Why the worst places always get better
This is the city from the film. Every junction has a fixed level of danger; crashes each year are that danger plus luck. Pick the worst junctions after year one, give them a camera, and see what happens in year two. Start with a camera that does nothing at all.
The rules, all of them. Five thousand junctions. Each has its own long-run average number of injury crashes a year, which never changes (most are fairly safe, a few are bad; two a year on average). Each year's count is random around that average. After year one every junction at or above the threshold is picked. A camera cuts a junction's danger by the percentage you set; at 0% it does nothing.
The fall you see at the picked junctions with a useless camera is regression to the mean: they were picked partly for being unlucky, and luck does not repeat. The page shows the exact expected fall beside the measured one. In a coin-toss trial the two halves had the same luck, so the difference between them is the camera and nothing else. What it leaves out: traffic moving to other roads, crashes that go unrecorded, and danger that changes over time.