# Method Atlas — Full Site Content for LLM Context

> An interactive educational platform teaching causal inference from first principles.
> Created by Saerom (Ronnie) Lee, Assistant Professor of Management, The Wharton School, University of Pennsylvania.
> URL: https://methodatlas.vercel.app

Method Atlas is a freely accessible guide for PhD students and empirical researchers learning causal inference through interactive lessons, downloadable code (R, Python, Stata), hands-on labs, and curated paper libraries. The site emphasizes identification — not just estimation — and the credibility revolution in empirical social science.

---

## Foundations (8 chapters)

1. **Why Causal Inference?** — The fundamental problem: why comparing treated and untreated units is not enough.
2. **The Anatomy of a Research Design** — The pipeline from question to credible answer.
3. **Selection Bias and Confounding** — The single biggest threat to empirical research; the reason every method exists.
4. **The Language of Identification** — Estimand, estimator, estimate. ATE, ATT, LATE. Precise vocabulary for credible research.
5. **DAGs for Beginners** — Directed Acyclic Graphs as a visual tool for thinking about causal relationships.
6. **A Taxonomy of Identification Strategies** — The full toolkit: design-based, model-based, experimental, quasi-experimental.
7. **Working with Data** — Loading, cleaning, reshaping, constructing variables.
8. **The Credibility Revolution** — How empirical economics transformed itself and what it means for your research.

URL: https://methodatlas.vercel.app/foundations/{slug}

---

## Methods (27)

### Established

| Method | Slug | Category | Difficulty | Key Assumption |
|---|---|---|---|---|
| Experimental Design (RCT) | experimental-design | Design-Based | Beginner | Random assignment, SUTVA, excludability |
| OLS (Robust SEs, Clustering) | ols-regression | Model-Based | Beginner | E[u\|X] = 0 (exogeneity) |
| Logit / Probit | logit-probit | Discrete Choice | Intermediate | Correct link function + exogeneity |
| Poisson / Negative Binomial | poisson-negative-binomial | Count Models | Intermediate | Conditional mean exp(X'β) correctly specified |
| Fixed Effects (Two-Way FE) | fixed-effects | Panel | Intermediate | Strict exogeneity; no time-varying confounders |
| Random Effects | random-effects | Panel | Intermediate | Unit effects uncorrelated with regressors |
| DiD (Canonical 2×2) | difference-in-differences | Design-Based | Intermediate | Parallel trends; no anticipation; SUTVA |
| RDD — Sharp | regression-discontinuity-sharp | Design-Based | Intermediate | Continuity at cutoff; no manipulation |
| RDD — Fuzzy | regression-discontinuity-fuzzy | Design-Based | Advanced | Continuity + monotonicity; strong first stage |
| Matching (PSM, CEM, NN) | matching-methods | Model-Based | Intermediate | Conditional independence + overlap |
| IV / 2SLS | instrumental-variables | Design-Based | Intermediate | Relevance (F>10) + exogeneity + exclusion restriction |
| Interrupted Time Series (ITS) | interrupted-time-series | Design-Based | Intermediate | No concurrent events; functional form for trend correctly specified |
| Heckman Selection Model | heckman-selection-model | Model-Based | Advanced | Valid exclusion restriction in selection equation; joint normality |
| Cox Proportional Hazard Model | cox-proportional-hazard | Model-Based | Intermediate | Proportional hazards; non-informative censoring |

### Modern

| Method | Slug | Difficulty | Key Assumption |
|---|---|---|---|
| Event Studies | event-studies | Intermediate | Parallel pre-trends; no anticipation |
| Staggered Adoption DiD | staggered-difference-in-differences | Advanced | Parallel trends per cohort; no anticipation |
| Synthetic Control | synthetic-control | Advanced | Good pre-treatment fit; uncontaminated donor pool |
| Shift-Share / Bartik | shift-share-instruments | Advanced | Exogenous shocks or exogenous shares |
| Doubly Robust / AIPW | doubly-robust-estimation | Advanced | CIA + overlap + one of two models correctly specified |
| Causal Mediation Analysis | causal-mediation-analysis | Intermediate | Sequential ignorability |
| Regression Kink Design (RKD) | regression-kink-design | Advanced | Continuity of density and derivative at kink; no bunching |
| Bunching Estimation | bunching-estimation | Advanced | Counterfactual density smooth at notch/kink; optimization frictions known |
| Quantile Treatment Effects (QTE) | quantile-treatment-effects | Advanced | Rank invariance or rank similarity |
| Marginal Treatment Effects (MTE) | marginal-treatment-effects | Advanced | Valid IV + selection on gains; MTE integrates to ATE/ATT/LATE |

