The Confound · Contagion or Homophily?

TWO PROCESSES · ONE NETWORK · THE SAME STATISTIC

Both panels use the same 167 people and the same 278 ties, and each draw places the same 24 gunshot victims. On the left, risk passes along ties. On the right, nothing passes between anyone. Try to tell which is which before pressing reveal.

CONTAGION · INDEPENDENT CASES60%
HOMOPHILY · TRAIT CLUSTERING5
PROCESS A
Risk transmits along ties, with some cases arriving independently from outside the network.
VICTIMS WITH A VICTIM NEIGHBOUR—
MEAN DISTANCE TO A VICTIM—
ODDS RATIO PER TIE—
PROCESS B
Nothing transmits. Risk follows a latent trait that is itself clustered across ties.
VICTIMS WITH A VICTIM NEIGHBOUR—
MEAN DISTANCE TO A VICTIM—
ODDS RATIO PER TIE—
ESTIMATED ODDS RATIO PER TIE OF DISTANCE · 0 DRAWS EACH0.20.30.50.7541.01.52no associationBoston estimateCONTAGIONHOMOPHILY

Press new draw to begin. Each draw refits the same logistic regression that network-exposure studies report.

CONTAGION
HOMOPHILY
THE CONFOUND
WHAT WOULD SETTLE IT
ABOUT THIS SIM

None of this says the concentration finding is wrong. Gunshot injury really does pool in small parts of a network, and that is useful whatever produces it. What the two panels show is that the pooling by itself does not license the causal reading. Transmission and a shared, clustered cause leave the same footprint in cross-sectional data, so anything that turns on which one is at work needs evidence the footprint cannot supply.

Simulated data on a synthetic network, built as a companion to Five Handshakes. Victim prevalence is set above the empirical rate so a single draw is legible.

References: Shalizi & Thomas (2011) · McPherson, Smith-Lovin & Cook (2001) · Papachristos, Braga & Hureau (2012) · Papachristos & Wildeman (2014) · Green, Horel & Papachristos (2017) · Papachristos (2009)