技能 数据科学 比较数据分析方法论指南

比较数据分析方法论指南

v20260724
cps-data-analysis
本指南为比较政治学(CPS)领域提供了一套严谨的数据分析标准和最佳实践。它旨在解决跨国数据中常见的测量误差、集群效应和虚假相关性等陷阱,确保研究结果具有极高的透明度、可重复性和稳健性,符合学术顶级期刊的审稿要求。
获取技能
420 次下载
概览

Data Analysis (cps-data-analysis)

Once the design is fixed (cps-research-design), this skill governs how the analyses are run and reported so a CPS reviewer trusts them. Comparative data bring distinctive hazards: few clusters (countries), cross-national measurement error, missing data that differ by regime, and the temptation to over-read a panel correlation as causal. The standard is modern, transparent, and replication-ready.

When to trigger

  • Estimating the main results, robustness, and heterogeneity
  • A reviewer questioned standard errors, specification, measurement, or fragility of the result
  • Deciding what goes in the main text vs. the supplementary/online appendix
  • Triangulating quantitative estimates with case evidence

Analysis priorities (in order)

  1. Main estimate that matches the design. The headline specification should be the one the identification argument justifies — not the one with the biggest coefficient or most stars.
  2. Honest uncertainty. Cluster at the assignment level (usually country / country-year); with few countries use wild-cluster bootstrap or randomization inference. Report CIs, not just stars.
  3. Measurement transparency. Name the source and coding of each comparative variable (e.g., V-Dem, Polity, CSES, Manifesto Project); show robustness to alternative codings of the key construct.
  4. Robustness as a coherent story. Alternative specifications, samples, codings, and estimators that probe the threats named in the design — not a scattershot table of every variant.
  5. Heterogeneity by theory. Subgroups/scope conditions pre-specified by the mechanism (cps-theory-building), not data-mined; adjust for multiple comparisons.
  6. Mechanism evidence. Tie the quantitative result to the mechanism — mediation cautiously, or case evidence in a multi-method design.

Comparative-data hazards to address explicitly

Hazard Symptom Fix
Few clusters (countries) over-rejection, tiny SEs wild-cluster bootstrap / randomization inference
Cross-national measurement error results flip across codings show robustness to V-Dem/Polity/alt scales
Differential missingness sample changes by regime type report attrition; multiple imputation with caution
Time-series confounding spurious trend correlations unit + period FE; over-time placebo

Failure-mode audit

Run this audit before interpreting the main coefficient:

  • Concept equivalence: Does the key variable mean the same thing across regimes, languages, regions, or institutions? If not, report measurement-invariance checks, alternative codings, or scope limits.
  • Selection into observation: Are only more democratic, richer, more peaceful, or better-measured cases observed? Report the observation process and show how estimates change under credible sample restrictions.
  • Temporal dependence: Are observations mechanically persistent across years? Use lag structure, unit trends, event-time plots, or placebo leads to avoid re-labeling persistence as effect.
  • Cluster leverage: Does one country, region, election, conflict, or reform episode drive the result? Show leave-one-cluster-out or influence diagnostics for claims that hinge on few cases.
  • Subgroup multiplicity: If theory predicts heterogeneity, pre-specify the dimensions and report how many comparisons were examined.

The output should connect each failure mode to a design threat. Do not add a robustness table unless it answers a named threat in cps-research-design.

Execution bridge (StatsPAI / Stata MCP)

Run the battery, don't just enumerate it. Full map: execution-with-mcp. CPS is comparative politics — cross-national and sub-national designs; emphasize identification and clustered / multiway inference.

  • Many outcomes / specifications: romano_wolf (step-down FWER) or benjamini_hochberg — report the adjusted threshold.
  • OVB sensitivity: oster_delta / sensemakr.
  • Inference: wild_cluster_bootstrap (few clusters), twoway_cluster / conley; multilevel data → cluster at the right level.
  • Re-fit off one handle: audit_result(result_id) lists the missing checks and the exact suggest_function for each.
  • Exhibits: etable / did_summary_to_latex from the handle — no retyped numbers.

Keep the decisive checks in the body and the exhaustive battery in the supplement. See the executed chain in the JF execution walkthrough.

Checklist

  • Headline specification = the one the design justifies
  • SEs clustered at the assignment level; few-cluster correction where needed
  • Every comparative variable's source and coding named; key construct robust to alt codings
  • Robustness probes the design's named threats; not a kitchen sink
  • Heterogeneity pre-specified by theory; multiple testing addressed
  • Main text vs. appendix split is deliberate; every appendix result is referenced
  • All results reproduce from the script destined for the CPS Dataverse

Anti-patterns

  • Treating a cross-national panel correlation as causal without the design to back it
  • Default OLS SEs with 20 countries (massively over-rejects)
  • Cherry-picking the coding of the key variable that gives significance
  • Robustness theater — many variants that never test the actual threat
  • Data-mined subgroups reported as confirmed heterogeneity
  • Results in the paper that the deposited code does not reproduce

Output format

【Headline result】estimate + CI, with the design it rests on
【Inference】clustering level + few-cluster correction if any
【Measurement】sources/codings + alt-coding robustness
【Failure-mode audit】concept equivalence / observation selection / temporal dependence / cluster leverage / multiplicity
【Robustness】the design-threats probed
【Heterogeneity】theory-driven subgroups + multiple-testing fix
【Reproducible?】script regenerates every exhibit [Y/N]
【Next】cps-tables-figures

Supplementary resources

信息
Category 数据科学
Name cps-data-analysis
版本 v20260724
大小 6.68KB
更新时间 2026-07-28
语言