ft50-empirical-writing — скилл для Claude Code: суть и запуск
Скилл ft50-empirical-writing помогает готовить эмпирические статьи для журналов списка FT50 по менеджменту, финансам и бухучёту: от стратегии идентификации до тестов устойчивости.
MIT ★ 240 обновлён 2026-09-18
npx skills add franklee16/academic-research-skills --skill ft50-empirical-writing -a claude-codeЧто делает ft50-empirical-writing
Скилл структурирует подготовку академических рукописей под требования журналов FT50. Он ведёт автора через цепочку от формулировки исследовательского вопроса и теории к эконометрической идентификации эффекта, проверкам устойчивости и механизмам. В центре внимания находится именно причинно-следственный вывод, а не подбор предсказательных моделей.
Материал содержит рекомендации по профильным изданиям: бухучёт (JAR, JAE, TAR), финансы (JF, JFE, RFS) и менеджмент (AMJ, ASQ, SMJ). Описаны стандарты для квазиэкспериментов: разность разностей (DID) с несколькими периодами введения стимула, разрыв регрессии (RDD), инструментальные переменные и псевдоэксперименты. Без чёткого обоснования экзогенности данных и кластеризации стандартных ошибок рецензенты журналов первого эшелона отклонят рукопись.
Какие задачи решает
- Выбрать журнал из списка FT50 под конкретный дизайн исследования и тип данных
- Спроектировать стратегию идентификации причинно-следственной связи через DID, RDD или IV
- Подготовить серию проверок устойчивости результатов к альтернативным спецификациям
- Описать эмпирические механизмы через анализ гетерогенности, медиацию или модерацию
Примеры запросов агенту
- Помоги выбрать журнал FT50 по финансам для статьи с методом разности разностей.
- Спланируй проверки устойчивости для модели с инструментальными переменными.
- Оформи раздел с механизмами и анализом гетерогенности для журнала JAR.
Как установить ft50-empirical-writing
1Через Skills CLI (Vercel)
npx skills add franklee16/academic-research-skills --skill ft50-empirical-writing -a claude-code2Вручную: положить файлы
~/.claude/skills/ft50-empirical-writing-main/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/ft50-empirical-writing-main/
1Через Skills CLI (Vercel)
npx skills add franklee16/academic-research-skills --skill ft50-empirical-writing -a codex2Вручную: положить файлы
.agents/skills/ft50-empirical-writing-main/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add franklee16/academic-research-skills --skill ft50-empirical-writing -a cursor2Вручную: положить файлы
.cursor/skills/ft50-empirical-writing-main/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install franklee16/academic-research-skills paper-writing/ft50-empirical-writing-main/SKILL.md2Вручную: положить файлы
.github/skills/ft50-empirical-writing-main/SKILL.md1Через Skills CLI (Vercel)
npx skills add franklee16/academic-research-skills --skill ft50-empirical-writing -a gemini-cli2Командой платформы
gemini skills install https://github.com/franklee16/academic-research-skills --path paper-writing/ft50-empirical-writing-mainВручную: положить файлы
.devin/skills/ft50-empirical-writing-main/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Claude.ai
Упакуйте папку в zip и загрузите в Settings → Capabilities → Skills
По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов
Вопросы о ft50-empirical-writing
Как установить ft50-empirical-writing в Claude Code?
Выполните команду npx skills add franklee16/academic-research-skills --skill ft50-empirical-writing-main -a claude-code. Конфигурация запишется в ~/.claude/skills/ft50-empirical-writing-main/SKILL.md.
Что делает ft50-empirical-writing?
Скилл направляет написание эмпирических научных статей под критерии 50 ведущих журналов по бизнесу и менеджменту. Он помогает выстроить причинно-следственную идентификацию и аргументацию.
Какие дисциплины и журналы охватывает руководство?
В материалах описаны стандарты для учёта (JAR, JAE, TAR), финансов (JF, JFE, RFS) и менеджмента (AMJ, ASQ, SMJ). Охвачены квазиэкспериментальные методы и панельные регрессии.
