科研 Skill
实证经济学写作助手
empirical-economics-writing
规划实证经济学研究全流程
Skill 说明
实证经济学研究与论文写作端到端工作流,支持帮我选回归变量、指标选取、变量设计、研究假设设计、CSMAR/Wind/Stata/Excel/CSV/DTA 数据清洗、面板数据构造、OLS/固定效应/DID/IV/RDD 回归、基准回归表、稳健性/异质性/中介/内生性检验和中英文论文/报告/docx | 输入:经济学题目 / 研究问题 / 数据包 / 选变量需求 | 输出:可复现实证工作流、回归代码、论文表格和实证经济学论文/报告
SKILL.md
Empirical Economics Writing
Use this skill to behave like a careful empirical-economics RA and thesis-writing assistant: start from the topic, turn it into a testable research design, make defensible variable/model/sample/test choices, then produce reproducible code, standard tables, and a paper/report that can survive expert review.
Conditional References
Load detailed guidance only for the stage that needs it:
- When drafting, revising, outlining, or reviewing paper prose, read
references/paper-architecture.mdand follow its section-function and argument-arc contracts. - Before creating, formatting, interpreting, or checking any table or figure, read
references/tables-and-figures.mdand follow its output contracts and editorial checks. - When formatting DOCX/PDF/LaTeX/Markdown, selecting a citation style, building the bibliography, or creating cross-references, read
references/formatting-and-citations.mdand follow its typography, citation-integrity, and automatic-numbering contracts. - When exporting a completed LaTeX empirical manuscript to DOCX, read
references/latex-and-docx-export.md, start fromassets/tex/mainbody-zh.texfor a Chinese paper unless an existing manuscript or user template governs, and use the bundled standalone exporter, templates, CSL files, and structural validation gate. The TeX asset is a compatibility base, not a prescribed paper outline. - When searching economics literature for variable measurement, hypotheses, identification, mechanisms, heterogeneity, or benchmark evidence, read
references/economics-literature-search.md. Hard ordering rule: the first literature-related tool call must be browser-tool discovery/activation or an actual browser call; never useacademic_search_*,academic-search_*, PaperSearch, scholarly aggregators, snippet-only paper APIs, or generic web-search tools before a real browser branch has been attempted and its outcome recorded. Do not use runtime in-app browser tools (in_app_browser_*) for literature search: they do not carry the user's authenticated browser session. Discover and use the external remote-browser MCP tools (mcp_remote_browser_*) first. After that remote-browser branch genuinely returns no usable candidates or hits an access blocker,academic_search_*may be used only as a supplementary discovery fallback; verify any retained candidate on a real source page through Browser. This fallback never replaces or completes a required CNKI branch. Treat Browser as a capability category rather than a literal tool name, rediscover and activate deferred browser tools before declaring that capability unavailable, use CNKI as the primary Chinese-literature source, and persist results throughscripts/bibliography_library.py. - Before final delivery of a full empirical paper, read
references/output-audit.md, create the audit manifest, and runscripts/empirical_output_audit.py empirical_audit.json --source mainbody.tex. This must cross-check registered tables against physical LaTeX tables and automatic prose references. Treat audit failures as delivery blockers unless the corresponding rule has a written, defensible exception. - For data cleaning, merging, variable construction, or regression work that does not produce paper prose or figures/tables, do not load these references.