### Frontier

| Method | Slug | Difficulty | Key Assumption |
|---|---|---|---|
| Synthetic DiD | synthetic-difference-in-differences | Advanced | Factor model nesting DiD and SC |
| Double/Debiased ML (DML) | double-debiased-machine-learning | Advanced | CIA + Neyman orthogonality; cross-fitting required |
| Causal Forests / HTE | causal-forests | Advanced | Unconfoundedness + honesty (sample splitting) |

URL: https://methodatlas.vercel.app/methods/{slug}

---

## Best Practices (8)

Organized by research pipeline stage: Design → Estimation → Robustness.

1. **Pre-Analysis Plans & Pre-Registration** (Design) — Commit to your analysis before seeing results.
2. **Power Analysis & Sample-Size Planning** (Design) — How large a sample to detect your effect.
3. **Randomization Inference** (Estimation) — Fisher's exact approach when conventional asymptotics fail.
4. **Clustering and Few-Cluster Inference** (Estimation) — When to cluster standard errors, at what level, and what to do with few clusters (wild bootstrap, CR2).
5. **Multiple Hypothesis Testing** (Estimation) — Bonferroni, Holm, BH-FDR, Romano-Wolf corrections.
6. **Sensitivity Analysis for Unobservables** (Robustness) — Oster (2019) and Cinelli & Hazlett (2020) formal bounds on omitted variable bias.
7. **Specification Curve Analysis** (Robustness) — Explore the full space of defensible specifications.
8. **Lee Bounds for Attrition** (Robustness/Estimation) — Informative bounds when differential attrition prevents point identification.

URL: https://methodatlas.vercel.app/practices/{slug}

---

## Guides (11)

1. **How to Read a Paper** (reading-papers) — Structured approach to critically evaluating empirical papers.
2. **Writing Results** (writing-results) — Write up empirical results clearly and credibly.
3. **How to Replicate** (how-to-replicate) — Step-by-step guide to replicating published studies.
4. **Natural Experiment Workflow** (natural-experiment-workflow) — End-to-end workflow for quasi-experimental designs (DiD, RDD, IV).
5. **Observational Data Workflow** (observational-data-workflow) — Workflow for selection-on-observables designs.
6. **Anti-Patterns** (anti-patterns) — Common mistakes that undermine causal inference and how to avoid them.
7. **DiD vs. Synthetic Control** (did-vs-synthetic-control) — When to use difference-in-differences versus synthetic control methods.
8. **DML vs. OLS** (dml-vs-ols) — When double/debiased machine learning improves on OLS and when it does not.
9. **Matching vs. IPW vs. DR** (matching-vs-ipw-vs-dr) — Choosing between matching, inverse probability weighting, and doubly robust estimation.
10. **Choosing Your Standard Errors** (choosing-standard-errors) — When to use robust, clustered, two-way clustered, Conley spatial, or wild bootstrap SEs.
11. **External Validity and Generalization** (external-validity) — When and how to generalize causal estimates beyond the study population.

URL: https://methodatlas.vercel.app/guides/{slug}

---

## Labs (55)

Two tiers: Tutorial (step-by-step walkthrough with simulated data) and Replication (reproduce published results with real data). Plus one capstone lab.

