538 offers an insightful comment on biases in polling that has drawn attention especially recently due to the oddness of the LAT/USC polls.
The bottom line is that, as long as the polls are methodologically sound, they all offer useful insights, and in absence of knowing what “the truth” is, we don’t know just how biased they really are. The big table of the in-house biases of many different polls the web page offers is interesting but at the same time misleading: all it means is that they are biased vs. the “average” which itself is likely to be “wrong” anyways.
I suspect that the “how did you vote in 2012” question is not introducing a bias due to voters “misremembering”: if voters say that they voted for Romney in 2012, even if they didn’t, that indicates a useful information about their current proclivities. The real problem might be that there are many voters (especially among the young who are very hostile to Trump–in some polls, Trump is in low 20s among the voters in their 20s) who will be systematically excluded from the sample for this reason.
Still, it will be interesting to see how different sub demographics vary in terms of their biases. I don’t care much for the aggregate data, which reflects a lot of reweighing of the data that reflects art more than science. I am interested in how different subdemographics change over time–which reveals, among others, that Trump’s problem is especially severe with the regular Republican voters who don’t like him much. So who does USC poll have Trump doing better with than other polls? Where, under the hood, does the Trump bias in their polls come from?