Final Project

If you prefer a theoretical project, conduct a structured literature review instead of empirical replication.

How to do a snowball search

1. Pick a seed paper from class (e.g., Mullainathan & Spiess 2017 on ML in policy; Gu, Kelly & Xiu 2020 on asset pricing).
2. Backward search: Read the bibliography of your seed paper to find foundational work.
3. Forward search: Use Google Scholar “Cited by” to find recent 2025–2026 developments.
4. Categorize: Group papers by prediction (E[Y|X]) vs. causal inference vs. policy.
5. Synthesize: Don't just list papers — compare them. How do newer ML methods improve on traditional econometrics?
6. Identify the gap: What is missing? Structural interpretation? Selection bias handling? External validity?

Suggested Structure

Grading Rubric

See the Writing & LaTeX Practical Guide slides for detailed section-by-section guidance, examples, and templates. Below is a quick reference.

Abstract (150–200 words)

One paragraph covering: (i) original paper, (ii) method, (iii) replication finding, (iv) extension & result. No citations.

1. Introduction (1–2 pages)

Reverse pyramid: broad motivation → specific paper → your replication → your extension → roadmap. Your own work should occupy at least half the section.

2. Literature (1–2 pages)

Group thematically (method, applications, China-related). Use transition sentences. End with "Our paper contributes by..."

3. Data (1–2 pages)

State source, sample size, unit of analysis, and exclusions. Include Table 1 (summary statistics) with only variables you use.

4. Empirical Strategy (2–3 pages)

Write the estimating equation, define every symbol, state identification assumptions, and explain standard errors.

5. Replication Results (3–4 pages)

Reproduce the core result, compare to the original, and run 2–3 robustness checks. Include figures (event-study, CATE, etc.).

6. Extension (2–3 pages)

Motivate the new question, explain the method difference, present 1–2 tables/figures, and interpret magnitudes economically.

7. Conclusion (0.5–1 page)

One sentence each for replication and extension findings, one explicit limitation, and one future direction. Synthesize, don't restate.

Tables & Figures Checklist

• Every table has a caption above it; every figure has a caption below it
• Axis labels are readable (no tiny fonts)
• Error bars / confidence intervals are shown
• Regression tables report N, R², and fixed effects
• Stars defined: * p<0.1, ** p<0.05, *** p<0.01
• At least 1 table and 1 figure in the main text

LaTeX Starter Template

\documentclass[11pt]{article}
\usepackage[margin=1in]{geometry}
\usepackage{graphicx, amsmath, natbib, booktabs, hyperref}

\title{Replication and Extension of [Paper Title]}
\author{Your Name \\ ECON6083}
\date{\today}

\begin{document}
\maketitle

\begin{abstract}
\noindent Brief summary of paper, replication approach,
key findings, and extension. (150--200 words)
\end{abstract}

\section{Introduction}
\section{Literature}
\section{Data}
\section{Empirical Strategy}
\section{Replication Results}
\section{Extension}
\section{Conclusion}

\bibliographystyle{aer}
\bibliography{references}

\appendix
\section{Additional Results}
\section{Code}
\end{document}

See the Writing & LaTeX Practical Guide slides for detailed section-by-section guidance, examples, and templates. Below is a quick reference.

1. Introduction (1–2 pages)

Broad motivation → scope of review → research question → roadmap. Be specific about the economic subdomain.

2. Conceptual Framework (2–3 pages)

Build a 2×2 taxonomy (e.g., prediction vs. causal × cross-sectional vs. panel). Define key terms and highlight trade-offs.

3. Literature Review (6–8 pages)

Organize thematically (not chronologically): foundational methods, applications, and criticisms. Compare, don't list.

4. Gap & Future Directions (2–3 pages)

Identify methodological, empirical, and external-validity gaps. Propose 2–3 concrete research projects to fill them.

5. Conclusion (1 page)

Summarize 2–3 takeaways, restate the most important gap, and end with a forward-looking sentence.

Each option below pairs the original US/EU paper with a concrete China extension using the datasets above. All regressions use the panel structure of the data.

Option 1 — DML → CFPS Internet Access

Question: What is the causal effect of internet access on household income?

• Data: cfps_panel.csv — treatment internet, outcome lnincome
• Model: Partially linear DML
• Extension: Compare Lasso vs. Random Forest nuisance functions; estimate CATE by education.

Option 2 — Causal Forests → CFPS Health Insurance Heterogeneity

Question: Who benefits most from medical insurance?

• Data: cfps_panel.csv — treatment medsure_dum, outcome health
• Model: Causal forest for CATE estimation
• Extension: Plot heterogeneity by age × education; use SHAP for feature importance.

Option 3 — DiD → Low-Carbon City Pilots (LCCP)

Question: Did low-carbon city pilots reduce SO₂ emissions?

• Data: china_city_panel_with_policies.csv — policy lccp_treat
• Model: Staggered DiD with city and year fixed effects
• Extension: Event-study plot; Callaway-Sant'Anna group-time ATT; compare with Sponge City / CBEC / Smart City effects.

Option 3b — DiD → Sponge City Pilots

Question: Did sponge city investment increase fiscal expenditure?

• Data: china_city_panel_with_policies.csv — policy sponge_treat
• Model: Staggered DiD
• Extension: Compare TWFE vs. Borusyak et al. imputation estimator; placebo leads.

Option 3c — DiD → CBEC Pilot Zones

Question: Did CBEC zones boost exports and FDI?

• Data: china_city_panel_with_policies.csv — policy cbec_treat
• Model: Staggered DiD
• Extension: Heterogeneous effects by region (East vs. West); dynamic effects with event study.

Option 3d — DiD → Smart City Pilots

Question: Did smart city construction improve environmental outcomes?

• Data: china_city_panel_with_policies.csv — policy smart_treat
• Model: Staggered DiD
• Extension: Test green innovation channel (use sci_exp as mediator); spatial spillovers.

Option 3e — DiD → LTCI Pilots

Question: Did long-term care insurance affect fiscal health spending?

• Data: china_city_panel_with_policies.csv — policy ltci_treat
• Model: Staggered DiD
• Extension: Compare 2016 vs. 2020 batch effects; test for pre-trends with placebo leads.

Option 4 — SCM → Any Single-Treated City

Question: What is the effect of a specific policy on a single city?

• Data: china_city_panel_with_policies.csv
• Model: Synthetic control method
• Extension: Placebo test on donors; leave-one-out sensitivity; compare SCM estimate with DiD.

Option 5 — Policy Learning → CFPS Digital Inclusion

Question: Who should be targeted for digital-inclusion subsidies?

• Data: cfps_panel.csv — treatment internet, outcome lnincome
• Model: Empirical welfare maximization
• Extension: Budget-constrained tree (subsidize 30%); compare naive (poorest) vs. learned rule.

Option 6 — Text/EPU → China EPU + Macro

Question: How does policy uncertainty affect industrial production?

• Data: china_epu.csv + china_pmi.csv
• Model: Time-series regression of PMI on EPU
• Extension: Lagged EPU(t−1) → PMI(t); BERT-based sentiment vs. keyword measure; fiscal/monetary sub-indices.