Download skills/11-James-Traina-compound-science/skills/causal-inference/SKILL.md from the publisher, or copy the path and open the repository. PlugMCP lists the skill and does not host the file.
Where does causal-inference live?
The publisher keeps it at https://github.com/brycewang-stanford/Auto-Empirical-Research-Skills.
Which agents can use causal-inference?
Claude, Cursor, and Codex can read a SKILL.md you add from the publisher's repository.
>-
name: causal-inference
argument-hint: ""
description: >-
This skill covers causal inference methods in observational and quasi-experimental settings. Use when the user is implementing, choosing between, or debugging causal identification strategies — including instrumental variables, difference-in-differences, regression discontinuity, synthetic control, or matching estimators. Triggers on "causal effect", "identification strategy", "instrumental variable", "2SLS", "GMM", "difference-in-differences", "DiD", "staggered treatment", "regression discontinuity", "RDD", "synthetic control", "matching", "propensity score", "IPW", "AIPW", "doubly robust", "LATE", "ATT", "ATE", "parallel trends", "exclusion restriction", "first stage", "weak instruments", or "endogeneity".
Causal Inference
Reference for implementing causal inference methods: from identification strategy to estimation to diagnostics and robustness. Covers the major quasi-experimental and observational methods used in applied economics and quantitative social science.
When to Use This Skill
Use when the user is:
Choosing an identification strategy for a causal question
Implementing IV/2SLS, DiD, RDD, synthetic control, or matching
Implementing a specific method? Go directly to that method's section below
Need full code? See references/method-implementations.md for complete implementations
Frameworks
Two complementary frameworks underpin all causal inference:
Potential Outcomes (Rubin): Define Y(1), Y(0) as potential outcomes under treatment and control. The causal effect is τ = Y(1) - Y(0). The fundamental problem: we never observe both for the same unit. All methods are strategies for constructing valid counterfactuals.
DAGs (Pearl): Graphical models encoding conditional independence assumptions. Use d-separation to determine what must be conditioned on (and what must NOT be conditioned on) to identify causal effects. Particularly useful for reasoning about bad controls (colliders, mediators), overcontrol bias, and which instruments satisfy the exclusion restriction.
Quick Reference: Methods at a Glance
Method
Key Assumption
Target Parameter
Key Package
IV/2SLS
Exclusion restriction, monotonicity
LATE
linearmodels (Py), fixest (R), ivregress (Stata)
DiD
Parallel trends
ATT
fixest (R), reghdfe (Stata), linearmodels (Py)
RDD
No manipulation, local continuity
LATE at cutoff
rdrobust (all)
Synthetic Control
Weights reproduce pre-treatment trends
ATT (single unit)
Synth/augsynth (R)
Matching/AIPW
Selection on observables
ATE or ATT
econml (Py), MatchIt/WeightIt (R)
Target Parameters
Be precise about what parameter you are estimating:
Parameter
Definition
Estimated by
ATE
E[Y(1) - Y(0)]
Randomized experiment, IPW, AIPW
ATT
E[Y(1) - Y(0) | D=1]
DiD, matching, selection-on-observables
LATE
E[Y(1) - Y(0) | compliers]
IV/2SLS (Imbens-Angrist 1994)
ATT(g,t)
Group-time specific treatment effect
Staggered DiD (Callaway-Sant'Anna)
Common mistake: IV estimates LATE, not ATE. DiD estimates ATT, not ATE. This matters for policy interpretation.
Instrumental Variables (IV/2SLS)
Key idea: Find a variable Z that shifts D (first stage) but affects Y only through D (exclusion restriction).
from linearmodels.iv import IV2SLS
result = IV2SLS.from_formula(
'lwage ~ 1 + exper + expersq + [educ ~ nearc4 + nearc2]',
data=df
).fit(cov_type='robust')
# Always check first-stage F > 10; report LIML as robustness with weak instruments
For R/Stata implementations, weak instrument corrections (LIML, Anderson-Rubin), and overidentification tests, see references/method-implementations.md.
IV Diagnostics Checklist:
First-stage F > 10 (or Olea-Pflueger effective F for robust inference)
Reduced form significant (regress Y directly on Z)
Overidentification test reported if over-identified
Compare OLS vs 2SLS — direction and magnitude as expected?
Report LATE interpretation — who are the compliers?
Difference-in-Differences (DiD)
Key idea: Compare changes over time between treated and control groups, assuming they would have followed parallel trends absent treatment.
import statsmodels.formula.api as smf
# Standard 2x2 DiD
result = smf.ols('y ~ treated + post + treated:post', data=df).fit(
cov_type='cluster', cov_kwds={'groups': df['state']}
)
# Coefficient on treated:post is the DiD estimate
Staggered treatment timing: With staggered adoption, TWFE can produce sign-reversed estimates due to negative weights. Use:
Callaway-Sant'Anna (did R package): Most flexible aggregation, doubly robust
Sun-Abraham (fixest::sunab): Integrates directly into feols; simpler for event studies
Bacon decomposition (bacondecomp R): Diagnose how much weight TWFE puts on contaminated comparisons
For full staggered DiD code (C-SA, Sun-Abraham, BJS24, de Chaisemartin-D'H), see references/staggered-did.md. For event study code and HonestDiD pre-trend sensitivity, see references/method-implementations.md.
