spb/wp4_uqo

1% ==========================================================================2\section{Robustness Checks}3\label{sec:robustness}4% ==========================================================================56This section evaluates the sensitivity of the main hedonic estimates to7alternative sample definitions, variable transformations, model8specifications, and subsample restrictions.910% ------------------------------------------------------------------11\subsection{Outlier Treatment}12\label{subsec:winsorize}13% ------------------------------------------------------------------1415We assess sensitivity to alternative outlier treatments through five16specifications: (1)~raw OLS, (2)~Winsorized dependent variable at the171st/99th percentiles, (3)~trimmed sample dropping observations outside18the 1st--99th range, (4)~semi-log specification, and (5)~median19regression (LAD). Across all five, the sign, significance, and20approximate magnitude of the key hedonic coefficients remain stable.21The adjusted $R^2$ ranges from 0.083 to 0.098, confirming that the22results are not driven by extreme prices.2324% ------------------------------------------------------------------25\subsection{Subsample Analysis by Property Type}26\label{subsec:room_type}27% ------------------------------------------------------------------2829We re-estimate the full model separately for each property type.30Table~\ref{tab:robustness_type} summarizes the results.3132\begin{table}[!htbp]33\centering34\caption{Hedonic Model by Property Type}35\label{tab:robustness_type}36\begin{threeparttable}37\begin{tabular}{lrrr}38\toprule39Property Type & $N$ & $R^2$ & Adj.\ $R^2$ \\40\midrule41Apartment & 77 & 0.186 & 0.082 \\42Cabin/Chalet & 691 & 0.084 & 0.068 \\43Condo & 237 & 0.104 & 0.065 \\44House & 679 & 0.098 & 0.083 \\45Other & 554 & 0.130 & 0.112 \\46Rental unit & 1,174 & 0.095 & 0.085 \\47\bottomrule48\end{tabular}49\begin{tablenotes}50\footnotesize51\item \textit{Notes:} OLS estimates with HC1 robust standard errors.52Each row reports the model estimated on the corresponding property type53subsample. Zero-variance regressors are dropped within each subsample.54\end{tablenotes}55\end{threeparttable}56\end{table}5758The Apartment subsample achieves the highest $R^2$ (0.186) despite the59smallest sample size. Cabin/Chalet and House show lower $R^2$ values,60consistent with greater heterogeneity depending on unobserved61experiential characteristics.6263% ------------------------------------------------------------------64\subsection{Price Segment Heterogeneity}65\label{subsec:price_segment}66% ------------------------------------------------------------------6768We partition the sample into five price segments (Budget, Moderate,69Premium, Luxury, Ultra-Luxury) and re-estimate the full model within70each. In the Budget segment, structural variables carry larger71standardized coefficients. In the Luxury and Ultra-Luxury segments,72amenity indices and reputation variables gain relative importance,73consistent with the quantile regression findings.7475% ------------------------------------------------------------------76\subsection{Alternative Location Controls}77\label{subsec:alt_location}78% ------------------------------------------------------------------7980We test five progressively richer location specifications: (1)~no81controls, (2)~city/neighbourhood fixed effects (top 30 by count),82(3)~latitude and longitude as continuous controls, (4)~spatial cluster83fixed effects ($K = 20$), and (5)~spatial lag (5-NN mean price). City84fixed effects raise the adjusted $R^2$ from 0.094 to 0.164. Under all85specifications, the non-locational hedonic coefficients remain stable86in sign and significance.8788% ------------------------------------------------------------------89\subsection{Host Type Stratification}90\label{subsec:professional}91% ------------------------------------------------------------------9293We stratify by host type: all hosts (baseline), Superhosts only94($R^2 = 0.077$), and multi-listing hosts. The lower $R^2$ for95Superhosts suggests that, conditional on meeting the quality threshold,96other observables explain less of the remaining price variation.9798\medskip99In summary, the main findings---the dominant role of bedrooms and100structural capacity, the significant review score premium, the concave101bedroom--price relationship, competitive density effects, and positive102local price contagion---are robust across alternative sample definitions,103outlier treatments, property type subsamples, price segments, location104controls, and host type stratifications.105