1/22
What is econometrics?
The application of statistical models to economic data
Statistical models are mathematical tools for reasoning under uncertainty (e.g. regression, ANOVA, time-series models, survival models). Economic data come from experimental or observational sources and can be cross-sectional, time series or panel data.
What are the two sources of economic data?
Experimental (the researcher controls/randomises X) and observational (X is recorded as it happens)
Experimental data come from a controlled experiment, such as a randomised trial. Observational data come from watching behaviour without assigning treatment, as in most economic data (surveys, admin records). Observational data make causal claims harder, because people choose their X.
What are the three types of economic data?
Cross-sectional: many units at one point in time. Time series: one unit over many periods. Panel: many units, each over many periods
Examples: wages of 1,000 workers in 2025 (cross-section), UK GDP quarterly 1990 to 2025 (time series), the same 1,000 workers tracked every year 2015 to 2025 (panel). Stock & Watson Ch. 1.3 covers this. Semester 2 focuses on time series.
What are the three things we use econometrics for?
Describe (what is going on?), explain (what causes what?), predict (what is likely to happen?)
Describe: how much more do graduates earn? Explain: does a degree RAISE earnings? Predict: what will unemployment be next year? In Semester 1, Part I covers regression as description and Part II covers regression and causality. A key course objective is telling a statistical relationship apart from a causal effect.
What is the intuitive definition of a regression?
It describes how a dependent variable Y tends to change as one or more independent variables X₁,…,X_M change
Y = dependent variable (outcome). X₁,…,X_M = independent variables (regressors). A regression on its own describes association; it does not establish causation.
What is the formal definition of a regression, and which feature does this course focus on?
A model of how a feature of the conditional distribution of Y given X₁,…,X_M depends on those Xs; the focus is the conditional mean E[Y | X₁,…,X_M]
Regression models one feature of Y's conditional distribution. Usually that feature is the conditional mean: the average Y for given values of X. This links to the revision notes, where the conditional distribution of Y given X is the most we can know about how X affects Y.
Match each question to its regression model and estimation method. Wages and education → ?
Linear (log wages), estimated by OLS
The full table from the slide: wages and education → linear (log wages), OLS. Whether having children changes whether a woman works → non-linear (probit), MLE. Distance and trade between countries → non-linear (exponential mean), Poisson PML. Spending and income when richer households vary more → linear, WLS. The Semester 1 focus is OLS.
Simple linear regression (SLR): write the model and name each term.
Y = β₀ + β₁X + u. Y = outcome (dependent variable), X = regressor, β₀ = intercept, β₁ = slope, u = error
Example: wage = β₀ + β₁·education + u. It is called 'simple' because there is only one regressor. Wooldridge Ch. 2.1 covers the model.
How do you interpret the slope β₁ in Y = β₀ + β₁X + u?
The average change in Y when X rises by one unit
β₁ = ΔE[Y|X]/ΔX, so always state the units. Example: β₁ = 0.72 means a one-unit rise in X is associated with Y being 0.72 units higher on average. β₀ (the intercept) is the predicted Y when X = 0.
What is the error term u?
Everything else, apart from X, that affects Y
u is unobserved and belongs to the population model. In a wage equation it includes ability, experience, luck and so on. Don't confuse it with the residual ûᵢ (= eᵢ), which is computed from the sample.
Define the fitted value ŷᵢ and the residual eᵢ (also written ûᵢ).
ŷᵢ = β̂₀ + β̂₁xᵢ is the line's prediction at xᵢ; eᵢ = yᵢ − ŷᵢ is the gap between the observed point and the line
yᵢ is the observed point. The residual is observed minus predicted, so a positive eᵢ means the point lies above the line. The textbooks write the residual as ûᵢ.
Why can't we choose the line by making the residuals sum to zero (Σeᵢ = 0)?
Every line through (x̄, ȳ) has Σeᵢ = 0 whatever its slope. One equation cannot pin down two unknowns (β₀, β₁)
Positive and negative residuals cancel out, so infinitely many lines satisfy Σeᵢ = 0. The fix is to square the residuals: Σeᵢ² is smallest for exactly one line. That is OLS.
What objective function does OLS minimise?
S(β₀, β₁) = Σᵢ₌₁ⁿ (yᵢ − β₀ − β₁xᵢ)², the sum of squared residuals
OLS = Ordinary Least Squares. The data (xᵢ, yᵢ), i = 1,…,n, are fixed numbers, and β₀, β₁ are the two choice variables ('knobs'). The values that minimise S get hats: β̂₀, β̂₁. Picture it as the line that makes the total area of the squares smallest.
What are the four steps for deriving the OLS estimators?
Write the objective S(β₀, β₁); differentiate with respect to β₀ and β₁; set both derivatives to 0; solve the two equations
There are two first-order conditions (FOCs) for two unknowns, so the system has a unique solution, provided Sxx ≠ 0.