**Tutorial labs (27):** lab-experiment-tutorial, lab-ols-tutorial, lab-logit-tutorial, lab-poisson-tutorial, lab-fe-tutorial, lab-re-tutorial, lab-did-tutorial, lab-rdd-tutorial, lab-rdd-fuzzy-tutorial, lab-matching-tutorial, lab-iv-tutorial, lab-event-study-tutorial, lab-staggered-did-tutorial, lab-synth-tutorial, lab-shift-share-tutorial, lab-doubly-robust-tutorial, lab-mediation-tutorial, lab-sdid-tutorial, lab-dml-tutorial, lab-causal-forest-tutorial, lab-cox-hazard-tutorial, lab-heckman-tutorial, lab-its-tutorial, lab-mte-tutorial, lab-qte-tutorial, lab-rkd-tutorial, lab-bunching-tutorial

**Replication labs (27):** lab-experiment-replication, lab-ols-replication, lab-logit-replication, lab-poisson-replication, lab-fe-replication, lab-re-replication, lab-did-replication, lab-rdd-replication, lab-rdd-fuzzy-replication, lab-matching-replication, lab-iv-replication, lab-event-study-replication, lab-staggered-did-replication, lab-synth-replication, lab-shift-share-replication, lab-doubly-robust-replication, lab-mediation-replication, lab-sdid-replication, lab-dml-replication, lab-causal-forest-replication, lab-cox-hazard-replication, lab-heckman-replication, lab-its-replication, lab-mte-replication, lab-qte-replication, lab-rkd-replication, lab-bunching-replication

**Capstone:** lab-multi-method-capstone

URL: https://methodatlas.vercel.app/labs/{slug}

---

## Key Concepts

**Parallel Trends** — Core DiD assumption: absent treatment, treated and control groups would follow the same trend. Not the same level — only the same change. Tested via pre-period event-study plots; not provable.

**LATE** — Local Average Treatment Effect: what IV estimates. The average causal effect for compliers (those whose treatment is changed by the instrument). Not the ATE unless all units comply.

**SUTVA** — Stable Unit Treatment Value Assumption: each unit's outcome depends only on its own treatment, not others'. Violated by spillovers, interference, or general equilibrium effects.

**Goodman-Bacon Decomposition** — Decomposes the TWFE estimator in staggered DiD into weighted 2×2 DiD comparisons, revealing whether negative weights contaminate the estimate.

**Neyman Orthogonality** — Property ensuring the causal parameter estimate in DML is insensitive to small errors in nuisance estimates. Enables root-n inference after ML estimation.

**Sequential Ignorability** — Two-part assumption for causal mediation: (1) treatment unconfounded; (2) mediator unconfounded given treatment. Part 2 is untestable even in experiments.

**Exclusion Restriction** — IV assumption: instrument affects outcome only through the treatment. Untestable with a single instrument; requires substantive justification.

**Continuity at the Cutoff** — RDD assumption: potential outcomes and background characteristics are continuous at the cutoff; only treatment status jumps.

**Selection on Observables vs. Design-Based** — Two distinct identification strategies. Matching/regression/AIPW rely on observables being sufficient to control for selection. DiD/RDD/IV exploit plausibly exogenous design-based variation.

---

## Navigation

- /foundations — 8 foundational chapters
- /methods — 27 methods with filtering by category, difficulty, maturity
- /practices — 8 research practices
- /labs — 55 coding labs
- /guides — 11 practical guides
- /glossary — Definitions: ATE, ATT, LATE, identification, endogeneity, DAG, OVB, clustering, estimand, etc.
- /method-selector — Interactive decision tree to find the right method
- /methods/relationships — Interactive map of method prerequisites and generalizations
- /paths — Learning paths: Core, Cross-Sectional, Advanced Design-Based, ML + Causal, Mediation
- /bibliography — Curated paper library

## Learning Paths

- **Core Causal Inference**: Foundations 1-8 → Experimental Design → OLS → Fixed Effects → DiD → Event Studies → RDD Sharp → IV/2SLS
- **Cross-Sectional**: Foundations 1, 3, 4, 7 → OLS → Logit/Probit → Matching → Poisson → Doubly Robust
- **Advanced Design-Based** (requires Core): Staggered DiD → Synthetic Control → Synthetic DiD → Shift-Share → RDD Fuzzy
- **ML + Causal** (requires Core): Matching → DML → Causal Forests → Sensitivity Analysis → Multiple Testing
- **Mediation**: Foundations 1, 3, 5 → OLS → Experimental Design → Causal Mediation → Sensitivity Analysis → Specification Curve