1\section{Methodological heterogeneity}2\label{sec:methodological}34Table~\ref{tab:metareg} reports the three-level meta-regressions, with cluster-robust standard errors, for the \nMetaRegModels{}5characteristic--family combinations with at least fifteen studies, and Figure~\ref{fig:moderators} plots the moderator effects for6six of them. Of the \nModTerms{} moderator coefficients tested, \nModTermsSig{} are significant at the five percent level, which is7about what chance would produce at that level and is the first result of this section: once estimates are weighted by precision8and the dependence within studies is modelled, the observable features of a study explain little of the between-study variance in9implicit prices. The exceptions are informative. The apartment indicator is significant in \nSigApt{} of \nTestedApt{} models and10raises the living-area semi-elasticity (\mrLivingAreaSemiElasticityApt{} percentage points per square metre, $p=\mrLivingAreaSemiElasticityAptP$),11as expected when the dwelling is smaller and the marginal square metre dearer. The spatial-estimator indicator is significant in12\nSigSpatial{} of \nTestedSpatial{} models, and in both cases it lowers the living-area effect slightly (elasticity13\mrLivingAreaElasticitySpatial{}, $p=\mrLivingAreaElasticitySpatialP$), which is the expected direction when a spatial lag absorbs14part of the size premium that is correlated across neighbours. Time fixed effects matter in \nSigTimeFe{} models; location fixed15effects, the working-paper indicator, the rent indicator and the log sample size are insignificant almost everywhere. The16exploratory machine-learning meta-regressions, trained on all moderators with study-grouped cross-validation, have negative17out-of-study $R^2$ for both random forests and gradient boosting: the moderators do not predict the implicit price of a new18study, and the permutation importances, which rank the effect-size family, the data period and the sample size first, describe19the training data rather than a transferable structure. The answer to RQ5 is therefore that spatial and conventional estimators20produce similar implicit prices for structural characteristics in the pooled evidence, with a small attenuation under spatial21models for living area; machine-learning marginal effects are too few and too heterogeneous in construction to be pooled with22regression coefficients, and are discussed in Section~\ref{sec:ai}.2324\begin{table}[H]25\centering26\caption{Three-level meta-regressions of harmonised effects on study and model characteristics (coefficient, cluster-robust SE;27$^{*}$ $p<0.05$ with Satterthwaite degrees of freedom)}28\label{tab:metareg}29\scriptsize30\resizebox{\textwidth}{!}{\input{tables/tab08a_meta_regression}}\\[1.2em]31\resizebox{\textwidth}{!}{\input{tables/tab08b_meta_regression}}32\end{table}3334\begin{figure}[H]35\centering36\includegraphics[width=\textwidth]{fig11_moderator_effects.pdf}37\caption{Moderator effects with cluster-robust 95 percent confidence intervals for six characteristic--family combinations.}38\label{fig:moderators}39\end{figure}40