# WP21 — Limitations (final audit, 2026-10-20)

1. **Retrieval bound.** Only 18.8 % of the 2,667 stage-one hedonic/unclear studies could be obtained as legal open-access full texts;
   the extracted sample (297 assessed, 242 eligible) over-represents recent, open-access and non-US work. The paywalled 1980s–2000s
   North American journal literature is under-represented.
2. **Discovery bound.** OpenAlex was budget-capped (no API key); SSRN, Google Scholar, IDEAS/EconPapers, CORE, DOAJ and repositories
   were not harvested; multilingual queries returned mostly English records.
3. **Agent-based screening and extraction.** Independent LLM agents applied written protocols; validation on 20 papers shows 95 %
   eligibility agreement and 100 % numerical agreement on 647 matched coefficients, but 78–79 % recall/precision on which auxiliary
   rows were recorded; the two-human-reviewer standard was not met.
4. **Harmonisation choices.** Construction-year variables, standardised betas, interactions, squared terms and un-harmonisable
   units are excluded from pooling (about a quarter of rows); Kennedy correction assumes normality; semi-elasticities per m² pool
   houses and apartments.
5. **Heterogeneity.** I² ≈ 100 % in almost every family; prediction intervals span zero for most characteristics; moderators explain
   little; ML meta-regressions have negative out-of-study R².
6. **Publication selection.** Egger asymmetry in 46/67 families is not interpretable as selection alone under this heterogeneity and
   dependence; PET-PEESE and selection models reported as sensitivity only; some selection models produce implausible values.
7. **Temporal/crisis questions (RQ4).** Pre-2000 evidence thin; no powered pre/post contrast for 2008 or COVID-19.
8. **AI synthesis (RQ6).** 40 retrievable studies; metrics not comparable across studies; SHAP effects not pooled with coefficients.
9. **Sirmans comparison.** The 2005/2006 full texts were not accessible; the comparison uses abstract-level findings only, and the
   "125 studies" figure could not be verified.
10. **Reproducibility.** Raw PDFs not redistributed; agent outputs are stored artefacts, but re-running the agents is not
    deterministic in the strict sense; all downstream computations are deterministic (seed 20261020).