1# WP38 — Literature review notes and contribution map (Phase 1)23*Prepared 2026-11-01. Every key cited below is in `literature/references_verified.csv` with a Crossref DOI (raw JSON in `crossref_cache/`) or a fetched URL. One seed could not be verified and is therefore not cited: `kempbenedict2014` (Kemp-Benedict, "An interpretation and critique of the Method of Reflections", MPRA 2014 — not retrievable from MPRA or IDEAS searches). Corrections to the mandate's reference list, as returned by Crossref: Couture–Faber–Gu–Liu is *AER: Insights* 2021 (not *Econometrica*); Gorodnichenko–Sheremirov–Talavera is *JEEA* 2018 (not *AEJ: Macro*); Kaplan–Menzio–Rudanko–Trachter is *AEJ: Micro* 2019 (not *JME*); Zentner–Smith–Kaya is *Management Science* 2013; Jo–Matsumura–Weinstein is an NBER working paper (w26506); "Cavallo 2023 on online geography" was interpreted as Cavallo–Kryvtsov, *JIE* 2023 (stockouts in online micro data) — if a different paper was meant it must be added and verified.*45## Stream 1 — Economic geography, specialization, agglomeration67**Established.** Industries are localised far beyond a random allocation of plants across space (`ellison1997`, `duranton2005`, `mori2005`); the degree of localisation is heterogeneous across industries and scales (`rosenthal2001`), and this holds in Canada with Statistics Canada plant microdata (`baldwin2008`). The micro-foundations are sharing, matching and learning (`duranton2004`; Marshall's external economies, `marshall1920`), with density elasticities of productivity in the 2–8 % range but demanding identification (`rosenthal2004`, `combes2015`, `duranton2020`). Whether local specialisation or diversity favours growth is unresolved (`glaeser1992`, `henderson1995`), with the related/unrelated-variety entropy decomposition as the modern operationalisation (`frenken2007`, `content2016`). Market size and love of variety (`krugman1980`, `krugman1991`, `fujita1999`) and the skill-ordered comparative advantage of cities (`davis2020`) motivate size-related assortment breadth. Urban-scaling work shows that the count of distinct business types in a city grows roughly with the logarithm of population, and that activities appear in a nested order across city sizes (`bettencourt2007`, `youn2016`, `hong2020`; local online content scales with city size in `sinai2004`). For Quebec and Canada, Polèse and Shearmur document that distance to metropolitan areas and the urban hierarchy still structure employment location and that peripheral regions decline (`polese2004`, `polese2006`, `shearmur2007`, `shearmur2008`). Spatial-unit choice (MAUP) matters moderately for such estimates (`briant2010`).89**Gap relevant to WP38.** All of this measures production or employment. No study in this stream uses product catalogues of online merchants as the object, and none covers Quebec's 17 administrative regions in the 2020s.1011## Stream 2 — Economic complexity, RCA, relatedness1213**Established.** The Balassa index (`balassa1965`) is skewed and unstable at the top; symmetric and normalised variants exist (`hinloopen2001`, `yu2009`, `laursen2015`). The product space (`hidalgo2007`) and the principle of relatedness (`hidalgo2018`, `boschma2017`) are robust empirical regularities: regions diversify into activities related to what they already do, documented with industry (`neffke2011`), patent (`kogler2013`, `rigby2015`, `balland2017`), occupation (`muneepeerakul2013`) and firm/industry data (`gao2018`); the EU smart-specialisation framework combines relatedness density and complexity (`balland2019`). ECI/PCI (`hidalgo2009`, `hausmann2011`, `hausmann2014`) and fitness–complexity (`tacchella2012`, `cristelli2013`, `servedio2018`) summarise bipartite matrices; ECI is formally a spectral clustering of a region-similarity matrix (`mealy2019`), rankings are sensitive to algorithm and sample choice (`morrison2017`, `sciarra2020`), and the state of the art and open problems are surveyed in `hidalgo2021` and `balland2022`. Structure in bipartite matrices (nestedness, specialisation) must be tested against degree-preserving null models (`saracco2015`, `hausmann2011`), in the spirit of the dartboard (`ellison1997`).1415**Gap.** Complexity tools have been applied to exports, patents, industries, occupations and firms — all production-side bases — but not to online retail assortments. Nothing is known about their behaviour on a 17 × 20 matrix built from listings, which is why WP38 must treat H7 as an indicator-validation exercise rather than a ranking.1617## Stream 3 — Online retail, long tail, concentration, geography of online demand1819**Established.