natural-experiments¶
Pack: 100xOS shared skills
Category:
analysisField: economics
License:
private (curator-owned)Updated: 2026-05-20
Stages:
data-analysisCurator-private skill — copy text from 100xOS/shared/skills/causal-inference/natural-experiments.md.
Natural Experiments: Taxonomy and Evaluation¶
What Makes a Natural Experiment¶
An exogenous event, institutional rule, or policy change assigns units to treatments as-if randomly, without researcher control. Credibility depends on: - Clear identification of the source of exogenous variation - Plausible independence from potential outcomes - Institutional knowledge of the assignment mechanism - Empirical support (balance tests, placebo tests, density tests)
Categories¶
| Category | Source of Variation | Key Concern |
|---|---|---|
| Policy changes | Cross-jurisdictional or over-time law/regulation variation | Policy endogeneity, anticipation effects |
| Geographic discontinuities | Borders, boundaries, spatial features | Endogenous sorting near boundaries |
| Weather/environmental | Rainfall, temperature, natural disasters | Multiple channels (exclusion restriction) |
| Lotteries | Draft, school admission, housing voucher | Imperfect compliance (fuzzy design, LATE) |
| Institutional assignment | Judge/examiner, date cutoffs, queue position | Conditional random assignment validity |
| Historical variation | Colonial institutions, historical infrastructure | Long causal chains, exclusion restriction |
Evaluating Natural Experiments¶
Internal Validity¶
- Is the variation truly exogenous? Can agents sort or select?
- Does the identifying assumption have testable implications?
- Are there confounding contemporaneous changes?
- Is the treatment well-defined or a bundle of changes?
External Validity¶
- Does the design estimate a LATE for a specific subpopulation?
- How representative is the affected population?
- Does the setting generalize to other contexts?
Reporting Standards¶
- Clearly state the source of identifying variation.
- Present the first stage (if IV) or treatment-control comparison.
- Show balance on pre-treatment covariates.
- Present reduced-form evidence.
- Discuss threats explicitly.
- Conduct robustness and placebo analyses.