Causal inference mixtape

Causal Inference: The Mixtape — Code Skill

Practitioner-oriented causal inference skill adapted from the upstream Causal Inference Mixtape repository. It covers 10 identification strategies with illustrative Python, R, and Stata patterns. Every pattern is an adapt-and-verify scaffold, not a ready-to-run or automatically valid analysis.

Before using a pattern, reconstruct the study's assignment mechanism and estimand, adapt the sample and variance estimator, check the installed package API, and validate the adapted code on a known-answer example or simulation. A syntactically valid template does not establish identification.

---

Methods Covered

Method	Python	R	Stata	Reference
OLS / Regression	statsmodels	estimatr	reg/reghdfe	Method patterns §1
Difference-in-Differences	statsmodels + C()	lfe/fixest	xtreg/reghdfe	Method patterns §2
Event Study (Dynamic DiD)	manual lead/lag	estimatr	reghdfe	Method patterns §3
Staggered DiD / TWFE	statsmodels	bacondecomp	bacondecomp	Method patterns §4
Regression Discontinuity	statsmodels local regression	rdrobust/rddensity	rdplot/rdrobust/rddensity	Method patterns §5
Instrumental Variables	linearmodels IV2SLS	AER/ivreg	ivregress 2sls	Method patterns §6
Synthetic Control	rpy2 → R Synth	Synth + SCtools	synth	Method patterns §7
Matching / PSM / IPW	manual logit + weights	MatchIt + Zelig	teffects/cem	Method patterns §8
DAGs / Collider Bias	dagitty (conceptual)	dagitty/ggdag	—	Method patterns §9
Randomization Inference	explicit assignment draws	ri2	ritest	Method patterns §10

---

Core Workflow

Implement a Causal Method

1. Identify the method from the table above
2. Load the appropriate illustrative pattern from references/method-patterns.md
3. Adapt variable names, fixed effects, and clustering to the user's data
4. Add design-appropriate diagnostics (parallel trends for DiD, density and continuity checks for RDD, and first-stage plus weak-instrument diagnostics for IV)

Choose the Right Language

Scenario	Recommendation
ML pipeline integration	Python (statsmodels + linearmodels)
Synthetic Control	R (Synth package) or Stata (synth) — Python lacks mature implementation
Bacon decomposition	R (bacondecomp) or Stata — no Python equivalent
Publication-ready tables	Stata (outreg2/esttab) or R (stargazer/modelsummary)
Coarsened Exact Matching	Stata (cem) or R (MatchIt) — no Python equivalent
Quick prototyping	Python with statsmodels

Cross-Language Equivalents

Task	Python	R	Stata
OLS with robust SE	smf.ols().fit(cov_type='HC1')	lm_robust()	reg y x, robust
Cluster SE	fit(cov_type='cluster', cov_kwds={'groups': g})	`felm(y ~ x	0
Two-way FE	C(id) + C(time) in formula	`felm(y ~ x	id + time)`
IV / 2SLS	IV2SLS.from_formula('y ~ 1 + exog + [endog ~ inst]')	`ivreg(y ~ exog	inst)`
DiD	C(treat)*C(post)	treat:post in formula	did_multiplegt or interaction

---

Key Python Patterns

DiD with Cluster-Robust SE

import statsmodels.formula.api as smf

model = smf.ols('y ~ C(treated)*C(post) + controls', data=df)
results = model.fit(cov_type='cluster', cov_kwds={'groups': df['firm_id']})

Event Study (Lead/Lag)

# Omit event time -1 from estimation and use it as the reference period.
def event_term(k):
    return f'rel_m{abs(k)}' if k < 0 else f'rel_{k}'

event_times = [k for k in range(-4, 5) if k != -1]
rel_cols = [event_term(k) for k in event_times]
for k, col in zip(event_times, rel_cols):
    df[col] = (df['relative_time'] == k).astype(int)

formula = 'y ~ ' + ' + '.join(rel_cols) + ' + C(id) + C(year)'

IV / 2SLS

from linearmodels.iv import IV2SLS

model = IV2SLS.from_formula('y ~ 1 + exog + [endog ~ instrument]', data=df)
results = model.fit(cov_type='clustered', clusters=df['cluster_var'])

---

Design and Robustness Check Patterns

Method	Checks to adapt to the design
DiD	Parallel trends (event study plot), placebo treatment dates
RDD	Sorting-risk diagnostics (which may include a density-discontinuity test), bandwidth sensitivity, local-polynomial specification checks, and covariate continuity
IV	Design-appropriate first-stage diagnostics, weak-instrument-robust inference, a design/mechanism argument for exclusion with falsifiable implications, and cautious interpretation of over-identification tests when applicable
Synthetic Control	Pre-treatment RMSPE, placebo distribution, leave-one-out
Matching	Covariate balance table, caliper sensitivity

---

Common Pitfalls

1. TWFE under staggered adoption — variation in treatment timing alone does not prove bias. With heterogeneous or dynamic treatment effects, forbidden comparisons involving already-treated units can make conventional TWFE differ from the target ATT. Diagnose its comparisons and weights; then use an estimator such as Sun–Abraham or Callaway–Sant'Anna when its assumptions match the design.
2. Synthetic Control with many treated units — the Synth package handles one treated unit. For multiple, use augmented synthetic control or stacked approach.
3. RDD sorting diagnostics — a density discontinuity can be a signal of sorting or manipulation risk, but it is not a test of researcher or participant intent and is not informative in every design. Choose among McCrary-style, rddensity, heaping/mass-point, institutional, and covariate-continuity checks as the running variable and assignment process warrant.
4. IV weak instruments — do not treat a first-stage F-statistic cutoff of 10 as universal proof of instrument strength. Match the diagnostic to the number of endogenous regressors, covariance estimator, and design; report weak-instrument-robust confidence sets or tests when appropriate.
5. Python Synth gap — no mature Python Synth package exists. Use rpy2 to call R's Synth from Python.
6. IV diagnostic overclaims — exclusion is primarily a design and mechanism claim. A placebo is interpretable only when its outcome or covariate is predetermined, and failure to reject an over-identification test does not prove that every instrument is valid.
7. Randomization inference mechanism — redraw assignments from the original assignment mechanism. The Python pattern below covers only complete randomization with a fixed treated count; blocked, stratified, paired, or clustered designs need corresponding constrained draws.

---

Additional Resources

Reference Files

• references/method-patterns.md — Illustrative patterns for all 10 methods; adapt and verify each one
• references/r-stata-comparison.md — Cross-language package comparison and method coverage gaps
• references/third-party-notices.md — Direct upstream provenance, local-change disclosure, and the MIT license required for redistribution

Prompt Files

• prompts/01-implement-method.md — Adaptation prompt for drafting a causal-method workflow
• prompts/02-robustness-checks.md — Adaptation prompt for drafting design-specific robustness checks