Core Workflow
Unless the user asks for only one local step, proceed in this order:
Citation-Style Gate (run before literature search)
If the user has not explicitly supplied a journal/template citation requirement, call the runtime ask_user tool once before literature search or drafting. Do not use a generic citation-style menu. Render the header, question, option labels, descriptions, and examples in the user's current conversation language and provide exactly these two options. For an English conversation, ask Which in-text citation format should this economics paper use? (header: Citation format). For a Chinese conversation, ask 这篇经济学论文采用哪种正文引文格式? (header: 引文格式) and use the exact option labels 著者—年份制(经济学常用,推荐) and GB/T 7714 顺序编码制(右上角 [1]), with Chinese descriptions containing the same examples below:
Author-year (economics common; Recommended; e.g. Test et al. (2026))—Conconi et al. (2016),饶品贵等(2017); map to the backward-compatible keycitation_style: gb-t-7714-author-date. Its bibliography uses Chinese economics-journal author-year formatting rather than GB/T type markers. In DOCX, Pandoc creates internal links from these citations to the corresponding bibliography entries.GB/T 7714 sequential numeric (superscript [1]; e.g. …[1])— render[1]at the upper right of the cited text, as in…as prior research finds[1].; map tocitation_style: gb-t-7714-numeric.
Record the selected value in writing_constraints.json. Do not offer Chicago, a separate generic “numeric” option, or any other citation style in this skill's gate. If the user has already specified a style, record and apply it without asking.
- Clarify the research question, sample unit, and data level.
- Select Y, core X, controls, mediators, heterogeneity variables, and robustness alternatives with support from literature or database definitions.
- Acquire and audit data: prioritize user-provided datasets; use web data only when needed and allowed.
- Clean, merge, and construct variables with reproducible scripts, preserving a sample-selection log.
- Choose the empirical model from the research question and data structure.
- Run descriptive statistics, correlations, baseline regressions, heterogeneity, robustness, mediation/mechanism, and endogeneity checks.
- Write the paper/report with consistent variable notation, tables, code, and conclusions.
- Run quality gates before delivery.
Workspace And Output Boundary
Use the session workspace ($WORK) as the default root for every newly created manuscript, bibliography, script, cache, table, figure, audit file, and final deliverable. User-supplied input directories, including data/ directories, are read-only by default. Write outside $WORK only when the user explicitly names that exact output location; an input path alone is not output authorization.
For common Chinese undergraduate or master's empirical papers, expect:
topic -> indicator selection and method review -> raw data -> Python/Stata processing -> Stata regression analysis -> tables and paper writing.
Use Python for messy data engineering, file conversion, OCR/table extraction, web/API data, and reproducibility checks. Use Stata for standard empirical-economics regressions when available; if Stata is unavailable, use R or Python statistical packages and state the substitution.
Whiteboard Iteration Flow
For end-to-end cases, do not write a one-pass report. Follow the significance/non-significance loop shown in the whiteboard:
- Input the topic.
- Check whether X and Y have a theoretical or empirical relationship. If not, stop or ask the user to change the topic.
- Determine X/Y measurement and find data. If measurement lacks support from literature, database definitions, or statistical standards, revise the measurement first.
- Process the data.
- Test whether core X significantly affects Y. If not significant, loop back to measurement and data processing: check alternative measures, sample range, data construction, or model specification. After reasonable attempts, if the result is still not significant, report the failure honestly and do not fabricate a positive result.
- Add controls and fixed effects stepwise. If the main effect disappears after controls, adjust only defensible items: add theoretically relevant controls, replace poorly measured controls, correct data construction, or adjust the model. If the model still cannot satisfy the research requirement, abandon or redesign the topic.
- Once the main model passes, run the four extension gates: heterogeneity, mediation/mechanism, robustness, and endogeneity.
- Finish only when the main model and required extension gates are supported by tables, model equations, and coherent interpretation.
Decision rules for the four extension gates:
- Heterogeneity: if a grouping test is not significant or contradicts the theoretical story, switch only to a defensible grouping such as ownership, industry, region, firm size, or marketization. Do not search groups without theory.
- Mediation/mechanism: if the mediator fails, replace it only with a variable supported by the X -> Z -> Y mechanism and measurable in the data. Define the mediation model before reporting results.
- Robustness: try reasonable checks such as alternative X/Y measures, shorter samples, excluding subsamples, adding controls, changing fixed effects, or changing standard errors. The conclusion should not rely on one fragile specification.
