Another election, another massive screw up by the “best” pollsters in the business. Yes, excuses will be made about how margins of error are only 1/20 and while rare 1/20 events do happen occasionally (it’s roughly like getting two pair in a five card poker hand on your initial deal), they really don’t happen twice in a row.
So any attempts to ascribe this to “margins of error” is complete BS. Clearly these polls are not (and didn’t in 2016) use random samples. The magic didn’t happen because the groundwork wasn’t there. No random samples, no magic.
Here’s another way of thinking about this. If they did use random samples in each election, you could also safely assume the polling result errors in each election are independent events. Which means by the “multiplication of probabilities” result I have talked about before, the odds of two screwups in a row would be >= 1/400 and that just isn’t believable.
So something clearly is wrong in their methodology: they aren’t getting random samples. You can come up with conjectures as to why. For example, a common method to get random samples in polls is to do random dialing of phone numbers. If a certain group is always more likely to hang up on you than another group, randomness ain’t happening. You can attempt to weigh your results to take this into effect, and pollsters certainly claimed they would do that after the 2016 debacle, but whatever they added to their secret sauce in 2020 just made the dish taste worse, not better – that is for sure.
My conclusion after this year’s debacle: absent some methodological innovation, political polling isn’t worth paying much attention to going forward and so neither is the effort put into it – by a lot of really smart people.