# WP21 — Phase A: Scientific foundation and the inventory of competing reviews

*Prepared 2026-10-20. Every bibliographic claim below rests on a Crossref (or DataCite) record saved in
`reports/verification/`; every number taken from a text is marked by the access level actually obtained
(**full text**, **abstract**, or **secondary**). Retrieval log: `reports/foundations_sources/SOURCES.md`.*

## 1. What was read, and at what level

Three works define the programme that WP21 extends. Rosen (1974, *JPE* 82(1), 34–55; DOI 10.1086/260169) is
closed (JSTOR/UChicago) and carries no abstract; its content is described from the standard account, cross-checked
against the two full-text reviews below. Sirmans, Macpherson and Zietz (2005, *JREL* 13(1), 1–44;
DOI 10.1080/10835547.2005.12090154) is closed with no repository copy (OpenAlex and Unpaywall both report
`closed`); only its abstract was obtained. Sirmans, MacDonald, Macpherson and Zietz (2006, *JREFE* 33(3), 215–240;
DOI 10.1007/s11146-006-9983-5) is flagged green open access through SSRN 902880, but SSRN returned HTTP 403 to every
request; the abstract, keywords and the ten-item reference list were read from the Springer landing page. The
follow-up by Khoshnoud, Sirmans and Zietz (2023, *JREL* 31(1), 1–47; DOI 10.1080/09277544.2023.2201020) was found
during verification and is central to the comparison; abstract only.

Four full texts were read and archived under `reports/foundations_sources/`: Malpezzi's selective review
(2002 draft of the 2003 Blackwell chapter, recovered from the Internet Archive capture of the author's university
URL listed by RePEc), Herath and Maier's WU discussion paper (2010, DataCite DOI
10.57938/e55da0fe-d130-415d-9a5d-7bbc07c329a1), the Tinbergen working-paper version of Debrezion, Pels and Rietveld
(2004, published 2007), and Moreno-Foronda et al. (2025, CC-BY). The remaining competing reviews were characterised
from abstracts retrieved through Crossref, OpenAlex, the Semantic Scholar batch API, PubMed, IDEAS/EconPapers and
publisher landing pages; MDPI full texts could not be fetched automatically (HTTP 403) and must be downloaded by hand
in Phase B.

## 2. Rosen (1974): the theoretical warrant

Rosen's contribution is the equilibrium interpretation of a price schedule for differentiated goods. A house is a
bundle of objectively measured characteristics; the hedonic price function is the envelope of consumers' bid
functions and producers' offer functions, so its gradient at any point is the implicit (marginal) price of a
characteristic. Two consequences govern everything WP21 does. First, the function is an equilibrium locus, not a
demand or supply curve: a first-stage hedonic coefficient is a reduced-form marginal price, and recovering
willingness-to-pay parameters requires a second stage whose identification problems (prices and quantities chosen
jointly along a nonlinear schedule) are the subject of Malpezzi's and Herath and Maier's reviews and of Kuminoff,
Parmeter and Pope (2010). Second, Rosen "had little to say about functional form" (Malpezzi) beyond the expectation of
nonlinearity, which is why the empirical literature has oscillated between linear, semi-log, log-log, Box–Cox and,
later, nonparametric and machine-learning surfaces. A meta-analysis therefore pools *reduced-form marginal prices
measured under heterogeneous functional forms*, and must harmonise them before it can compare them.

## 3. The Sirmans programme reconstructed

**Sirmans, Macpherson and Zietz (2005).** The abstract describes "a review of recent studies that have used hedonic
modeling to estimate house prices" and reports results as directions of effect: positive for slanted versus flat roof,
sprinkler system, garden bath, separate shower stall, double oven and gated community; negative for the absence of
attic space, earthquake zone, proximity to a hog farm, a landfill or high-voltage lines, corporate ownership, the
percentage of Black or Hispanic residents, and a flood-insurance requirement. The tabulation is therefore a
sign-and-significance count by characteristic, not a pooled effect size. The figure of **125 studies** frequently
attributed to the paper could **not be verified** from any accessible text, abstract or secondary source consulted
here (Khoshnoud et al. 2023 describe the 2005 paper only as the work they "extend"); WP21 should cite the count only
after reading the full text through the UQO library. The paper reviews predominantly US work published in real-estate
journals (secondary, inferred from the 2023 extension's sampling frame), and its reference-list page range (3–43 in the
2006 paper's own citation versus 1–44 in Crossref) should be reported as 1–44.