- Endogeneity: use feasible methods such as instrumental variables, lagged X, placebo tests, parallel trends, or excluding high reverse-causality samples. A completed check must appear in a standalone registered results table rather than prose only. If no credible method exists, record a substantive
endogeneity_evidence_missingexception instead of pretending it is solved.
Research Design
Before running regressions, turn the topic into a precise design:
- Define the sample unit: firm-year, city-year, province-year, country-year, policy-region-year, transaction-level, etc.
- Define Y as the outcome most directly tied to the research question.
- Define X as the theoretical treatment or core explanatory variable; do not mix highly overlapping substitutes as if they were separate main variables.
- If X is a policy variable, construct
treat,post, andtreat_postor equivalent DID variables from policy timing and treated units; do not assume they already exist. - Select at least 5 controls for lightweight work. For full thesis-style work, use at least 8 baseline controls by default. Controls must be justified by theory and literature, not selected only because they are significant. If theory, data coverage, sample size, or identification permits fewer than 8, record the reason as a written exception.
- If mechanism/mediation analysis is requested, design at least 2 mediators and explain the X -> Z -> Y logic and model in the research design section.
- Plan heterogeneity before seeing results. A full thesis-style paper should include at least 3 theoretically meaningful dimensions by default, commonly ownership, industry, and East/Central/West region. If no defensible third dimension is measurable, record the reason as a written exception rather than inventing a grouping.
- Plan robustness before the conclusion. Common checks include replacing Y, replacing X, shortening or shifting the sample period, excluding special regions/industries/samples, changing winsorization, adding controls, changing fixed effects, and changing standard errors.
Every variable choice needs a measurement source: cited literature, a database definition file, a statistical yearbook definition, or a transparent construction rule.
Maintain a variable_registry with one canonical symbol for every variable and optional equivalent_group labels for alternative measures of the same construct. After a symbol is defined, reuse it exactly in equations, tables, figures, prose, and appendices. A baseline specification may contain at most one member of an equivalence group unless a written justification establishes that the variables measure distinct economic concepts in that design.
Data Workflow
Build a reproducible data pipeline and keep raw data immutable.
- Inventory every data file: format, rows, columns, years, units, key fields, data source, and variable definitions.
- Read database description files before guessing variables. Preserve official variable names, units, and coding rules.
- Normalize merge keys before merging: stock-code padding, date/year extraction, province/city names, industry codes, report type, currency/unit, and duplicates.
- Company-level data often come from CSMAR, Wind, or user-provided downloads; do not invent proprietary data that require accounts.
- City, province, and national data may come from statistical yearbooks, official statistics, or reliable databases; many city/province sources are PDFs, images, or web tables and may require OCR plus manual checks.
- Keep a sample-selection table or log: initial rows, each filter, rows dropped, final observations, firm/region counts, years, and the reason for each filter.
- Reconcile the sample arithmetically:
initial observations - sum(rows dropped at each step) = final observations. When the final panel is unbalanced, state whyunits x periodsdiffers from final observations. - Explain final sample size arithmetically. If the panel is unbalanced, do not claim
units x years = observations; explain IPOs, delistings, missing variables, failed region matches, fiscal-date filters, and other losses. - For financial statement data, state filters such as consolidated statements only, annual
12-31reports, excluding financial firms, and dropping missing key variables. - Winsorize continuous variables only after defining the analysis sample, usually at 1%/99%; state the rule.
Never hide sample loss. For example, if the final sample is 476 firms and 4501 firm-year observations over 11 years, explain that it is an unbalanced panel rather than implying it equals 476 * 11.
Model Selection
Choose the model from causal structure and data type:
- OLS/fixed effects: continuous Y in cross-sectional or panel data.
- Logit/Probit/LPM: binary Y; report marginal effects when useful.
- DID: policy shocks with treated/control units and pre/post periods.
- Event study / parallel trends: DID validity and dynamic effects.
- Spatial econometric models: when theory and data imply spatial spillovers.