** Online variety created consumer surplus and shifted sales toward niche titles (`brynjolfsson2003`, `brynjolfsson2011`, `zentner2013`, `waldfogel2017`; popularised by `anderson2006`), but hits remain dominant and the tail is thin in revenue terms (`elberse2008`, `tan2017`); search-cost theory predicts both long tails and superstars (`barisaac2012`). E-commerce reallocates retail share toward large firms (`goldmanis2010`) and platform owners/complementors (`zhu2018`, `einav2016`). The geography of online demand is systematic: people buy more online where local supply is thin and when they are "preference minorities" (`forman2009`, `choi2011`, `brynjolfsson2009`), and across-market demand heterogeneity limits the value of online variety for products with local tastes (`quan2018`). Distance still matters online (`hortacsu2009`, `blum2006`, `lendle2016`). Gains from e-commerce are larger for remote consumers (`dolfen2023`), rural access can be expanded experimentally (`couture2021`), and e-commerce changes relative prices across space (`jo2019`, `lieber2012`, `goldfarb2019`). Platform design shapes price dispersion (`dinerstein2018`, `ellison2009`, `baye2004`, `brynjolfsson2000`, `clay2001`, `baye2006`).2021**Gap.** The entire stream is demand- or transaction-based (sales ranks, purchases, auctions). The supply-side question — what independent local merchants put online, and how that varies across regions of one province — is not addressed.2223## Stream 4 — Regional inequality, rural digital access, e-commerce adoption2425**Established.** Firms in large cities adopted the commercial internet first (`forman2005`), the wage gains from internet investment accrued to already-advantaged counties (`forman2012`), and the overall picture is one of complementarity between the internet and cities (`sinai2004`, `forman2018`). Broadband effects on local and rural growth are positive but modest and identification is difficult (`kolko2012`, `whitacre2014`), and firm-level broadband productivity effects are heterogeneous (`haller2015`, `haller2019`). Adoption and usage divides differ (`goldfarb2008`); Canadian household divides are documented with CIUS microdata (`haight2014`). Measurement frameworks: `oecd2019`; Statistics Canada's CIUS (`statcan_cius`, `statcan2023cius`) and SDTIU (`statcan_sdtiu`); the Quebec NETendances programme (`atn_netendances`).2627**Where peer-reviewed evidence is thin.** We found no peer-reviewed study of regional e-commerce *supply* (merchant presence or catalogues) in Canada or Quebec. Quebec-specific evidence is grey literature (ISQ, ATN NETendances, government pages). This stream is the weakest in verified academic sources, and the paper should say so.2829## Stream 5 — Product matching, hedonic prices, geographic price dispersion3031**Established.** Record linkage rests on Fellegi–Sunter (`fellegi1969`, `christen2012`). Scraped online prices are a valid but selected source: online and offline prices coincide in most multi-channel retailers (`cavallo2017`), scraped data have specific frequency and stockout properties (`cavallo2018`, `cavallo2023`), and the Billion Prices Project set the standards for their use (`cavallo2016`); online price-setting facts and cross-border integration are documented in `gorodnichenko2017`, `gorodnichenko2018`, `cavallo2014`. Price dispersion is mostly relative (store × good) and persistent (`kaplan2015`, `kaplan2019`); US chains price almost uniformly across locations (`dellavigna2019`); spatial price gaps grow with distance (`atkin2015`); variety and prices differ across cities (`handbury2015`, `handbury2021`); product entry drives unequal gains (`jaravel2019`). Hedonic theory and practice: `rosen1974`, `pakes2003`, `bajari2005`. Web-scraped data are entering official CPIs (`konny2022`).3233**Gap.** Cross-merchant dispersion among *independent* small online stores in one province, with product fixed effects from fuzzy-matched titles, has no direct precedent; the identification caveat that region and merchant effects are nearly collinear is specific to WP38.3435## Stream 6 — Measurement error, selection, representativeness3637**Established.** For non-probability big data, error is governed by the data-defect correlation rather than by sample size (`meng2018`, applied in `bradley2021`; the Google Flu warning, `lazer2014`). Inference from non-probability samples requires explicit selection models (`elliott2017`, `buelens2018`, `heckman1979`, `groves2006`, `little2019`); official statistics agencies, including Statistics Canada, are explicit about the conditions (`daas2015`, `beaumont2020`, `rao2021`). Measurement error in regressors and outcomes (`hausman2001`, `bound2001`) and the MAUP (`briant2010`) are additional threats.3839## Institutional block (Quebec/Canada)4041Canadian Business Counts derive from the Business Register (payroll, $30,000 revenue or incorporation rules; counted by location; Statistics Canada warns against time-series use) — `statcan2026cbc`, `statcan_brdef`, `statcan_br`. Regional statistics for the 17 administrative regions and 104 MRCs: `isq2024`, `quebec_regions`. Origin certification: food marks (`quebec_aliments`, `alimentsqc`); non-food marks by Les Produits du Québec — "Produit du Québec" requires at least 85 % of direct production and manufacturing costs incurred in Quebec plus final substantial transformation there, "Fabriqué au Québec" only the final substantial transformation, "Conçu au Québec" a Quebec-based design workforce (`produitsqc`, fetched 2026-11-01). These are the only admissible origin evidence for H6.4243## The eight