원클릭으로 Manus에서 모든 스킬 실행

did-causal-inference

스타4

포크0

업데이트2026년 6월 9일 07:16

Use when the user needs difference-in-differences (DID) for causal inference, including DID treatment effect estimation, the parallel trends assumption, two-period/two-group (2x2) designs, policy or intervention/program evaluation, natural experiments, event studies for dynamic effects, two-way fixed effects (TWFE) regressions, pre-trend checks, staggered adoption caveats (Goodman-Bacon, Callaway-Sant'Anna, Sun-Abraham), ATT estimation, and clear reporting for panel data or repeated cross-section treatment-effect studies in Python, R, Stata, or SQL-backed analytics workflows.

설치

Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.

Manus에서 실행

출처

Zhongyang-ai

Zhongyang-ai/causal-inference-skill-for-marketing

GitHub 저장소 열기 Creator 저장소 보기

다운로드

Manus에서 실행

Difference-in-Differences (DID)

Use this skill for treatment effect questions where the user observes outcomes before and after an intervention for a treated group and a comparison group, and wants to isolate the causal effect by differencing out time trends and fixed group differences.

Workflow

Define the estimand before code:
- Treatment (the policy/intervention), outcome, unit of analysis, and the timing of the intervention.
- Identify the treated group (exposed to the intervention) and the control group (not exposed but observed over the same periods).
- Identify the pre period and post period relative to the intervention date.
- Target estimand: typically ATT (effect of treatment on the treated). State the eligible population and time window.
- Data structure: panel (same units observed over time) or repeated cross-section (different units, same groups/periods).
Check the parallel trends assumption:
- Parallel trends requires that, absent treatment, treated and control groups would have evolved with the same outcome trend.
- It is not directly testable; support it with pre-treatment evidence.
- With multiple pre periods, plot and test pre-trends: treated and control should track each other before the intervention.
- With only two periods, rely on domain reasoning, placebo groups/periods, and similarity of baseline trends.
Estimate the effect:
- 2x2 DID by hand: DID = (Ybar_treated_post - Ybar_treated_pre) - (Ybar_control_post - Ybar_control_pre).
- Regression form: Y = b0 + b1*Treat + b2*Post + b3*(Treat x Post) + e, where b3 is the DID estimate (ATT under parallel trends).
- Multi-period / many-unit designs: two-way fixed effects (TWFE) with unit and time fixed effects, treatment indicator switching on post-adoption.
- Add pre-treatment covariates only if they are not affected by treatment and improve the comparability of trends.
Get the uncertainty right:
- Cluster standard errors at the level of treatment assignment (commonly the unit or group), not the observation, to handle serial correlation.
- With few clusters, use cluster-robust corrections, wild cluster bootstrap, or report the limitation.
- Report the DID estimate with a confidence interval, and a practical effect size, not just a p-value.
Run an event study for dynamic effects:
- Interact treatment with leads and lags of event time, omitting one pre period as the baseline.
- Lead (pre-treatment) coefficients near zero support parallel trends; lag coefficients trace the dynamic treatment path.
- Use the event study to detect anticipation effects and effect ramp-up/decay.
Stress test the conclusion:
- Placebo tests: fake treatment dates or untreated groups should yield null effects.
- Vary the control group and the estimation window.
- Under staggered adoption, check whether naive TWFE is biased and use a robust estimator (Callaway-Sant'Anna, Sun-Abraham, de Chaisemartin-D'Haultfoeuille) or Goodman-Bacon decomposition.
- State residual risks from confounding shocks that hit one group only.

Reporting Checklist

Estimand (usually ATT), treated vs control definition, intervention timing, pre/post windows.
Data structure (panel or repeated cross-section) and unit of analysis.
Parallel trends evidence: pre-trend plot/test or the reasoning when untestable.
DID estimate from the 2x2 means and/or the interaction regression, shown to agree.
Standard errors with the clustering level stated, plus confidence interval.
Event-study results for dynamic effects and anticipation.
Placebo/robustness checks and staggered-adoption handling.
Caveats: untestable parallel trends, composition change, spillovers.