**Sirmans, MacDonald, Macpherson and Zietz (2006).** The abstract is explicit about scope and method. The object is
"the nine housing characteristics that appear most often in hedonic pricing models for single-family housing: square
footage, lot size, age, bedrooms, bathrooms, garage, swimming pool, fireplace, and air conditioning". The method is
meta-regression of *estimated coefficients* on moderators: geographical location, time, type of data, model
specification, and wealth (median household income). The reference list confirms the lineage: Stanley and Jarrell
(1989, 1998), Jarrell and Stanley (1990), Card and Krueger (1995), Doucouliagos (1995), Higgins and Thompson (2004) on
spurious meta-regression findings, Bateman and Jones (2003) on multilevel meta-analysis, and Smith and Huang (1995) as
the hedonic precedent. Findings (abstract): coefficients differ significantly by location for square footage, lot
size, age, bathrooms, swimming pool and air conditioning; "controlling for time shows that the effects of these housing
characteristics on house price have not changed over time"; data type matters for bathrooms; income has no significant
effect; the lot-size coefficient falls when square footage is controlled; the size of the hedonic model affects the
square-footage coefficient. The number of studies and estimates per characteristic, the handling of functional form
(whether semi-log coefficients were converted to percentages), the weighting scheme and the treatment of multiple
estimates per study are **not accessible** from the abstract and are left blank in `competing_reviews.csv`. The
keywords ("selling price; time on the market; meta regression; moderator variables") show that the same template was
re-used for the 2010 time-on-the-market meta-regression (Sirmans, MacDonald and Macpherson 2010, *JHR* 19(2)), whose
abstract reports sensitivity of the TOM coefficient to year of sale, income, model size and specification but not to
location.

**Khoshnoud, Sirmans and Zietz (2023).** The extension reviews 252 studies published 2005–2021 in real-estate journals
(150 in *JREFE*, *JRER* and *REE*; 42 in *JHR*), finds the field "global in both authorship and data usage, although most
of the studies use U.S. data", and documents the arrival of spatial autoregressive, difference-in-differences and
quantile models. Like the 2005 paper it is descriptive; it does not pool effect sizes.

**Limitations of the programme — explicit and implicit.** The 2006 abstract itself concedes two: it covers only
single-family housing and only nine structural attributes. The others follow from the design and from the verified
descriptions: (i) the evidence base is drawn from a handful of US real-estate journals, so the "no change over time"
and "varies by location" results are statements about US regions and publication years rather than data periods or
countries; (ii) coefficients are compared as reported, without conversion into a common effect-size family or
inverse-variance weighting, and the 2005 tabulation counts signs, which cannot distinguish genuine heterogeneity from
sampling error; (iii) multiple estimates per study are not treated as nested, so inference ignores within-study
dependence (Bateman and Jones 2003 is cited, but the abstract gives no sign that multilevel estimation was used);
(iv) no publication-selection diagnostics are mentioned; (v) location, environmental and neighbourhood attributes,
rents, listing prices, condominiums and non-US markets are outside the pooled analysis; (vi) estimation is OLS
throughout — spatial and machine-learning estimators post-date the sample.

## 4. Other general reviews and methodological benchmarks

Malpezzi (2002/2003, full text) is the canonical applied review: theory offers "not much" guidance on specification;
the semi-log form is preferred for interpretability and heteroskedasticity, with the Halvorsen–Palmquist and Kennedy
corrections for dummies; coefficients are "not robust with respect to omitted variables" (Butler 1982; Ozanne and
Malpezzi 1985) even when predictions are; submarket F-tests "usually reject the null"; and spatial econometrics is
singled out as the most promising extension. Herath and Maier (2010, full text) classify the 471 most-cited ISI papers:
321 empirical, 134 methodological; 178 on neighbourhood characteristics but only 16 on structural characteristics, which
they call under-researched, as are crime (3 papers) and racial composition (11). Boyle and Kiel (2001) survey
environmental externalities; Hill (2013) provides a taxonomy of hedonic indexes; Bishop et al. (2020) codify best
practice for environmental hedonic designs; Bowen et al. (2001) argue that spatial diagnostics should be routine;
Jayantha and Oladinrin (2020) map 269 Scopus records 1970–2019 bibliometrically. None pools effect sizes.

