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.
identification argument justifies — not the one with the biggest coefficient or most stars.
countries use wild-cluster bootstrap or randomization inference. Report CIs, not just stars.
Polity, CSES, Manifesto Project); show robustness to alternative codings of the key construct.
probe the threats named in the design — not a scattershot table of every variant.
(cps-theory-building), not data-mined; adjust for multiple comparisons.
evidence in a multi-method design.
| 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 |
Run this audit before interpreting the main coefficient:
institutions? If not, report measurement-invariance checks, alternative codings, or scope limits.
cases observed? Report the observation process and show how estimates change under credible sample
restrictions.
unit trends, event-time plots, or placebo leads to avoid re-labeling persistence as effect.
Show leave-one-cluster-out or influence diagnostics for claims that hinge on few cases.
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.
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.
romano_wolf (step-down FWER) orbenjamini_hochberg — report the adjusted threshold.
oster_delta / sensemakr.wild_cluster_bootstrap (few clusters), twoway_cluster / conley;multilevel data → cluster at the right level.
audit_result(result_id) lists the missing checks and theexact suggest_function for each.
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.
【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
../../resources/code/ — clean → estimate → robustness → tables skeleton (Stata + Python)../../resources/external_tools.md — estimation and inference packages (R / Stata / Python)