Finally, our paper positions knowledge as an economic good. It is valuable but scarce, just like data. So we can study the trade-off between knowledge and data using the same tools economists use to study other goods.

You can find the paper on arXiv: arxiv.org/abs/2509.09170

Our paper also highlights how humans and machines differ. Humans know how to "ask the right questions" and can learn from limited data. In contrast, machines rely on pattern recognition and need lots of data. So the less data are available, the better it is to have humans collect and interpret them.

The paper provides theoretical support for our empirical work in Uganda: We gave some farmers both knowledge and data, and others data only. Farmers with deeper knowledge made more profitable decisions and more accurate predictions.

See Anirudh's page for details: sites.google.com/view/anirudh...

3. People with deeper knowledge are better off and need less data.

The more you know about which features matter, the more you can learn from a given amount of data, and so the less you need to reach a given welfare goal.

2. The value of conceptual knowledge depends on how much data you can collect.

If you can't collect much, then it's valuable to know which features matter most. But if you can collect lots of data, then you don't need to know which features matter because you can "let the data speak."

1. Conceptual knowledge is more valuable when states are more "reducible."

If only one feature matters, then collecting data on it helps you learn about many states at once. But if many features matter, then you can't do much better than learn about each state on its own.

*How much better* is a quantity we define in our paper. We call it "the value of conceptual knowledge." It equals the welfare gain from knowing how states relate and using that knowledge to collect better data.

We study the value of conceptual knowledge mathematically. It has three key properties:

If you know about states' features, then you can collect data on them; if you don't, then you have to collect data on the states themselves.

It's better to collect data on features because it makes use of structural relationships.

We generalize the farming example:

Imagine you want to learn some unknown "states" (e.g., fertilizer effects). Conceptual knowledge tells you how the states relate. It allows you to represent them as combinations of structural "features" (e.g., nutrient effects).

The farmer's knowledge of nitrogen is conceptual: it's in his mind, rather than his data. Nonetheless it allows him to collect better data. How much better? That's the question our paper answers!

Consider a farmer testing fertilizers. If he views them as black boxes, then he has to test them separately. But if he knows they all contain nitrogen, then he can combine them to estimate the "nitrogen effect." This is better than testing one-by-one: he can learn about many fertilizers at once.

This paper uses a decision theoretic framework to study when and why conceptual models help you run more informative experiments:

The intro says "[m]en tended to work in smaller teams than women, but co-authored more papers and so had more co-authors overall."

The paper uses NBER working papers from 1973--2019. @paulgp.com's data include the past five years

Figure 1 of this paper shows that papers with female authors tend to have more co-authors: