Methods Catalog
20 causal inference methods, from the established workhorse regressions to frontier machine learning approaches.
Showing 20 of 20 methods
Experimental Design
The benchmark for causal inference — random assignment eliminates selection bias by design.
OLS (Robust SEs, Clustering)
The workhorse of empirical research — linear regression with modern standard error corrections.
Logit / Probit
Models for binary outcomes — when your dependent variable is yes/no, pass/fail, or adopt/don't adopt.
Poisson / Negative Binomial
Models for count outcomes — patents filed, citations received, number of acquisitions.
Fixed Effects (Two-Way FE)
Removes time-invariant unobserved confounders by exploiting within-unit variation over time.
Random Effects
A more efficient alternative to fixed effects when the unobserved effect is uncorrelated with regressors.
Difference-in-Differences (Canonical 2×2)
Estimates causal effects by comparing changes over time between treated and control groups.
Regression Discontinuity Design – Sharp
Exploits a sharp cutoff in treatment assignment to estimate causal effects near the threshold.
Regression Discontinuity Design – Fuzzy
When crossing the cutoff changes the probability of treatment (not a guarantee), use fuzzy RDD — essentially IV at the cutoff.
Matching (PSM, CEM, NN, Weighting)
Reduces selection bias by comparing treated units to similar control units based on observed characteristics.
Instrumental Variables / 2SLS
Uses an external source of variation (instrument) that affects treatment but not the outcome directly.
Event Studies (Dynamic Treatment Effects)
Visualize how treatment effects evolve over time — and test whether pre-trends support the parallel trends assumption.
Staggered Adoption DiD (Modern Estimators)
Under staggered adoption with heterogeneous effects, traditional TWFE can produce biased estimates — modern estimators correct for this.
Synthetic Control
Constructs a weighted combination of control units that best approximates the treated unit's pre-treatment trajectory.
Shift-Share / Bartik Instruments
Uses national-level shocks interacted with local-level exposure to construct instruments for endogenous variables.
Doubly Robust / AIPW Estimation
Combines outcome modeling and propensity score weighting — consistent if either model is correctly specified.
Causal Mediation Analysis
Goes beyond 'does the treatment work?' to ask 'through which pathway does it work?' — extends the Baron-Kenny framework by addressing its identification challenges.
Synthetic Difference-in-Differences
Combines the strengths of DiD (parallel trends) and synthetic control (matching on pre-treatment trajectory) into a single estimator.
Double/Debiased Machine Learning (DML)
Uses machine learning for nuisance parameter estimation while preserving valid inference on the causal parameter of interest.
Causal Forests / Heterogeneous Treatment Effects
Estimates how treatment effects vary across individuals — who benefits most and who benefits least.