This working paper by Andy Gelman, Doug Rivers, ad others is a truly excellent paper, even if their title is, perhaps, a bit misleading. The basic premise corresponds closely to my take on how to approach polling data: vast majority of people already hold fairly well-defined views on politics and their choices are unlikely to change much. Seemingly big swings reported in short term polling results are rarely the product of people changing their minds as much as peculiar selection biases that crop up at different times in polls. This applies especially for the “partisan voters”: people who say they are partisans, at least in today’s political environment, do not change their views so readily. So it is worthwhile to post-stratify the polling data on the basis of the partisan characteristics of the respondents, and more generally, pay close attention to how the winds are shifting among those who identify themselves as overt partisans.
I did not generally take a more technically sophisticated approach to crunching poll numbers than looking at variability within each politico-socio-economic demographic across multiple cross-tabs, for which I guess I should be rather ashamed. Still, the point that kept coming up was unmistakable: hardly anything was changing within each demographic for most part: for example, Trump enjoyed considerable and unwavering support among working class white Republican men, somewhat more variable support among the white Republican working class women, and quite variable support among the affluent white Republican men, and wildly changing support levels among the affluent white Republican women. Potentially similar phenomenon exists on the Democratic side: while the “average” Sanders voter is probably enough of a liberal that there is no chance of his/her swinging to the Republican side, with or without a Trump, but a sizable minority might. We have only an incomplete picture of the latter group, due to lack of access to the detailed polling data, but, in some states, e.g. West Virginia, they were unusually heavily concentrated, providing the glimpse of how misleading it is to characterize the entire lot by their average.
Swing voters are, of course, the voters who exhibit a very high level of intertemporal variance in a given election season. Their relative scarcity means that a lot of campaign resources are wasted on essentially deaf ears, but their existence, if only in small numbers, does provide an opportunity for campaigns–especially if their conditional variances are better understood.
There is no reason to believe that the swing voters should necessarily stay constant from election to election: in today’s setting, the potential swing voters among the Republicans seem to be the affluent white female voters–who normally would have been very low variance Republican voters. The so-called Rendell strategy, however much it might be derided among the liberals, is founded on a solid premise, for better or worse, given the way current election is shaping up. But one might say the same about the potential plot being hatched up by the Trump campaign, aided and abetted by the Clinton camp, to capture whatever working class white voters that are left that might still lean Democratic (and voted for Sanders in the primaries.)