- Mediation models: estimate X -> Z and X+Z -> Y; define the equations before presenting results.
- Instrumental variables: use only with a credible instrument; explain relevance, exclusion restriction, first stage, and weak-instrument checks.
Write model equations clearly and compactly. In baseline formulas, use Controls_it or ControlVars_it to represent controls when that improves readability. Unless there is a clear theoretical reason, do not include two equivalent X variables in the same baseline model, such as GDP_growth and ln_GDP.
Fixed effects and standard errors must match the design:
- Firm-year panels often need year and industry/firm fixed effects, with standard errors clustered at the firm, city, province, or policy-unit level depending on treatment variation.
- DID usually needs unit and year fixed effects, clustered at the treatment-assignment level.
- Table rows for fixed effects should say "Yes" only when the fixed effect is actually included in that column.
Tables and Regression Sequence
Every regression claim needs a table. Do not describe a regression that is not tabulated.
Every regression specification and every prose claim about it must resolve to an exact stable table identifier and column identifier. Related specifications may share one table, and secondary results may appear in an appendix, but prose-only coefficient or significance reporting is not allowed.
Before rendering the first table, apply the mandatory table-style gate in references/tables-and-figures.md: if no style or target-journal template is already specified, use the runtime's ask-user or structured user-input tool to ask the user to choose either standard three-line tables or economics-journal tables with internal horizontal and vertical separators. Wait for the answer, then use the selected style consistently throughout the paper.
When the user has selected the formal Chinese empirical-paper presentation exemplified by a thesis or journal reference, retain the selected border family and apply the chinese-empirical-thesis presentation profile: centred bold captions, model-number row followed by the dependent-variable row, coefficients and parenthetical statistics on separate lines, left-aligned variable names, centred estimates, and a left-aligned note. This profile is compatible with the standard three-line choice; it does not silently add vertical rules.
Unless field conventions or the user request differ, baseline tables should proceed as follows:
- Column (1): core X only, no controls, no fixed effects; if fixed effects are required by design, explain why.
- Column (2): add controls.
- Column (3): add fixed effects. This is usually the main baseline column to interpret.
- Optional later columns: add extra controls or alternative baseline specifications; do not mix heterogeneity, mediation, or robustness into the baseline table.
In a Chinese paper, a variable-definition/description/measurement subsection must contain one independent captioned variable-definition table with a stable tab: identifier; prose-only variable lists are not acceptable. Use Symbol | Variable role | Definition | Data source by default, put units and transformations in the definition or note, and list X and Y only once. Robustness alternatives belong in the robustness section, not as duplicate "core variables" in the main variable-definition table.
Table discipline:
- Put the dependent-variable symbol in the table title, header, or first row.
- Reuse exactly the symbols defined in the variable-definition table in all later tables.
- Prefer standard abbreviations from databases or literature; do not invent unusual symbols.
- For regression tables, report coefficients, standard errors in parentheses, significance stars, observations, R-squared or pseudo R-squared, fixed effects, and clustering notes. Correlation matrices report coefficients only, with no significance stars.
- If R-squared, pseudo R-squared, signs, zeros, or standard errors look abnormal, check logs and commands before interpreting.
- If correlation analysis appears in the text, include a correlation table; otherwise omit it.
- A correlation table must contain Y, the main X, and every baseline control used in the main specification.
- Heterogeneity tables should use grouped regressions or interaction models with clear group labels; unless justified, controls and fixed effects should match the baseline.
- Use one analytical question or dimension per main-text table. A robustness subtype, endogeneity method, or mechanism is a separate table unless its columns answer the same narrowly defined question. Do not make a catch-all extension table.
- Each pre-specified heterogeneity dimension must have its own table. For the default ownership, industry, and regional dimensions, produce three tables; each may use a two-level
group label -> model numberheader (or a three-column regional header) but do not combine the dimensions into one wide table. - Robustness sections should include tables for each regression family, not prose alone.
