Causal Inference

When two things occur at the same time, that doesn’t yet mean one causes the other. In countries with higher chocolate consumption, there are more Nobel laureates — that’s not a coincidence, but it’s also not an explanation. Both are presumably linked to prosperity, not to each other. Causal inference is the tool for unmasking such spurious correlations and finding real causes. It asks: What would happen if one deliberately intervened in a system? In doing so, it goes well beyond merely observing patterns in data.

Correlation and causation — the crucial difference

Conventional data analysis reveals correlations: two quantities that change together. Does an app sell more in-app purchases in summer when temperatures rise? That’s a correlation. Whether the heat triggers the purchases or whether more free time in summer is the cause is something the correlation alone doesn’t reveal.

For businesses, doctors, and policymakers, this is a real problem. Anyone who knows only correlations can draw the wrong conclusions — and then intervene in the wrong place. A pharmaceutical company that merely observes that patients taking a medication recover faster might overlook the fact that these patients were already younger and healthier to begin with. Causal inference provides methods to avoid exactly this kind of confusion.

Interventions, counterfactuals, and control groups

The cleanest tool of causal inference is the randomized controlled experiment, often known as an A/B test. A group is randomly split into two halves. One half receives a treatment or change, the other does not. Because the assignment is random, all other differences between the groups are balanced out on average. Whatever difference remains at the end must be due to the treatment.

Often, however, such an experiment isn’t possible — for ethical, practical, or financial reasons. That’s where mathematical models come into play. A central concept here is the counterfactual: What would have happened if the cause had not occurred? A patient took the medication and got better — but what would have happened without the medication? This second course of events cannot be observed directly, but it can be estimated using models.

Another technique is called the instrumental variable method. Here, one looks for a factor that influences the presumed cause but doesn’t directly affect the outcome. Such detours make it possible to work out causality even from purely observational data — provided one finds the right detour.

Causal inference in AI, medicine, and policy

In medicine, causal inference is the standard for drug approval. Clinical trials are nothing other than large-scale randomized experiments. They are meant to ensure that an active ingredient actually helps — and isn’t just being used on patients who were already on the path to recovery anyway.

In AI research, the topic is gaining significant importance. Classical language models and image recognition systems are good at recognizing patterns. But they cannot understand causes — they don’t know why something happens, only that it often occurs together. Researchers like Turing Award winner Judea Pearl have argued for years that true intelligence only emerges once systems can think causally: not merely observing, but predicting what will happen given a targeted intervention.

In economic policy, causal inference became known to a broader public through the 2021 Nobel laureates — Angrist, Card, and Imbens. They developed methods for drawing causal conclusions from natural chance occurrences in reality. For example, David Card used differing minimum wage laws in neighboring US states as a natural experiment to measure the real effect of the minimum wage on employment.