Stat Paper Plan

Stat Paper Plan: From Research Results to Paper Outline

Generate a structured outline for a statistics, applied statistics, or ML theory paper from: $ARGUMENTS

Constants

• CLAUDEREVIEWERMODEL = claude-opus-4-6 — Claude model invoked via the Agent tool as a subagent for fast internal review.
• CODEXREVIEWERMODEL = gpt-5.5 — External LLM invoked via Codex MCP for senior-statistician external review. Uses model_reasoning_effort: xhigh.
• REVIEW_MODE = both — Options: claude (internal only, fast), codex (external only, deep), both (claude first, then codex on final draft). Default both for high-stakes submissions; switch to claude for quick iteration.
• TARGET_VENUE = AOS — Default venue. Supported:
• Theory/methodology venues: AOS (Annals of Statistics), JASA (JASA Theory & Methods), JRSSB, BIOMETRIKA, BERNOULLI, EJS, STATSINICA, MSL
• Application venues: AOAS (Annals of Applied Statistics), JASA_ACS (JASA Applications and Case Studies, formerly JASA_APP), BIOSTATISTICS, STATMED (Statistics in Medicine), JCGS, JABES
• ML theory conferences: COLT, ALT
• PAPER_TYPE = auto — Auto-detect from content. Options:
• theory — main contribution is theoretical results (rates, bounds, characterizations)
• methodology — main contribution is a new method with theory + empirics
• application — main contribution is solving a real scientific problem; data and findings drive the paper
• auto — infer from the input narrative and target venue
• MAX_PAGES — Venue-dependent working norm. AOS: about 25 pages in standard journal style before overflow moves to supplement. BERNOULLI / EJS: no hard limit encoded here; size to contribution. AOAS: most published papers will not exceed 20 pages in standard journal style. JASA T&M / JASA ACS: 35 double-spaced pages as the operative limit, roughly 26-27 lines per page; verify counting boundary at submission time. JRSSB: 25-35. BIOMETRIKA: regular and synthesis papers are normally fewer than 20 pages; Miscellanea max 8 pages. BIOSTATISTICS: no fixed page rule encoded here; follow recent issues and use supplement aggressively. STATMED: 15-25. JCGS: 20-30. COLT: 30 main-body pages. ALT: 20-25 main-body pages. COLT/ALT: appendix unlimited.
• Caveat: these are skill defaults reflecting working norms at the dates the venue entries in stat-venue-checklists.md were last checked. Always re-check the current journal author guidelines before submission.

Inputs

The skill expects one or more of these:

1. NARRATIVE_REPORT.md or STORY.md — research narrative with claims, theorems, evidence
2. Theorem statements and proof outlines — formal results with sketches
3. Simulation results — JSON files, tables, figures from experiments
4. IDEA_REPORT.md — from idea-discovery pipeline
5. Existing draft — partial LaTeX or notes to restructure
6. For application papers:

• DATA_DESCRIPTION.md — dataset description, source, variables, exploratory analysis
• APPLICATION_REPORT.md — analysis results, scientific findings, domain interpretation
• EDA outputs — exploratory figures, tables, summary statistics

If none exist, ask the user to describe the paper's main theorem, method, or application in 3-5 sentences.

Detecting Paper Type

If PAPER_TYPE = auto, infer from inputs:

• Application paper signals: a specific dataset is named in the narrative; the scientific question precedes the method; substantive findings are emphasized; target venue is AOAS, JASA_ACS, BIOSTATISTICS, STATMED, JCGS, or JABES
• Theory paper signals: focus is on rates, bounds, or characterizations; data is secondary or absent; target venue is AOS, BERNOULLI, EJS, COLT, ALT, or MSL
• Methodology paper signals: focus is on a new method with both theory and empirics; data analysis is illustrative rather than central; target venue is JASA T&M, JRSSB, BIOMETRIKA, STATSINICA

If signals are mixed, ask the user explicitly. The paper type determines section structure, theory depth, and length allocation.

