\begin{table}[!ht]
\caption{Sample construction: from the raw border panel to the estimation sample}\label{table:referee_flow}
\centering
\begin{tabular}{lccc}\hline\hline
Step & Pairs & Neighbours & Focals \\ \hline
Raw focal--neighbour border panel (1999--2014) & 12{,}716 & 3{,}224 & 3{,}224 \\
\quad after panel-length filter ($\ge 7$ years) & 12{,}716 & 3{,}224 & 3{,}224 \\[2pt]
Clean-comparison restriction (kept) & --- & 1{,}854 & 1{,}158 \\
\quad dropped: multi-exposure neighbours & --- & 1{,}370 & --- \\
\hline
\multicolumn{4}{l}{\emph{Estimation sample (neighbour level):}}\\
\quad Treated neighbours (finite $G_i$) & & 886 & \\
\quad Comparison neighbours ($G_i=\infty$) & & 968 & \\
\hline
\multicolumn{4}{l}{\emph{Treated neighbours by cohort $G_i$:}}\\
\multicolumn{4}{l}{\quad 2002: 166 \;\; 2003: 30 \;\; 2004: 21 \;\; 2005: 100 \;\; 2006: 43}\\
\multicolumn{4}{l}{\quad 2007: 16 \;\;\, 2008: 235 \;\, 2009: 80 \;\; 2010: 37 \;\;\, 2011: 158}\\
\hline\hline
\end{tabular}\par
\begin{tablenotes}\footnotesize
\item The unit of the raw panel is the ordered focal--neighbour pair sharing a contiguous border, observed over 1999--2014 ($12{,}716$ pairs $\times\,16$ years $=203{,}456$ pair-years) among the $3{,}224$ municipalities (below $50{,}000$ inhabitants, around the first seven FPM thresholds) that appear as a focal or neighbour. The panel-length filter is non-binding here. The \emph{clean-comparison restriction} keeps neighbours whose focal partners collectively record at most one qualifying upevent and no qualifying downevent; it drops $1{,}370$ multi-exposure neighbours ($958$ with $\ge 2$ qualifying upevents and $689$ with at least one qualifying downevent, categories that overlap). After the restriction the panel is collapsed to one observation per neighbour-year, leaving $886$ treated and $968$ comparison neighbours. The treated cohort $G_i$ is the calendar year of the neighbour's identifying focal partner's qualifying upevent; $63\%$ of treated neighbours fall in the three census-reset cohorts (2002, 2008, 2011).
\end{tablenotes}\end{table}
