I think believe the shift (from mid August to early September 2016) is mostly due to which polls have dropped (there are clear house effects) and sampling error.
The reason is that it doesn't appear at all in the USA LA-Times panel poll. http://www.latimes.com/politics/
They poll the same people again and again. They weighted so that self reported 2012 voting corresponds to actual 2012 votes. This means that they underweight people who claim they voted for Obama (systematically more claim to have voted for the winner than actually voted for the winner). This makes the level of the poll uninteresting.
But changes show how actual people have changed their minds, and are interesting