- The added-controls robustness specification should contain at least 11 controls by default. If fewer than 11 defensible controls are available, record a written exception explaining the theoretical or data constraint.
Paper Structure
Use a standard empirical-economics structure:
- Title and abstract: research question, data, method, main coefficient, mechanism/heterogeneity, contribution.
- Introduction: motivation, literature gap, identification intuition, contribution, main findings.
- Literature review and hypotheses: cite high-quality literature and connect theory to X, Y, controls, mediators, and heterogeneity.
- Research design: data sources, sample filters, variable definitions, model specification, fixed effects, standard errors.
- Empirical results: descriptive statistics, necessary correlations, baseline regressions, economic magnitude.
- Heterogeneity analysis: pre-specified groups and model.
- Robustness checks: variable replacement, sample adjustment, added controls, alternative fixed effects/standard errors/models.
- Mechanism or mediation analysis: only when variables and models have been defined earlier.
- Endogeneity or identification tests: report at least one feasible check in an automatically referenced standalone table; prose-only “discussion” is not a test. Use an explicit limitation only when no credible method is available.
- Conclusion and implications: summarize evidence without making causal claims beyond the identification design.
- References and appendices: cite data and literature; include code and table notes when needed.
Interpret results in two paragraphs:
- Paragraph 1 explains statistical meaning: coefficient direction, magnitude, significance, table/column, and whether the hypothesis is supported.
- Paragraph 2 explains theoretical or economic meaning.
In the baseline-results section, focus on X -> Y. Discuss controls in detail only when a result is surprising, exactly zero, theoretically central, or useful for diagnosing the model.
Do not put heterogeneity or interaction interpretation in baseline results unless the table is explicitly an interaction table. Do not put robustness substitutes in baseline tables or the main variable-definition table.
Deliverables
Depending on the user request, produce some or all of:
data_audit.md: raw files, candidate variables, merge keys, sample filters, final observations.papers.json: canonical structured literature library for browser-search-backed projects.references.bib: generated from the records actually cited in the paper.empirical_audit.json: variable registry, equivalence groups, sample flow, table/column map, specifications, claims, heterogeneity dimensions, and justified exceptions used by the output auditor.writing_constraints.json: recorded language, table style, and citation style used consistently across exports.build_panel.py/build_panel.do: cleaning and merge script.main_reg.do/main_reg.R/main_reg.py: regression script.- regression logs and machine-readable results.
- paper-ready tables.
- reproducible figures plus a
figuresregistry inempirical_audit.jsonrecording the figure ID, type, caption, theory basis or data source, and producing script. - final
.docxor Markdown report.
For DOCX delivery, use an auditable LaTeX manuscript as the source and run only scripts/export_paper.py according to references/latex-and-docx-export.md. For a new Chinese manuscript, copy assets/tex/mainbody-zh.tex into $WORK/mainbody.tex first, then adapt its content freely without restoring a TeX \tableofcontents, a manual bibliography, or unsupported package/macro complexity. Do not create a project-local export_docx.py, fall back to a whole-paper python-docx build, or hand-patch a failed output. The exporter discovers Pandoc through PANDOC_PATH, PATH, and the runtime-bundled pypandoc binary; it never installs Pandoc, edits the user's PATH, or invokes LibreOffice/soffice. If none is available, stop DOCX delivery and retain the TeX and analysis outputs. The skill owns conversion, native-OMML display equations, one standard Word TOC field using Word's complex-field form, bookmark hyperlinks for table/figure/equation/bibliography cross-references, five-point empirical table layout, centred figures, isolated bibliography, GB/T bilingual cleanup, Times New Roman enforcement for every Latin/number run, and structural OOXML validation including Word property ordering. A user- or journal-supplied reference DOCX overrides the bundled language template except where the user explicitly requires the Chinese-paper Latin font contract. Treat exporter or audit failures as blockers. A TOC placeholder in a preview client is not a broken output: on delivery, explicitly tell the user to open the DOCX in Microsoft Word, right-click the TOC, choose Update Field → Update entire table, and save; this normal one-time TOC refresh creates final page numbers and clickable entries. Do not ask the user to update any other fields. If the exporter reports a DOCX structural or cross-reference failure, do not add, remove, or rewrite manuscript citations, table/figure references, or equation references to satisfy a count; report the technical failure after one unchanged retry.
