UQO Working Paper No. 21· v1.1· November 2026· 85 pages
The Global Evidence on Housing Price Determinants
A Meta-Analysis of Hedonic Pricing Models, 1970–2026
Across 140 open-access hedonic studies the living-area elasticity is 0.62 [0.50, 0.73] and a bathroom is associated with 8.62 percent, yet prediction intervals span zero in all 52 characteristic families.
A systematic search of 103,011 records yields 242 extracted studies and a primary pool of 4,139 harmonised estimates, pooled with a three-level model and cluster-robust inference. No moderator survives multiplicity adjustment, and clustered funnel tests detect asymmetry in 7 of 52 families. The search and taxonomy are global, but the pooled evidence is the open-access literature: 875 of 2,667 sought full texts were obtained, and the pooled implicit prices are associations in the published literature, not causal effects.
Simon-Pierre Boucher — contact@spboucher.ai
0.62 [0.50, 0.73]
pooled elasticity of price to living area (k = 46 studies), three-level model with cluster-robust inference
+8.62% / +5.03%
price association of an additional bathroom and of an additional bedroom in the primary pool
−0.61% / −3.30%
association with a year of age and with a kilometre of distance from the centre (k = 23); the distance value is conditional on the cap applied to the steepest gradients
4.96–13.03%
premiums of eight signature amenities (garage, pool, air conditioning, view, fireplace, basement, terrace, garden), each with a cluster-robust interval that excludes zero; a waterfront location is associated with 30.71 percent but rests on k = 8 studies and is reported as secondary
52 of 52
characteristic families whose prediction interval spans zero: every implicit price is heterogeneous beyond sampling error
0 of 144
moderator terms that survive a false-discovery adjustment (18 are nominally significant); 56 of the tests rest on fewer than four effective clusters, so this is also an absence of power
13 of 17
comparisons with a held-out test set in which machine-learning models predict better than hedonic benchmarks; too few to support claims about the general accuracy of machine-learning valuation
Abstract
What does the hedonic housing literature say about the implicit prices of dwelling characteristics, and how much of it can be pooled? A systematic search of 103,011 records, agent screening of 6,996 candidates and open-access retrieval of 875 of 2,667 sought full texts (32.8 percent) yield 242 extracted studies; 4,139 harmonised estimates from 140 studies in 52 characteristic–family combinations form the primary pool. Under a three-level model with cluster-robust inference, the living-area elasticity is 0.62 [0.50, 0.73]; a bathroom is associated with 8.62 percent, a year of age with -0.61 percent and a kilometre from the centre with -3.30 percent; the premiums of eight signature amenities lie between 4.96 and 13.03 percent. Prediction intervals span zero in every family, no moderator survives multiplicity adjustment, and clustered funnel tests detect asymmetry in 7 of 52 families. The search and taxonomy are global; the pooled evidence is the open-access literature.
JEL: R31, R21, C13, C80, C45
Key Results
A harmonised synthesis across the whole characteristic taxonomy
Existing quantitative meta-analyses are attribute-specific and the general successors of the Sirmans programme are descriptive; this paper harmonises coefficients across functional forms, weights by precision, models within-study dependence and reports prediction intervals, to the extent the open-access literature allows.
Pooling restricted to a common estimand
The synthesis separates a descriptive account (which characteristics are studied, with what sign) from a pooled analysis of 4,139 estimates in 52 families; band, category, cohort and time-dummy sets are withdrawn from pooling and one quantile per model is kept.
Heterogeneity dominates every implicit price
Structural characteristics and signature amenities have pooled associations whose intervals exclude zero and whose signs match the hedonic canon, but no observable feature of a study, including the estimator, is detectably associated with the implicit price once multiplicity is accounted for.
Small-study patterns are not where expected
Under clustered inference, funnel asymmetry appears in 7 of 52 families, in the living-area elasticity rather than in the amenity premiums the protocol expected, and is not interpretable as selection alone; no correction is applied.
A measured account of what the literature allows
The paper accounts for every record: 32.8 percent of sought full texts obtainable by an automated fetcher, a further share retrievable but not yet assessed, and reporting gaps in units, dummy references and transformations that limit cumulative knowledge.
Measured reliability of automated screening and extraction
An independent double extraction and a blind re-read of 20 papers from the original PDFs find no transcription error in the coefficients but a 7.68 percent error rate in the recorded dependent-variable transformation; taxonomy assignment agrees in 86.0 percent of matched pairs.
Quality control
This paper went through a multi-stage quality control: 7 quality-control reports and 117 points checked. The version presented on this page is the revised manuscript. Scores are on a 0–100 scale and are a diagnostic of the manuscript, not an acceptance probability.
Score after revision
69.0/100
55.2 before quality control · +13.8
Quality control dated November 7, 2026
- Quality-control decision
- major revision
- Scientific integrity
- PASS
- Reproducibility
- R1reproduced from documented analysis data
- Maturity
- Bone revision cycle from submission
- Quality-control reports
- 7
- Points checked
- 117
Scores by dimension
Before quality controlAfter revision
- Originality & contribution67.9 → 70.0
- Scientific rigor47.9 → 68.0
- Identification & methodology50.0 → 65.0
- Data & reproducibility56.4 → 70.0
- Robustness & validation48.6 → 72.0
- Writing67.9 → 73.0
- Figures & tables48.6 → 72.0
- Literature positioning57.1 → 68.0
Revised manuscript85 pages13 figures37 tables2 equations111 references38,174 words
What quality control changed
Quality control found that the first version had averaged rows that do not share a common estimand and had not reported its headline numbers from the model its protocol declared primary. The revised paper restricts pooling to a primary pool, reports every headline number from the three-level model with cluster-robust inference, and states that the pooled evidence is the open-access literature.
