ft50-empirical-writing

Imported from liyuanbo1024/ft50-empirical-writing (SKILL.md). Install upstream with npx skills add liyuanbo1024/ft50-empirical-writing. Copyright stays with the author.

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

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

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

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

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

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)

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:

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

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

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

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

E. Synthetic Control Method (SCM)

Specification

The synthetic control is a weighted average of control units:

$$\hat{\tau}{1t} = Y{1t} - \sum_{j=2}^{J+1} w_j^* Y_{jt}$$

where $w_j^*$ are weights chosen to minimize the pre-treatment gap in outcomes and predictors between the treated unit and the synthetic control.

Checklist

• Single (or few) treated unit(s) clearly identified (SCM designed for small number of treated units)
• Donor pool of control units described with inclusion/exclusion rationale
• Predictor variables listed (pre-treatment outcomes + covariates); their weights reported
• Unit weights reported in a table (which control units matter and with what weight)
• Pre-treatment fit visually demonstrated (gap plot); RMSPE reported
• In-space placebo tests: run SCM for each control unit as if treated; compare treatment effect to placebo distribution
• Ratio of post/pre RMSPE reported (Abadie, Diamond & Hainmueller, 2010)
• Leave-one-out robustness (remove each donor unit with positive weight one at a time)
• Robustness to predictor set (alternative weighting schemes)
• Inference: p-value from rank of post/pre RMSPE ratio among all units
• Generalized SCM (Xu, 2017) or matrix completion (Athey et al., 2021) for multiple treated units with staggered adoption

Common Pitfalls

F. Matching Methods (PSM, CEM, Entropy Balancing)

Specification

Propensity Score Matching (PSM):

1. Estimate propensity score: $p(X_i) = \Pr(T_i = 1 | X_i)$ via logit/probit
2. Match treated to control units on $\hat{p}(X_i)$ (1:1 nearest neighbor, with or without caliper)
3. Estimate ATT: $\hat{\tau}{ATT} = \frac{1}{N_T} \sum{i: T_i=1} (Y_i - \sum_{j: T_j=0} w_{ij} Y_j)$

Coarsened Exact Matching (CEM): Temporarily coarsen each variable into bins, exact match on bins, then estimate treatment effect on matched data.

Entropy Balancing: Re-weight control observations to achieve exact balance on specified moments of covariates.

Checklist

• Matching is used as a supplementary method to regression, not as a standalone solution to endogeneity
• Selection on observables assumption (CIA/unconfoundedness) explicitly stated and defended
• Variables in the propensity score: only pre-treatment covariates unaffected by treatment
• Common support demonstrated (overlapping histograms of propensity scores)
• Covariate balance after matching reported (standardized differences < 0.1 or 0.25; variance ratios near 1)
• Matching procedure transparently described (algorithm, caliper, replacement, ratio)
• Outcome analysis on matched sample: standard errors should account for matching uncertainty or use Abadie-Imbens SE
• Sensitivity to unobserved confounding: Rosenbaum bounds (for continuous outcomes) or Oster/Imbens bounds
• CEM or entropy balancing preferred over PSM due to reduced researcher degrees of freedom

Common Pitfalls

Robustness Standards for FT50

FT50 reviewers expect a comprehensive robustness framework that systematically addresses threats to internal validity. Robustness is not a collection of ad-hoc checks — it is an organized, theoretically motivated battery of tests.

The Robustness Pyramid (Ordered by Priority)

Level 5: Sensitivity to Unobservables (Oster Bounds, Rosenbaum Bounds, Imbens Bounds)
    ↑
Level 4: Placebo / Falsification Tests (In-Time, In-Space, In-Group)
    ↑
Level 3: Alternative Specification / Estimator (Poisson, Logit, Quantile, LASSO controls)
    ↑
Level 2: Alternative Sample (Restricted/Balanced Panel, Excluding Outliers, Trimming)
    ↑
Level 1: Alternative DV and IV (Re-measurement, Alternative Operationalizations)
    ↑
Baseline: Main Specification

