This is from PPP's own description of their methodology (except the bold, my emphasis):
So one should expect their partisan samples to vary somewhat randomly from poll to poll, with occasional extremes to either side.
That sounds like a flawed methodology. In the end, partisan affiliation trumps all the rest when it comes to polls. Pollsters say folks change their minds about their partisan affiliation on the edges, all the time, on a day to day basis, or week to week or whatever, so no worries. The sample is pure, but the minds of this vacillating cohort of voters are mush, that's the rap. But a change of this magnitude? Really? This is a controversial topic in the polling industry. Traditionally the orthodox view in statistical theory is that one should only reweight for demographics where the distribution among the population is independently known from a non-polling source. By this standard party ID should not be used, since it is only known from other polls which may be just as biased demographically. (Party
registration is known in a little over half the states, but generally pollsters are asking a standard question about party ID in all states that isn't applicable across different partisan registration regimes).
For this reason the traditional large polls generally do not re-weight by partisanship. In addition to PPP, none of Gallup, Pew, Quinnipiac, or the national network polls do this. Rasmussen by contrast does so, as do many smaller outfits.