Guardrails

Parallel trends cannot be proven, only supported; passing a pre-trend test is necessary but not sufficient.
Under staggered adoption (units treated at different times) with heterogeneous effects, naive TWFE can be biased and even wrong-signed because already-treated units act as controls; use a modern estimator.
Composition change (the units making up a group differ across periods, especially in repeated cross-sections) can masquerade as a treatment effect.
DID assumes no spillovers: the control group must not be affected by the treatment (SUTVA). Contaminated controls bias the estimate.
A level difference between groups is not an effect; only the differential change after treatment is.
DID controls for time-invariant unobserved confounders but not for time-varying confounders or shocks that hit only one group.
Do not interpret the ATT as an ATE unless the design and homogeneity assumptions support it.

이 저장소의 다른 Skills

같은 저장소

aipw-causal-inference

Zhongyang-ai/causal-inference-skill-for-marketing

Use when the user needs augmented inverse probability weighting (AIPW) or doubly robust (DR) estimation for causal inference from observational data. AIPW combines an outcome regression model with propensity score weighting so the estimator is consistent if EITHER the propensity model OR the outcome model is correctly specified. Covers ATE and ATT estimation, the efficient influence function and its use for variance and confidence intervals, cross-fitting / sample splitting (TMLE-adjacent and double machine learning, DML, ideas) to control overfitting when using flexible ML learners, robustness to model misspecification, overlap/positivity checks, and clear reporting for treatment effect studies in Python, R, or SQL-backed analytics workflows.

2026-06-094

causal-inference

Zhongyang-ai/causal-inference-skill-for-marketing

Use as the entry point for any causal inference or treatment-effect question when the right method is not yet decided. This router inspects the user's data shape (cross-sectional vs panel, randomized vs observational, sample size) and business goal (average effect vs who-to-target), then recommends and dispatches to the appropriate sub-skill — PSM (psm-causal-inference), IPW (ipw-causal-inference), AIPW/doubly-robust (aipw-causal-inference), DID (did-causal-inference), or Uplift modeling (uplift-modeling). Triggers include 评估某个动作/投放/政策的因果效果, incrementality, lift, treatment effect, ATE, ATT, CATE, 该对谁投放 (who-to-target), A/B test analysis with confounding, observational treatment-effect studies, policy evaluation, and questions like 用哪种因果推断方法 / which causal method should I use. Prefer this skill first whenever the user uploads data and asks for a causal effect without naming a specific method.

2026-06-094

ipw-causal-inference

Zhongyang-ai/causal-inference-skill-for-marketing

Use when the user needs inverse probability weighting (IPW) or inverse probability of treatment weighting (IPTW) for causal inference from observational data, including propensity-score weighting, stabilized weights, ATE or ATT estimation via weighting, an introduction to marginal structural models, positivity/overlap diagnostics, weight trimming and truncation, effective sample size, Horvitz-Thompson and Hajek estimators, and covariate balance after weighting using weighted standardized mean differences. Applies to Python, R, Stata, or SQL-backed analytics workflows for observational treatment effect studies.

2026-06-094

psm-causal-inference

Zhongyang-ai/causal-inference-skill-for-marketing

Use when the user needs propensity score methods for causal inference, including propensity score matching (PSM), inverse probability weighting (IPW), overlap/common support checks, caliper or nearest-neighbor matching, covariate balance diagnostics using standardized mean differences, ATT/ATE estimation, sensitivity analysis, and clear reporting for observational treatment effect studies in Python, R, Stata, or SQL-backed analytics workflows.

2026-06-094

uplift-modeling

Zhongyang-ai/causal-inference-skill-for-marketing

Use when the user needs uplift modeling to estimate individual or conditional treatment effects (CATE/ITE) and heterogeneous treatment effects, i.e. who-to-target / treatment targeting questions rather than a single average effect. Covers segmenting a population into persuadables, sure things, lost causes, and sleeping dogs; meta-learners (S-learner, T-learner, X-learner) and uplift/causal trees; ranking-based evaluation with the Qini curve, uplift curve, and AUUC/Qini coefficient instead of ordinary classification or regression metrics; and marketing campaign targeting, retention, and budget allocation. Best suited to settings with a treatment flag and an outcome, ideally from a randomized A/B test or quasi-random/de-confounded assignment, where the goal is individual-level heterogeneity. Applies to Python, R, or SQL-backed analytics workflows.

