OK, here's one fairly explicit comparison to the polling world. It's usually going to be better to do a meta-analysis of all studies, including thinking about biases that could systematically affect all of them, than to spend a huge amount of energy picking apart any one study.
There are at least ~a dozen studies already that attempt to measure seroprevalence (how many people were infected) for COVID-19 in various parts of the world. Will likely be dozens more in a few weeks. I'm seeing analysis of e.g. the Santa Clara Co study that take it in a vacuum.
Taken as a whole, those studies seem to suggest it's possible—but hardly certain—that the ratio of total infections to detected cases is higher than the ~10x ratio commonly cited. But the thing is there's not just going to be one universal ratio; it will vary from place to place.
Testing capacity probably varies by something like 10x across different major Western countries. It also probably varies by something like 3x-5x across different US states. So you could easily have some places that are detecting 1.5% of their infections, and some that detect 15%.
