Lingo

Causal Inference & LATE

Moving beyond correlation to understand what actually causes what.

What is it?

Causal inference is the set of methods for estimating cause and effect from data—answering “what happens if we change X?” rather than merely observing that X and Y move together. The 2019 and 2021 Nobel Prizes recognized this work. A central concept is the Local Average Treatment Effect (LATE): the causal effect only for the specific subpopulation whose treatment status was actually changed by the instrument.

Why correlation is not causation

Confounding: A third variable may cause both X and Y. Wealth causes both ice cream purchases and better health, so the two correlate—yet ice cream does not cause health.

Reverse causation: The arrow may point the other way. Stress and long work hours correlate, but which causes which?

Selection bias: The people we observe were not randomly assigned. Those who choose a treatment differ from those who don’t in ways that confound the comparison—this is why naive comparisons of treated and untreated groups are misleading.

Core methods

Randomized controlled trials: The gold standard. Random assignment balances confounders on average, so the difference in outcomes is attributable to treatment.

Regression discontinuity: Compare units just above and just below a cutoff (e.g., a test threshold for a scholarship). Those near the threshold are nearly identical, so any outcome jump reveals the treatment effect.

Difference-in-differences: Compare the change over time for a treated group against the change for an untreated comparison group, netting out common trends.

Instrumental variables: Use an exogenous variable (the instrument) that affects treatment but not the outcome directly, isolating the causal path.

LATE and its limits

What LATE measures: An instrumental-variable estimate recovers the effect only for the compliers—those induced to take treatment by the instrument. It is local, not global.

Why it matters: The LATE for “college attendance due to distance from campus” may differ from the effect for the average person. The finding only generalizes with care.

Interpretation: A LATE is the right answer to a precisely delimited question, but the wrong answer to a broad one (“what’s the effect of college for everyone?”). Researchers must be explicit about which population they identify.

References

  1. Wikipedia: Causal Inference— Survey of methods for estimating cause and effect from observational and experimental data.
  2. Nobel Prize: 2021 (Card, Angrist, Imbens)— Nobel Prize page recognizing contributions to natural experiments and causal inference methods.
  3. Nobel Prize: 2019 (Banerjee, Duflo, Kremer)— Nobel Prize page recognizing experimental approaches to alleviating global poverty.

Examples