Identification Strategies

Routing Summary

This folder covers quasi-experimental methods from MHE Chapters 4-6, plus Bayesian and advanced synthetic control methods, DAG-based causal identification, and Bayesian propensity score weighting. Contains 16 notes plus the Propensity Score Matching sub-topic.

Sub-topics

Sub-topicNotesCovers
Propensity Score Matching5Propensity score & the balancing property (Rosenbaum-Rubin), matching methods & distance measures (nearest-neighbor, caliper, Mahalanobis, full/optimal), covariate balance diagnostics, common support / overlap — Stuart (2010)

Concept Map

ConceptNoteTypeDepends OnKey Result
2SLS, Wald estimator, exclusion restrictionInstrumental VariablesconceptOmitted Variables Bias, Regression and the CEF, The Selection Problem, Conditional Independence AssumptionIV estimates causal effect for compliers when CIA fails
LATE theorem, compliers/always-takers/never-takersLocal Average Treatment EffectsconceptInstrumental Variables, The Selection Problem, The Experimental IdealIV estimates LATE, not ATE — only for compliers
Fixed effects, common trends, Card & KruegerDifferences-in-DifferencesconceptThe Selection Problem, Regression and the CEF, Conditional Independence Assumption, Omitted Variables BiasDiD removes time-invariant unobserved confounders
Sharp and fuzzy RD, bandwidth choiceRegression Discontinuity DesignsconceptInstrumental Variables, Local Average Treatment Effects, Regression and the CEF, The Selection ProblemRD exploits threshold discontinuities for local causal effects
Bayesian DiD, posterior over treatment effectBayesian Difference in DifferencesconceptDifferences-in-Differences, The Selection Problem, The Experimental Ideal, Regression and the CEFFull posterior over counterfactual and treatment effect
Synthetic control, donor pool, convex weights, placebo inferenceSynthetic ControlconceptDifferences-in-Differences, The Selection Problem, The Experimental IdealWeighted combination of untreated units estimates counterfactual for single treated aggregate
Linear factor model, bias bound, sparsity theorem, V-matrixSynthetic Control Bias TheoryconceptSynthetic Control, The Selection ProblemBias bounded by pre-treatment fit; convex hull projection guarantees sparsity
RMSPE ratio, permutation p-value, backdating, leave-one-outSynthetic Control Inference and DiagnosticsconceptSynthetic Control, Synthetic Control Bias TheoryPermutation inference valid without asymptotic theory; RMSPE ratio detects poor donors
5 contextual + 3 data requirements, when not to use SCSynthetic Control RequirementsconceptSynthetic Control, Synthetic Control Bias TheorySC needs large effect, clean comparison group, and pre-treatment fit within convex hull
Penalized SC, bias-corrected SC, elastic net, matrix completionSynthetic Control ExtensionsconceptSynthetic Control, Synthetic Control Bias Theory, Synthetic Control Inference and DiagnosticsMultiple extensions handle multiple treated units, sparse donors, and matrix completion
IFE model, ATT estimand, 3-step GSC, bootstrap inferenceGeneralized Synthetic Control MethodconceptSynthetic Control, Differences-in-Differences, The Selection ProblemGSC unifies DiD and SC via interactive fixed effects; valid for multiple treated units
DAGs, forks/chains/colliders, backdoor criterion, d-separationDAGs and Causal IdentificationconceptThe Selection Problem, The Experimental IdealValid adjustment set blocks all backdoor paths; do-calculus enables causal inference from observational data
Bayesian IPTW, Liao-Zigler marginalization, Rubin’s rules SEBayesian Propensity Score WeightingconceptDAGs and Causal Identification, The Selection Problem, Bayesian Linear RegressionMarginalize over posterior propensity scores to incorporate treatment-model uncertainty

Notes

  • Mostly Harmless Econometrics - Overview — CONTAINS: Full book structure map, key concepts by chapter, cross-links to all identification strategy notes
  • Instrumental Variables — CONTAINS: 2SLS estimation, Wald estimator, exclusion restriction, quarter-of-birth and draft lottery examples, first-stage F-test
  • Local Average Treatment Effects — CONTAINS: LATE theorem, complier/always-taker/never-taker/defier taxonomy, characterizing compliers, external validity
  • Differences-in-Differences — CONTAINS: Fixed effects regression, common trends assumption, Card & Krueger minimum wage, event studies
  • Regression Discontinuity Designs — CONTAINS: Sharp and fuzzy RD, Lee incumbency example, Maimonides’ Rule, bandwidth selection, McCrary test
  • Bayesian Difference in Differences — CONTAINS: PyMC implementation, posterior over treatment effect delta, explicit counterfactual prediction, parallel trends as model constraint
  • Synthetic Control — CONTAINS: Formal potential outcomes setup, OLS weights (overfitting), constrained convex weights, scipy optimisation, California Proposition 99 cigarette tax example, placebo/permutation inference, Fisher’s exact test p-value
  • Abadie 2021 - Overview — CONTAINS: Full paper structure map, German reunification running example, 5 key takeaways, links to all derived notes
  • Synthetic Control Bias Theory — CONTAINS: Linear factor model (Eq. 10), bias bound theorem, sparsity theorem (convex hull projection), V matrix cross-validation, comparison with regression weights
  • Synthetic Control Inference and Diagnostics — CONTAINS: RMSPE definition, RMSPE ratio (Eq. 12), permutation p-value, backdating definition, leave-one-out robustness
  • Synthetic Control Requirements — CONTAINS: 5 contextual requirements (effect size, comparison group, no anticipation, no interference, convex hull), 3 data requirements, when not to use
  • Synthetic Control Extensions — CONTAINS: Penalized SC for multiple treated units (Eq. 13), uniqueness/sparsity theorem, bias-corrected SC (Eq. 15–16), elastic net SC (Eq. 17–18), matrix completion methods
  • Xu 2016 - Overview — CONTAINS: Paper overview, contribution summary, GSC vs DID/IFE/SC comparison table, caveats
  • Generalized Synthetic Control Method — CONTAINS: IFE model Assumption 1, strict exogeneity Assumption 2, ATT estimand, 3-step GSC estimator, LOO cross-validation Algorithm 1, parametric bootstrap Algorithm 2, Monte Carlo performance (Table 1), EDR voter turnout example
  • DAGs and Causal Identification — CONTAINS: Forks, chains, colliders, conditioning rules, d-separation formal definition, backdoor path definition, backdoor adjustment formula, valid adjustment sets, Simpson’s paradox
  • Bayesian Propensity Score Weighting — CONTAINS: IPTW formula, pseudo-population interpretation, why Bayesian propensity scores are problematic (likelihood incompatibility), Liao-Zigler marginalization integral, brms R implementation, Rubin’s rules SE combination

Sources

See Also