Econometric Concepts

Fits a line that minimizes squared residuals: best linear predictor of Y from X under the Gauss–Markov conditions.

Report uncertainty: use robust standard errors when errors are heteroskedastic; high R² ≠ causal.

Controls soak up confounders so the remaining variation in X is closer to as-good-as-random.

Bad controls (post-treatment or mediators) bias estimates; watch functional form.

Use Z that moves X but affects Y only through X (exclusion). First-stage relevance is essential.

Weak instruments → big bias; always show first-stage F and justify exclusion in words.

Identifies effects from differential changes over time; needs a parallel-trends assumption.

Check pretrends; group composition and staggered timing can matter.

Local randomization at a cutoff yields local treatment effects near the threshold.

Bandwidth choice, balance around cutoff, and manipulation checks (McCrary) matter.

Within-unit comparisons remove time-invariant bias; cluster SEs at the unit level.

Serial correlation inflates precision if not clustered; beware dynamic FE bias with short panels.