The two cultures, revisited

There’s an old division in quantitative work between two cultures. The framing loosely echoes Leo Breiman’s 2001 essay (his cut was between statistical modeling traditions, but the spirit is similar). One asks why: what’s the causal structure, which parameter has policy meaning, can I identify an effect that would survive a counterfactual world. The other asks whether it works: does the system perform under the conditions it’ll face, does the prediction hold out of sample, does the policy produce the outcome when deployed. Economists tend to live in the first culture. Operations researchers and ML people tend to live in the second. The cultures talk past each other often enough that it’s worth taking the difference seriously.

I’ve spent time in both, and I find the contrast genuinely interesting. My senior thesis lives squarely in the second culture: an optimization framework that compares three allocation policies head to head on a supply chain problem in Botswana. The headline finding is a fact about how the policies respond to the same uncertainty. Earlier work as an economics RA pulled in the other direction. The questions there are different in a way that took me a while to register. Not “does this method outperform that one on a benchmark,” but “if we changed this policy, would the outcome change, and how much, and for whom.” The methodological apparatus of instruments, discontinuities, and difference-in-differences exists because the answer matters and you usually can’t run the experiment.

What I’ve come to think is that the cultures aren’t really competing. They’re doing different things, and both things are worth doing.

Prediction culture is at its best when you have a system you can build, a benchmark that reflects deployment conditions, and a comparison that’s internally valid by construction. The test is empirical performance. If the system holds up, the system holds up. Most operations research, most applied ML, most engineering with a working artifact at the end sits in this register, and there’s something deeply satisfying about it. You build a thing, you measure how it does, you can point at a working system and say “look, it works.” The rigor lives in the honesty of the comparison.

Inference culture is at its best when you can’t build the system, or when the question is fundamentally counterfactual. Did this policy reduce mortality? What’s the effect of cash transfers on schooling? Would universal antimicrobial access change the trajectory of resistance? You can’t benchmark these against alternatives in a closed loop, because the alternatives aren’t observable. You have to argue from data to a world that didn’t happen, and the argument has to be defensible. The credibility revolution in economics produced some of the most careful empirical work of the last few decades, precisely because the methodological standards force you to be specific about what you’re claiming. There’s a different kind of satisfaction in pinning down a parameter that actually means what you say it means.

What I think gets missed in the usual fight is that the two cultures map onto different tolerances for uncertainty. Prediction people are comfortable saying “I don’t know if it’ll generalize to a different setting, but I know it works here.” Inference people are comfortable saying “I don’t know if this will work in any setting until I deploy it, but I know what the parameter means.” Each tolerance feels alien to the other, and the methodological disagreement often turns out to be a temperamental one wearing more formal clothes. Once you see this, the arguments become a lot easier to follow, and a lot less personal.

The most interesting projects, to me, are the ones where you need both. Policy-adjacent work usually does. You want a system that holds up and you want to know that the parameters you used to design it actually mean something. You want benchmarks and you want identification. The project, when both are in play, is to put them next to each other honestly, without pretending they’re the same thing or that one substitutes for the other.

Figuring out which question you’re actually facing turns out to be most of the work. And there’s something clarifying about realizing that the cultures aren’t a fight to be won. They’re more like two different lenses, each clearer on what the other can’t quite see.