closest empirical precedents44451. **Quan & Williams (2018, RAND)** — `quan2018`. Footwear sales across US markets with a demand model: the value of online variety is overstated when local tastes differ across markets. *Comparison:* both papers are about spatial heterogeneity in what online retail offers/needs; Quan–Williams identify demand heterogeneity from transactions, WP38 observes only supply heterogeneity in listings and cannot say whether regional assortments reflect regional tastes.462. **Youn et al. (2016, J. R. Soc. Interface)** and **Hong et al. (2020, Science Advances)** — `youn2016`, `hong2020`. Business-type diversity across US metros scales with population; industries enter in a nested order. *Comparison:* H2 and the nestedness checks of H7 are direct analogues on online category diversity across 17 regions; WP38 has far fewer spatial units, administrative rather than functional, and listings rather than establishments.473. **Neffke, Henning & Boschma (2011, Econ. Geogr.)** and **Balland et al. (2019, Reg. Stud.)** — `neffke2011`, `balland2019`. Regional diversification follows relatedness density in multi-decade panels. *Comparison:* WP38 can build a category-relatedness matrix and relatedness density, but with a single snapshot H4 is at best a cross-sectional co-occurrence description; entry regressions would conflate catalogue onboarding with economic entry.484. **Cavallo (2017, AER)** and **Gorodnichenko & Talavera (2017, AER)** — `cavallo2017`, `gorodnichenko2017`. Online prices as a measurement instrument; dispersion and integration across stores and borders. *Comparison:* H5 inherits their matched-product logic but across many independent stores rather than a few multi-channel retailers or one platform; WP38 has one snapshot and no stickiness dimension.495. **Hortaçsu, Martínez-Jerez & Douglas (2009, AEJ: Micro)** — `hortacsu2009`. Distance and same-city effects in eBay/MercadoLibre transactions. *Comparison:* WP38 has seller locations only; it describes where online supply sits, not where it is shipped — gravity is out of reach.506. **Couture, Faber, Gu & Liu (2021, AER: Insights)** — `couture2021`. RCT of e-commerce access in rural China: gains concentrated among a minority of households. *Comparison:* WP38's rural/peripheral regions appear only as merchant locations; no treatment, no welfare.517. **Polèse & Shearmur (2004, 2006)** — `polese2004`, `polese2006`. Canadian location patterns by distance class; declining peripheral Quebec regions. *Comparison:* WP38 asks whether the same centre–periphery gradient appears in online catalogue breadth in 2026; descriptive, with the aggregator's merchant selection as the main confounder.528. **Forman, Ghose & Goldfarb (2009, Mgmt. Sci.)** — `forman2009`. Online purchases respond to local store availability. *Comparison:* supply-side mirror again; WP38 can relate regional online supply to Business Counts of local retail establishments but cannot observe purchases.5354## Contribution map5556**Would be novel (and defensible):**57- A documented, validated *supply-side* dataset of online product listings of Quebec merchants by administrative region, with explicit data-quality ledgers (duplicates, variants, missing prices, geocoding quality) — no precedent in streams 1–4.58- Revealed assortment advantage (RAA) and entropy/HHI diversity measures on listing counts with merchant-count-preserving permutation nulls and merchant-weighted variants — a transparent adaptation of RCA/dartboard logic to a new base (C01, C07, C12).59- A region–category bipartite "digital assortment" indicator whose eigenvector interpretation is validated against configuration-model nulls and bootstraps by merchant, reported as a family (QDACI components) rather than a headline score (C09, C10, C12).60- Cross-merchant dispersion of matched products among independent online stores in one province, with matched-product fixed effects and explicit collinearity diagnostics between merchant and region effects (C18, C22, C23).61- A representativeness audit of an online aggregator against Business Counts and CIUS/SDTIU, framed with the data-defect-correlation logic (C25).6263**Would not be novel, and must not be claimed:**64- Any new method: RCA, entropy decomposition, relatedness density, ECI/fitness, hedonic fixed effects and Fellegi–Sunter matching are all established.65- Evidence on agglomeration economies, related diversification over time, long-tail *sales* concentration, gravity in online trade, or welfare gains from e-commerce — WP38 has no productivity, panel, sales, flow or welfare data.66- A "production complexity" or "regional economic complexity" index for Quebec: without verified origin, the indicator describes observed online catalogues only.67- A local-origin price premium (H6) unless certification marks (`produitsqc`, `quebec_aliments`) can be attached to listings; otherwise H6 is untestable.6869**Honest framing for the manuscript:** descriptive and associational, single cross-section (2026-11-01 snapshot), listings not production, merchant location not product origin, non-probability sample.70