Исходный текст
SKILL.md · MIT
---
name: ft50-empirical-writing
description: Use when writing or revising empirical research papers targeting FT50 journals across all management disciplines (Accounting, Finance, Management, Marketing, Operations/IS, Economics). Covers the full empirical pipeline — Research Question → Theory → Identification → Data → Results → Robustness → Mechanism → Implications — with discipline-specific journal guidance, identification strategy design, robustness frameworks, mechanism evidence, and writing patterns. Also use when stuck at any stage of an FT50-targeted empirical paper and needing stage-specific guidance.
---
# FT50 Empirical Research Paper Writing
## Overview
The FT50 (Financial Times 50) is the definitive list of 50 premier journals used by the Financial Times to rank business schools globally. Publishing in FT50 journals requires papers to meet the **highest standards of empirical rigor, identification credibility, and theoretical contribution** across all management disciplines.
This skill covers the full empirical pipeline for causal-inference-oriented research: **Research Question → Theory → Identification → Data → Results → Robustness → Mechanism → Economic Significance → Implications**. Unlike MS&E papers that prioritize mathematical modeling or ML papers that emphasize benchmark performance, FT50 empirical papers are evaluated on the **credibility of the causal identification strategy** and the **compelling nature of the research design**.
**Core Principle**: An FT50 empirical paper tells a story of **causal identification**. The reader must be convinced — through a combination of institutional knowledge, theoretical reasoning, and econometric design — that the estimated effect is not merely a correlation but a genuine causal relationship, and that the magnitude is economically meaningful.
---
## When to Use This Skill
- Drafting an empirical paper targeting any FT50 journal across Accounting, Finance, Management, Marketing, Operations/IS, or Economics
- Designing an identification strategy (DID, RDD, IV, RCT, Synthetic Control, matching)
- Building a robustness framework that anticipates reviewer concerns
- Developing mechanism evidence (mediation, moderation, heterogeneity analysis)
- Selecting the right FT50 journal given a research design and contribution profile
- Writing the identification and robustness sections of an empirical paper
- Revising based on reviewer feedback from an FT50 journal (especially on causal claims)
- Converting a working paper or dissertation chapter into an FT50-submittable manuscript
- Assessing whether a research design is strong enough for FT50-level scrutiny
---
## Journal Selection: The FT50 Landscape
### The FT50 List and Discipline Mapping
The FT50 spans six broad management disciplines. Journal selection is a function of your paper's **topic domain**, **identification strategy rigor**, **theoretical contribution type**, and **practical relevance**.
#### Accounting
| Journal | Abbreviation | Core Identity | Emphasis |
|---------|-------------|--------------|----------|
| Journal of Accounting Research | JAR | Archival-empirical, capital markets | Causal identification in accounting settings; disclosure, auditing, regulation |
| Journal of Accounting and Economics | JAE | Positive accounting theory, contracting | Theory-empirical integration; compensation, governance, debt contracting |
| The Accounting Review | TAR | Broad accounting, experimental and archival | Mix of analytical, archival-empirical, and behavioral experimental |
| Contemporary Accounting Research | CAR | Broad international scope | Welcomes non-US settings; diverse methods including qualitative |
| Review of Accounting Studies | RAST | Valuation, disclosure, text analysis | Financially oriented; NLP/textual analysis increasingly common |
**Accounting FT50 Methodology Expectations**: Strong archival-empirical tradition. Panel data with firm and year fixed effects is the baseline. Difference-in-differences and regression discontinuity are standard. IV estimation is expected when endogeneity is a first-order concern. Natural experiments are highly valued. Behavioral experiments accepted at TAR and CAR. Capital market studies require event study methodology with appropriate benchmark models. Standard errors must be clustered at the level of treatment assignment.
