In one line
When too few installs sit behind a postback, the returned values could re-identify someone. So SKAN gives back less detail at low volume, and the bar it uses is called crowd anonymity.
What disappears
Two things. The precise conversion value (fine, 0–63) and the granular source identifier. What remains is the three-level coarse value and the coarsest campaign identifier.
The paradox specific to SKAN
Analysis usually improves as you slice data finer. Here it inverts: splitting campaigns shrinks each slice's volume, drops its tier, and coarsens what you receive.
Port an Android campaign structure straight to iOS and you can flatten your iOS data wholesale. On iOS, deliberately consolidating campaigns to build volume often buys more measurement than the structure costs.
Go deeper
How the threshold interacts with each measurement window is covered in the iOS ATT and SKAN measurement guide.