UQO Working Papers · WP21 · v1.0
The Global Evidence on Housing Price Determinants: A Meta-Analysis of Hedonic Pricing Models, 1970–2026
- Year
- 2026
- First release
- October 2026
- Latest revision
- October 2026
- Publication status
- Working paper
The hedonic core is global, but no implicit price is homogeneous
Across 242 studies from 55 countries, living area carries an elasticity of 0.65 and a bathroom a 5.91% premium, yet prediction intervals span zero for most characteristics and observable study features explain little of the dispersion.
Abstract
This paper synthesises the global hedonic housing literature from 1970 to 2026 with a reproducible pipeline: 103,011 records identified from Semantic Scholar, Crossref, arXiv and OpenAlex, 56,727 unique records screened by rules and by independent agents under a PRISMA-style protocol, 2,249 hedonic studies included at stage one, 501 open-access full texts retrieved, and 242 eligible studies from 55 countries extracted into 18,831 coefficient rows with a validated double-extraction sample. Coefficients are harmonised into percentage premiums, semi-elasticities per common unit and elasticities, and pooled with random-effects, three-level and cluster-robust models, with publication-selection diagnostics and Bayesian and machine-learning cross-checks.
The structural core of the hedonic model is global: the elasticity of price to living area is 0.65, a bathroom adds 5.91 percent, a year of age subtracts 0.24 percent and a kilometre from the centre 3.08 percent, with similar means across continents and decades. Signature amenities carry premiums between four and eleven percent and a waterfront location 32.50 percent. Every implicit price is nonetheless heterogeneous beyond sampling error, prediction intervals span zero for most characteristics, and observable study features, including spatial estimators, explain little of the heterogeneity.
Funnel asymmetry is widespread but cannot be interpreted as selection alone. Machine-learning models predict better than hedonic benchmarks in most of the forty retrievable comparisons, while AI-derived text and image features add modest accuracy and some interpretable characteristics. The synthesis extends Sirmans et al. (2005, 2006) in coverage, harmonisation, inference and scope, and is bounded by the open-access retrieval rate that it documents.
Key findings
103,011
bibliographic records identified, 1970–2026
242
eligible hedonic studies from 55 countries
18,831
coefficient rows extracted and harmonised
0.65
pooled elasticity of price to living area
+5.91%
price premium per additional bathroom
+32.50%
waterfront premium, the largest signature amenity
Scientific contributions
- The largest harmonised hedonic synthesis to date
- From 103,011 records to 56,727 unique, 2,249 included hedonic studies and 242 extracted from 55 countries, yielding 7,693 usable harmonised estimates from 149 studies across 67 characteristic–family combinations.
- The structural core is global
- Living area (elasticity 0.65, k = 46), bathrooms (+5.91%), age (−0.24% per year) and distance to the centre (−3.08% per km) have precisely estimated means that are similar across continents and decades once specification is controlled.
- Heterogeneity dominates every implicit price
- Median I² is close to 100% and prediction intervals span zero for most characteristics; only 12 of 182 moderator terms are significant under robust variance estimation and machine-learning meta-regressions have negative out-of-study R².
- Selection is detectable but not correctable
- Egger, PET-PEESE and step selection models find widespread funnel asymmetry, suggestive for amenity premiums, but the asymmetry cannot be read as publication selection alone.
- Machine learning predicts better, AI features add little
- In most of the 40 retrievable same-test-set comparisons tree ensembles beat the hedonic benchmark (median MAPE 9.0 vs 12.4), and nine studies report modest incremental gains from text and image features.
- Validated agent-based extraction
- An independent double extraction of 20 papers reached 100% numerical agreement on 647 matched coefficients, supporting the LLM-assisted extraction of 297 full texts.
Methodology
Methodology (5)
PRISMA-style systematic search and screening
Discovery across four bibliographic services, deduplication, rule-based and agent screening with coded reasons, and a PRISMA flow generated directly from the logs.
Harmonisation of implicit prices
Coefficients converted to percentage premiums (Kennedy correction), semi-elasticities per common unit, elasticities and partial correlations, with quality flags for each estimate.
Multilevel meta-analysis
REML random-effects, three-level and cluster-robust models (metafor rma.uni/rma.mv with clubSandwich CR2) for 67 characteristic–family combinations, with meta-regressions on study moderators.
Publication-selection diagnostics
Egger tests, PET-PEESE and step selection models applied to each characteristic, reported as diagnostics rather than corrections.
Bayesian and machine-learning cross-checks
Gibbs-sampled Bayesian pooling and grouped random-forest and gradient-boosting meta-regressions tested for out-of-study prediction.
Research data
- Bibliographic corpus
- 103,011 records retrieved on 2026-10-20 from the Semantic Scholar bulk API, Crossref, arXiv and OpenAlex with 67 multilingual queries; 56,727 unique after DOI and normalised-title deduplication, screened by rules and by independent agents with coded reasons.
- Harmonised coefficient database
- 501 open-access full texts (18.8% retrieval rate) via Unpaywall, Semantic Scholar and arXiv; 297 papers extracted, 242 eligible, 2,022 model columns and 18,831 coefficient rows harmonised into Kennedy-corrected premiums, unit-harmonised semi-elasticities and elasticities (7,693 usable estimates from 149 studies).
- AI and machine-learning comparison set
- 40 studies with retrievable same-test-set comparisons between machine-learning models and hedonic benchmarks, plus studies reporting incremental gains from text and image features.
Reproducibility
- Makefile pipeline (screen, effects, models, tables, figures, paper) reproduces the paper from stored records and logs without live queries
- Preregistered protocol (protocol/preregistration.md, eligibility and search strategy) and a documented extraction protocol with a 20-paper double-extraction validation
- 14 figures, 23 table fragments and 913 numeric macros regenerated by scripts/visualization/; pytest suite in tests/
- Full LaTeX source of the 51-page paper with sections, appendices and bibliography; open-access PDFs not redistributed but re-downloadable from the retrieval log
Research artifacts
Version history
| Version | Date | Description | File |
|---|---|---|---|
| v1.0current | October 2026 | Current manuscript as published on www.spboucher.ai (revised edition after the 2026-10-17 scientific review for papers 1–17). | PDF · 51 p. |