Workflow

Step 1: Extract Claims, Theorems, Data, and Evidence

Read all available documents. What to extract depends on paper type.

For theory and methodology papers, extract:

1. Main theorems (1-3 primary theoretical results for theory; 2-4 for methodology)
2. Supporting results (lemmas, propositions, corollaries)
3. One-sentence contribution (what the paper proves or proposes)
4. Assumptions needed for each result
5. Evidence: which simulations verify which theorems, which real data demonstrates practical value
6. Known limitations
7. Rate comparisons with existing literature

For application papers, additionally extract (often this is the dominant input):

1. The scientific question (what substantive question is the paper answering)
2. The dataset: source, size, structure, variables, time period, collection method
3. Key data characteristics: features of the data that pose statistical challenges (dependence, missingness, dimension, heterogeneity, censoring, mixed types)
4. Domain context: what the field currently knows or assumes
5. Limitations of current domain practice: what standard methods are used and why they are inadequate
6. Substantive findings: what the analysis revealed about the scientific question
7. Comparison methods: which methods practitioners use that should be benchmarked
8. Validation strategy: holdout, cross-validation, sensitivity analyses
9. Reproducibility: data availability, code, computational requirements
10. Light theoretical results (typically 1-2 main theorems max)

Build a Theorems-Evidence Matrix:

| Result | Statement (informal) | Assumptions | Verified by | Status |
|--------|---------------------|-------------|-------------|--------|
| Thm 1 (Upper bound) | Rate n^{-2s/(2s+d)} for estimator | (A1)-(A3) | Sim 1, Fig 2 | Proved |
| Thm 2 (Lower bound) | Minimax lower bound matches | Model only | N/A (information-theoretic) | Proved |
| Thm 3 (CLT) | Asymptotic normality | (A1)-(A4) | Sim 2 (coverage) | Proved |
| Prop 1 (Computation) | O(n log n) algorithm | — | Sim 3 (timing) | Proved |

Also build a Claims-Evidence Matrix (parallel to the existing paper-plan style):

| Claim | Evidence | Status | Section |
|-------|----------|--------|---------|
| Minimax optimal rate | Thm 1 + Thm 2 | Supported | §3 |
| Weaker assumptions than prior work | Assumption comparison | Supported | §2, Remark 1 |
| Practical computational efficiency | Prop 1 + Sim 3 | Supported | §4, §5 |
| Works on real data | Real data analysis | Partially supported | §6 |

For application papers, additionally build a Findings-Evidence Matrix (this is often the most important planning artifact for application papers):

| Finding | Statistical evidence | Validation | Domain significance | Section |
|---------|---------------------|------------|--------------------|---------|
| [Finding 1: specific scientific claim] | [estimates with CI from §6] | [holdout/CV result] | [why this matters to the field] | §6.2 |
| [Finding 2: contrast with prior view] | [comparison with standard method] | [sensitivity analysis] | [implications for practice] | §6.3 |
| [Finding 3: novel pattern] | [exploratory finding with confirmatory test] | [cross-validation] | [opens new questions] | §6.4 |

And a Data-Challenges Matrix:

| Data feature | Statistical challenge | Methodological response |
|--------------|----------------------|------------------------|
| [e.g., spatial dependence] | Standard iid assumption fails | Hierarchical spatial model in §3.1 |
| [e.g., informative missingness] | Complete-case bias | Joint modeling of missingness in §3.2 |
| [e.g., high dimensionality with sparsity] | Curse of dimensionality | Regularized estimation in §3.3 |

These matrices ensure that the application paper's structure flows from data → challenge → method → finding, rather than from method → demonstration.

Step 2: Determine Paper Type and Structure

Before committing to a structure, read ../stat-shared-references/stat-writing-principles.md for the narrative arc.

