\section{Global evidence on housing characteristics}
\label{sec:evidence}

\subsection{The evidence base}

The coefficient synthesis rests on \nStudiesExtracted{} eligible studies from \nCountriesExtracted{} countries,
\nModelsExtracted{} model columns and \nCoefExtracted{} coefficient rows, of which \nUsableEstimates{} harmonised estimates from
\nUsableStudies{} studies enter the pooled analysis (Table~\ref{tab:studychar}). Europe (\nStudiesEurope{} studies), Asia
(\nStudiesAsia{}) and North America (\nStudiesNorthAmerica{}) dominate; South America (\nStudiesSouthAmerica{}), Africa
(\nStudiesAfrica{}) and Oceania (\nStudiesOceania{}) are thinly covered (Figure~\ref{fig:map}, Table~\ref{tab:geography}). The
United States accounts for \shareUSextracted{} percent of extracted studies, against the near-totality of the Sirmans programme.
Condominiums and apartments (\nStudiesCondo{}) are as numerous as single-family houses (\nStudiesSingleFamily{}), a direct
consequence of the growth of the Asian and European listing-data literature; \nStudiesTransaction{} studies use transaction
prices, \nStudiesListing{} listing prices, \nStudiesRent{} rents and \nStudiesAppraisal{} appraised values. Ordinary least squares
provides \shareOlsEst{} percent of usable estimates, spatial estimators \shareSpatialEst{} percent and quantile regressions
\shareQuantileEst{} percent. The median study has \medianNobs{} observations. Figure~\ref{fig:heatmap} and
Table~\ref{tab:most_studied} show which characteristics are studied where: living area, bedrooms, bathrooms, age and distance to
the centre are reported almost everywhere, while most interior, market and AI-derived characteristics appear in fewer than ten
studies (Table~\ref{tab:gaps}).

\begin{table}[H]
\centering
\caption{Characteristics of the extracted studies}
\label{tab:studychar}
\small
\input{tables/tab03_study_characteristics}
\end{table}

\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{fig04_variable_frequency_heatmap.pdf}
\caption{Variable-frequency heatmap: number of extracted studies reporting each characteristic, by continent (thirty most
frequent characteristics).}
\label{fig:heatmap}
\end{figure}

\begin{table}[H]
\centering
\caption{Most frequently studied characteristics (extracted studies; share of positive printed coefficients is a sign count
across all models and is not a pooled estimate)}
\label{tab:most_studied}
\footnotesize
\input{tables/tab05_most_studied}
\end{table}

\subsection{Pooled implicit prices}

Tables~\ref{tab:premiums} to~\ref{tab:elasticities} and Figures~\ref{fig:forest_premium} to~\ref{fig:forest_elasticity} report
the pooled estimates for the \nPooledFamilies{} characteristic--family combinations with at least five independent studies. Three
regularities organise the results. First, the structural core of the hedonic model is remarkably stable across the world: the
elasticity of price with respect to living area is \estLivingAreaElasticity{} (95 percent confidence interval
\loLivingAreaElasticity{} to \hiLivingAreaElasticity{}, $k=\kLivingAreaElasticity$ studies), each additional bathroom adds
\estBathroomsSemiElasticity{} percent and each bedroom, conditional on area, \estBedroomsSemiElasticity{} percent, a year of age
subtracts \estAgeSemiElasticity{} percent, and a kilometre of distance from the centre subtracts \estCbdDistanceSemiElasticity{}
percent, with the lot-size elasticity at \estLotSizeElasticity{} once living area is controlled. Second, the signature amenities
carry premiums of the order documented in the attribute-specific literature: a garage \estGaragePctPremium{} percent, a waterfront
location \estWaterfrontPctPremium{} percent ($k=\kWaterfrontPctPremium$), air conditioning, a view, a basement, a terrace or balcony
and a swimming pool between four and eleven percent, and proximity to hazardous or industrial sites a discount of about six
percent (Table~\ref{tab:premiums}). Third, every pooled family is heterogeneous beyond sampling error: the median $I^2$ is
\medianIsq{} percent and the 95 percent prediction intervals, which describe the implicit price a new study would be expected to
find, span zero for most characteristics even where the confidence interval of the mean excludes it. The mean implicit price of a
characteristic is therefore estimated with precision, but it is the centre of a wide distribution, not a parameter that transfers
across markets. The three-level estimates, which weight studies rather than estimates, and the cluster-robust standard errors,
which are typically two to four times the naive random-effects standard errors, are reported alongside.

