Causal Inference for Public Policy

Original title (in Spanish): Inferencia, Causalidad y Políticas Públicas.

All course materials — syllabus, lecture slides, problem sets, replication code, and reference papers — are hosted on GitHub:

  github.com/eduard-martinez/causal-inference   Full syllabus (PDF)

Course description

Impact evaluation aims to estimate the causal effect of programs, projects, and public policies in order to guide evidence-based decisions. In a context of limited resources, it is essential to identify the most cost-effective interventions, anticipate unintended effects, and maximize benefits for the population.

The course is highly applied, with an emphasis on Public Policy and Development Economics, drawing on applications from Colombia and other developing countries. Each method is paired with an anchor paper from the empirical literature, and every session combines theory, replication exercises in R, and a student-led discussion of a related paper.

Learning outcomes

By the end of the course, students will be able to:

• Apply the potential outcomes framework and identify the conditions under which OLS, IV, RCT, RD, fixed effects, DiD, event studies, and matching methods deliver causal estimates.
• Formally derive identification and unbiasedness properties of the main causal estimators (mean-difference decomposition, Wald estimator, identification under continuity, parallel trends, etc.).
• Critically appraise empirical studies in the economics literature and propose applications to real-world policy contexts.
• Design and implement impact evaluations end-to-end in R — from data preparation through estimation and interpretation.

Schedule

Module 1 — Foundations of Causal Inference

Week 1 — Foundations of Causal Inference. Correlation vs. causation; Rubin Causal Model; fundamental problem of causal inference; endogeneity, selection bias and omitted-variable bias; DAGs; causal estimands (ATE, ATT, ATC, ITT, LATE). Readings: Cunningham (2021), ch. 1 & 4; Angrist & Pischke (2009), ch. 1. Slides — week 1

Module 2 — Experimental and quasi-experimental designs

Week 2 — Randomized Controlled Trials (RCT). Randomization by design, internal validity, the three properties of RCTs, SUTVA, stratification and clusters, statistical power, ATE / ITT / TOT under noncompliance, threats to validity (attrition, spillovers, John Henry and Hawthorne effects). Anchor paper: Duflo, Hanna & Ryan (2012, AER) — Incentives Work: Getting Teachers to Come to School. R activity: Partial replication of Tables 1–3 with haven, fixest, broom. Slides — week 2 · Replication code — week 2

Week 3 — Instrumental Variables (IV). Endogenous variables and noncompliance; relevance, independence, exclusion; Wald estimator and 2SLS; heterogeneous effects (compliers, never-takers, always-takers, defiers); monotonicity and the LATE theorem (Imbens & Angrist, 1994); weak-instrument diagnostics. Anchor paper: Angrist (1990, AER) — Lifetime Earnings and the Vietnam Era Draft Lottery. R activity: IV with AER::ivreg and fixest::feols. Slides — week 3 · Replication code — week 3

Week 4 — Sharp Regression Discontinuity (RD). Quasi-random variation around a threshold; identification under continuity (Hahn, Todd & Van der Klaauw, 2001); local polynomial estimation; bandwidth selection (CCT, IK); robust inference (Calonico, Cattaneo & Titiunik, 2014); McCrary density test and placebo cutoffs. Anchor paper: Lee (2008, JoE) — Randomized Experiments from Non-random Selection in U.S. House Elections. R activity: rdrobust and rddensity. Slides — week 4 · Replication code — week 4

Week 5 — Fuzzy Regression Discontinuity. From Sharp to Fuzzy; Fuzzy RD as local IV with Z = 1{X ≥ c}; identification of the local LATE; 2SLS with local polynomial regression; interpretation as effect on compliers at the threshold. Anchor paper: Pop-Eleches & Urquiola (2013, AER) — Going to a Better School: Effects and Behavioral Responses. R activity: Fuzzy RD with rdrobust and manipulation diagnostics. Slides — week 5 · Replication code — week 5

Module 3 — Panel data methods

Week 6 — Panel Data and Fixed Effects. Panel structure and unobserved heterogeneity; within estimator and the demeaning transformation; unit and time fixed effects; cluster-robust standard errors; limits — observed heterogeneity that varies over time, dynamic effects. Readings: Cunningham (2021), ch. 8; Angrist & Pischke (2009), ch. 5. R activity: Panel with fixest::feols. Slides — week 6 · Replication code — week 6

Week 7 — Difference-in-Differences (DiD). Canonical 2×2 DiD; parallel-trends assumption; DiD as panel with fixed effects; multi-period DiD; triple differences; limitations of TWFE under staggered adoption (introduction to Goodman-Bacon, 2021). Anchor paper: Card & Krueger (1994, AER) — Minimum Wages and Employment (NJ–PA). R activity: DiD with fixest and a simple event study. Slides — week 7 · Replication code — week 7