Robustness Checklist by Level

Level 1: Alternative Dependent Variables and Key Independent Variables

• Alternative measurement of the dependent variable (log vs. level, per-capita vs. total, alternative data source)
• Alternative measurement of the key independent variable (continuous vs. binary, alternative threshold, alternative source)
• If using a constructed index: show components separately; demonstrate robustness to index construction method
• If using a ratio: show numerator and denominator separately; demonstrate relationship is not driven by denominator
• If using textual/sentiment-based measure: robustness to alternative dictionaries, algorithms, or human-coding validation

Level 2: Alternative Sample

• Restricted sample (e.g., exclude financial crisis years, exclude largest 1% of firms, exclude early/late adopters)
• Balanced panel (if main sample is unbalanced)
• Excluding influential observations (DFBETA, Cook's distance; report with and without)
• Jackknife: exclude one unit/country/industry at a time
• Sample split by expected effect magnitude (e.g., high vs. low predicted compliance)
• Winsorization robustness: alternative thresholds (1%/99%, 5%/95%, no winsorization)

Level 3: Alternative Specification and Estimator

• Alternative fixed effects structure (e.g., adding industry×year FE, region×year FE)
• Alternative functional form (log-linear, linear-log, Poisson pseudo-maximum likelihood for non-negative outcomes)
• Alternative estimator: OLS vs. Poisson vs. Negative Binomial for count outcomes; linear probability vs. logit/probit for binary outcomes
• Quantile regression (effect at median, 25th, 75th percentiles)
• LASSO/post-double-selection for high-dimensional controls (Belloni, Chernozhukov & Hansen)
• Standard error robustness: cluster at alternative level, Conley spatial SE, wild cluster bootstrap (when clusters < 50)
• Methods robust to heteroskedasticity: HC2, HC3 standard errors
• If panel: robustness to including/excluding unit-specific linear time trends

Level 4: Placebo and Falsification Tests

• In-time placebo: Move treatment date earlier (e.g., 2, 3, 4 periods before actual treatment); no significant "effect"
• In-space placebo: Assign treatment to units that were never treated; no significant "effect"
• In-group placebo: Test treatment on an outcome that should not be affected (theoretical justification required)
• Alternative dependent variable that should not be affected: Demonstrate discriminant validity
• Randomization inference: Shuffle treatment assignment; compare actual estimate to distribution of placebo estimates

Level 5: Sensitivity to Unobservables

• Oster (2019) bounds: $\delta$ for $\beta = 0$ — how large must selection on unobservables be relative to selection on observables to nullify the result? Rule of thumb: $\delta > 1$ is reassuring.
• Altonji, Elder & Taber (2005) ratio: Ratio of coefficient movement with and without controls; larger ratio = less sensitivity.
• Rosenbaum bounds (for matching designs): $\Gamma$ — by what factor must the odds of treatment differ between treated and control (due to an unobservable) to change inference? Report $\Gamma$ at which p-value crosses 0.05.
• Imbens (2003) / Conley, Hansen & Rossi (2012): "Plausibly exogenous" IV bounds; report treatment effect CI under various degrees of exclusion restriction violation.

Robustness Reporting Standards

Mechanism Evidence

FT50 papers increasingly require not just demonstrating that X causes Y, but why X causes Y. Mechanism evidence strengthens causal interpretation and theoretical contribution.

Three Types of Mechanism Evidence

1. Mediation Analysis (Causal Pathway)

The classical Baron & Kenny (1986) approach:

1. $Y = \alpha_1 + c X + \varepsilon_1$ — Total effect ($c$ significant)
2. $M = \alpha_2 + a X + \varepsilon_2$ — X affects mediator ($a$ significant)
3. $Y = \alpha_3 + c' X + b M + \varepsilon_3$ — Direct ($c'$) and indirect ($a \times b$) effects

Modern standards for FT50:

• Bootstrapped confidence intervals for indirect effect ($a \times b$) with 5,000-10,000 resamples (Preacher & Hayes, 2004, 2008)
• Report bias-corrected and percentile bootstrap CIs; if CI excludes zero, mediation is supported
• For experimental designs: use parallel mediation models (PROCESS macro) or structural equation modeling (SEM)
• For observational designs: Acknowledge that mediators are endogenous; use causal mediation analysis (Imai, Keele & Tingley, 2010) or IV-based mediation (Dippel et al., 2020)
• Key limitation: Mediation analysis in observational data requires the same identification assumptions for the mediator as for the treatment

