The link above is to a well-done New York Times article analyzing the results highlighted on the website for the USC Daybreak Poll that has made things look so much more favorable for Donald Trump than other polls.
Let me emphasize that the underlying data for the Daybreak poll are extremely valuable. Having a panel makes it possible to answer many questions that cannot be answered well with a repeated cross-section. The problem is with the calculation for the highlighted comparison between Donald Trump and Hillary Clinton support.
The most important problem with the graph highlighted on the Daybreak Poll website is the weighting by the candidate a poll respondent claimed to have voted for in the last election. Nate Cohn is good at talking about the biases that introduces because people underreport voting for the loser. Thus, forcing the weights to make the reports of who people voted for equal to the actual shifts the weights too much toward the sort of people who might have actually voted for the loser. Many self-reported “Obama” or minor-candidate voters were really Romney voters. People who admitted voting for Romney are more Republican than the overall set of people who actually voted for Romney. So inflating the weights of people who reported voting for loser Romney up to equal the fraction of those who actually voted for Romney makes things look more favorable for Trump than they should be.
To me, the main way the data on voting in the last election should be used is in correcting for each demographic group the difference between the percent chance they said they would vote and whether they actually voted or not. It is not clear that this needs to use the self-reported voting after the fact at all; exit polls should provide good evidence on actual voting percentages by demographic group that can be compared to the probabilities people said in advance in each demographic group in this kind of data collection in 2012 (which I know was done on RAND’s American Life Panel in 2012).