#### Finance
| Journal | Abbreviation | Core Identity | Emphasis |
|---------|-------------|--------------|----------|
| Journal of Finance | JF | Top flagship, broad scope | Highest bar for causal identification; asset pricing, corporate finance |
| Journal of Financial Economics | JFE | Corporate finance, governance, contracting | Theory-empirical integration; institutional detail matters |
| Review of Financial Studies | RFS | Asset pricing, fintech, intermediation | Methodological innovation; structural estimation increasingly common |
| Journal of Financial and Quantitative Analysis | JFQA | Quantitative methods, portfolio theory | Methodologically rigorous; welcomes computational methods |
**Finance FT50 Methodology Expectations**: The "identification revolution" has fully arrived. Causal claims demand quasi-experimental or experimental designs. DID with staggered treatment (using Sun & Abraham, Callaway & Sant'Anna, or Borusyak et al. estimators), RDD, IV, and natural experiments. Standard errors clustered at firm or state level. Placebo tests mandatory. Economic magnitude emphasized alongside statistical significance. Matched sample designs common in corporate finance (PSM, CEM, entropy balancing). Event studies must address confounding events and use appropriate estimation windows.
#### Management
| Journal | Abbreviation | Core Identity | Emphasis |
|---------|-------------|--------------|----------|
| Academy of Management Journal | AMJ | Empirical, strong theory, organizational behavior | Compelling stories grounded in rich theory; multi-level research |
| Academy of Management Review | AMR | Pure theory and conceptual development | No empirical data; must propose new theory or substantially extend existing |
| Administrative Science Quarterly | ASQ | Organizational theory, qualitative and quantitative | Deep theoretical contribution; ethnographic, longitudinal, and archival |
| Strategic Management Journal | SMJ | Strategy, competitive advantage, firm performance | Causal identification increasingly required; natural experiments valued |
| Organization Science | OS | Organization theory, formal and informal | Welcomes formal modeling alongside empirical; computational methods |
| Journal of Management | JOM | Broad management, reviews, meta-analyses | Comprehensive reviews with theoretical integration; large-scale empirical |
| Journal of Management Studies | JMS | Broad management, critical and pluralistic | Welcomes heterodox approaches; critical management studies |
| Journal of International Business Studies | JIBS | International business, cross-border strategy | Multi-country empirical designs; comparative institutional analysis |
| Human Relations | HR | Organizational behavior, critical perspectives | Sociologically informed; qualitative and mixed methods valued |
| Human Resource Management | HRM | HR practices, employment relations | Practitioner-relevant but theoretically grounded |
| Leadership Quarterly | LQ | Leadership theory and measurement | Multi-level, multi-source designs; endogeneity addressed in leadership-outcome studies |
**Management FT50 Methodology Expectations**: Methodological pluralism — qualitative (ethnography, grounded theory, multiple case studies), quantitative (survey, archival panel), and mixed methods all accepted. For quantitative studies: multi-level modeling, panel data methods, and increasing demand for causal identification (IV, DID, natural experiments). Survey-based studies must address common method bias, measurement validity (CFA/SEM with appropriate fit indices), and endogeneity. ASQ and AMJ value rich qualitative data and process theorizing. SMJ increasingly requires identification strategies comparable to economics. Theory sections must be extensive and contribute to management theory — purely empirical contributions without theoretical advance are desk-rejected.
#### Marketing
| Journal | Abbreviation | Core Identity | Emphasis |
|---------|-------------|--------------|----------|
| Journal of Marketing | JM | Broad marketing, strategy-focused | Substantive contribution to marketing practice; field experiments valued |
| Journal of Marketing Research | JMR | Quantitative, consumer behavior, methods | Methodologically rigorous; experiments (lab, field, online), econometrics |
| Journal of Consumer Research | JCR | Consumer behavior, psychological foundations | Behavioral experiments; theoretical depth in consumer psychology |
| Marketing Science | MktSci | Quantitative modeling, structural estimation | Analytical models with empirical estimation; structural econometrics |
| Journal of the Academy of Marketing Science | JAMS | Broad marketing, strong theory | Conceptual and empirical; managerial relevance required |
| Journal of Consumer Psychology | JCP | Consumer psychology, experimental | Theory-driven experiments; psychological mechanisms |
**Marketing FT50 Methodology Expectations**: Structural econometric modeling for MktSci (demand estimation with BLP, dynamic structural models). Field and lab experiments for JMR, JCR, JCP. JM values field experiments and large-scale empirical work. JAMS requires strong theoretical contribution with empirical validation. Standard experimental design considerations: power analysis (a priori), randomization checks, manipulation checks, attention checks, demand effects, experimenter bias. Marketing papers must demonstrate both internal validity and external validity (generalizability). Mediation analysis (bootstrapped indirect effects; PROCESS models) is the dominant mechanism-evidence approach in consumer behavior. Moderation analysis to demonstrate boundary conditions is standard. Multi-study packages (3-5 or more studies per paper) are common in JCR and JMR.
