There should be some way to formalize this into a hypothesis. Testable with the same data I mean. If you were correct we should see more clustering then we do with a random sample.
You'd get weird mixture model type systems. There are certain subclasses: school-based vs non-school-based. You'd have to control for this since school-based shootings are highly temporally correlated. But then maybe there's crossover too. Like everything, estimating these things with more complex models is hard to do right, and findings are usually overstated.
