\begin{table}[!ht]
\caption{Netting out the neighbour's own FPM channel: spillover ATTs on own-bracket-fixed neighbours}\label{table:referee_ownfixed}
\centering
\begin{tabular}{lcc}\hline\hline
Neighbour outcome (log, \%) & Full sample & Own-bracket-fixed \\
                            & ATT (s.e.)  & ATT (s.e.) \\ \hline
\emph{Neighbour's own FPM revenue} & 1.08 (0.43)$^{**}$ & 0.42 (0.44) \\[2pt]
Current expenditure         &  1.08 (0.66)        &  0.58 (0.92) \\
Education expenditure       &  3.71 (1.01)$^{***}$ &  4.43 (1.65)$^{***}$ \\
Public-sector employment    &  2.67 (1.29)$^{**}$  &  1.47 (1.63) \\
Private-sector employment   & $-1.05$ (1.53)      & $-2.11$ (2.24) \\
Number of firms             & $-1.56$ (1.00)      & $-0.83$ (1.59) \\
Firms (services)            & $-1.50$ (1.05)      & $-0.76$ (1.70) \\
Government transfers (ex-FPM) & 4.45 (2.98)       &  1.59 (4.47) \\
Population                  &  0.45 (0.31)        &  0.60 (0.33)$^{*}$ \\
\hline
Treated / comparison neighbours & 886 / 968 & 404 / 468 \\
\hline\hline
\end{tabular}\par
\begin{tablenotes}\footnotesize
\item Overall (simple-weighted) Callaway--Sant'Anna ATTs in percent (log coefficients $\times 100$), with neighbour-clustered influence-function standard errors in parentheses. The \emph{Own-bracket-fixed} column restricts both treated and comparison neighbours to those whose own FPM bracket (and hence own FPM coefficient) never changes over 1999--2014, so the neighbour's own-FPM channel is shut off by construction; $404$ of the $886$ treated neighbours satisfy this restriction. The top row is a validation: on this subsample the neighbour's own FPM revenue is flat ($0.42\%$, not significant), confirming the restriction removes the own-coefficient response documented in Table~\ref{table:referee_neighfpm}. The education-expenditure spillover survives --- indeed strengthens --- and remains significant at $1\%$ despite the halved sample, so it is not an artefact of the neighbour's own FPM. The current-expenditure, public-employment and transfer responses attenuate, reflecting both the removal of the own-FPM channel and the loss of power; the private-side estimates remain negative and imprecise. $^{*}\,p<0.10$, $^{**}\,p<0.05$, $^{***}\,p<0.01$.
\end{tablenotes}\end{table}