Assumptions cheatsheet
Identifying assumptions for every method, with testability indicators and the standard diagnostic each assumption calls for.
Difference-in-differences
Estimand: ATTAbsent treatment, treated and control groups would have followed the same outcome trend.
Units do not alter behavior before treatment in response to expected future treatment.
No interference between units and a single version of treatment.
Staggered DiD (Callaway–Sant'Anna)
Estimand: ATT(g,t)For each cohort g, trends would have matched those of the not-yet-treated (or never-treated) units, conditional on covariates.
Cohort g does not respond before period g.
Already-treated units are not used as the comparison group for later-treated cohorts.
Instrumental variables (2SLS)
Estimand: LATEThe instrument is correlated with treatment. Weak instruments inflate 2SLS bias toward OLS.
The instrument affects the outcome only through treatment, no direct path from Z to Y.
The instrument is as-good-as-randomly assigned, independent of potential outcomes and treatment compliance types.
No defiers, if Z moves D for any unit, it moves it in the same direction for all. Ensures complier pool is non-empty.
Sharp RDD
Estimand: ATE (local)Potential outcomes are smooth functions of the running variable at the cutoff. Any discontinuity in observed outcomes is attributed to treatment.
Units cannot precisely sort above or below the cutoff. Assignment is effectively random in a neighborhood of c.
Observed covariates are continuous at the cutoff, ruling out discontinuities that could confound the estimate.
Fuzzy RDD
Estimand: LATE (local)Same as sharp RDD: both potential outcomes continuous in the running variable at c.
Same as sharp RDD.
Treatment probability weakly increases (or decreases) at the cutoff for all units, no defiers.
Matching & IPW
Estimand: ATT or ATEAll common causes of D and Y are observed in X. Unobserved confounding would invalidate the estimate.
Every covariate profile has a nonzero probability of treatment and control. Propensity scores of 0 or 1 violate overlap.
No interference between units; a single version of treatment.
Causal forest
Estimand: CATENo unobserved confounders given X.
Propensity bounded away from 0 and 1 uniformly in X.
Each tree uses separate subsamples for splitting and for estimation. Required for valid confidence intervals on CATEs.
AIPW (augmented IPW)
Estimand: ATE or CATENo unobserved confounders given X.
Propensity bounded away from 0 and 1.
Consistent if either the outcome model or the propensity model is correctly specified, not both required. Achieves semiparametric efficiency when both converge fast enough.
For valid inference, the cross-fitted outcome and propensity models must converge fast enough. Most flexible learners (lasso, forests) satisfy this.