SynthPopCan

SynthPopCan helps researchers create, inspect, and validate modelled Canadian households and people from Census-derived data. It supports reproducible work without requiring researchers to expose private source data or become software developers first.

SynthPopCan is an independent research-software project. It is not affiliated with, endorsed by, or sponsored by Statistics Canada or the Government of Canada.

Near-term goals:

1. Provide a Python library and CLI that can create synthetic populations through IPF from Statistics Canada margin/control tables.
2. Maintain explicit 2016 and 2021 Census microdata workflows for household- and person-level synthetic populations using tree-based generation plus calibration, while keeping source adapters and metadata appropriately census-vintage-specific.
3. Maintain a local web app for configuring runs, inspecting controls, generating from prepared models, validating outputs, and downloading results.

Version 1.0.0 establishes a stable, explicitly bounded interface for research scripts and artifacts. Its packaged 1.x compatibility contract freezes the documented CLI, curated Python API, callable signatures, and supported persisted schemas while keeping later additive extension possible. It carries forward the bounded calibration and small-area evidence from 0.9.0 without turning that evidence into a universal representativeness, privacy, or fitness claim.

It also includes a complete 2016/2021 field-eligibility inventory, a bilingual Quebec 2021 case study, dated preservation and software-management records, and an accepted open-by-default prepared-model policy. All 32 public Census-derived model concepts now point to verified, non-overwriting versions with embedded scoped licensing and preserved source provenance. The release boundary and post-1.0 research tracks are recorded in PLANS.md.

Detailed documentation is published at https://synthpopcan.readthedocs.io/. The source files live under docs/; start with docs/index.rst for task-based navigation to the web app, IPF from StatCan margin tables, generated-from-model workflows, the beginner Python API, and advanced microdata/model-training material.

The development branch and the latest documentation may describe the next release before it reaches PyPI. The PyPI badge above identifies what pip install synthpopcan currently installs; use the stable documentation for that published version. Maintained Can-FED and ODEF adapters are included in 0.7.2 and later.

Project planning and research notes are tracked separately:

• PLANS.md: current roadmap, open work, and release sequencing.
• adr/: accepted architecture decisions, alternatives, and consequences.
• NOTES.md: research synthesis from local materials and external literature.
• CHANGELOG.md: public release notes and completed release highlights.
• CORRECTNESS.md: tested correctness claims, evidence, limitations, and reproducibility commands.
• docs/stewardship.md: supported environments, maintenance expectations, preservation identifiers, and dated FAIR4RS and software-management records.

Quick Start

These commands use a Unix-style shell. Windows readers should first complete the WSL setup in the Installation guide, then run the same commands inside the Ubuntu terminal.

Install the published package from PyPI:

Network required for installation.

python3 -m pip install synthpopcan

The base installation is enough to generate from portable frequency or CART models. Starting with 0.7.0, only researchers training new CART models need python3 -m pip install "synthpopcan[model-build]".

For a guided first look, start the local browser workbench:

synthpopcan serve

It opens forms, previews, run history, and downloads on this computer only. The local web app guide includes fictional teaching data and explains what stays on disk. Use the command line below when the work needs to be scripted or repeated exactly.

Then inspect the command line:

synthpopcan --help

Run a small offline smoke test with the bundled fictional model:

Runnable teaching example. Enter both commands in order.

synthpopcan models generate demo-linked-household-person \
  --households 10 \
  --condition "geo=Demo North" \
  --out synthpopcan-quickstart \
  --random-seed 42
synthpopcan validate linked synthpopcan-quickstart

This verifies linked household/person generation, but it does not create a representative Canadian population. The documentation explains how to choose research sources, controls, and model packages.

For a one-off command without installing the tool into the current environment, use uvx:

Network and uv required. See the Installation guide before using this route if uvx --version is not already available.

uvx synthpopcan --help
uvx synthpopcan guide ipf
uvx synthpopcan guide model
uvx synthpopcan guide small-area

From a source checkout for development:

Source checkout, network, Git, and uv required.

git clone https://github.com/dlq/synthpopcan.git
cd synthpopcan
uv sync
uv run synthpopcan --help

For installation details, see docs/installation.md.

