Marist conduct the NBC/WSJ *state* polls, not the national one.
National is done by Hart Research (D) and Public Opinion Strategies (R). They're assuming a demographic similar to 2012 and a slightly lower turnout.
Apologies, my recall failed me. I actually went and found their LV screen to make up for it:
Likely Voters defined as voted in either 2012 or 2014 or 18-24 year olds with
self-reported interest in the election of 5 to 10 on a 1-10 scale.
The first is pretty standard. When I worked at YouGOV, we found that voting in the previous election of the same format gave you an 80% likelihood of voting in the next election - elections are habit-forming (incidentally, this is the greatest gift from Obama to the Democratic Party; high-enthusiasm wins have big aftershock effects on voting cohorts). People who didn't vote in the previous election of the same format only voted 15% of the time. What's more, people who didn't vote in previous elections but do end up voting in the election in question tend to like demographically similar to those who had voted in past elections anyway (probably community effects); so it's a pretty useful screen. It gets a little inaccurate for the 25-30 demographic and the 65+ demographic, overstating the former and understating the latter, but you can't do perfect. I've seen variations that add additional socioeconomic weighting and the like, or test for whether your participated in local elections (as an example, people who vote in council elections in the UK are over 95% likely to vote in national elections), but this is a good one.
The second I think is a bit ambitious. People who rated themselves as 5 out of 10 interested in the election still only had a 60% turnout rate. What's worse is that it isn't demographically consistent - if someone aged 20 said they were 5 out of 10 interested, they were less likely to vote than someone aged 40 who said they were 5 out of 10 interested. So if they're setting 5 out of 10 as the bar for millennial inclusion, they've actually included a fair number of people who might be as low as 30% or 40% likely to vote. So I think they're being quite optimistic about millennials still - they're assuming that Clinton will be able to excite them; or more likely they're just using Obama's figures as a benchmark and adjusting the bar to produce the same amount. The last one a lot of pollsters do because even though it normally fucks up, you can point it and say "yeah but it fucked up in a way we could never have predicted, if this had worked like the last election this never would have happened!" etc. It gives them plausible deniability.
Just thought this might be an interesting post about LV from someone who worked on it. I will say that if they're finding lower turnout even with a relatively permissive screen, that's a bit worrying for the Democrats. That said, they remark on that themselves:
Potential threats to Mrs. Clinton include signs that support and enthusiasm among African-American and Hispanic voters are waning. Those voters aren’t gravitating to Mr. Trump, suggesting they could stay home or vote for third-party candidates, whose impact on the race remains a wild card.
The poll also found that younger voters, a key Democratic constituency, are far less interested in the election than are other age groups and voting blocs. Asked to rate their interest in the election on a scale of 1 to 10, only half the voters under 35 ranked their interest as a 9 or 10. Overall, 68% of all registered voters ranked their interest that high.
So: get Kaine to Florida, get Sanders to Ohio. Don't let them leave.