IMPORTANT: Section count is flexible (5-9 sections). Choose what fits the content.

Theory paper (main contribution is theorems/bounds):

1. Introduction (2-3 pages)
2. Problem Setup and Assumptions (2-3 pages)
3. Main Results (3-5 pages)
   — theorem statements, rate comparisons, discussion
4. Estimation / Methodology (1-2 pages)
   — the estimator/test/procedure if not already in Setup
5. Simulation Studies (2-3 pages)
6. Discussion (1-2 pages)
— Supplement: full proofs, additional simulations

Theory paper with real data:

1. Introduction (2-3 pages)
2. Problem Setup and Assumptions (2-3 pages)
3. Main Results (3-5 pages)
4. Simulation Studies (2-3 pages)
5. Real Data Analysis (2-3 pages)
6. Discussion (1-2 pages)
— Supplement: full proofs, additional results

Methodology paper:

1. Introduction (2-3 pages)
2. Background and Notation (1-2 pages)
3. Proposed Method (3-4 pages)
   — algorithm, implementation, tuning
4. Theoretical Properties (2-3 pages)
   — consistency, rates, efficiency
5. Computation (1-2 pages)
   — algorithm complexity, practical considerations
6. Simulation Studies (3-4 pages)
7. Application / Real Data Analysis (2-3 pages)
8. Discussion (1-2 pages)
— Supplement: proofs, additional simulations and data analysis

Application paper (AOAS, JASA ACS, Biostatistics, JCGS):

1. Introduction (2-3 pages)
   — scientific question, data, statistical challenges,
     findings preview, contributions
2. Data and Scientific Background (2-4 pages)
   — domain context, dataset description, EDA figures,
     statistical challenges revealed by the data
3. Methodology (3-4 pages)
   — model formulation, estimation, inference (scoped to problem)
4. Theoretical Properties (1-2 pages, LIGHT)
   — 1-2 main theorems only (consistency, asymptotic distribution)
   — heavy theory in supplement
5. Simulation Studies (2-3 pages)
   — DGPs informed by the real data
   — comparison with domain-standard methods
6. Application / Real Data Analysis (4-6 pages) ← LARGEST SECTION
   — analysis setup
   — main analysis with substantive findings
   — comparison with existing approaches in the field
   — validation (holdout, CV, sensitivity)
   — substantive interpretation
7. Discussion (1-2 pages)
   — substantive findings summary
   — methodological summary
   — practical recommendations for practitioners
   — limitations
   — extensions and open questions
— Supplement: detailed data description, full proofs,
  additional simulations, additional analyses, software docs

Application paper alternative (heavily validated):

1. Introduction
2. Data and Background
3. Methodology
4. Theoretical Analysis (light)
5. Computation
6. Simulation Studies
7. Main Application
8. Sensitivity and Validation
9. Discussion

Use the alternative when validation is itself a major contribution (e.g., demonstrating robustness across multiple datasets or settings).

Read ../stat-shared-references/stat-application-writing.md before locking in the application paper structure.

COLT/ALT theory paper:

1. Introduction (2-3 pages)
2. Problem Setup and Preliminaries (2-3 pages)
3. Main Results (4-6 pages)
4. Proof Techniques / Analysis (3-5 pages)
   — more proof detail in main body since page limit is generous
5. Experiments (2-3 pages, optional for COLT)
6. Discussion and Open Problems (1-2 pages)
— Appendix: complete proofs

Step 3: Section-by-Section Planning

For each section, specify:

### §0 Abstract
- **Problem**: [the specific statistical problem addressed]
- **Gap**: [what is unknown or suboptimal]
- **Main result**: [theorem in plain language, with rate]
- **Approach**: [proof technique or method, in one sentence]
- **Verification**: [simulation or data confirmation]
- **Implication**: [what this enables]
- **Estimated length**: 150-250 words
- **Self-contained check**: understandable without the paper?