2026-06-094

name

did-causal-inference

description

Difference-in-Differences (DID)

Workflow

Define the estimand before code:
- Treatment (the policy/intervention), outcome, unit of analysis, and the timing of the intervention.
- Identify the treated group (exposed to the intervention) and the control group (not exposed but observed over the same periods).
- Identify the pre period and post period relative to the intervention date.
- Target estimand: typically ATT (effect of treatment on the treated). State the eligible population and time window.
- Data structure: panel (same units observed over time) or repeated cross-section (different units, same groups/periods).
Check the parallel trends assumption:
- Parallel trends requires that, absent treatment, treated and control groups would have evolved with the same outcome trend.
- It is not directly testable; support it with pre-treatment evidence.
- With multiple pre periods, plot and test pre-trends: treated and control should track each other before the intervention.
- With only two periods, rely on domain reasoning, placebo groups/periods, and similarity of baseline trends.
Estimate the effect:
- 2x2 DID by hand: DID = (Ybar_treated_post - Ybar_treated_pre) - (Ybar_control_post - Ybar_control_pre).
- Regression form: Y = b0 + b1*Treat + b2*Post + b3*(Treat x Post) + e, where b3 is the DID estimate (ATT under parallel trends).
- Multi-period / many-unit designs: two-way fixed effects (TWFE) with unit and time fixed effects, treatment indicator switching on post-adoption.
- Add pre-treatment covariates only if they are not affected by treatment and improve the comparability of trends.
Get the uncertainty right:
- Cluster standard errors at the level of treatment assignment (commonly the unit or group), not the observation, to handle serial correlation.
- With few clusters, use cluster-robust corrections, wild cluster bootstrap, or report the limitation.
- Report the DID estimate with a confidence interval, and a practical effect size, not just a p-value.
Run an event study for dynamic effects:
- Interact treatment with leads and lags of event time, omitting one pre period as the baseline.
- Lead (pre-treatment) coefficients near zero support parallel trends; lag coefficients trace the dynamic treatment path.
- Use the event study to detect anticipation effects and effect ramp-up/decay.
Stress test the conclusion:
- Placebo tests: fake treatment dates or untreated groups should yield null effects.
- Vary the control group and the estimation window.
- Under staggered adoption, check whether naive TWFE is biased and use a robust estimator (Callaway-Sant'Anna, Sun-Abraham, de Chaisemartin-D'Haultfoeuille) or Goodman-Bacon decomposition.
- State residual risks from confounding shocks that hit one group only.

Reporting Checklist

Estimand (usually ATT), treated vs control definition, intervention timing, pre/post windows.
Data structure (panel or repeated cross-section) and unit of analysis.
Parallel trends evidence: pre-trend plot/test or the reasoning when untestable.
DID estimate from the 2x2 means and/or the interaction regression, shown to agree.
Standard errors with the clustering level stated, plus confidence interval.
Event-study results for dynamic effects and anticipation.
Placebo/robustness checks and staggered-adoption handling.
Caveats: untestable parallel trends, composition change, spillovers.

Guardrails

Parallel trends cannot be proven, only supported; passing a pre-trend test is necessary but not sufficient.
Under staggered adoption (units treated at different times) with heterogeneous effects, naive TWFE can be biased and even wrong-signed because already-treated units act as controls; use a modern estimator.
Composition change (the units making up a group differ across periods, especially in repeated cross-sections) can masquerade as a treatment effect.
DID assumes no spillovers: the control group must not be affected by the treatment (SUTVA). Contaminated controls bias the estimate.
A level difference between groups is not an effect; only the differential change after treatment is.
DID controls for time-invariant unobserved confounders but not for time-varying confounders or shocks that hit only one group.
Do not interpret the ATT as an ATE unless the design and homogeneity assumptions support it.