#### Operations & Information Systems
| Journal | Abbreviation | Core Identity | Emphasis |
|---------|-------------|--------------|----------|
| Manufacturing & Service Operations Management | M&SOM (MSOM) | Empirical OM, behavioral operations | Data-driven operations; field experiments, structural estimation |
| Operations Research | OR | Mathematical optimization, stochastic models | Proofs, optimality guarantees, computational experiments |
| Information Systems Research | ISR | IS economics, behavioral IS, design science | Econometric identification and analytical modeling both valued |
| MIS Quarterly | MISQ | IS management, digital transformation | Broad IS phenomena; qualitative, quantitative, design science |
| Journal of Operations Management | JOM (Ops) | Empirical OM, supply chain, sustainability | Strong identification, multiple data sources, practice-grounded |
| Production and Operations Management | POM | Broad OM, data-driven, analytics | Empirical, analytical, and behavioral; strong practical relevance |
| Management Science | MS | Broad management science, high rigor | Cross-disciplinary; optimization, econometrics, behavioral experiments |
**Operations & IS FT50 Methodology Expectations**: MS, OR, and MSOM span analytical modeling (optimization, stochastic processes, game theory) and empirical work (DID, IV, structural estimation, field experiments). ISR and MISQ accept econometric identification, analytical modeling, and design science. JOM (Ops) and POM increasingly emphasize empirical identification with multiple data sources, natural experiments, and field experiments. OM empirical papers must demonstrate intimate knowledge of the operational context — "institutional detail" in OM replaces what "theory" does in management. IS papers must address digital technology phenomena with appropriate identification strategies (instrumental variables for endogenous technology adoption, DID for policy/program evaluations). Computational social science and machine learning methods are increasingly accepted in ISR and MISQ.
#### Economics (FT50-listed)
| Journal | Abbreviation | Core Identity | Emphasis |
|---------|-------------|--------------|----------|
| Econometrica | ECTA | Top theory and econometrics | Highest bar for formal theory and methods |
| American Economic Review | AER | Top general interest, short papers | Important, broadly accessible contributions; ~40 page limit in AER |
| Journal of Political Economy | JPE | Chicago tradition, deep theory | Long papers; deep theoretical foundation with empirical validation |
| Quarterly Journal of Economics | QJE | Long empirical papers, applied micro | Institutional detail, natural experiments, credible designs |
| Review of Economic Studies | REStud | Top European-based general interest | Theory and applied work; newer methods welcomed |
| Review of Economics and Statistics | REStat | Applied econometrics, empirical | Methodological innovation in applied settings |
| Journal of Econometrics | JoE | Econometric theory and methods | New estimators, tests, and identification results |
| Journal of Labor Economics | JOLE | Labor economics | Causal identification in labor markets; natural experiments |
| Rand Journal of Economics | RAND | Industrial organization, regulation | IO theory and empirical; structural estimation |
| Journal of Business and Economic Statistics | JBES | Statistical methods for economics | Time series, forecasting, machine learning in economic settings |
**Economics FT50 Methodology Expectations**: The "credibility revolution" originated here. Causal identification is non-negotiable. Papers are evaluated on: (1) identification strategy credibility, (2) institutional detail that motivates the design, and (3) economic significance. Standard designs: DID (with event-study plots showing pre-trends), RDD (with McCrary density tests, donut-hole robustness), IV (with first-stage F-statistic >> 10, plausibly exogenous instruments), RCT (with pre-analysis plan, balance tables, attrition analysis), synthetic control (with in-space placebo tests). Standard errors: cluster-robust at treatment level. Multiple hypothesis testing corrections (Bonferroni, Romano-Wolf, sharpened q-values). P-hacking concerns addressed transparently. Pre-analysis plans and replication packages increasingly required. The QJE event-study plot (coefficient-by-year with confidence intervals) is the gold standard for visualizing DID results.
