Simple estimates work surprisingly well for a lot of things because a lot of the 'unquantifiable' complexity being ignored behaves like noise. When you have dozens of factors pulling in different directions—some developers multitask better, some lose flow more easily, some codebases are more memory-hungry, and so on—it all tends to just average out, and the result is reasonably accurate. Accurate enough that it's useful data to make a decision with, at least.