start with no model, or may hold a prior subjective model. The paper treats two kinds of receivers. A naive receiver accepts any model that is consistent with the disclosed data. A sophisticated receiver accepts a model only when the data conclusively identifies the causal link.
The paper gives precise conditions for when persuasion succeeds. One main finding is that, when the sender’s suggested causal link is the true one, the sender often needs to disclose only one or two well‑chosen additional variables to convince a sophisticated receiver. The authors show this with formal theorems and an illustrative example about MBA degrees and earnings. In that example, education and earnings are actually correlated because of hidden traits like ability and social skills. Revealing a single extra variable that ties to experience can make the observed pattern consistent only with the school’s causal story, and so persuade the receiver.
The results also show a strong asymmetry. It is often much easier to establish that X causes Y than to prove there is no causal link. To persuade a receiver there is no relationship, the sender must expose every common cause that could explain the correlation. In practice there can be many such confounding factors. The paper notes that debunking claims like “immigration causes crime” or “police funding causes crime” would require accounting for every common cause of the two variables, which can be prohibitively demanding.