\begin{table}[H]
\centering
\caption{Pooled percentage premiums of binary attributes (log-price models; Kennedy-corrected; REML random effects, three-level
estimate and cluster-robust SE)}
\label{tab:premiums}
\scriptsize
\input{tables/tab06_pooled_premiums}
\end{table}

\begin{table}[H]
\centering
\caption{Pooled semi-elasticities of continuous attributes (percent change in price per harmonised unit: m$^2$ for areas, km for
distances, years for age, counts for rooms, dB for noise, $\mu$g/m$^3$ for pollution)}
\label{tab:semielasticities}
\scriptsize
\input{tables/tab07_pooled_semi_elasticities}
\end{table}

\begin{table}[H]
\centering
\caption{Pooled elasticities (log-log models)}
\label{tab:elasticities}
\scriptsize
\input{tables/tab07b_pooled_elasticities}
\end{table}

\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth,height=0.88\textheight,keepaspectratio]{fig05_forest_pct_premium.pdf}
\caption{Forest plot of pooled percentage premiums of binary attributes. Circles: REML random-effects estimates with 95 percent
confidence intervals; squares: three-level estimates; shaded bars: 95 percent prediction intervals.}
\label{fig:forest_premium}
\end{figure}

\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth,height=0.88\textheight,keepaspectratio]{fig06_forest_semi_elasticity.pdf}
\caption{Forest plot of pooled semi-elasticities (percent per harmonised unit).}
\label{fig:forest_semi}
\end{figure}

\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{fig07_forest_elasticity.pdf}
\caption{Forest plot of pooled elasticities from log-log models.}
\label{fig:forest_elasticity}
\end{figure}

\subsection{Comparison with the Sirmans programme}

Table~\ref{tab:sirmans} places the pooled estimates for the nine characteristics of \citet{sirmans2006value} next to the
qualitative findings that could be verified from that paper's abstract and references. The signs agree: living area, lot size,
bathrooms, garage, pool, fireplace and air conditioning are positive and age is negative; bedrooms, which Sirmans et al.\ found
of mixed sign, are positive on average here but with a prediction interval that includes large negative values, which is the
expected consequence of conditioning on living area in most specifications. What the present synthesis adds is the magnitude
with its uncertainty, the distinction between semi-elasticities and elasticities, and the global prediction interval; what it
cannot do is reproduce the Sirmans meta-regressions on raw coefficients, which are not comparable across functional forms.

\begin{table}[H]
\centering
\caption{The nine Sirmans et al.\ (2006) characteristics: pooled estimates of this synthesis and the qualitative findings of the
earlier meta-analysis (abstract-level; the full text was not accessible)}
\label{tab:sirmans}
\scriptsize
\resizebox{\textwidth}{!}{\input{tables/tab14_sirmans_comparison}}
\end{table}

\subsection{Summary}

Figure~\ref{fig:summary} assembles the pooled percentage premiums and semi-elasticities in one view, coloured by taxonomy group.
Structural characteristics and signature exterior amenities have precisely estimated positive means; environmental disamenities
(hazardous sites, highways, noise, pollution) have negative means that are less precisely estimated and rest on fewer studies;
locational distances have negative semi-elasticities whose magnitudes depend heavily on the market's spatial structure. The
prediction intervals are the message for applied work: an appraiser or a policy analyst transferring an implicit price from the
literature to a particular market should expect the local value to lie anywhere within these intervals unless local evidence
narrows them.

\begin{figure}[H]
\centering
\includegraphics[width=\textwidth,height=0.88\textheight,keepaspectratio]{fig14_summary_evidence.pdf}
\caption{Summary of the pooled evidence: REML estimates with 95 percent confidence intervals for all pooled percentage premiums
and semi-elasticities, coloured by taxonomy group.}
\label{fig:summary}
\end{figure}