### Journal Selection Decision Flowchart
```
What is the disciplinary home of your phenomenon?
├── ACCOUNTING (auditing, disclosure, tax, governance, contracting)
│ ├── Strong causal identification + capital markets → JAR, JAE
│ ├── Behavioral/experimental → TAR, CAR
│ ├── International setting → CAR
│ └── Theoretical-empirical integration → JAE, RAST
│
├── FINANCE (asset pricing, corporate, banking, fintech)
│ ├── Top causal identification + broad interest → JF
│ ├── Corporate finance + institutional detail → JFE
│ ├── Asset pricing + structural methods → RFS
│ └── Quantitative/computational methods → JFQA
│
├── MANAGEMENT (strategy, OB, OT, HR, IB, leadership)
│ ├── Strong causal identification + strategy → SMJ
│ ├── Rich theory + quantitative or qualitative → AMJ, ASQ
│ ├── Pure theory (no data) → AMR
│ ├── Formal modeling + empirical → OS
│ ├── Cross-border/cross-cultural → JIBS
│ ├── Leadership focus → LQ
│ └── HR/employment focus → HRM, HR
│
├── MARKETING (strategy, CB, quantitative, advertising)
│ ├── Structural econometric modeling → MktSci
│ ├── Field experiment + strategy → JM
│ ├── Consumer behavior experiments → JCR, JCP
│ ├── Quantitative methods + experiments → JMR
│ └── Conceptual + empirical + managerial → JAMS
│
├── OPERATIONS / IS (supply chain, service ops, digital, analytics)
│ ├── Analytical modeling (optimization/stochastic) → OR, MS
│ ├── Empirical OM (DID, field experiments) → MSOM, JOM (Ops), POM
│ ├── IS economics + econometric identification → ISR
│ ├── IS management + mixed methods → MISQ
│ └── Cross-disciplinary + high rigor → MS
│
└── ECONOMICS (labor, IO, public, development, econometrics)
├── Top general interest → AER
├── Deep theory + empirical → JPE, ECTA
├── Applied micro + natural experiments → QJE, REStud
├── Applied econometrics → REStat, JoE, JBES
├── Labor → JOLE
└── IO + regulation → RAND
```
---
## The FT50 Empirical Paper Structure
Unlike the IMRAD format in biomedical sciences or the model-analysis-algorithm structure in MS&E, an FT50 empirical paper follows a **causal-identification-centric** structure:
| Section | Typical Length | Key Content |
|---------|---------------|-------------|
| Abstract | 100-250 words | Research question, identification strategy, key estimate, mechanism, economic significance |
| S1 Introduction | 3-5 pages | Motivating puzzle, research question, institutional context preview, identification approach, preview of findings, contributions (numbered, 3-5) |
| S2 Institutional Background / Theory | 3-6 pages | Institutional details that motivate the design; theoretical framework generating testable predictions; conceptual model |
| S3 Data & Measurement | 3-5 pages | Data sources, sample construction (with attrition table), variable definitions, descriptive statistics, balance/correlation table |
| S4 Identification Strategy / Research Design | 3-6 pages | Identification challenge (endogeneity source), empirical specification (estimating equation), identification assumptions (stated and defended), graphical evidence (reduced-form plots) |
| S5 Main Results | 4-8 pages | Baseline specification results, coefficient of interest interpretation, economic magnitude discussion (back-of-envelope calculations), falsification/placebo tests |
| S6 Robustness | 4-8 pages | Alternative DV, alternative IV, alternative sample, alternative specification, alternative estimator, placebo in time/space, bounding (Oster, Altonji et al.), sensitivity analysis |
| S7 Mechanism Evidence | 2-5 pages | Mediation analysis, moderation analysis, heterogeneity across theoretically motivated dimensions, ruling out alternative mechanisms |
| S8 Discussion & Implications | 2-4 pages | Summary of findings, theoretical contribution, practical/managerial/policy implications, connection to prior literature |
| S9 Conclusion | 0.5-1 page | Summary, limitations, future research directions |
| References | — | 40-80 papers typical for FT50 empirical papers |
| Online Appendix | Unlimited | Additional robustness, alternative specifications, additional mechanism tests, data construction details, replication package |
**Key structural differences from other writing traditions**:
1. **The identification section is the centerpiece.** Reviewers read this before anything else after the introduction. If the identification is not credible, the paper is rejected regardless of the results.
