Skills Data Science Building Reproducible Academic Replication Packages

Building Reproducible Academic Replication Packages

v20260724
jleo-replication-package
This guide details how to assemble a comprehensive, reproducible replication package for academic manuscripts, especially in fields relying on complex institutional data (e.g., court records, contracts). It mandates creating a one-command master script, documenting the entire data pipeline from raw sources to final exhibits, and providing detailed protocols for hand-coded variables and data provenance. The goal is to ensure that any researcher can perfectly reproduce every finding.
Get Skill
134 downloads
Overview

Replication Package (jleo-replication-package)

When to trigger

  • A coauthor or referee needs to reproduce every table and figure from raw institutional data
  • The paper relies on hand-collected institutional data (court records, contracts, legislative votes, agency rulings) whose construction is undocumented
  • Some institutional data are proprietary or confidential (firm contracts, sealed court records) and access terms must be stated
  • The code runs on one machine but the path from raw source to final exhibit is not reproducible
  • You need a Data Availability Statement and a README before submitting or at the revision stage

Reproducibility for institutional data

JLEO is an OUP economics journal; OUP and COPE expect transparency, and the field increasingly expects a working replication package even where OUP does not run a dedicated data-editor check (verify the current JLEO data-and-code policy on the official OUP page — 待核实). The distinctive challenge at JLEO is that the data are often institutional artifacts — court dockets, procurement contracts, committee assignments, constitutional provisions — that require careful, documented construction from primary sources. Build the package so a stranger reproduces every number.

Build the package

  1. One-command build. A master script runs raw → cleaned → analysis → every exhibit in order, with a fixed random seed where any randomness enters.
  2. Document the institutional data construction. For hand-coded institutional variables (an asset-specificity index, a judicial-independence score, a governance-form classification), provide the coding protocol, the source documents, and inter-coder reliability if human coding was involved. This is where JLEO replication most often fails.
  3. State data provenance and access. For each source: origin (court system, regulator, commercial provider), access date, license/terms, and whether others can obtain it. Confidential firm or court data: state the access procedure and provide the code plus a synthetic or restricted-use path.
  4. Map exhibits to scripts. A table lists, for each table/figure in the paper, the script and line that produces it.
  5. Pin the environment. Software versions, packages, and (for proprietary software) the exact commands; note OS if results are sensitive.
  6. Data Availability Statement. A clear DAS naming which data are public, which are restricted, and how to request access, consistent with OUP/COPE expectations.

Checklist

  • A master script reproduces every exhibit end-to-end, raw → final, with seeds fixed
  • Hand-coded institutional variables have a documented coding protocol and source documents
  • Inter-coder reliability reported where institutional variables were human-coded
  • Each data source has provenance, access date, and license/terms stated
  • Confidential institutional data have a stated access procedure and a code-only or synthetic path
  • An exhibit→script map lets a reader find the code behind any table or figure
  • Environment (software/package versions) pinned; a Data Availability Statement drafted

Anti-patterns

  • A zip of scripts with no master file and no order — the reviewer cannot tell what to run
  • Hand-coded institutional indices with no coding protocol, so the key variable cannot be reconstructed
  • "Data available on request" with no procedure, for institutional data others genuinely cannot get
  • Hard-coded absolute paths and unpinned package versions that break on another machine
  • Treating the package as an afterthought for institutional data that took months to build by hand

Worked vignette (illustrative)

A paper builds a governance-form variable by reading 3,000 procurement contracts and classifying each as arm's-length, hybrid, or integrated. The replication risk is the classification, not the regression. The package therefore includes: the coding manual with decision rules, a sample of coded contracts, a second coder's classifications on a 10% subsample with the agreement rate (say κ = 0.84, illustrative), and the script that turns the coded file into the analysis dataset — so the institutional measure, not just the estimation, is reproducible.

Referee / editor concern mapped to the package fix

  • "How did you classify governance form? I cannot reproduce the key variable." → Ship the coding manual, source-document samples, and inter-coder reliability — not just the final dataset.
  • "Your court/contract data are confidential; how can this be replicated?" → State the exact access procedure and provide a code-only path plus a synthetic dataset that runs the full pipeline.
  • "Which script produces Table 4?" → Include the exhibit→script map and a one-command master build with seeds fixed.
  • "What is the data-and-code policy here?" → Confirm the current JLEO/OUP policy live (待核实) and draft a Data Availability Statement consistent with COPE expectations.

Output format

【Master build】one command raw→exhibits, seeds fixed? [Y/N]
【Institutional data construction】coding protocol + sources documented? [Y/N]
【Inter-coder reliability】reported where human-coded? [Y/N/NA]
【Provenance & access】per-source origin/date/license stated? [Y/N]
【Confidential data path】access procedure + code/synthetic path? [Y/N/NA]
【Exhibit→script map】present? [Y/N]
【DAS + environment】drafted and pinned? [Y/N]
【Next skill】jleo-referee-strategy
Info
Category Data Science
Name jleo-replication-package
Version v20260724
Size 5.79KB
Updated At 2026-07-28
Language