1# WP21 — Methodological decisions log23| Date | Decision | Reason | Consequence |4|---|---|---|---|5| 2026-10-20 | OpenAlex demoted from primary to verification source | Without an API key OpenAlex now caps requests to a shared daily IP budget (HTTP 401 after ~200 records on 2026-10-20); no key could be obtained autonomously | Discovery relies on Semantic Scholar bulk search (abstracts, OA links, citations), Crossref (coverage, no abstracts for most publishers) and arXiv |6| 2026-10-20 | Semantic Scholar queried unauthenticated with patient backoff (15 s → 300 s) and 3 s between pages | Shared rate limit returns HTTP 429 frequently | Slow but complete; every query and its hit count logged in `data/metadata/search_log.csv` |7| 2026-10-20 | Crossref relevance search capped at 4,000 records per query | Totals are in the hundreds of thousands because `query.bibliographic` is fuzzy; beyond the first few thousand the relevance is negligible | Recorded as a coverage limitation; Crossref records mostly lack abstracts so stage-1 screening on them is title-based |8| 2026-10-20 | SSRN, Google Scholar, IDEAS/EconPapers not harvested | No compliant bulk interface; SSRN blocks automated access | Grey literature partly covered through Semantic Scholar (which indexes SSRN/RePEc metadata) and arXiv; limitation reported |9| 2026-10-20 | Stage-1 screening = rule-based keyword scoring on title + abstract, then LLM-assisted screening of candidates | Volume (≈40–60k unique records) | Rules are in `scripts/screening/keywords.yaml`; every decision and reason stored per record |10| 2026-10-20 | Full texts restricted to legal open-access copies (Unpaywall best OA location, Semantic Scholar OA PDF, arXiv) | No institutional proxy; no scraping of paywalled content | Expected retrieval rate well below 50 %; the eligible-but-not-retrieved set is reported in the PRISMA flow and compared on metadata to the retrieved set |11| 2026-10-20 | LLM-assisted extraction by independent agents working from page dumps of candidate tables (PyMuPDF), with a JSON schema and a written protocol; 15 % double extraction | GROBID not available; regex table parsing unreliable for regression tables | Only `high`/`medium` confidence rows enter the primary synthesis; `low` rows and disagreements are excluded until checked |12| 2026-10-20 | Kennedy-corrected percentage premiums as the primary binary-attribute family, uncorrected as sensitivity | Protocol §3 | Both stored in the effects table |13| 2026-10-20 | Partial correlations computed with df = n_obs − k_params when k_params is reported, else n_obs − 10 (flagged) | Many papers do not report the number of regressors | Sensitivity with df = n_obs − 2 reported |14| 2026-10-20 | Bayesian model fitted to study-level aggregated effects (inverse-variance within study) | Independence of the hierarchical likelihood; three-level dependence handled in the frequentist `rma.mv` instead | Reported as a cross-check, not the primary estimate |15| 2026-10-20 | Hedonic price-index papers estimated on micro data are included at stage 1 (coefficients usually reported); pure repeat-sales / aggregate index methods excluded (E2) | Screeners diverged on this class | Applied in round 2; round-1 "unclear" index papers re-coded consistently |16| 2026-10-20 | Aggregate units of analysis (city, tract, district, project means) are excluded from the coefficient synthesis at full text; community/estate-level Chinese studies flagged for human review | Protocol I1 requires dwelling-level prices | Extractors mark such papers ineligible with the reason; the unclear estate-level cases are kept with a flag |17| 2026-10-20 | Airbnb / short-term nightly rates excluded (E8); long-run effects of STR on sale prices or rents included | Not residential housing prices in the protocol sense | — |18| 2026-10-20 | ML papers reporting SHAP/PDP marginal effects coded ELIGIBLE_HEDONIC at stage 1 but their marginal effects are NOT pooled with regression coefficients; they enter the AI synthesis only | SHAP values are not implicit prices (CLAUDE.md §9) | — |19| 2026-10-20 | GWR/MGWR local-coefficient summaries (mean/median) extracted with se = null and confidence medium; excluded from the pooled synthesis (no SE) but counted in the descriptive synthesis | Local coefficient distributions have no single SE | — |20| 2026-10-20 | Standard errors printed as 0.000 (rounding) are stored as null, not zero; p-only and stars-only tables keep p / stars; SEs recovered only from t, CI or exact p (formula stored) | Protocol §5 | Stars-only estimates are unusable in the inverse-variance synthesis |21