2. **The institutional background section is essential for quasi-experimental designs.** It justifies why the natural experiment or instrumental variable is valid. Without institutional detail, the identification strategy lacks credibility.
3. **Robustness is not an afterthought.** In FT50 papers, the robustness section is often longer than the main results section. It must systematically address every conceivable threat to validity.
4. **Mechanism evidence is increasingly required**, even in papers where the reduced-form relationship is sufficient for policy/practice implications. Knowing *why* an effect exists strengthens causal interpretation.
5. **Economic significance** must be discussed alongside statistical significance. A statistically significant coefficient with a tiny magnitude may not warrant publication.
---
## Identification Strategies
The credibility of the identification strategy is the single most important factor in FT50 empirical papers. Each strategy has specific specification requirements, diagnostic tests, and common pitfalls.
### A. Difference-in-Differences (DID)
#### Specification
The canonical two-way fixed effects (TWFE) specification:
$$Y_{it} = \alpha_i + \lambda_t + \beta \cdot Treat_i \times Post_t + \varepsilon_{it}$$
For staggered treatment timing, the event-study specification:
$$Y_{it} = \alpha_i + \lambda_t + \sum_{k \neq -1} \beta_k \cdot D_{it}^k + \varepsilon_{it}$$
where $D_{it}^k$ are indicators for $k$ periods relative to treatment.
**Modern Staggered DID Estimators** (required for settings with heterogeneous treatment effects and staggered adoption):
- Sun & Abraham (2021): `eventstudyinteract` in Stata, `sunab`
- Callaway & Sant'Anna (2021): `csdid` in Stata, `did` in R
- Borusyak, Jaravel & Spiess (2024): `did_imputation` in Stata, `didimputation` in R
- de Chaisemartin & D'Haultfœuille (2020): `did_multiplegt` in Stata, `DIDmultiplegt` in R
- Gardner (2022): `did2s` in Stata and R
#### Checklist
- [ ] Event-study plot (coefficient-by-period with 95% CI; reference period at $t=-1$) showing pre-trends
- [ ] No evidence of differential pre-trends (joint F-test of pre-treatment coefficients = 0, with p-value reported)
- [ ] For staggered designs: use at least one modern estimator that is robust to heterogeneous treatment effects
- [ ] Treatment and control groups defined before analysis; no ad-hoc group construction
- [ ] Standard errors clustered at the level of treatment assignment (state, firm, industry)
- [ ] Composition of treatment and control groups stable over time (balanced panel or demonstrate no selective attrition)
- [ ] Parallel trends assumption defended with institutional arguments, not just statistical tests
- [ ] Mean of dependent variable in control group reported (for economic magnitude interpretation)
- [ ] Number of treated and control units reported
- [ ] Never-treated, not-yet-treated, and already-treated clearly distinguished in staggered designs
#### Common Pitfalls
| Pitfall | Consequence | Remedy |
|---------|------------|--------|
| TWFE with staggered heterogeneous effects | Negative weights; biased $\hat{\beta}$ | Use modern staggered DID estimators |
| Clustering at wrong level | Severely understated standard errors | Cluster at treatment assignment level; consider wild cluster bootstrap with few clusters |
| Interpreting pre-trend insignificance as proof of parallel trends | Statistical power issue | Provide economic argument for parallel trends; use pre-trend magnitude relative to treatment effect |
| Selective attrition correlated with treatment | Biased estimates | Show attrition by treatment status; bounding approaches |
| Treatment effect heterogeneity hidden by aggregation | Misleading average effect | Report treatment effect heterogeneity by cohort/duration |
| Event-study with binning of endpoints | Arbitrary endpoint choice | Report full dynamic specification; justify endpoint binning |
| Contamination from other concurrent treatments | Confounded estimates | Document all concurrent policy changes; consider synthetic control as robustness |
### B. Regression Discontinuity Design (RDD)
#### Specification
Sharp RDD:
$$Y_i = \alpha + \beta \cdot D_i + f(X_i - c) + \gamma \cdot D_i \cdot f(X_i - c) + \varepsilon_i$$
where $D_i = \mathbf{1}[X_i \geq c]$, $f(\cdot)$ is a flexible polynomial (linear or quadratic) of the running variable centered at the cutoff $c$, and the interaction allows different slopes on each side.