Where To Start

Most readers should start in the Sphinx documentation rather than in this README:

Build the documentation locally with:

Source checkout required. Contributor setup is documented in CONTRIBUTING.md.

uv run sphinx-build -W -b html docs docs/_build/html

Developer Benchmarks

Benchmark fixtures are developer tooling, not normal user workflows:

Source checkout required. The province-scale timing run can also require substantial time and memory.

uv run python scripts/benchmarks.py ipf
uv run python scripts/benchmarks.py small-area
uv run python scripts/benchmarks.py small-area --province-scale
uv run python scripts/benchmarks.py ipf-backends

Use --seed-records for smaller or larger IPF runs. The province-scale small-area timing check is opt-in because it depends on the machine. Optional SciPy CSR and Polars comparisons remain benchmark probes rather than runtime backends, and full-data tree-model smoke tests remain outside the default suite. The CLI, beginner Python API, and local web app call the same Python domain implementations; IPF also shares its file-backed workflow orchestration. Browser guidance and durable artifacts do not define a separate computational tier. Scale limits come from the selected workflow, model, output size, and available machine resources.

Data Policy

Large, raw, private, or access-controlled data are not tracked in git.

• data/raw/ is a local ignored cache for authoritative public inputs, organized by provider, product family, and vintage.
• data/derived/ is a local ignored cache for reproducible conversions, subsets, and durable model artifacts; it must not be mistaken for an authoritative source.
• data/work/ is disposable project-local scratch space for model builds, experiments, and other restartable intermediate files.
• data/private/sources/ is a local ignored cache for access-controlled or sensitive source datasets; generated artifacts do not belong there.
• references/ is a local ignored cache for copied papers, proposals, and legacy code references.

Public geography, school, healthcare, road, and environmental layers should generally be fetched from authoritative public sources such as Statistics Canada, open.canada.ca, donneesquebec.ca, and municipal/provincial open-data portals rather than stored in this repository.

Local-only manifests may exist inside ignored data directories to document what is present on a development machine.

Model Packages

Reviewed model packages may be distributed with the project when they are explicitly intended as public research artifacts. The installed package should stay small: only the tiny demo model is bundled. Larger published models are downloaded on demand with synthpopcan models fetch MODEL_ID.

The public catalogue includes parallel 2016 and 2021 linked packages for Canada, supported provinces, and the five CMAs identified in the hierarchical PUMF. List the exact IDs with synthpopcan models list; for example:

synthpopcan models fetch quebec-2021-all-fields
synthpopcan models generate quebec-2021-all-fields \
  --households 1000 --out quebec-2021-population/

Release assets are gzip-compressed to keep downloads small. The CLI handles decompression and stores a normal JSON model package in the local cache.

Bundled model packages are not raw Census microdata. They should still be treated as derived research artifacts with provenance, disclosure-risk checks, and limitations. A model package being marked as a publishable candidate means it passed the project's current checks; it is not a claim of official approval, legal privacy certification, or fitness for every research use.

Before publishing a new model package, review docs/data.md, docs/tree.md, PLANS.md, and CONTRIBUTING.md.

How To Cite

Releases and prepared model packages are archived on Zenodo, so cite whichever matches what your work actually depended on.

For reproducibility, prefer the version DOI for the release and the model package DOI for each package you generated from: together they pin the exact code and the exact artifact, and every model record publishes the checksums needed to verify the file you downloaded.

Citation metadata lives in CITATION.cff, which GitHub renders as a ready-made citation from the sidebar.

The source history through v0.9.0 is also independently preserved in Software Heritage. Use its SWHIDs to identify source objects; continue to use the release DOI and artifact checksums to identify the software and model bytes used in research.

Prepared model packages are derived from Statistics Canada public use microdata files under the Statistics Canada Open Licence. The catalogue, archive records, and current registered package bytes carry the required attribution and machine-readable rights contract. Historical package versions remain available under their original identifiers and checksums for reproducibility; the registry now selects the verified non-overwriting v1.0.0-rights.1 versions. See docs/data.md for the full attribution and licensing terms. The accepted open-by-default policy in ADR-0014 offers CC BY 4.0 only for original prepared-model rights the package author owns or controls. That scoped grant is cumulative with the continuing Statistics Canada conditions; it does not cover source Information, facts, or unprotectable results, and it does not relax privacy or provenance safeguards. The fail-closed existing-record correction implementation passed independent review, and all 32 metadata corrections, 32 non-overwriting versions, remote byte checks, and 32 registry updates completed on 2026-08-16. The durable correction record preserves sanitized evidence. External review remains welcome but is not a 1.0.0 gate.

Development Acknowledgement

Development of SynthPopCan has been supplemented by the use of large language models for tasks including code generation, review, testing, documentation, and research assistance. All resulting contributions remain subject to human review and the project's automated correctness and quality checks. Responsibility for the project and its releases rests with its maintainer.