FOC 1: what does ∂S/∂β₀ = 0 give?
Σeᵢ = 0, which leads to β̂₀ = ȳ − β̂₁x̄
Derivation: ∂S/∂β₀ = −2Σ(yᵢ − β̂₀ − β̂₁xᵢ) = 0. Divide by −2 to get Σûᵢ = 0. Distribute the sum: Σyᵢ = nβ̂₀ + β̂₁Σxᵢ. Divide by n: ȳ = β̂₀ + β̂₁x̄, so β̂₀ = ȳ − β̂₁x̄. For any slope, the intercept anchors the line at the means.
FOC 2: what does ∂S/∂β₁ = 0 give?
Σxᵢeᵢ = 0, which leads to β̂₁ = Σ(xᵢ − x̄)(yᵢ − ȳ) / Σ(xᵢ − x̄)²
Derivation: ∂S/∂β₁ = −2Σxᵢ(yᵢ − β̂₀ − β̂₁xᵢ) = 0, so Σxᵢûᵢ = 0. Substitute β̂₀ = ȳ − β̂₁x̄ to get Σxᵢ(yᵢ − ȳ) = β̂₁Σxᵢ(xᵢ − x̄). Deviations from the mean sum to zero, so you can swap xᵢ for (xᵢ − x̄) for free: Σxᵢ(yᵢ − ȳ) = Σ(xᵢ − x̄)(yᵢ − ȳ) and Σxᵢ(xᵢ − x̄) = Σ(xᵢ − x̄)². This is the condition that tells the possible slopes apart.
What are the OLS slope and intercept formulas?
β̂₁ = Σ(xᵢ − x̄)(yᵢ − ȳ) / Σ(xᵢ − x̄)² = Sxy/Sxx; β̂₀ = ȳ − β̂₁x̄
Slope = sample covariance of X and Y ÷ sample variance of X, i.e. Covₙ(x,y)/Varₙ(x). This requires Sxx = Σ(xᵢ − x̄)² ≠ 0, so X must vary in the sample. More spread in X gives a more precisely pinned-down slope.
Worked example: Σ(xᵢ − x̄)(yᵢ − ȳ) = 36, Σ(xᵢ − x̄)² = 50, x̄ = 4, ȳ = 6. What are β̂₁ and β̂₀?
β̂₁ = 0.72, β̂₀ = 3.12
β̂₁ = 36/50 = 0.72. β̂₀ = ȳ − β̂₁x̄ = 6 − 0.72 × 4 = 6 − 2.88 = 3.12. Check: the fitted line at x̄ gives 3.12 + 0.72 × 4 = 6 = ȳ. ✔
Which three algebraic properties of OLS follow automatically from the first-order conditions?
Σûᵢ = 0; Σxᵢûᵢ = 0 (residuals uncorrelated with X in the sample); the fitted line passes through (x̄, ȳ)
Σûᵢ = 0 comes from FOC 1, and Σxᵢûᵢ = 0 from FOC 2. Proof that the line passes through the means: ŷ(x̄) = β̂₀ + β̂₁x̄ = (ȳ − β̂₁x̄) + β̂₁x̄ = ȳ. These are derived results that hold in every OLS sample; they aren't assumptions.
Why do the estimates have hats? (population vs sample)
β₀ and β₁ (no hats) are fixed but unknown population parameters. β̂₀ and β̂₁ (hats) are estimates computed from a sample
Population = everyone we could have surveyed. In the population, Y = β₀ + β₁X + u, and β₀, β₁ are the truth we are after. Sample = the n people we did survey (e.g. n = 28 gave β̂₀ = 1.40, β̂₁ = 0.72). A different sample gives different estimates, so β̂ is a random variable with a sampling distribution.
The true values are β₀ = 1.60 and β₁ = 0.68, but a sample containing only low-X individuals gives β̂₁ = 0.48. What does this show?
Estimates vary across samples. Good samples hover around the truth, but a bad (unrepresentative) sample can mislead
Sampling explains why estimates vary. The population model defines the target. This is why the next topics cover the sampling properties of OLS (unbiasedness, variance) and inference.
Why does more variation in X give a more precise slope estimate?
Sxx = Σ(xᵢ − x̄)² sits in the denominator. With more spread in X, the line is pinned down more tightly, and if Sxx = 0 the slope isn't defined at all
If every xᵢ is the same, you can't estimate how Y changes with X: Sxx = 0 and β̂₁ is undefined. The slide calls this 'A3' (the assumption of sample variation in X, covered with the SLR assumptions).
What is the difference between a statistical relationship and a causal effect?
A statistical relationship is an association (Y tends to move with X). A causal effect means changing X itself changes Y, all else equal
Telling these apart is a core course objective. With observational data, β̂₁ describes how the average Y differs across X. It is causal only if the other factors in u are unrelated to X (covered in Part II on endogeneity and causality). Example: graduates earn more (description) vs a degree raises earnings (cause).