Fuzzy RDD:
$$D_i = \alpha_1 + \pi \cdot Z_i + f_1(X_i - c) + \nu_i \quad \text{(first stage)}$$
$$Y_i = \alpha_2 + \beta_{FRD} \cdot \widehat{D}_i + f_2(X_i - c) + \varepsilon_i \quad \text{(second stage)}$$
where $Z_i = \mathbf{1}[X_i \geq c]$ is the instrument.
**Bandwidth selection**: MSE-optimal bandwidth (Calonico, Cattaneo & Titiunik, CCT), coverage error rate (CER) optimal bandwidth. Use `rdrobust` in Stata/R.
#### Checklist
- [ ] McCrary (2008) density test: no manipulation of the running variable at the cutoff
- [ ] Baseline covariate balance at the cutoff (regress each predetermined covariate on treatment at the cutoff)
- [ ] Robustness to bandwidth choice: plot $\hat{\beta}$ as a function of bandwidth ($h \in [0.5 h^*, 2 h^*]$)
- [ ] Robustness to polynomial order: linear, quadratic, cubic
- [ ] Donut-hole RDD (excluding observations very close to the cutoff)
- [ ] Placebo cutoffs at other values of the running variable
- [ ] Conventional, bias-corrected, and robust confidence intervals reported (CCT)
- [ ] Optimal bandwidth reported with justification
- [ ] Local-linear specification preferred over higher-order polynomials (Gelman & Imbens, 2019)
- [ ] Number of observations above and below the cutoff reported
- [ ] Graphical evidence: binned scatterplot with global polynomial fit (or local means)
#### Common Pitfalls
| Pitfall | Consequence | Remedy |
|---------|------------|--------|
| High-order global polynomials | Noisy estimates, overfitting | Use local-linear within optimal bandwidth |
| Running variable manipulation | Invalid design | McCrary test; institutional argument why manipulation impossible |
| Donut-hole without justification | Excluding informative observations | Transparently report; justify if heaping at cutoff |
| Inference with discrete running variable | Standard errors incorrect | Use randomization inference with discrete running variable (Cattaneo et al.) |
| Extrapolation: LATE interpretation | Not generalizable | Interpret as LATE at cutoff; caution about external validity |
| Ignoring bandwidth uncertainty | Understated standard errors | Use robust confidence intervals (CCT) |
### C. Instrumental Variables (IV)
#### Specification
First stage: $X_i = \pi_0 + \pi_1 Z_i + \Pi_2 W_i + \nu_i$
Second stage (structural): $Y_i = \beta_0 + \beta_1 \widehat{X}_i + \beta_2 W_i + \varepsilon_i$
Reduced form: $Y_i = \gamma_0 + \gamma_1 Z_i + \gamma_2 W_i + \eta_i$
Where $Z_i$ is the instrument, $X_i$ is the endogenous regressor, $W_i$ are exogenous controls.