2. Moderation Analysis (Boundary Conditions)

Moderation demonstrates when or for whom the treatment effect varies:

$$Y = \alpha + \beta_1 X + \beta_2 Z + \beta_3 X \times Z + \varepsilon$$

• Continuous moderator: Simple slopes analysis at $\pm 1$ SD of $Z$; Johnson-Neyman regions of significance
• Categorical moderator: Subgroup analysis with Chow test for difference in coefficients
• Reporting: Interaction plot with confidence bands; marginal effects plot
• Caution: Subgroup analyses must be pre-specified in PAPs for experimental designs; ex-post subgroup fishing inflates Type I error
• FT50 expectations: Moderation results should be theoretically motivated, not data-mined; discuss why the moderator should matter based on theory or institutional context

3. Heterogeneity Analysis (Treatment Effect Heterogeneity)

Heterogeneity analysis examines how treatment effects vary across theoretically relevant dimensions:

$$Y = \alpha + \beta_1 D + \beta_2 D \times Z + \gamma Z + \varepsilon$$

• Approaches:
  • Split-sample estimation (by firm size, industry, region, time period)
  • Interaction terms (continuous treatment heterogeneity)
  • Causal forest / generalized random forest (Athey, Tibshirani & Wager, 2019)
  • Quantile treatment effects (effect at different points of the outcome distribution)
• Levels of heterogeneity relevant by discipline:
  • Economics: regional, demographic, time period heterogeneity
  • Finance: firm size, leverage, governance quality, institutional ownership
  • Accounting: Big N vs. non-Big N auditors, high vs. low analyst following
  • Management: industry dynamism, firm capabilities, leadership characteristics
  • Marketing: product category, consumer segment, channel type
  • Operations: process type, supply chain complexity, firm size
• Reporting: Coefficient plots by subgroup; forest plots; interaction coefficients with confidence intervals

Mechanism Evidence Best Practices

Common FT50 Rejection Reasons

Understanding why FT50 papers are rejected is essential for prevention. The following are based on editorial reports, reviewer comments, and published editorials across FT50 journals.

0-to-Draft Pipeline: Five-Stage Workflow

This is the core workflow of this skill. Each stage builds on the previous. Stages 1-3 are sequential; Stages 4-5 can partially overlap. At each stage, ask the user for their current materials before proceeding.

Stage 1: Question Formulation & Positioning

Goal: Define a compelling research question, identify the identification challenge, establish the theoretical/conceptual framework, select a target FT50 journal, and position the paper's contribution.

Input needed from user: Research area, preliminary phenomenon of interest, any existing data access or institutional knowledge, any preliminary literature review.

Agent actions:

1.1 Understand the phenomenon and research question
    - Ask: What is the economic/managerial phenomenon? What is the outcome of interest?
    - What treatment/policy/shock/variation does the paper exploit?
    - Is the research question causal ("does X cause Y?"), descriptive, or predictive?
    - Clarify: What would ideal identification look like? What is feasible?
    - Output: One-paragraph research question with clear outcome and treatment/variation

1.2 Identify the core endogeneity challenge
    - Omitted variable bias? Simultaneity/reverse causality? Measurement error? Selection?
    - Self-selection into treatment? Policy endogeneity?
    - What confounds the relationship between X and Y?
    - Output: explicit endogeneity diagnosis

1.3 Theoretical / conceptual grounding
    - Does the paper build on an existing theory (agency theory, transaction cost economics, institutional theory, etc.)?
    - Or does it develop a conceptual framework from institutional knowledge?
    - What are the testable predictions?
    - Management field: identify the theoretical conversation the paper joins
    - Economics/Finance: identify the behavioral/institutional mechanism being tested
    - Output: theoretical framework summary with predictions

1.4 Literature gap analysis
    - Search for 15-20 closest papers using paper-lookup or research-lookup
    - Map literature along dimensions: research question, method/identification, data, key finding, theoretical lens
    - Build a structured Gap Table
    - Identify: what does this paper do that no prior paper has done?
    - Output: draft gap table + explicit differentiation statement

1.5 FT50 journal selection
    - Use the Journal Selection guide and Decision Flowchart above
    - Match: disciplinary home + identification strategy + contribution type → journal(s)
    - Consider: Which FT50 journals have recently published on similar topics? What is the "conversation" your paper joins?
    - Output: recommended primary journal + 1-2 fallback options with rationale