DiD Diagnostics Checklist:
Pre-trends: Event study shows no significant pre-treatment coefficients
Parallel trends sensitivity: Rambachan-Roth or similar analysis
Staggered timing: If varies, use C-SA or S-A — NOT naive TWFE
Clustering at level of treatment assignment (typically state/county)
Anticipation: Check period just before treatment
Bacon decomposition if using TWFE
Regression Discontinuity (RDD)
Key idea: Units just above and below a threshold are locally comparable; the jump at the threshold identifies the causal effect.
from rdrobust import rdrobust, rdbwselect, rdplot
# Basic sharp RD with bias-corrected robust CI
result = rdrobust(y=df['outcome'], x=df['running_var'], c=0)
# Reports: point estimate, robust CI, MSE-optimal bandwidth, N left/right
# Density test for manipulation
from rddensity import rddensity
density_test = rddensity(X=df['running_var'], c=0)
For fuzzy RDD, bandwidth sensitivity tables, and R/Stata implementations, see references/method-implementations.md.
RDD Diagnostics Checklist:
McCrary density test: No manipulation of running variable at cutoff
Covariate balance: Run RD on predetermined covariates as placebo outcomes
Local linear (p=1) is standard — avoid high-order polynomials
Donut hole: Drop observations very close to cutoff
Placebo cutoffs: Run RD where no effect should exist
Synthetic Control
When to use: Single or very few treated units, long pre-treatment series, no obvious comparison group. SC constructs a synthetic counterfactual as a weighted average of donor units.
Diagnostics: Pre-treatment RMSPE (fit quality), permutation/placebo tests across donor units, leave-one-out stability, time placebo at earlier date.
For full implementation (Synth setup, augsynth, permutation tests), see references/synthetic-control.md and references/method-implementations.md. The identification-critic agent can evaluate SC identification assumptions.
Matching and Weighting
Key idea: Reweight control group to match treated group on observed characteristics. Only valid under selection-on-observables (no unobserved confounders).
AIPW (doubly robust) is the recommended default — consistent if either the propensity score model or the outcome model is correctly specified.
from econml.dr import LinearDRLearner
from sklearn.ensemble import GradientBoostingRegressor, GradientBoostingClassifier
# Doubly robust learner
dr = LinearDRLearner(
model_regression=GradientBoostingRegressor(),
model_propensity=GradientBoostingClassifier()
)
dr.fit(Y=df['y'], T=df['treatment'], X=df[covariates], W=None)
ate = dr.ate(df[covariates])
For propensity score estimation, IPW/Hajek estimators, and manual AIPW implementation, see references/method-implementations.md.
Matching Diagnostics Checklist:
Covariate balance: Standardized mean differences < 0.1 after weighting
Common support: Substantial overlap in propensity score distributions
Sensitivity analysis: Rosenbaum bounds
No post-treatment covariates in the propensity model
Trim if propensity scores near 0 or 1
Method Selection Guide
Scenario
Recommended Method
Key Assumption
Random assignment with imperfect compliance
IV/2SLS
Exclusion restriction, monotonicity
Policy change at a threshold
RDD
No manipulation, local continuity
Policy change at a time point, treated and control groups
DiD
Parallel trends
Staggered policy adoption across units
Staggered DiD (C-SA, S-A)
Parallel trends (conditional)
Single treated unit, long pre-period
Synthetic control
Weights reproduce pre-treatment
Treatment assignment based on observables
Matching/IPW/AIPW
Selection on observables
Decision heuristic:
Is there a sharp threshold? → RDD
Is there an instrument? → IV
Is there a clean pre/post + treated/control? → DiD
Only one treated unit? → Synthetic control
Rich observables, selection on observables plausible? → AIPW
None of the above → structural model may be needed
Common Anti-Patterns
Anti-Pattern
Problem
Better Approach
TWFE with staggered timing and heterogeneous effects
Negative weights, biased estimates
Use Callaway-Sant'Anna or Sun-Abraham
Reporting 2SLS without first-stage F
Reader cannot assess instrument strength
Always report first-stage F (and LIML as robustness)
High-order polynomial in RDD
Overfitting, poor boundary properties
Use local linear (p=1) with rdrobust
Matching on post-treatment variables
Conditioning on outcome of treatment
Only match on pre-treatment covariates
Claiming parallel trends hold because pre-event coefficients are insignificant
Low power; absence of evidence ≠ evidence of absence
Use Rambachan-Roth sensitivity analysis
IPW with extreme propensity scores (near 0 or 1)
Huge variance, unstable estimates
Trim, use normalized/Hajek weights, or switch to AIPW
Reporting only one bandwidth in RDD
Cherry-picking concern
Show results across bandwidth range
Cluster-robust SEs with few clusters (< 30-40)
Poor finite-sample coverage
Wild cluster bootstrap (Cameron, Gelbach, Miller 2008)
Integration with compound-science
econometric-reviewer — Reviews identification strategy, standard errors, and asymptotic properties
identification-critic — Evaluates exclusion restrictions, support conditions, and identification completeness
references/method-implementations.md — Full IV/2SLS, DiD event study, RDD, and matching/AIPW implementation code
references/staggered-did.md — Full implementation code for Callaway-Sant'Anna, Sun-Abraham, BJS24, de Chaisemartin-D'Haultfoeuille, and Bacon decomposition
references/synthetic-control.md — Standard SC optimizer, permutation/placebo test code, augmented SC, diagnostics checklist
How do I install causal-inference?
Download skills/11-James-Traina-compound-science/skills/causal-inference/SKILL.md from the publisher, or copy the path and open the repository. PlugMCP lists the skill and does not host the file.
Where does causal-inference live?
The publisher keeps it at https://github.com/brycewang-stanford/Auto-Empirical-Research-Skills.
Which agents can use causal-inference?
Claude, Cursor, and Codex can read a SKILL.md you add from the publisher's repository.