## 5. What a modern global synthesis can add

1. **Global coverage.** Every verified general review is US-centred; the 2023 extension still relies mostly on US data.
   WP21 targets all countries, 1970–2026, and reports continent, income group and urbanisation as moderators.
2. **Effect-size harmonisation.** Sirmans et al. compare raw coefficients. WP21 converts to percentage premiums
   (100[exp(β)−1]), elasticities, semi-elasticities and partial correlations, with unit conversions and propagated
   uncertainty, pooling only within families (preregistration §3).
3. **Multilevel dependence.** Estimates nested in models, samples and studies; three-level REML models with
   cluster-robust (CR2) inference, and Bayesian hierarchical models with convergence diagnostics (§4 M2, M5).
4. **Publication selection.** Funnel asymmetry, PET-PEESE and selection models, and the published-versus-working-paper
   contrast — absent from the Sirmans programme though present in Mohammad et al. (2013), Beltrán et al. (2018) and
   Guignet et al. (2022).
5. **Spatial, ML and AI estimators.** Separate interpretable marginal effects from predictive performance; test whether
   spatial or ML-derived marginal effects differ from OLS once moderators are controlled (H3), and synthesise the
   incremental information of text, image and other AI-derived characteristics (RQ6).
6. **Temporal evolution by data period, not publication year.** The "no change over time" result of 2006 is defined on
   publication years; WP21 uses data-period midpoints and tests crisis, monetary-policy and COVID-19 windows (RQ4).
7. **Rents versus prices, transaction versus listing, property types.** Outcome families kept separate and compared
   rather than restricted to single-family transaction prices.
8. **Neighbourhood, environmental and location attributes** are brought into the same framework as structural ones,
   with a hierarchical taxonomy, so that the attribute-specific meta-analyses below become subsets of one database.
9. **Reproducibility.** Public meta-dataset, pipeline and validation statistics, following the standard set by
   Guignet et al. (2022).

## 6. Competing reviews: coverage versus the WP21 targets

Fifty verified items (49 reviews plus Kuminoff et al. 2010 as a methodological benchmark) are inventoried in `data/metadata/competing_reviews.csv` (counts only where the abstract
or full text states them). Three patterns matter for the "largest synthesis" claim.

*Scope.* No existing meta-analysis pools structural, neighbourhood, environmental and location attributes together.
The general reviews (Sirmans 2005/2006/2023; Malpezzi; Herath and Maier; Boyle and Kiel; Hill; Jayantha and
Oladinrin) are descriptive, and the largest of them covers 252 studies (Khoshnoud et al.) or 471 bibliometric records
(Herath and Maier). Quantitative pooling exists only for single attributes.

*Size.* The largest attribute-specific meta-datasets are Guignet et al. (2022: 36 studies, 656 estimates, water
clarity, US), Beltrán et al. (2018: 37 works, 364 estimates, flood risk), Simons and Saginor (2006: 75 articles, ~290
observations, contamination), Heberling et al. (2024: 29 studies, 290 estimates), Cespedes-Lopez et al. (2019: 66
documents, 173 estimates, EPC), Braden et al. (2011: 46 studies, 129 estimates, waste sites), Mohammad et al. (2013:
23 studies, 102 observations, rail), Debrezion et al. (2007: 102 results collected, 73 used, rail stations), Nelson
(2004: 20 studies, 33 estimates, airport noise) and Zhang and Yen (2020: 23 studies, BRT). WP21's targets (≥5,000
records screened, ≥1,000 full texts, several hundred eligible studies) exceed every single inventoried review by an
order of magnitude in studies, but the comparison must be made per characteristic: for water quality or flood risk the
attribute-specific meta-analyses may remain larger than the WP21 subset, and the paper should say so.

*Method.* Random-effects or multilevel meta-regression with publication-bias diagnostics appears in Mohammad et al.,
Zhang and Yen, Brander and Koetse (multilevel), Beltrán et al., Guignet et al. and Heberling et al.; the Sirmans
papers and Smith and Huang (1995) use OLS/MAD meta-regression of raw coefficients; most reviews of school quality,
crime, wind turbines, power lines, trees and greenways are vote-counting or structured narrative. The ML/AVM reviews
(Wang and Li 2019: 104 papers; Geerts et al. 2023: 93; Tekouabou et al. 2024: 70; Moreno-Foronda et al. 2025: 23;
Valier 2020; Ja'afar et al. 2021; Naz et al. 2024) compare predictive accuracy and never synthesise marginal effects,
which is exactly the gap RQ5–RQ6 address. Pérez-Rave et al. (2019) was checked and is a methods paper, not a review.