🗺️ The Economic Geography of Digital Retail Supply
UQO Working Paper No. 38 — Product Diversity, Regional Specialization, and the Hidden Structure of Quebec's Online Economy
TL;DR — A read-only snapshot (1 Nov 2026) of Fabri-Ka, an aggregator of Quebec online stores, gives 370 054 cleaned listings from 1 557 merchants in 17 administrative regions. Regional specialization is real at the level of which merchants sell online (merchant-equal Cramér's V 0.112 vs a store-reshuffle null of 0.086, p < 0.001, in every sample/labelling/discovery-source variant and at MRC level) but is indistinguishable from chance once listing volumes are counted (trimmed-listing p = 0.43, a low-power comparison), so the preregistered joint criterion for H1 is not met. Catalogue sizes are extremely skewed (Gini 0.85; top 10 % of stores hold 79 % of listings). Total regional variety rises with size; rarefied (15-store) diversity does not (ρ = 0.04, CI [−0.65; 0.62], n = 16). Matched products (precision 0.98–1.00) differ across merchants by a median 4.7 log points; resellers post 11.6 log points below producers (p = 0.004), while the joint regional test does not reject equality under a correct wild cluster bootstrap (p = 0.098; one region drives the Wald). The ECI-type complexity indicator fails bootstrap validation at 17 × 20 and is withheld; the QDACI is published as components only. H6 is untestable. Results are reported under a validated classifier labelling (primary, A-002) and the platform's rule labels (46 % audited precision). Three adversarial referee reports and the response log are included.
📌 Status
| Item | Value |
|---|---|
| Version | 1.0 (November 2026) |
| Snapshot | fabrika.db backup 2026-11-01T17:38:30Z, sha256 8d963ba9…98fb, commit 689c7bb |
| Preregistration | research/preregistration.md (sha256 b7cd3f17…), amendments A-001, A-002 in research/decisions.md |
| SSRN / journal | not yet submitted |
🗂 Repository layout
CLAUDE.md mandate (verbatim)
config/settings.yaml paths, seed 20261101, snapshot identity, regions
research/ preregistration (frozen), decision log + amendments
reports/ 00_source_audit, phase2_data_audit, sample_flow, data_dictionary, model_inventory,
hypothesis_registry_{rule,model}, limitations, referee_1_econometrics, referee_2_geography, referee_3_data_engineering, referee_responses, reproducibility
literature/ seed references, Crossref-verified metadata (142/143), evidence matrix, review notes, references.bib
data/raw/ snapshot (not committed; provenance in README.md) · data/external/ public covariates + manifest
data/interim/, data/processed{,_model}/ parquet/csv intermediates per category labelling (not committed)
src/ingest src/clean src/features src/models src/validation src/reports src/literature src/audit
experiments{,_model}/ JSON summary of every step; hypothesis registry
paper/ main.tex, sections/, tables{,_model}/, figures{,_model}/, references.bib, main.pdf
tests/ pytest invariants of the cleaned panel
logs/ run ledger (jsonl, every API call with usage), output manifest (sha256), pipeline log▶️ Reproduce
make env # uv venv (Python 3.12) + pinned deps
make ingest clean match validate # snapshot → parquet → samples → matching → audits (LLM labels are cached CSVs)
make features models robust tables figures paper # rule labelling (preregistered)
WP38_LABEL=model make features models robust tables figures # classifier labelling (A-002)
make testAccess requirements: the authorised snapshot in data/raw/fabrika-snapshot-20261101/ (owner export; not redistributed) and, only to regenerate annotations, OPENAI_API_KEY / ANTHROPIC_API_KEY in .env. External covariates re-download with make external from the URLs in data/external/manifest.json.
✅ Hypotheses — final status (classifier labelling primary; rule labelling identical in status)
| H | Statement | Status | Key statistic |
|---|---|---|---|
| H1 | Regional specialization beyond merchant counts | not supported (preregistered joint criterion); merchant-equal dependence robust | merchant-equal V 0.112 vs null 0.086 (p < 0.001); listing-weighted p 0.20; trimmed p 0.43 |
| H2 | Diversity ~ regional scale | inconclusive | ρ(rarefied entropy, log pop) 0.04 [−0.65; 0.62], n = 16; unrarefied ρ 0.52 |
| H3 | Skewed catalogue sizes | supported | Gini 0.849 [0.811; 0.878]; top-10 % 78.9 % |
| H4 | Related diversification | exploratory | LPM 0.102 (SE 0.031), placebo p 0.005; 1 region-level entry (untestable) |
| H5-T1 | Geographic price structure | not supported | Wald(14) 164.2, wild-boot p 0.098 (1 064 groups, 143 stores; Outaouais = 65 of 164) |
| H5-T2 | Competitive price structure | supported | reseller −0.116 (SE 0.027), p 0.004, Holm 0.016 |
| H6 | Local-origin premium | untestable | no product-level provenance |
| H7 | Complexity indicator | not validated | bootstrap rank ρ 0.68 (< 0.8); LOO predictive ρ 0.01 |
| QDACI | Composite | withheld | median agreement across weightings 0.29; components published |
Generated registries: reports/hypothesis_registry_model.md, reports/hypothesis_registry_rule.md.
⚠️ What this is not
Listings are posted offers of merchants located in a region; they are not sales, employment, output or product origin. The indicators describe observed online catalogues.