Correlation is not causation, I tell my team. But what exactly is the missing ingredient?
I build credit-scoring models for a mobile-money company in Nairobi. Our models find patterns: people who top up airtime at regular times repay loans more often. Everyone on my team can recite "correlation is not causation". Nobody can say what causation adds.
It matters practically. If regular top-ups merely correlate with repayment, then a customer who starts topping up regularly to improve her score hasn't become more creditworthy, and our model will be fooled. If the habit causes repayment, perhaps by reflecting stable income or discipline, then encouraging it might genuinely help.
The view that has helped me most is interventionism, developed by James Woodward and, in a more mathematical form, by Judea Pearl. Roughly: X causes Y if an intervention that changed X, and nothing else, would change Y. That turns the question into one about counterfactual experiments. It also explains why randomised trials work: randomisation is an intervention.
But I have two worries. First, interventionism seems to explain how we test causation, not what it is. Second, it struggles where interventions are impossible or meaningless. What would it be to intervene on someone's gender, or their birthplace, while holding everything else fixed? Yet those are exactly the variables fairness questions are about.
I'd like to hear from people who use causal reasoning in their work, and from people who think Hume was right that the missing ingredient is only in our heads.