Let's jump off from a recent Pew Research Center report looking at why their polling failed to adequately predict the size of the Republican margin in contested corners of the 2014 elections. Their conclusions say something about how it all works.
Those of us who follow polling sometimes suspect that the pollsters aren't getting a sample that is genuinely representative of people who vote. Perhaps they can't find enough people willing to answer their questions? Or they don't call cell phones? Or they canvass too few people of color so views in those communities are not represented?
Pew does not attribute polling miscues to these failings. They contend their samples are adequately mathematically adjusted to overcome these sources of bias. Instead, they conclude that the questions they have been asking are not very good at determining how likely respondents are to vote. Most of us like to think of ourselves as good citizens, so saying we'll vote reinforces our self-satisfaction. We tend to say we will vote whether we actually do or not.
Pew went back and checked their respondents against the publicly available voter file. Who we vote for is secret, but whether we voted is a matter of public record. And what Pew discovered about the 2014 results told them why they underestimated the Republican vote:
As an campaign junkie, I find it easy to respond "Well, duh ..." Doesn't everyone know Democrats are less likely to vote, especially in off-year elections, than Republicans? The difference in turnout also correlates with age, income, and minority status quite closely, as does party identification.
In the past, polls have built their samples from randomized telephone numbers and considered whether their respondents will actually vote as a secondary question. Pew now questions whether these queries get realistic answers and is considering whether there might be other ways to predict who will actually vote.
Campaigns approach building lists of voters to contact from the opposite direction. They look at the history of all registered voters. Voting is a habitual activity. They can assume that most people who have voted regularly in similar elections in the past will vote in the next one. Then there are people who have never yet voted or have voted occasionally -- these can be desirable targets for efforts to get them to vote. Perhaps they can be persuaded to participate ...
Of course campaigns are not interested in voting for citizenship's sake. They want to turn out supporters to vote for their candidate. In our current deeply polarized context, party registration -- mostly Democratic or Republican -- is strongly predictive of how a registered voter will cast her ballot. So if a state records a party preference (31 do at present, but varying numbers of voters "decline to state"), that is a major factor in choosing who to target. Publicly available birthdates (in some registration systems) reveal voters' ages. And data files created and massaged by the political parties do a better job every year of attaching information on the race, religious affiliations, and even some attitudes of individual voters.
Given the improving quality of available voter files, the real problem for campaigns is whether they can organize themselves to use whatever money and contact methods are available to reach those valuable needles in the registration haystack: people who would vote their way if only they actually voted.
This is laborious, expensive and takes smarts. Most turnout campaigns fail to maximize their resources for this purpose; it is simply too hard to do enough to have much result. They settle for making random noise -- or default to pushing robocalls to people who would cast a ballot anyway.
People (like me) who've worked in this field think that at best turnout campaigns can increase their candidate's vote by perhaps 1-3 percent. Of course, in a truly close election, that is the difference between winning and losing.