Back on topic, a quick look at the data, compared with approximate state populations (from 2014 census estimates) added in, indicates a pretty clear progression in terms of % of population represented through those categories (caucus, no paper trail, paper trail). Assuming about 70% of population is eligible to vote, and about 48% are democrat-oriented (based on http://www.gallup.com/poll/15370/party-affiliation.aspx), let’s compare how many votes were counted vs the democrat-oriented eligible voting population.
Caucuses ranges from 0.13% (Iowa) up to 10.45% (Minnesota), with average 3.80% and median 4.00%.
No-paper-trail ranges from 15.83% (Texas) up to 38.47% (Pennsylvania), with average 24.82% and median 22.76%.
The paper trail category ranges from 18.08% (Arizona) up to 63.98% (Vermont) with average 37.97% and median 35.04%.
Compared to https://en.wikipedia.org/wiki/File:U.S._Vote_for_President_as_Population_Share.jpg it looks as though the paper-trail primaries are pretty close to representative of % of population who vote in modern-day elections since 1940, while the no-paper trail ones are about as representative as the women’s suffrage era in the 1920s, and the caucus states approximate the era when only wealthy white male landowners could vote prior to the 1820s.
Assuming the data is valid, an easy hypothesis is that paper-trail states are potentially significantly more representative than no-paper-trail states and substantially significantly more representative than caucus states. (I’ll leave it to somebody else to figure out the exact statistical significance.)
That doesn’t really address the question about the differences between polls and votes, but it lines up with the idea that there is a big difference in those categories and a pattern of discrepancies specifically in states where the vote is not representative and is designed to not be auditable would be somewhat suspicious.