Machine Learning in Economics

How economists use machine learning to measure poverty from satellite images, predict outcomes and find patterns, and why prediction is not the same as causation.

Economists increasingly use machine learning, computer methods that learn patterns from large amounts of data. These tools open new possibilities, but they also have limits.

Prediction versus causation

Economists Sendhil Mullainathan and Jann Spiess explained in 2017 that machine learning excels at prediction: guessing an outcome from patterns in data. Traditional econometrics focuses on causation: what happens if we change something.

Both matter, but for different questions.

Measuring the unmeasured

In many poor countries, data on income and poverty are scarce. Researchers used machine learning with satellite images:

A 2016 study by researchers at Stanford, including Neal Jean and Marshall Burke, trained models on daytime satellite images and survey data to predict poverty levels in African villages.
Night-light data, showing how brightly areas are lit, help estimate economic activity.

These methods help target aid and track development where surveys are rare.

Prediction problems in policy

Some policy decisions depend on predictions:

Bail decisions: a study by Jon Kleinberg and colleagues found that machine learning predictions of which defendants would commit crimes could, in principle, reduce crime or jail populations compared with judges’ decisions.
Targeting benefits: predicting which households are poorest.
Tax audits: predicting which returns are likely to be fraudulent.
Hospital care: predicting which patients need urgent attention.

Helping causal research

Machine learning can also support causal studies, for example by finding which groups benefit most from a programme, known as heterogeneous treatment effects.

Risks

Bias: models trained on biased data can repeat discrimination.
Opacity: complex models can be hard to explain.
Overfitting: models may find patterns that don’t generalise.
Misuse: prediction can be mistaken for explanation.

Poverty from space

An aid agency needs to know which villages in a country are poorest, but the last survey was years ago. Researchers use satellite images showing roofs, roads and farmland, combined with limited survey data, to create a detailed poverty map. The agency targets its programmes more precisely.

Thinking a good prediction explains why something happens

Machine learning can predict well without revealing causes. Policy needs to know what will change outcomes.

Key takeaways

Machine learning excels at prediction; econometrics focuses on causation.
Satellite images and night lights help measure poverty and activity.
Predictions can improve decisions such as bail and targeting.
Bias, opacity and overfitting are key risks.