Cps research design

Research Design (cps-research-design)

CPS is methodologically pluralist but demanding about each tradition. The design must credibly connect the comparative argument (cps-theory-building) to evidence and rule out the leading rival (cps-literature-positioning). This skill is mode-aware: pick the section that matches your work and defend the comparative leverage — the variation across cases or time that identifies the claim.

When to trigger

• Specifying the identification strategy, case selection, or experimental design
• A reviewer questioned causal claims, case choice, external validity, comparability, or a confound
• Designing a cross-national panel, a subnational comparison, or a natural experiment across borders
• Justifying why the design adjudicates the rival account, not just shows an association

Comparative-causal toolkit (cross-national / panel)

• Identification first. State the estimand and the assumptions that license a causal reading (parallel trends, exclusion, continuity, ignorability). Defend them; don't assert them.
• Designs: cross-national panels with unit and period fixed effects; DiD/event study around reforms (use modern staggered-adoption estimators, not naive TWFE); RD around institutional thresholds; IV (first-stage strength, exclusion, weak-IV-robust inference); survey experiments fielded comparatively.
• Comparability. Defend that the units are measured the same way across countries (V-Dem vs. Polity coding, harmonized surveys); address country-level confounding and cross-national measurement error.
• Inference: cluster at the level of treatment assignment (often country or country-year); few-cluster corrections (wild bootstrap) when the number of countries is small; multiple-comparison adjustment.
• Sensitivity: how strong must an unobserved country-level confounder be to overturn the result?

Case-based / qualitative comparison

• Comparison logic: most-similar (control on shared traits, vary the cause) or most-different (shared outcome despite different contexts) — justified by design, not convenience.
• Case selection: typical, deviant, most/least-likely, paired comparison. Say what each case is a case of and avoid selecting on the outcome.
• Process tracing with explicit tests (hoop, smoking-gun, straw-in-the-wind); state what evidence would have disconfirmed the mechanism.
• Source transparency: archives, interviews, fieldnotes — plan documentation (see cps-transparency-and-data).

Experiments (survey / field, fielded comparatively)

• Preregister design and primary analyses; report power/MDE; pre-specify subgroups and the comparison.
• For cross-country survey/conjoint experiments: equivalence of instruments and treatment realism across contexts; sampling frames; what the comparative contrast licenses about generalization.

Multi-method linkage

• Use the quantitative estimate for the average comparative effect and the case evidence for the mechanism; state how each method covers the other's blind spot, not as decoration.

The adjudication test (CPS-specific)

For the single strongest rival, write one sentence: "If the rival were true rather than my argument, the cross-case/over-time pattern would look like ___; instead it looks like ___." If you cannot, the design does not yet identify the comparative contribution.

Execution bridge (StatsPAI / Stata MCP)

Estimate and audit the design, don't only describe it. Full map: execution-with-mcp. CPS is comparative politics — cross-national and sub-national designs; emphasize identification and clustered / multiway inference.

• detect_design → recommend → fit with as_handle=true → audit_result.
• Observational causal claims: staggered DiD (callaway_santanna / sun_abraham + bacon_decomposition + honest_did_from_result); IV (effective_f_test + anderson_rubin_ci); RDD (rdrobust + mccrary_test).
• Experiments: randomization-based inference, romano_wolf for many-outcome family-wise control, and mediate for mediation (not naive controlling-away).
• Sensitivity: oster_delta / sensemakr for observational claims.

Report the effect size in interpretable units; route the full battery to the appendix/supplement. A run end-to-end (synthetic data, real returns) is in the JF execution walkthrough.

Anti-patterns

• Naive TWFE on staggered reforms; clustering at the wrong level; ignoring the small-number-of-countries problem
• "Causal" language on a design that only supports cross-national correlation
• Convenience or selecting-on-the-outcome case selection dressed up as most-similar design
• Cross-national survey experiment with non-equivalent instruments across countries
• A design with no comparative leverage — one snapshot that cannot distinguish your argument from the rival

Output format

【Mode】comparative-causal / case-based / experiment / multi-method
【Comparative leverage】the across-case / over-time variation that identifies the claim
【Estimand or claim】what is being identified/shown
【Key assumption(s)】and how each is defended (incl. comparability)
【Rival ruled out】the adjudication sentence
【Robustness/sensitivity】planned checks
【Next】cps-data-analysis

Supplementary resources

• ../../resources/external_tools.md — comparative datasets, identification packages, and CAQDAS for qualitative work
• ../../resources/code/ — staggered-DiD / IV / RDD / DML command chain to adapt