Normalize document language to zh or en before choosing templates and labels. For Chinese output, enforce localized abstract/contents/references headings, 一、 / (一) / 1、 heading numbering, bold SimHei 10.5 pt table captions, SimSun 10.5 pt figure captions with Times New Roman Latin text, and CJK font-audit receipts for raster figures. Keep the bibliography heading unnumbered at outline level 1 so it enters the TOC as plain 参考文献. Render bibliography entries at 10.5 pt and 1.5 spacing, justified without bullets or hanging indents. Under author-year style, list every bibliography author: English entries use A and B, A, B, and C, and the same final-and rule for longer lists; Chinese entries separate all authors with 、. Use et al. / 等 only for in-text citations with three or more authors. Format Chinese journal articles as 作者,年份,《题名》,《期刊》第X期。; format English articles as Author, Year, “Title”, Journal, Volume(Issue), Pages. with the journal italicized and no [J] / [J/OL] marker.
When producing any formatted paper, apply references/formatting-and-citations.md. Do not ask the user to choose font sizes: use its language-specific typography default unless a supplied target-journal template overrides it. If the citation style is not already explicit, complete its mandatory citation-style gate before formatting citations or the bibliography. Use automatic cross-references for tables, figures, equations, and sections rather than hard-coded numbers.
Quality Gates
Before delivering, verify:
- Research design: X, Y, controls, mediators, heterogeneity, robustness, and endogeneity are defined in the correct sections.
- Literature support: variable measurement and model choice have credible sources.
- Sample construction: final N, firm/region counts, years, and panel balance can be explained from the data log.
- Variable notation: each variable has one symbol, reused exactly in later tables.
- Baseline table: columns progress from X-only, to controls, to fixed effects, unless justified.
- Fixed effects: "Yes/No" rows match the actual commands.
- Controls: full thesis-style work uses at least 8 baseline controls and at least 11 controls in the added-controls robustness specification by default; any reduction needs a written exception.
- Heterogeneity: a full thesis-style paper uses at least 3 theoretically meaningful dimensions by default; any reduction needs a written exception.
- Extension-table scope: each heterogeneity dimension maps to exactly one standalone table; unrelated robustness, mechanism, and endogeneity subtypes are not compressed into one table unless a written exception explains why.
- Robustness: try to include variable replacement, sample restriction/exclusion, and added-control checks.
- Mediation/mechanism: if requested, include at least two mediators and define the models before results.
- Endogeneity: do not only say "future research"; implement feasible checks or explain why they cannot be done.
- Tables: every described regression has a table; every table has the dependent variable, observations, fixed effects, standard error/clustering note, and clear variable symbols.
- Manuscript evidence: every table ID in
empirical_audit.jsonresolves to a physical labeled table inmainbody.texand an automatic\autoref/\refin prose; manifest-only tables block delivery. - Table and figure contracts: load
references/tables-and-figures.md; verify that the chosen table style is consistent, prose references resolve, required fields and notes are present, and event-study or placebo figures are reproducible from data or estimation output. - Figure plan: record every included figure; include a conceptual mechanism diagram only when grounded in the stated theory, and quantitative descriptive trends, maps, coefficient plots, event studies, or placebo figures only when backed by saved data and a producing script. Do not add a DID-only diagnostic to a non-DID paper.
