\section{Robustness}
\label{sec:robustness}

Table~\ref{tab:robustness} re-estimates the market-wide
difference-in-differences at $h\in\{5,30\}$ and the network differential at
$h=5$ under eleven designs; Tables~\ref{tab:design_checks}
and~\ref{tab:network_components} report the design diagnostics discussed
above. We describe only checks that were executed; analyses that the
replication package does not contain are listed as such in the Online
Appendix and in the concluding section.

\emph{Tick regime.} Pre-2001 minute returns embed eighth- and sixteenth-tick
bid--ask bounce. Restricting events to 2002 onward leaves the market-wide
estimates unchanged ($-1.03$ and $-3.04\times10^{-4}$) and the network
differential at $65.8\times10^{-4}$. \emph{Time of day.} Restricting events
to 10:00--15:30 ET, which removes the opening and closing periods where
seasonal scaling is most demanding, and separately excluding the 10:00 and
14:00--14:30 ET windows that host most scheduled macroeconomic releases,
changes no estimate by more than one standard error. \emph{Market state.}
Excluding 2008--2009 and 2020 leaves both objects unchanged; within the
crisis years alone the market-wide estimate is imprecise
($2.7\pm5.1\times10^{-4}$) and the network differential is, if anything,
larger ($68\times10^{-4}$). \emph{Source magnitude.} Raising the source
threshold to $|z|\ge8$ and $|z|\ge12$ moves the market-wide point estimates
toward zero or slightly positive, consistent with the co-arrival pattern of
Table~\ref{tab:design_checks}, and lowers the network differential from
$64$ to $50$ and $34\times10^{-4}$. The decline with magnitude is the one
systematic pattern in the table; it is consistent with the largest moves
being firm-specific news that peers correctly do not price, and also with
the largest moves being the ones most often accompanied by halts or
erroneous prints that bars cannot flag. \emph{Universe composition.}
Excluding source securities flagged as likely non-common by the
ticker-shape heuristic leaves every estimate within rounding of the
baseline; ETFs cannot be identified and are not excluded. \emph{Inference.}
Month-cluster resampling raises the standard error of the network
differential from $1.8$ to $3.4\times10^{-4}$ and ticker-cluster resampling
to \NetDiffFiveTickerSE$\times10^{-4}$, leaving a $t$-statistic above
fourteen in every case; the market-wide standard errors are insensitive to
the clustering unit. \emph{Placebo network.} The neighbor-minus-other
contrast computed on placebo events alone is $-2.7\pm2.8\times10^{-4}$ over
the full sample. As Table~\ref{tab:network_components} shows, this average
conceals decade-level offsets of opposite signs, which is why every network
statistic in the paper is a real-minus-placebo difference and never a
real-arm contrast on its own.

\input{tables/table8_robustness}

Several checks that a complete treatment would include were not executed
and are not claimed. The pipeline stores events at the $|z|\ge6$ threshold
and the receiver threshold $|z|>2$ only, so lower source thresholds
($|z|\ge4,5$), a higher receiver threshold ($|z|>3$), and the top-1000 and
top-3000 universes would require re-running the detection stage on the raw
panel. Neighbor sets built on 5- or 30-minute residual correlations, or on a
transaction-time covariance estimator, and the pre-event ($t-5$) network
contrast, are the two extensions most needed to settle the interpretation of
the secular increase in the network differential. Permuted-network placebos
that preserve industry and liquidity composition, a separate estimate of
propagation inside announcement windows, and a Hawkes-type model of the
co-jump minute series are natural further steps.