Respol methods

Methods (respol-methods)

When to trigger

• The design is a default (OLS + controls, or a single case) and a referee questions whether it can support the claim
• A patent or bibliometric indicator is used as if it transparently measures "innovation" or "knowledge"
• A policy-evaluation paper rests on before/after comparison without an identification strategy
• A survey reports correlations vulnerable to common-method bias or selection
• A mixed-methods paper is really two studies stapled together with no integration logic

The Research Policy methods bar: pluralism with construct discipline

RP privileges no single technique — econometrics, bibliometrics/patent analysis, surveys, case studies, and mixed methods are all first-class — but it demands that the method fit the innovation-studies claim and that constructs be defended. The recurring RP failure is not weak statistics; it is treating an indicator (patents, citations, R&D spend) as if it were the latent thing (invention, knowledge flow, innovation effort). State what each measure does and does not capture, and design around its known biases.

Design paths by method

Patent / bibliometric quantitative

• Construct validity first. Patents measure patentable invention propensity, not innovation; citations measure traceable knowledge linkage, not knowledge value. Justify the indicator for your claim and acknowledge truncation, sectoral propensity, and home/strategic bias.
• Citation hygiene. Distinguish examiner- vs. applicant-added citations; correct self-citations; handle citation truncation (fixed windows or quasi-structural correction).
• Endogeneity. Patent/co-patent measures of spillovers are endogenous to firm location and choices — defend with design (instruments, natural experiments) or bound the bias, don't assume exogeneity.
• Family/coverage. State the patent office(s), family definition (DOCDB/INPADOC), and the matching of patents to firms/regions.

Causal policy evaluation (R&D subsidies, missions, IP reform)

• Move beyond naive before/after: DID with a credible control (and, for staggered policy rollout, Callaway-Sant'Anna / Sun-Abraham / de Chaisemartin-D'Haultfœuille rather than plain TWFE); IV with a defended exclusion restriction; RDD at an eligibility threshold with density and covariate-smoothness checks.
• Tie the estimand to the innovation mechanism (additionality vs. crowding-out, direction vs. rate of innovation), and argue policy-invariance for any counterfactual.

Survey (firm innovation behavior, CIS-style)

• Report sampling frame, response rate, and non-response/selection checks; address common-method bias by design (separated sources, marker variables, Harman is necessary not sufficient).
• Establish measurement validity (construct definitions, reliability, discriminant validity) before structural claims; state the population the estimates generalize to.

Case study / qualitative

• Justify case selection on theoretical grounds (typical, extreme, polar); specify data sources and triangulation; pursue analytic (not statistical) generalization; show how rich data became constructs.

Mixed methods

• State the integration logic: does the qualitative work generate the mechanism the quantitative work tests, or do the strands triangulate a construct? RP rejects "two papers in one" with no sequencing rationale.

Execution bridge (StatsPAI / Stata MCP)

Estimate and audit the design, don't only describe it. Full map: execution-with-mcp. Research Policy is innovation studies — patent/firm panels with selection; foreground identification and the selection objection.

• 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 control.
• Sensitivity: oster_delta / sensemakr for observational claims.

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

Checklist

• The method is chosen to fit the innovation-studies claim, not by convenience or fashion
• Each patent/bibliometric indicator's construct meaning and known biases are stated
• Citation measures handle examiner/applicant, self-citation, and truncation
• Causal designs use a credible counterfactual and modern staggered-treatment estimators where TWFE would bias
• Surveys report response/selection and address common-method bias by design
• Qualitative work defends case selection and generalization type
• Mixed methods state an explicit integration logic

Anti-patterns

• "Patents = innovation" with no construct caveat
• Spillover claims from co-location/citations treated as exogenous
• Plain TWFE on a staggered policy rollout with no heterogeneity discussion
• Survey structural models with unaddressed common-method bias
• A single case framed as if it gave statistical generalization
• Mixed methods with no sequencing or integration rationale

Output format

【Journal】Research Policy
【Skill】respol-methods
【Method】patent-bibliometric / causal-policy-eval / survey / case / mixed
【Construct defense】what the measure captures and its biases
【Identification / validity】counterfactual / first-stage / selection / triangulation
【Innovation-mechanism link】how the design speaks to the mechanism
【What it does NOT support】...
【Verdict】pass / revise / reroute
【Next skill】respol-data-analysis