- Paper architecture: when prose is delivered, load
references/paper-architecture.md; verify that sections perform distinct argumentative roles and that contributions, result interpretations, and policy implications are tied to reported evidence. - Formatting and citations: for formatted outputs, load
references/formatting-and-citations.md; verify the fixed typography profile or supplied template, one consistent citation style, a canonical bibliography source, and resolved automatic cross-references with no hard-coded numbering. - DOCX export: when Word is delivered, load
references/latex-and-docx-export.md; for a new Chinese manuscript initialize fromassets/tex/mainbody-zh.tex; before the first figure useask_user/structured input to select and record grayscale or colour unless the user/template already specifies it. Verify Pandoc discovery/preflight, selected template/CSL, recorded table/citation/figure-colour styles, at least ten distinct academic sources actually cited in a full-paper body, exactly one native Word TOC field and internal cross-reference links, no non-TOC Word field code or external hyperlinks, native OMML equations, explicit centred table borders, Times New Roman Latin/number typography, valid OOXML, and no unresolved authoring markers. - Interpretation: distinguish percentage points from percent changes and compute economic magnitude correctly.
- Reproducibility: scripts can run from raw/intermediate data to final tables, and logs or outputs are saved.
- Executable audit:
scripts/empirical_output_audit.py empirical_audit.json --source mainbody.texexits successfully with no unresolved hard-rule failures.
Expert-Review Failure Patterns
Avoid these errors:
- Failing to explain why final observations do not equal
units x years. - Defining both the main variable and a robustness alternative as "core explanatory variables".
- Adding
ln_GDPor another equivalent variable to a baseline model that is supposed to focus on GDP growth, without a clear rationale. - Listing too many controls directly in the formula when a unified control term would be clearer.
- Writing correlation analysis without a correlation table.
- Mixing heterogeneity interactions or robustness variables into the baseline table.
- Marking fixed effects as "Yes" in a column where they are not included.
- Reporting abnormal R-squared without checking commands and sample.
- Over-explaining controls while under-explaining the core X coefficient.
- Claiming heterogeneity when grouped coefficients do not support the text.
- Discussing robustness, mediation, or IV without regression tables and model equations.
- Using informal labels such as "company age" in later tables after defining symbols such as
FirmAge.
相关科研 Skill
中药材分子鉴定用ITS2等DNA条形码鉴定药材基原物种,处理多基原与降解样品局限。适用于药材真伪鉴别研究,不替代法定检验。
中药网络药理学用TCMSP等数据库构建成分-靶点-通路网络并做富集分析,批判性讨论方法学局限。适用于中药机制假设生成与论文方法复核。
中医术语标准与知识图谱建立中医药术语与WHO/ICD-11标准的映射,设计本体并用Neo4j构建知识图谱。适用于数据整合与科研检索,不作临床决策系统。
方剂配伍规律挖掘对方剂做药名规范化、关联规则与网络聚类分析,发现候选核心药组。适用于用药规律挖掘,结果仅为待验证假设,不作处方依据。
证候与症状数据分析对症状量表数据做术语规范、潜在类别与因子分析,探索统计学证候模式。适用于证候分型研究,结果不用于个体辨证诊断。
泵用混合物热力学计算Calculates thermodynamic properties and mixture phase behavior for pump applications using the thermo Python library’s DIPPR/NIST correlations. Use for VLE or mixture property calculations.
常见问题
实证经济学写作助手能做什么?
实证经济学研究与论文写作端到端工作流,支持帮我选回归变量、指标选取、变量设计、研究假设设计、CSMAR/Wind/Stata/Excel/CSV/DTA 数据清洗、面板数据构造、OLS/固定效应/DID/IV/RDD 回归、基准回归表、稳健性/异质性/中介/内生性检验和中英文论文/报告/docx | 输入:经济学题目 / 研究问题 / 数据包 / 选变量需求 | 输出:可复现实证工作流、回归代码、论文表格和实证经济学论文/报告。
如何安装实证经济学写作助手?
在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「实证经济学写作助手」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。
实证经济学写作助手需要哪些权限?
该科研 Skill声明需要:本地文件。安装前会再次列出,由你确认后才会生效。
实证经济学写作助手经过官方审核吗?
是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。