\begin{table}[!ht]
\caption{Spatially-robust inference: neighbour- vs microregion-clustered standard errors}\label{table:referee_cluster}
\centering
\begin{tabular}{lccc}\hline\hline
Neighbour outcome (log, \%) & ATT & SE (neighbour) & SE (microregion) \\ \hline
Current expenditure         &  1.08 & (0.66)        & (0.68) \\
Education expenditure       &  3.71 & (1.01)$^{***}$ & (1.08)$^{***}$ \\
Public-sector employment    &  2.67 & (1.29)$^{**}$  & (1.35)$^{**}$ \\
Private-sector employment   & $-1.05$ & (1.53)      & (1.55) \\
Number of firms             & $-1.56$ & (1.00)      & (1.10) \\
Firms (services)            & $-1.50$ & (1.05)      & (1.14) \\
Government transfers (ex-FPM) & 4.45 & (2.98)       & (3.22) \\
Population                  &  0.45 & (0.31)        & (0.33) \\
\hline
Clusters                    &       & 1{,}854 neighbours & 502 microregions \\
\hline\hline
\end{tabular}\par
\begin{tablenotes}\footnotesize
\item The Callaway--Sant'Anna ATTs (in percent; log coefficients $\times100$) are identical across columns; only the clustering of the influence-function standard errors differs --- at the neighbour level (the headline) and at the microregion level. Microregion clustering allows arbitrary correlation among neighbours of the same focal that lie in the same IBGE microregion (the relevant spatial unit, with $502$ clusters here), addressing the concern that a single focal shock reaches several adjacent neighbours. It inflates the standard errors by only $5$--$9\%$ and leaves the significance pattern unchanged: the education-expenditure spillover remains significant at the $1\%$ level and public-sector employment at the $5\%$ level. $^{**}\,p<0.05$, $^{***}\,p<0.01$.
\end{tablenotes}\end{table}