### §1 Introduction
- **Problem motivation**: [why this problem matters, specific not generic]
- **Prior work and gap**: [what is known, what rates exist, what's open]
- **Contribution**: [main result(s) in plain language]
- **Proof technique overview**: [key insight, novel tools]
- **Contribution bullets**: [2-4 specific, verifiable claims]
- **Results preview**: [strongest result surfaced early]
- **Paper organization**: [roadmap of remaining sections]
- **Key citations**: [5-10 papers to position against]
- **Estimated length**: 2-3 pages

### §2 Problem Setup / Model / Assumptions
- **Model**: [formal model specification]
- **Notation**: [key symbols defined here]
- **Assumptions**: [list with labels (A1), (A2), ...]
- **Assumption discussion**: [which are standard, which are novel, comparison with prior work]
- **Estimated length**: 2-3 pages

### §3 Main Results
- **Theorem statements**: [list each theorem/proposition]
- **Rate comparison table**: [this paper vs prior work]
- **Proof sketches**: [1-2 paragraphs per main theorem]
- **Remarks and corollaries**: [optimality, extensions, special cases]
- **Estimated length**: 3-5 pages

### §4 Estimation / Method (if separate from §2-3)
- **Estimator/procedure definition**: [formal specification]
- **Algorithm**: [pseudocode if applicable]
- **Tuning parameters**: [how chosen, adaptive or oracle]
- **Computational complexity**: [time and space]
- **Estimated length**: 1-3 pages

### §5 Simulation Studies
- **DGP designs**: [list each data generating process]
  - DGP 1: [model, parameter values, purpose]
  - DGP 2: [model, parameter values, purpose]
  - DGP 3: [adversarial/challenging case]
- **Sample sizes**: [e.g., n = 100, 500, 2000, 10000]
- **Number of replications**: [500-1000]
- **Comparison methods**: [3-5 existing methods]
- **Metrics**: [MSE, coverage, power, computation time, etc.]
- **Rate verification**: [log-log plots planned?]
- **Figures planned**:
  - Fig X: [type, what it shows]
  - Table X: [what it shows]
- **Estimated length**: 2-4 pages

### §6 Real Data Analysis (if applicable)
- **Dataset**: [source, size, why appropriate]
- **Preprocessing**: [steps, variables used]
- **Comparison methods**: [same as simulations if possible]
- **Domain interpretation**: [what the results mean substantively]
- **Estimated length**: 2-3 pages

### §7 Discussion
- **Contribution summary**: [rephrased, not copied from intro]
- **Connections to prior work**: [how results relate or improve]
- **Limitations**: [assumptions that may be restrictive]
- **Open problems**: [specific technical questions remaining]
- **Extensions**: [natural generalizations]
- **Estimated length**: 1-2 pages

### Application Paper Section-by-Section Plan

For application papers (PAPER_TYPE = `application`), use the following section plan instead of (or in addition to) the templates above.

```markdown
### §0 Abstract (application paper)
- **Scientific question**: [what substantive question is being asked]
- **Dataset**: [which dataset enables it]
- **Statistical challenge**: [what data feature makes standard analysis inadequate]
- **Proposed approach**: [method developed or adapted]
- **Application finding**: [the most important discovery — name it]
- **Validation**: [how the finding was checked]
- **Implication**: [what this means for the domain]
- **Estimated length**: 200-280 words

### §1 Introduction (application paper)
- **Scientific question and stakes**: [first paragraph hook]
- **Dataset introduction**: [brief, with relevance to the question]
- **Statistical challenges**: [features that demand new methodology]
- **Limitations of current approaches**: [what the domain does today, why it falls short]
- **Proposed approach**: [high-level method d

…

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [gyf9712](https://github.com/gyf9712)
- **Source:** [gyf9712/stat-writing-skills](https://github.com/gyf9712/stat-writing-skills)
- **License:** MIT

Install and usage instructions live in the source repository linked above.