- Pool: 7,693 usable estimates from 149 studies in 67 characteristic–family combinations → 4,139 estimates from 140 studies in 52 families, after withdrawing band, category, cohort and time-dummy sets and keeping one quantile per model.
- Headline estimates: living-area elasticity 0.65 → 0.62 [0.50, 0.73]; bathroom 5.91 → 8.62 percent; year of age −0.24 → −0.61 percent; kilometre from the centre −3.08 → −3.30 percent.
- Waterfront premium: 32.50 percent presented as the largest signature amenity → 30.71 percent on k = 8 studies, reported as secondary; signature amenity premiums ‘between four and eleven percent’ → 4.96 to 13.03 percent for eight amenities.
- Publication selection: funnel asymmetry ‘widespread’ → clustered funnel tests detect asymmetry in 7 of 52 families, not in the amenity premiums the protocol expected.
- Machine learning: better prediction ‘in most of the forty retrievable comparisons’ → in 13 of the 17 comparisons with a held-out test set.
- Scope and validation: ‘the structural core is global’ → the search and taxonomy are global, the pooled evidence is the open-access literature (875 of 2,667 sought full texts); taxonomy agreement in the double extraction 100 percent → 86.0 percent, and a blind re-read finds a 7.68 percent error rate in the recorded dependent-variable transformation.
Limitations stated in the paper
- The pooled values describe the recent open-access literature: 32.8 percent of sought full texts were obtained, several hundred obtainable PDFs were not assessed, and the pre-2000 North American journal literature is under-represented.
- Screening and extraction were performed by language-model agents whose model version and decoding settings were not logged; the duplicate extractor and the re-reader are of the same model family, and no human-coded validation sample exists.
- Moderators are coarse and no term survives multiplicity adjustment, which is also an absence of power for families with fewer than twenty studies; publication-selection diagnostics have fewer than four degrees of freedom in 39 of 52 families and no correction was applied.
- The pooled implicit prices are associations in the published literature, not causal effects; the comparison with the Sirmans studies is at abstract level, and one study supplies at least half of the fixed-effect weight in 30 families.
Data
Bibliographic corpus
103,011 records identified on 2026-10-20 from the Semantic Scholar bulk search API, Crossref and arXiv, with OpenAlex used for verification (67 queries attempted, 66 returned records); 56,727 unique records after deduplication, 6,996 candidates screened on title and abstract, 2,249 included at stage one as hedonic coefficient studies.
Full-text retrieval
Of 2,667 full texts sought, 875 were obtained from open-access locations over two runs (32.8 percent); only the 501 obtained in the first run could feed the extraction, and 206 of those plus the 374 obtained in the second run were never assessed.
Coefficient database and primary pool
242 eligible studies from 55 countries, 2,022 model columns and 18,831 coefficient rows; 4,139 harmonised estimates from 140 studies in 52 characteristic–family combinations enter the primary pool.
Machine-learning and AI comparison set
The synthesis of machine-learning and AI-derived evidence rests on 40 studies and 17 defensible comparisons with a held-out test set between machine-learning models and hedonic benchmarks, together with what text- and image-derived characteristics add.
Methodology
Systematic search and screening
PRISMA 2020 principles adapted to economics, with a protocol written before any coefficient was extracted; rule-based scoring followed by screening with a coded reason for every decision, and a flow of records generated from the logs on a single base.
Harmonisation into effect-size families
Coefficients converted into percentage premiums with the Kennedy correction, semi-elasticities per common unit and elasticities, with a rule-based classification of premium rows, magnitude caps disclosed with their counts, and exclusion of rows that are not a common estimand.
Three-level meta-analysis with cluster-robust inference
The model the protocol declared primary: estimates nested in studies, cluster-robust intervals and prediction intervals for each family pooled at five studies or more, with moderator tests adjusted for multiplicity.
Publication-selection diagnostics and robustness
Clustered funnel tests and selection models reported as diagnostics rather than corrections; leave-one-out, high-confidence subsets, a Bayesian sampler and grouped machine-learning meta-regressions as cross-checks.
Validation of the extraction
Independent double extraction of 20 papers, blind re-read from the original PDFs, a re-screen of unclear and rule-excluded records, and propagation of the measured error rates through the bathrooms family (standard deviation of 0.52 percentage points, a scenario built from bounds).
Reproducibility
- The replication repository contains the complete pipeline: discovery, screening with coded decisions, retrieval with every attempt logged, extraction protocol and one record per paper with a provenance block, duplicate extractions and re-read records, harmonisation, econometrics, figures and tables.
- Every number quoted in the text is a macro produced from the result files; dependencies are pinned and random seeds fixed.
- A regeneration test re-runs the models and tables in a temporary copy and compares every result file and every macro with the committed ones; it completes in 3.4 minutes on a sixteen-core machine.
- The screening and extraction steps use language-model agents following written protocols; their outputs are stored artefacts of one run and would not be reproduced byte for byte, which is why their reliability is measured by validation exercises.
- Raw PDFs are not redistributed; the retrieval log allows them to be re-downloaded from the same open-access locations.
Cite this paper
Boucher, S.-P. (2026). The Global Evidence on Housing Price Determinants. UQO Working Paper No. 21, Université du Québec en Outaouais.
@techreport{boucher2026wp21,
author = {Simon-Pierre Boucher},
title = {The Global Evidence on Housing Price Determinants},
institution = {Université du Québec en Outaouais, Département des sciences administratives},
number = {21},
type = {UQO Working Paper},
year = {2026},
month = {nov},
url = {https://www.spboucher.ai/research/wp21},
}More in Real Estate & Urban Economics