Jay Kuo's The Polls Were Wrong quotes major polling site 538 about the problems with the polls:
Primary polling is notoriously off, but when it is consistently off in the same direction by massive amounts, something is wrong. Even polling aggregator 538 seemed to acknowledge this. Around midnight last night, G. Elliott Morris of 538 wrote up his thoughts on primary polling, and this caught my attention:
POTUS primary polls are by far the least accurate type of poll in 538’s pollster rating database going back to 1999. But error is different than bias, and when most of the polls are off in the same direction, something has gone awry.
That “something” is likely the difficulty in obtaining opinions of moderate Republicans from samples of “likely Republican primary voters.” Remember that fewer than one percent of people called for a poll actually complete the interview. That means the ones that do are statistical “weirdos” (excuse the technical language). Pollsters adjust for this by weighting their samples to known population benchmarks — like the percent of all adults who are white, over 65, have a college education etc. But in primaries, such benchmarks do not actually exist; pollsters are just making educated guesses about them.
My theory is that most of these primary polls pulling samples of voters from voter registration lists are missing moderate crossover partisans and first-time voters. Additionally, we know that people who are highly motivated to participate in polls (the “weirdos”) also happen to be the most politically and ideologically extreme Americans. That's a recipe for polling bias in primaries, where weighting to party, past vote and polarized demographic benchmarks does not control for the partisan consequences of overrepresenting politically engaged Americans.