科研 Skill

观察数据因果推断

alterlab-causal-inference

Estimates causal effects with DiD, IV, RDD, panel fixed effects, and propensity scores via statsmodels, linearmodels, DoWhy, and EconML. Use for observational or quasi-experimental causal estimation.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

Causal Inference — Name the Assumption, Estimate, Then Try to Break It

Skill type: ANALYSIS MODULE. Estimates a causal effect from data that was not fully randomized. The discipline is not the estimator — it is the identifying assumption the estimator relies on, stated before the fit and stress-tested after. If the design is not yet fixed, that belongs upstream in alterlab-ssci-design-gate.

Core Mission

EVERY CAUSAL ESTIMATE INHERITS AN ASSUMPTION. STATE IT, ESTIMATE UNDER IT, THEN REFUTE IT.

When to Use This Skill

  • "Estimate the effect with difference-in-differences / an event study."
  • "I have an instrument for the treatment — run instrumental variables / 2SLS."
  • "There's a cutoff score — run a regression discontinuity."
  • "Panel data with unit and time fixed effects — estimate the treatment effect."
  • "Give me the heterogeneous treatment effect / CATE across subgroups."

Does NOT Trigger

The request is really about…Route toWhy not this skill
Choosing the design & its identifying assumptionalterlab-ssci-design-gateDesign routing precedes estimation.
Plain OLS / descriptive or inferential stats (no causal identification)alterlab-statistical-analysis / alterlab-statsmodelsNo treatment-effect identification problem.
Latent-variable / SEM / factor structurealterlab-sem-psychometricsMeasurement model, not treatment effect.
Auditing whether the final claim is licensedalterlab-ssci-inference-gateClaim audit, downstream.

Hand-off rule. When the request matches a row above, open your answer by naming that skill by its exact id (for example alterlab-ssci-design-gate) and saying in one line why it fits better, then stop applying this skill's procedure. Name it even when you also need to ask for missing input. If only part of the request is out of scope, name the skill for that part and do only the in-scope part. A brief general pointer is fine; the named skill does the work.

Running the code. In Claude Code, run analysis code in a throwaway environment instead of installing packages: uv run --python ">=3.11" --with statsmodels --with "linearmodels>=6" --with pandas python analysis.py (add --with "pyfixest>=0.29", --with "dowhy>=0.12", --with "econml>=0.15" or --with "rdrobust>=1.3" as the design needs). This needs no install approval and leaves the user's environment untouched. If uv or the network is unavailable, say so and hand over the script; never report results you did not compute.

Estimator map (verified libraries, pinned)

DesignIdentifying assumptionEstimator (verified API)
DiD / event studyparallel trendspyfixest (v0.60): `pf.feols("y ~ i(year, treat, ref=2018)
Panel fixed effectsno time-varying confounderslinearmodels (v7): PanelOLS.from_formula("y ~ 1 + x + EntityEffects + TimeEffects", panel).fit(cov_type='clustered', cluster_entity=True).
Instrumental variablesexclusion restriction + relevancelinearmodels IV2SLS.from_formula("y ~ 1 + exog + [treat ~ z1 + z2]", df).fit(); check first-stage F (weak instrument).
Regression discontinuitycontinuity at the cutoff (no sorting)rdrobust (v2): rdrobust(y, x, c=cutoff), rdbwselect, rdplot; McCrary/density check for manipulation.
Selection-on-observablesconditional ignorabilityDoWhy (v0.14): CausalModel(df, treatment, outcome, graph).identify_effect() → estimate_effect(method_name="backdoor.propensity_score_matching") → refute_estimate(..., method_name="random_common_cause").
Heterogeneous effects (CATE)(as above) + overlapEconML (v0.17): LinearDML() / CausalForestDML() .fit(Y, T, X=X, W=W) then .effect(X).

A stdlib router that maps the design to the estimator + its assumption + the verified call: scripts/estimator_router.py. Fuller worked patterns and diagnostics: references/estimators.md.

The mandatory two steps around every estimate

  1. Before: write the identifying assumption in one sentence and how you will defend it (pre-trend plot for DiD, first-stage F and an exclusion argument for IV, density/McCrary test for RDD, overlap/common-support for PSM). No assumption, no causal estimate.
  2. After: run a refutation/robustness check — DoWhy.refute_estimate (random common cause, placebo treatment, data subset), a pre-trend/event-study plot, a bandwidth-sensitivity for RDD, or an E-value / sensitivity analysis for unmeasured confounding. Report it next to the estimate.

Output Template

DESIGN + ASSUMPTION: <e.g. DiD; parallel trends, defended by the pre-2019 event-study plot>
ESTIMATOR:           <library.call(...), pinned version>
ESTIMATE:            <point estimate, 95% CI, clustered SE — never a bare p-value>
DIAGNOSTIC:          <first-stage F / density test / pre-trends / overlap>
REFUTATION:          <placebo / random-common-cause / bandwidth sensitivity result>
CLAIM SCOPE:         causal IFF the assumption + diagnostics hold; else associational

Quality Standards

  • Report effect sizes with confidence intervals and the SE structure (clustered where relevant).
  • Never present a causal estimate without its diagnostic and at least one refutation.
  • Hand the final estimate + assumption to alterlab-ssci-inference-gate for the claim audit.
  • No fabricated flags: every call above is verified against the library's current docs; if a flag is unverified in your installed version, check --help/docs, do not guess.

References

  • references/estimators.md — per-design worked calls, diagnostics, and refutation menu.
  • scripts/estimator_router.py — stdlib design→estimator+assumption router.

Part of the AlterLab Academic Skills suite.

示例提示词

我有面板数据,想用双重差分和工具变量估计政策的因果效应。

所属插件

社会学、政治学与城市研究面向社会科学研究:问卷与访谈设计、质性编码、文本即数据、社会网络、ABM、比较政治与选举数据、RCT 与 DID、城市与区域分析。

安装该插件会一并获得同领域的 58 个 Skill / MCP。

同插件的其他科研 Skill

常见问题

观察数据因果推断能做什么?

Estimates causal effects with DiD, IV, RDD, panel fixed effects, and propensity scores via statsmodels, linearmodels, DoWhy, and EconML. Use for observational or quasi-experimental causal estimation。

如何安装观察数据因果推断?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「观察数据因果推断」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。

观察数据因果推断需要哪些权限?

该科研 Skill没有声明额外权限;执行具体操作时仍由 Agent 按自身规则请求确认。

观察数据因果推断经过官方审核吗?

是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。