Decision Tool
Which Method Should I Use?
Answer a few questions about your research design and data, and we will suggest a starting method. The selector is a guide, not a substitute for thinking carefully about your specific context.
Method Selector
Question 1 of 911%
Was the treatment randomly assigned?
This tool provides starting points, not definitive answers. Consult your advisor.
Method Comparison at a Glance
27 methods| Method | Key Assumption | Key Threat | Panel? | Outcome Type | Maturity | Difficulty |
|---|---|---|---|---|---|---|
| Experimental Design | Random assignment, SUTVA | Attrition, non-compliance, spillovers | No | Any | Established | Beginner |
| OLS | Exogeneity (E[u|X] = 0) | Omitted variable bias | No | Continuous | Established | Beginner |
| Logit / Probit | Correct distributional assumption | Omitted variable bias, separation | No | Binary | Established | Intermediate |
| Poisson / Negative Binomial | Conditional mean correctly specified | Overdispersion, excess zeros | No | Count | Established | Intermediate |
| Cox Proportional Hazard | Proportional hazards, non-informative censoring | PH violation, informative censoring, competing risks | No | Duration / Time-to-event | Established | Intermediate |
| Fixed Effects | Strict exogeneity (no feedback from outcomes to regressors) | Time-varying confounders, Nickell bias | Yes | Any | Established | Intermediate |
| Random Effects | Unit effects uncorrelated with regressors | Violation of RE assumption (correlated effects) | Yes | Any | Established | Intermediate |
| Difference-in-Differences | Parallel trends | Differential pre-trends, anticipation effects | Yes | Any | Established | Intermediate |
| Interrupted Time Series | Stable pre-trend, no concurrent events | Autocorrelation, history threats, short pre-period | No (single unit) | Any | Established | Intermediate |
| RDD -- Sharp | Continuity at cutoff, no manipulation | Running variable manipulation, specification sensitivity | No | Any | Established | Intermediate |
| Regression Kink Design | Smooth conditional expectations at kink | Bunching invalidates smoothness, weak first-stage kink | No | Any | Modern | Advanced |
| RDD -- Fuzzy | Continuity + monotonicity at cutoff | Weak first stage, running variable manipulation | No | Any | Established | Advanced |
| Matching | Selection on observables (CIA), overlap | Unobserved confounders, poor common support | No | Any | Established | Intermediate |
| Heckman Selection Model | Joint normality, exclusion restriction | No credible exclusion restriction, normality violation | No | Continuous (selected sample) | Established | Advanced |
| Instrumental Variables / 2SLS | Relevance, exogeneity, exclusion restriction | Weak instruments, exclusion restriction violation | No | Any | Established | Intermediate |
| Event Studies | No anticipation, parallel trends | Staggered timing contamination, underpowered pre-trends | Yes | Any | Modern | Intermediate |
| Staggered DiD | Parallel trends per cohort, no anticipation | Heterogeneous treatment effects + staggered timing | Yes | Any | Modern | Advanced |
| Synthetic Control | Good pre-treatment fit, no spillovers | Poor fit, contaminated donor pool | Yes (aggregate) | Any | Modern | Advanced |
| Shift-Share / Bartik | Exogenous shocks or shares | Correlated shocks, concentrated Rotemberg weights | Often | Any | Modern | Advanced |
| Bunching Estimation | Smooth counterfactual density | Polynomial misspecification, optimization frictions | No | Density / Elasticity | Modern | Advanced |
| Doubly Robust / AIPW | CIA + one of two models correctly specified | Both models misspecified, positivity violation | No | Any | Modern | Advanced |
| Quantile Treatment Effects | CIA (unconfoundedness), correct specification | Conditional vs unconditional confusion, rank invariance | No | Continuous | Modern | Advanced |
| Causal Mediation | Sequential ignorability | Mediator-outcome confounding, treatment-mediator interaction | No | Any | Modern | Intermediate |
| Synthetic DiD | Parallel trends or good pre-treatment fit | Few donors, anticipation effects biasing time weights | Yes | Any | Frontier | Advanced |
| Double/Debiased ML | CIA + Neyman orthogonality | Weak nuisance models, no cross-fitting | No | Any | Frontier | Advanced |
| Causal Forests / HTE | Unconfoundedness, honesty | Spurious heterogeneity in small samples | No | Any | Frontier | Advanced |
| Marginal Treatment Effects | Valid IV + threshold-crossing selection model | Limited propensity score support, parametric extrapolation | No | Any | Modern | Advanced |