Why the USC/LA Times tracking poll is a crock.

You may know that it doesn't use a random sample; it uses an online panel.

But the problems go even beyond that. For example, it weights the sample according to its respondents report of how they voted in 2012 -- a number that has often been shown to be inaccurate in the past. (More people always recall having voted for the winner.)

http://www.latimes.com/politics/la-na-pol-polling-differences-20160809-snap-story.html

http://www.nytimes.com/2016/08/09/upshot/a-favorable-poll-for-donald-trump-has-a-major-problem.html?rref=collection%2Fsectioncollection%2Fupshot&action=click&contentCollection=upshot®ion=rank&module=package&version=highlights&contentPlacement=4&pgtype=sectionfront

One factor that could be contributing to the panel’s tilt toward Mr. Trump is its decision to weight its sample according to how people say they voted in 2012.

The pollsters ask respondents whether they voted for President Obama or Mitt Romney. They then weight the sample so that Obama voters represent 27 percent of the panel and Romney voters represent 25 percent, reflecting the split of 51 percent to 47 percent between the two among actual voters in 2012. (The rest include newly eligible voters and those who stayed home.)

SNIP

With these figures in mind, the U.S.C./LAT poll’s decision to weight its sample to 27 percent for Mr. Obama and 25 percent for Mr. Romney is quite risky. If the panelists, like those in other surveys, are likelier to recall voting for the winner (Mr. Obama), then the poll is unintentionally giving extra weight to Republican voters. Or you can imagine a counterfactual: If the poll were weighted to 33 percent for Obama and 25 percent for Romney (per the NYT/CBS numbers), then Mrs. Clinton would hold a more comfortable lead.

SNIP

There are many other things that could be causing the difference between the U.S.C./LAT results and other surveys — like its unusual probabilistic measure of vote support (voters indicate how likely they are to vote for a candidate on a scale from 0 to 100) or the difficulties of recruiting and maintaining a panel.