Skip to content
Module 01 of 1240 min readIntermediate

What econometrics is for

Why we don't just compute correlations: the gap between description, prediction, and causation, and what each requires.

8%

Listen along

Read “What econometrics is for” aloud

Plays in your browser using on-device text-to-speech — nothing leaves the page.

Econometrics is statistics applied to questions economists actually care about. Most of those questions are causal: does raising the minimum wage reduce employment? Does a 1pp interest-rate hike cool inflation? Does microcredit lift household consumption?

Pure statistics gives you correlations. Econometrics gives you a framework for turning correlations into defensible causal claims — and, just as importantly, for spotting when a correlation cannot be safely interpreted as cause.

The three things you can ask of data

  • Description: what is the average loan size in this SACCO? (Easy. Just compute it.)
  • Prediction: given a member's history, how likely are they to default? (Hard, but well-defined. Optimise out-of-sample fit.)
  • Causation: if we raised the loan limit, how would defaults change? (Hardest. Requires assumptions about what would have happened in a world that didn't occur.)

Why causation is hard

You never observe the same person under both treatment and control. The 'counterfactual' — what would have happened — has to be inferred. Every credible econometric method is a strategy for constructing that counterfactual.

The fundamental problem of causal inference

For person i, define Y₁ᵢ as their outcome with treatment, Y₀ᵢ without. The causal effect on i is Y₁ᵢ − Y₀ᵢ. We never see both for the same person. So we estimate average effects across a population, where matching treatment and control groups gives us a credible counterfactual.

Random assignment (the gold standard) makes the two groups identical in expectation, so any post-treatment difference must be due to treatment. Most economic data is non-experimental — that's where the methods in this course do their work.

What this course covers

  • Linear regression: the workhorse, properly understood
  • Endogeneity: the four ways your regression coefficient ends up wrong
  • Instrumental variables: the cleanest correction, when you can find one
  • Difference-in-differences: leveraging policy variation across groups and time
  • Panel data: fixed effects and within-variation
  • Time series: why monthly data needs different machinery
  • Limited dependent variables: when y is binary or censored
  • Reading published empirical work critically

What you don't need

You don't need calculus beyond derivatives, matrix algebra beyond the basics, or statistical software beyond a willingness to learn one (R, Python's statsmodels, or Stata).

Exercise

Look up the Card-Krueger 1994 minimum wage paper. In one sentence, what counterfactual did they construct, and how?

Loading progress…
LeadAfrikPublic Economics Hub