#### Checklist
- [ ] Economic/institutional argument for instrument relevance (not just statistical)
- [ ] Economic/institutional argument for exclusion restriction (why $Z$ affects $Y$ only through $X$)
- [ ] First-stage F-statistic reported: $F > 10$ (Stock & Yogo critical values); Kleibergen-Paap F-statistic for clustered SE
- [ ] First-stage coefficients reported in a table (not just "results available upon request")
- [ ] Reduced-form results reported alongside IV results (transparent: show both)
- [ ] OLS results reported for comparison (bracket the IV estimate)
- [ ] Limited information maximum likelihood (LIML) if weak instruments or many instruments
- [ ] Overidentification test (Sargan-Hansen J-test) if multiple instruments, but acknowledge limitations
- [ ] "Plausibly exogenous" sensitivity analysis (Conley, Hansen & Rossi, 2012) if exclusion restriction is uncertain
- [ ] Interpretation: LATE vs. ATE discussion (whose treatment effect is identified?)
- [ ] Standard errors: cluster-robust if instrument varies at group level
#### Common Pitfalls
| Pitfall | Consequence | Remedy |
|---------|------------|--------|
| Weak instruments ($F < 10$) | Biased IV toward OLS; large SE | Use LIML; report weak-instrument robust CIs (Anderson-Rubin) |
| Exclusion restriction violation | Inconsistent estimates | Plausibly exogenous bounds (Conley et al.); never-happen tests |
| Bad controls in IV | Inconsistent estimates | Controls must be predetermined or exogenous; justify each control |
| Many instruments relative to sample | Overfitting first stage; biased IV | LIML instead of 2SLS; jackknife IV |
| Interpreting LATE as ATE | Overgeneralization | Discuss complier characteristics; compare to full sample |
| Not reporting reduced form | Hides weak/broken instrument story | Always report reduced form alongside IV |
| First-stage with FE and nonlinearity | Incorrect standard errors in manual 2SLS | Use `ivreghdfe` (Stata) or `feols` with IV (R fixest) |
### D. Randomized Controlled Trial (RCT)
#### Specification
Simple comparison of means (ATE):
$$\hat{\tau} = \bar{Y}_T - \bar{Y}_C$$
With covariate adjustment for precision:
$$Y_i = \alpha + \tau \cdot T_i + \beta X_i + \varepsilon_i$$
where $T_i$ is treatment assignment and $X_i$ are pre-specified baseline covariates.
#### Checklist
- [ ] Pre-analysis plan (PAP) registered (AEA RCT Registry, OSF, AsPredicted); link provided
- [ ] Power analysis reported (ex ante minimum detectable effect)
- [ ] Randomization procedure described in detail (stratification, blocking, clustering)
- [ ] Balance table: pre-treatment covariates compared across treatment and control; joint F-test p-value
- [ ] Attrition analysis: differential attrition by treatment status; bounding (Lee bounds) if asymmetric
- [ ] Intention-to-treat (ITT) as primary specification
- [ ] Treatment-on-the-treated (TOT/LATE) as secondary specification if imperfect compliance
- [ ] Multiple hypothesis testing correction (Romano-Wolf stepdown, sharpened q-values, or Westfall-Young)
- [ ] Heterogeneous treatment effects analysis (pre-specified subgroups)
- [ ] Standard errors: robust (HC2/HC3); cluster-robust if randomization is clustered
- [ ] CONSORT diagram (flow of participants through each stage)
#### Common Pitfalls
| Pitfall | Consequence | Remedy |
|---------|------------|--------|
| No pre-analysis plan | p-hacking concerns | Register PAP before data collection or accessing outcome data |
| Differential attrition | Biased ATE | Report attrition by arm; Lee bounds; inverse probability weighting |
| Multiple outcomes without correction | Inflated Type I error | PAP must specify primary outcome; correct for multiple testing |
| Spillovers / SUTVA violation | Biased ATE | Design to minimize spillovers; measure and model if expected |
| Non-compliance ignored | ATE vs....
© franklee16 · MIT · franklee16/academic-research-skills