Columbia University · Course · 2026–2027

What a shock does, one horizon at a time.

A local projection turns a causal question about a shock into a sequence of regressions, one for each horizon after it. This course builds the method from the first regression to a defended empirical result: the estimand, the identification, the inference, and the replication that lets someone else check the work. Fourteen lectures, each with a Stata practicum and a browser lab.

LP estimates β̂h, one regression each true response θh 95% pointwise band — a claim about repeated samples

Course schedule

The sequence follows the research workflow: define the dynamic estimand, argue identification, evaluate estimation and uncertainty, extend to heterogeneous and panel designs, then use and defend the evidence. Every lecture ships with a replication lab, a Stata problem set, and a browser lab that form one practicum.

01From an Economic Question to a Local ProjectionP01 — shock-to-response explorer
02Levels, Differences, Cumulative Responses & UnitsP02 — transformation workbench
03Identification, Controls & the Pre-Shock Information SetP03 — identification sandbox
04LP-IV & Cumulative MultipliersP04 — instrument & multiplier lab
05Pointwise Inference, Persistence & Lag AugmentationP05 — coverage simulator
06Inference About the Response PathP06 — claims about a curve
07LPs versus VARs: Estimands, Bias & VarianceP07 — bias–variance lab
08Smoothing & Restrictions Across HorizonsP08 — smoothing workbench
09State Dependence & Asymmetric ResponsesP09 — state-dependence lab
10What Linear LPs Mean in Nonlinear EconomiesP10 — causal-weight explorer
11Panel LPs & Common Macroeconomic ShocksP11 — where identification comes from
12LP Difference-in-DifferencesP12 — clean-control builder
13Sensitivity Analysis & Policy CounterfactualsP13 — counterfactual workbench
14Research Synthesis, Replication Audit & DefenseP14 — replication referee desk
The capstone

Replicate a published response. Extend it. Survive the audit.

Every student reproduces one published local-projection result from its replication package, adds one motivated extension with its identifying assumptions stated, and hands the whole package to a peer who tries to reproduce it in a clean directory. A null result can earn full credit; an unsupported causal claim cannot.

14 × 3
replication labs · Stata problem sets · browser labs

About

I built this for students and research assistants who need to use local projections in their own work and to read other people’s. The method is simple to run and easy to misread: the same regression can be a causal response or a correlation, a level or a cumulative effect, a reliable band or a decorative one. The course keeps those distinctions in front of the student at every step.

Fourteen lectures follow one research sequence, from the economic question to the defended result. Each pairs a chapter of notes with a replication of a published figure, a Stata problem set, and a browser lab built around a prediction the student must commit to before touching the controls. Ordinary least squares, basic instrumental variables, confidence intervals, and a little Stata are enough to begin. The setup page has a readiness exercise for the rest.