orgsci-data-analysis - SKILL.md Agent Skill

name: orgsci-data-analysis description: Use when analyzing and reporting results for an Organization Science manuscript across its eclectic methods — establishing trustworthiness for qualitative work, the right estimator for quantitative/multilevel data, transparency for simulation/formal models, and mechanism-supported inference where causal identification is impossible. Executes and reports the analysis; it does not design the study (orgsci-methods) or frame the contribution (orgsci-contribution-framing).

Data Analysis & Transparency (orgsci-data-analysis)

When to trigger

Data are collected (or a model built) and it is time to analyze and report
A reviewer says "the analysis does not support the inference"
You are unsure how to demonstrate rigor for qualitative, computational, or formal work
Your estimator may not match your design (nested, time-to-event, latent constructs)

Rigor is method-specific — report what your design demands

Because Organization Science is methodologically eclectic, "rigor" looks different by method. Report the standard that fits, transparently.

Qualitative / inductive. Validity becomes trustworthiness: a data structure (first-order codes → second-order themes → aggregate dimensions), an audit trail, and representative quotations so the path from raw data to constructs is traceable; state case-selection logic, saturation, and how disconfirming evidence was handled.
Quantitative / multilevel. Match the estimator to the design: multilevel/HLM for nested data (justify aggregation with ICC(1), ICC(2), r_wg), SEM for latent constructs and mediation, fixed/random effects and event-history for panel/founding-failure data; cluster SEs to the nesting structure.
Experimental. Report manipulation and attention checks, randomization balance, and effect sizes — not just p-values; pre-register where feasible.
Computational / simulation. Report parameter ranges, sensitivity analyses, seeds, and number of runs; show the qualitative pattern's robustness and make the mechanism visible.
Formal-analytical. State assumptions and prove or numerically verify the comparative statics; show which results are general and which are knife-edge.

Inference without identification

The venue holds that causal inference is valued but "not necessary and often impossible." Where clean identification is unavailable, support inference with mechanism evidence, theoretical logic, institutional knowledge, triangulation, and falsification/placebo logic, and rule out alternatives empirically where possible. Do not overclaim causality; do not abandon a credible mechanism for want of an instrument.

Transparency and replicability

Provide enough detail and references that others could replicate the work. Under the 2025 data and methods transparency policy, authors should be ready to share statistical code upon editor request during review, and accepted quantitative papers must publicly share data/code unless a documented exception and alternative transparency plan applies. Keep clean, regenerable scripts regardless of whether the underlying data can be public.

Execution bridge (StatsPAI / Stata MCP)

Run the battery, don't just enumerate it. Full map: execution-with-mcp. Org Science spans field studies, experiments, and computational/qualitative work; the chain below is for its empirical/causal lane — simulation and qualitative work are outside it.

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 appendix. See the executed chain in the JF execution walkthrough.

Anti-patterns

OLS on nested data (use HLM); causal-steps mediation instead of bootstrapped indirect effects.
Qualitative findings with no data structure or audit trail — thick description as "analysis."
Simulation from a single parameter set with no sensitivity analysis.
Causal language the design cannot support; specification fishing.

Evidence pass for Organization Science

Use this as a second-pass capability check. First lock a level map, a mechanism paragraph, and the cover-letter contribution statement; then test whether the manuscript addresses interdisciplinary organization reviewers who ask whether the mechanism travels across levels of analysis.

Primary move: Audit unit, comparison, uncertainty, missingness, sensitivity, and reproducibility before making any prose or submission recommendation.
Decision ledger: return claim / evidence / blocker / next edit rows so the next pass can patch the manuscript directly.
Neighbor test: compare against AMJ for empirical management framing, ASQ for organization-theory depth, Management Science for formal/quantitative operations; if the neighboring outlet has the stronger audience claim, recommend re-routing before polishing.
Submission-ready gate: before final advice, re-open resources/official-source-map.md for upload-week rules and name the one live-check item that could change the recommendation.

Output format

【Method】qualitative / multilevel / panel-EH / experiment / simulation / formal
【Rigor evidence】trustworthiness (data structure, audit trail) OR estimator + SE clustering OR sensitivity/seeds OR proofs
【Inference】identification? if not — mechanism + design logic + falsification; alternatives ruled out
【Replicability】detail/scripts ready; data/code sharing or exception plan ready
【Next step】orgsci-contribution-framing