\begin{table}[!ht]
\caption{Baseline balance, neighbour level (pre-event, one observation per neighbour)}\label{table:referee_balance}
\centering
\begin{tabular}{lccccc}\hline\hline
 & \multicolumn{2}{c}{\textbf{Comparison}} & \multicolumn{2}{c}{\textbf{Treatment}} & \textbf{Norm.} \\
\cline{2-3} \cline{4-5}
Variable & Mean & SD & Mean & SD & diff. \\ \hline
Population &  16952.64 &  10147.12 &  17384.96 &   9412.24 &  0.044 \\
Public-sector employment &    501.54 &    297.13 &    517.75 &    291.37 &  0.055 \\
Private-sector employment &   1316.65 &   1910.33 &   1486.32 &   1920.38 &  0.089 \\
Number of firms &    180.19 &    218.21 &    192.24 &    197.76 &  0.058 \\
Pop. growth (\%) &      1.01 &      1.69 &      0.98 &      2.91 & -0.011 \\
\hline Neighbours & \multicolumn{2}{c}{968} & \multicolumn{2}{c}{886} & \\ \hline\hline
\end{tabular}\par
\begin{tablenotes}\footnotesize
\item One observation per neighbour. Treatment neighbours are averaged over their pre-event window (event time $-4$ to $-2$, excluding the $-1$ anticipation year and all post-event years); comparison neighbours are averaged over their never-treated history. Each variable is purged of common calendar-year effects (year-demeaned and re-centred on its grand mean) before averaging, so the comparison is not contaminated by the different calendar composition of the two groups. The last column is the Imbens--Rubin normalised difference $(\bar x_T-\bar x_C)/\sqrt{(s_T^2+s_C^2)/2}$; $^{\dagger}$ marks $|\text{nd}|>0.25$. Any residual level gaps are differenced out by the within estimator; the sharper test of the identifying assumption is the absence of differential pre-trends, evident in the flat pre-event leads of the event studies.
\end{tablenotes}\end{table}