For polls what they will do is take a sampling. The data comes in. If they like the sampling results - they roll with the results. If they don't like the sampling results - they will keep the results they like & sample again. They will do this until the sampling results of the poll match up with their narrative.
So instead of "Let's take a poll and figure out what people think about Trump" -- they will say "Let's do a poll until the poll shows that Trump has an unfavorable rating at 45% or below". So first round - let's say Trump hits at 60 percent favorability. They will keep the Negative 40% on the books and keep 40 percent of the Positive on the books - then resample. Now their totals look like 55-45. 55 percent positive instead of 60. Resample again.
They will toss out things against their poll agenda - keep things go with their agenda for the poll, until they get to what looks like what they wanted... Then they release the poll.