Week 8 — Event Study. Event-study specifications with relative dummies; pre-trends as visual test of parallel trends; leads and lags; staggered-treatment bias (Goodman-Bacon, 2021; de Chaisemartin & D’Haultfœuille, 2020; Sun & Abraham, 2021; Callaway & Sant’Anna, 2021); modern robust estimators. Readings: Borusyak, Jaravel & Spiess (2024); Roth, Sant’Anna, Bilinski & Poe (2023). R activity: Event-study with fixest::sunab and the did package. Slides — week 8

Module 4 — Selection on observables

Week 9 — Propensity Score Matching (PSM). Selection on observables and unconfoundedness; the propensity score (Rosenbaum & Rubin, 1983); nearest-neighbor, kernel, stratification matching; inverse-probability weighting; common support and balance diagnostics; limits of relying on observables. Readings: Cunningham (2021), ch. 5; Imbens & Wooldridge (2009). R activity: MatchIt and cobalt.

Problem sets

• Problem set 1 — Foundations, RCT and IV (weeks 1–3): PDF · datasets (datos_1_observacional.csv, datos_2_piloto.csv, datos_3_sisben.csv)
• Problem set 2 — RD, Panel and DiD (weeks 4–7): PDF · datasets (datos_jornada_unica.csv)

Each problem set has a mathematical component (40%) — formal derivations of identification properties — and an applied component in R (60%) — Monte Carlo simulation plus partial replication of an anchor paper. Submitted as a single PDF plus a reproducible .R script.

Evaluation

Component	Weight
Article presentation (15 min)	10%
Problem set 1 — Foundations, RCT, IV	15%
Problem set 2 — RD, Panel, DiD	15%
Final project (proposal + mini-paper + oral defense)	60%

The final project is built around a policy-relevant research question chosen by the student. After the proposal is approved, the instructor provides a synthetic dataset and a policy context aligned with the question. Deliverables: proposal (10%), 10-page mini-paper with diagnostics and robustness (60%), oral defense in the last week (30%).

Five worked example proposals, each with its accompanying synthetic dataset, are available in the repo: final-project datasets (Propuesta_1–Propuesta_5 + Dataset_1–Dataset_5).

Methodology

Each class is organized in three parts: (1) theoretical exposition of the model and its identification assumptions; (2) guided R applications with prepared datasets and code, anchored on a canonical empirical paper; (3) student-led discussion of a related empirical paper from the bibliography (15-minute presentation covering the research question, identification strategy, and main results).

Reading library

The repository includes a curated library of empirical papers organized by method, used for the student-led presentations and as references for the problem sets and final project:

• Instrumental Variables (IV) — 12 papers (Angrist & Krueger 1991; Acemoglu et al. 2019; Deryugina et al. 2019; Finkelstein et al. 2012; Bhuller et al. 2020; Duranton & Turner 2011; …).
• Randomized Controlled Trials (RCT) — 9 papers (Duflo, Hanna & Ryan 2012; Miguel & Kremer 2004; PROGRESA; Banerjee et al. 2015; Haushofer & Shapiro 2016; Björkman & Svensson 2009; …).
• RCT + IV (noncompliance) — 8 papers (Angrist 1990 draft lottery; Kling, Liebman & Katz 2007 MTO; Crépon et al. 2015; Ashraf, Karlan & Yin 2006; …).
• Sharp Regression Discontinuity — 6 papers (Lee 2008; Pop-Eleches & Urquiola 2013; Ludwig & Miller 2007; Bronzini & Iachini 2014; …).
• Fuzzy Regression Discontinuity — 7 papers (Barrera et al. 2018; Brollo et al. 2013; Card, Dobkin & Maestas 2008; Khanna 2023; Battistin et al. 2009; …).

Core bibliography

• Cunningham, S. (2021). Causal Inference: The Mixtape. Yale University Press.
• Angrist, J. D., & Pischke, J.-S. (2009). Mostly Harmless Econometrics: An Empiricist’s Companion. Princeton University Press.
• Imbens, G. W., & Rubin, D. B. (2015). Causal Inference for Statistics, Social, and Biomedical Sciences. Cambridge University Press.
• Hernán, M., & Robins, J. (2020). Causal Inference: What If. Chapman & Hall/CRC.
• Gertler, P., Martinez, S., Premand, P., Rawlings, L., & Vermeersch, C. (2017). La evaluación de impacto en la práctica (2ª ed.). Banco Mundial.

Course copies of the three core texts (Cunningham; Angrist & Pischke; Gertler et al.) are in the repo’s books/ folder.

The full reference list — including the foundational papers for IV, RD, DiD, event studies and PSM — is in the course syllabus.