1.6 Contribution statement
    - Formulate the one-sentence contribution: "Using [identification strategy] in [empirical context], we find that [treatment] causes [outcome] through [mechanism], contributing to [literature stream(s)]."
    - Draft 3-5 numbered contributions: conceptual/theoretical (1-2), empirical (1-2), practical/managerial (1)
    - Output: contribution list

1.7 Writing plan
    - Outline the paper structure based on the FT50 Empirical Paper Structure (above)
    - Estimate page allocation per section for target journal
    - Identify: what data is needed? what identification strategy? what robustness tests?
    - Output: section-by-section writing plan

Stage 1 exit criteria: Research question precisely stated, endogeneity challenge identified, theoretical grounding established, gap table completed, target journal chosen, contribution list written, writing plan ready.

Stage 2: Identification Design

Goal: Design the identification strategy, select data sources, and plan the empirical framework — the centerpiece of an FT50 empirical paper.

Input needed from user: Research question and endogeneity challenge (from Stage 1), data sources (existing or planned), institutional context knowledge.

Agent actions:

2.1 Identification strategy selection
    - Given the empirical context and data structure, which design is feasible?
    - Match variation type → identification strategy:
        - Policy change over time across units → DID (staggered or classic)
        - Sharp cutoff in assignment rule → RDD
        - Plausibly exogenous variation source → IV
        - Randomized intervention → RCT
        - Single treated unit with many controls → SCM
        - Selection on observables (last resort) → matching + Oster bounds
    - Output: recommended identification strategy with feasibility assessment

2.2 Estimation equation specification
    - Write the formal estimating equation(s)
    - Define: dependent variable, key independent variable, control variables, fixed effects, standard error structure
    - Justify each control variable (predetermined; not "bad controls")
    - Specify standard error clustering level
    - Output: formal empirical specification

2.3 Identification assumptions documentation
    - For each assumption of the chosen strategy, state:
        - What the assumption is (in plain language)
        - Why it is plausible in this setting (institutional argument)
        - How it will be tested/defended (diagnostic test)
        - What happens if it fails (robustness or caveat)
    - Output: identification assumptions table

2.4 Data source assessment
    - Identify data sources: administrative, survey, commercial, scraped, hand-collected
    - Assess: sample size adequacy (power analysis), temporal coverage, unit coverage, variable availability
    - Document: data access, confidentiality, IRB requirements
    - Output: data plan with sources, variables, and coverage

2.5 Sample construction plan
    - Define inclusion/exclusion criteria
    - Plan sample construction flowchart / CONSORT diagram
    - Plan attrition analysis
    - Output: sample construction protocol

2.6 Robustness framework planning
    - Using the Robustness Pyramid, plan robustness tests at each level
    - Identify: which tests are diagnostic for the specific identification strategy?
    - Plan mechanism tests
    - Output: robustness plan document

Stage 2 exit criteria: Identification strategy selected and justified, estimation equations specified, identification assumptions documented and defended, data sources identified, sample construction planned, robustness framework planned.

Stage 3: Analysis & Results

Goal: Execute the analysis, generate results tables and figures, and verify the statistical integrity of all findings.

Input needed from user: Data (or data access), identification design (from Stage 2), statistical software.

Agent actions:

3.1 Data preparation
    - Execute sample construction protocol (from Stage 2.5)
    - Produce attrition table
    - Handle: missing values (document pattern; justify method), outliers (winsorization at defensible thresholds), measurement error
    - Output: final analysis sample + sample construction documentation

3.2 Descriptive evidence
    - Summary statistics table (N, mean, SD, min, max, percentiles)
    - Balance table / correlation table
    - Time series plot of key variables (if time series)
    - Graphical evidence: maps (if geographic), trends, distributions
    - Output: Tables 1-2 + descriptive figures

3.3 Identification diagnostics
    - Execute all diagnostic tests specified in Stage 2.3
    - DID: event-study plot with pre-trends
    - RDD: McCrary density test, covariate balance at cutoff, RD plot
    - IV: first-stage table with F-statistic, reduced form
    - RCT: balance table, attrition by arm
    - SCM: pre-treatment fit plot, unit weight table
    - Matching: standardized differences before/after matching (love plot)
    - Output: identification diagnostics section + diagnostic figures

