quant-analysis

name	quant-analysis
description	Use when the user needs quantitative social-science analysis, empirical modeling, causal inference, regression, panel models, DID, IV, RDD, PSM, SCM, DML, robustness checks, empirical tables, figure-ready model outputs, or replication-package preflight.

Quant Analysis

This skill handles quantitative social-science analysis. It is not a generic statistics shortcut: it must keep research question, measurement, identification, diagnostics, robustness, interpretation, and reproducibility tied together.

Required References

Before formal analysis or causal wording, read:

• <VELA_RUNTIME_ROOT>\skills\catalog\empirical_quant_workflow.json

• <VELA_RUNTIME_ROOT>\skills\catalog\quality_gates.json

• <VELA_RUNTIME_ROOT>\skills\catalog\publication_style_rules.json

First Decision

Declare the task type before choosing a model:

• descriptive: distributions, rates, group differences, trends, or summary statistics.

• associational: relationships or correlations without a causal design.

• predictive: prediction, classification, scoring, or forecasting.

• causal: effect, impact, mechanism, policy effect, treatment effect, or counterfactual claim.

If the task is causal, do not run straight to regression. First select a design family and complete the causal-identification check.

Minimum Inputs

Ask for or infer these, and mark missing items:

• research question;

• unit of analysis;

• sample, population, geography, and time window;

• outcome variable;

• main explanatory variable or treatment;

• control variables or covariate set;

• estimand, descriptive statistic, prediction target, or causal design;

• data path or model output path;

• preferred language or tool chain: Python, R, Stata, Excel, or supplied outputs.

Design Families

Use empirical_quant_workflow.json as the design contract. Common families:

• descriptive or correlational analysis;

• panel fixed effects;

• difference-in-differences and event study;

• instrumental variables;

• regression discontinuity;

• matching, weighting, and balance designs;

• synthetic control and synthetic DID;

• machine-learning-assisted causal estimation;

• shift-share and Bartik designs;

• triple difference;

• randomized or field experiments;

• dose-response and continuous-treatment designs;

• mechanism, mediation, distributional effects, CATE, and causal forests.

Do not force economics or AER conventions onto every social-science task. Use these families only when the user's question, data, field expectation, reviewer request, or intended journal makes them relevant. Descriptive, qualitative, humanities, text-interpretive, and ordinary Chinese social-science writing can stay with the local workflow without importing applied-microeconometrics defaults.

Modern causal defaults matter. For example:

• Staggered DID cannot rely on naive TWFE as the only estimator.

• Modern DID must consider cohort, imputation, interaction-weighted, Bacon decomposition, or HonestDID checks when timing is staggered.

• IV needs exclusion logic, reduced form, first stage, and weak-IV-aware inference when relevant.

• Shift-share work must state whether identification comes from baseline shares or external shocks, then audit Rotemberg weights or shock-level inference.

• RDD should avoid high-order global polynomials and must check density/manipulation and covariate continuity.

• Matching and weighting need overlap, balance, and pre-treatment covariate discipline.

• Machine-learning causal estimates still need identification; prediction performance is not causal evidence.

Stata Ecosystem

When a project uses Stata or asks for AER/AEJ-style applied microeconometrics output, treat Stata as an auditable tool chain rather than pasted results.

Do not switch a Python, R, Excel, or supplied-output project into Stata just because Stata support exists. Stata is enabled when the project already uses it, collaborators or reviewers require it, a replication package needs it, or the user asks for it.

Minimum Stata package and command map:

• FE and panel: reghdfe, ftools, xtserial, xttest3, xtcsd.

• IV and weak-IV checks: ivreg2, ranktest, weakivtest, ivreghdfe, Anderson-Rubin style inference when needed.

• DID and event study: csdid, drdid, did_imputation, did_multiplegt, did_multiplegt_dyn, eventstudyinteract, bacondecomp, honestdid.

• RDD: rdrobust, rddensity.

• Synthetic and matching: synth, synth_runner, synth2, scul, psmatch2, teffects, ebalance, cem.

• Robust inference and multiple testing: boottest, ritest, rwolf, wyoung, psacalc.

• Tables and figures: esttab, estout, outreg2, asdoc, asdocx, coefplot, binscatter, heatplot.

For formal delivery, require a traceable do-file structure: run_all.do, 00_globals.do, 00_install_packages.do, cleaning, descriptives, main models, robustness, heterogeneity/mechanism, tables, figures, logs, and a replication README. Reject Stata outputs that cannot be traced back to commands, logs, and output paths.

Workflow

1. Build an empirical question card: question, unit, sample, variables, task type, and target output.

2. Lock the measurement map: construct definitions, proxy limitations, missingness, outliers, transformations, weights, clustering, and fixed effects.

3. If causal, run causal_identification_checked:

   • estimand or target parameter;

   • design family;

   • identifying assumption in plain language;

   • assignment or treatment process;

   • required diagnostics;

   • red flags and fallback language.

4. Build the analysis plan:

   • data cleaning and merge checks;

   • descriptive table or Table 1;

   • model family and inference rule;

   • robustness, heterogeneity, mechanism, placebo, and sensitivity checks;

   • figure/table output map;

   • replication artifact map.

5. Run or review the analysis.

6. Hand formal figures and tables to figure-table-studio.

7. Hand formal manuscript prose to manuscript-writing-studio, while preserving the empirical boundaries.

8. Hand final code/data/package work to reproducibility-package or social-science-submission-packager when the project is submission-bound.

Academic De-AI Surface Audit

The local workflow adapts useful de-template writing checks from Awesome-Agent-Skills-for-Empirical-Research, but does not install those external humanizer skills as competing routes.

Use academic-humanization-studio only after claims, evidence, variables, estimates, uncertainty, and citations are locked:

• remove empty rhetoric, mechanical transitions, vague attribution, inflated novelty, and repetitive sentence rhythm;

• preserve all facts, sample definitions, variables, coefficients, p values, confidence intervals, model limits, citations, and uncertainty;

• for English journal prose, prefer precise claims, compact sentences, field-standard phrasing, and explicit limitations;

• for Chinese social-science prose, keep concept density, problem awareness, theoretical linkage, and PEEL closure;

• run detect-first, rewrite-second, and second-pass audit;

• never promise to bypass Turnitin, CNKI AMLC, GPTZero, or any detector, and never make a detector score the acceptance target.

If improving prose requires changing a claim, variable, result, citation, or limitation, stop and ask the user to approve a substantive revision.

Required Outputs

For non-trivial quantitative work, provide:

• empirical question card;

• variable and sample map;

• design-family or descriptive-boundary note;

• diagnostics checklist;

• robustness matrix;

• table/figure output map;

• interpretation note separating estimate, inference, mechanism, and limitation;

• replication checklist or blocker list.

Failure Rules

Stop or downgrade the output when:

• unit, sample, or variable definitions are ambiguous;

• the requested causal wording lacks a design and identifying assumptions;

• data or model outputs are missing but the user asks for concrete estimates;

• diagnostics fail and the text would overstate the result;

• the user asks for detector evasion rather than academic-quality cleanup.