Replication Package (imfer-replication-package)

When to trigger

• The analysis is settled and the data/code package must be assembled before submission or at acceptance
• Some inputs are restricted (IMF program data, central-bank micro data, proprietary flows) and cannot be redistributed
• A cross-country dataset stitches many sources (IFS, BOP, WEO, BIS, EPFR, national accounts) and the lineage is undocumented
• A referee or editor asks for a reproducibility check or a data-availability statement
• Code runs only on one machine; seeds, versions, and the build order are not pinned

The IMFER reproducibility reality

IMFER work leans on international-macro data that is often partly restricted — IMF surveillance data, central-bank confidential series, commercial flow data (EPFR), or program-specific files. The package must make everything reproducible in principle even when some inputs cannot be shipped: provide the build scripts, the exact source and vintage of each series, and clear instructions for obtaining the restricted inputs, so a replicator with access can rebuild the analysis dataset and regenerate every exhibit. Confirm the journal's current data-availability and deposit requirements on the official pages (检索于 2026-06；以官网为准).

The restricted-data spectrum (classify each input)

International-macro inputs are rarely all-public or all-secret; classify each on a spectrum and document accordingly. Fully public (IFS, WEO, BIS statistics, World Bank): ship the pull scripts and vintage. Public-but-licensed (Bloomberg, Refinitiv, EPFR, Datastream): ship cleaning code plus the license/access route, not the raw series. Restricted-by-agreement (central-bank confidential micro data, IMF surveillance files): ship the build script plus contact/access instructions and any aggregate that the agreement permits. Author-constructed (a hand-coded narrative classification, an event list): ship it in full with the coding rules. The data-availability statement is just this classification made explicit, input by input.

Packaging craft

1. Map every series to a source and vintage. Cross-country panels silently mix vintages (a WEO release, an IFS pull); record exactly which, because revisions change results.
2. Separate public from restricted up front. Write the data-availability statement first; it dictates what ships and what needs access instructions.
3. Script the build, do not hand-edit. Every transformation from raw to analysis dataset must be in code, so a replicator with the restricted input can reconstruct your sample.
4. Pin the environment. International-macro pipelines often span Stata, R, and Python; lock each so results do not drift with package updates.
5. Regenerate exhibits from scratch in a clean environment before submission — the most common failure is a figure that no longer matches the script.
6. Document country and currency conventions in the codebook; a replicator must know your USD/local, gross/net, deflator choices.

Checklist

• Data-availability statement: each input classified public/restricted with source, vintage, access path
• Restricted inputs never shipped; access instructions + build script provided instead
• Raw-to-analysis lineage fully scripted and ordered
• Numbered code regenerates every table and figure
• Environment pinned (versions + packages) across all languages used
• Codebook covers variables, country codes, units, currency/deflator/gross-net conventions
• Exhibit-to-script mapping table provided (every table/figure traceable to its code)
• Seeds set and reported for simulation/bootstrap/estimation
• Clean-environment rebuild verified; exhibits match the scripts
• Current deposit / data-availability rules confirmed on official pages or marked 待核实

Anti-patterns

• Shipping confidential IMF/central-bank micro data instead of access instructions
• A panel with no record of which data vintage was used (results not reconstructable after revisions)
• Hand-edited intermediate files that no script can reproduce
• Unpinned environment, so a referee's rerun drifts from the paper
• A README that assumes the author's exact machine and paths
• Treating reproducibility as an acceptance-time afterthought rather than building it in
• A data-availability statement that says "available on request" for inputs that have a real public source or access route

Worked vignette (illustrative)

A capital-flows paper merges EPFR fund flows (commercial, licensed), IFS balance-of-payments (public), and a central bank's confidential intervention log (restricted). The package ships the public IFS pulls and all build scripts, but for EPFR and the intervention log it ships only the cleaning code plus access instructions (how to license EPFR, whom to contact at the central bank). The data-availability statement classifies each input, records the IFS vintage (2024 Q1 release) and the EPFR pull date, and the README runs the public-data portions end to end. A replicator with the licenses can rebuild the full analysis dataset and regenerate every exhibit — reproducible in principle without redistributing restricted data.

Referee/editor pushback mapped to the package fix

• "Some inputs are restricted — is this reproducible?" → Provide build scripts plus access instructions for the restricted inputs; never ship them.
• "Which data vintage produced these numbers?" → Record source and vintage for every series in the data-availability statement and codebook.
• "Your rerun gives different numbers." → Pin the environment across Stata/R/Python and verify a clean-environment rebuild before submission.

Output format

【Journal】IMF Economic Review
【Skill】imfer-replication-package
【Data-availability】public vs restricted, with sources + vintages: ___
【Restricted handling】access instructions + build script (not shipped): ___
【Lineage】raw→analysis fully scripted? [Y/N]
【Code】numbered, regenerates all exhibits? [Y/N]
【Environment】versions/packages pinned across languages? [Y/N]
【Codebook】variables, country codes, currency/gross-net conventions? [Y/N]
【Clean rebuild】exhibits match scripts? [Y/N]
【Next skill】imfer-referee-strategy