3.4 Main results estimation
    - Estimate baseline specification
    - For DID: event-study specification (preferred over single Post coefficient)
    - For RDD: local-linear with optimal bandwidth
    - For IV: 2SLS (first stage, reduced form, and structural equation all reported)
    - Report: coefficients, standard errors, p-values, confidence intervals
    - Include control mean of DV (for economic magnitude context)
    - Output: Main results tables + key coefficient plot

3.5 Economic significance interpretation
    - Compute: standardized effect size (Cohen's d or equivalent)
    - Back-of-envelope calculation: what does the coefficient mean in practical terms?
    - Benchmark against related estimates from published literature
    - Output: economic significance discussion

3.6 Full robustness execution
    - Execute all planned robustness tests (Levels 1-5 of Robustness Pyramid)
    - Report robustness results in structured tables or coefficient plots
    - Identify any robustness check where results are fragile; discuss transparently
    - Output: robustness section with tables and figures

3.7 Mechanism analysis
    - Execute mechanism tests: mediation, moderation, heterogeneity
    - For causal mediation: implement Imai et al. (2010) or Dippel et al. (2020) approach
    - Report null results on alternative mechanisms
    - Output: mechanism evidence section

3.8 Results section drafting
    - Write results section following journal conventions
    - Tables: self-contained with clear titles, notes, variable definitions
    - Figures: publication-quality; colorblind-friendly palette; vector format
    - NO interpretation beyond stated results — save for Discussion
    - Output: §4 Results, §5 Robustness, §6 Mechanism sections

Stage 3 exit criteria: All analyses executed and verified, results tables and figures generated, all diagnostic tests passed, robustness framework complete, mechanism evidence assembled, results sections drafted.

Stage 4: Writing & Assembly

Goal: Write the full paper from Introduction through Conclusion; assemble all sections; check cross-section consistency; format for target journal.

Input needed from user: All prior stage outputs, target journal selection, any institutional knowledge.

Agent actions:

4.1 Introduction drafting
    - P1: Motivating puzzle or striking fact (see Writing Pattern 1)
    - P2: Generalize to broader economic/managerial significance; set stakes
    - P3: Institutional context preview — what variation does this paper exploit? (high-level, no detailed specification)
    - P4: Identification challenge — "Identifying the causal effect of X on Y is challenging because..."
    - P5: Our approach — "We overcome this challenge by exploiting [variation] in a [DID/RDD/IV/RCT] design..."
    - P6: Preview of findings — "We find that..." (key coefficient + economic magnitude)
    - P7: Contributions (numbered, 3-5 items)
    - P8: Paper organization
    - Output: §1 Introduction draft

4.2 Institutional background / theory section
    - If quasi-experimental: detailed institutional context that motivates and validates the design
    - Why was the policy/practice adopted? Who was affected? Why might effects be heterogeneous?
    - If theory-driven: develop conceptual framework generating testable predictions
    - Use numbered predictions or hypotheses (H1, H2, H3...)
    - Diagrams: institutional timeline, conceptual framework figure
    - Output: §2 Institutional Background / Theory draft

4.3 Data & measurement section
    - Data sources described with provenance
    - Sample construction: step-by-step with numbers (attrition table)
    - Variable definitions: precise, reproducible
    - Descriptive statistics table
    - Balance / correlation analysis
    - Output: §3 Data & Measurement draft

4.4 Identification / research design section
    - Statement of the identification problem
    - Estimation equation formally presented
    - Identification assumptions: stated, defended, tested
    - Diagnostic evidence (pre-trend plots, balance at cutoff, first-stage table)
    - Why this design is credible in this setting
    - Output: §4 Identification Strategy draft

4.5 Results, robustness, mechanism sections
    - Integrate drafts from Stage 3 into polished sections
    - Ensure flow: main results → robustness confirmation → mechanism revealed
    - Output: §5-7 polished drafts

4.6 Discussion & implications
    - Connect findings back to research question and theoretic

*Truncated - read the full file at https://github.com/liyuanbo1024/ft50-empirical-writing/blob/343d78d416b1121ea6d8da3b6e6fa93339906f61/SKILL.md.*