In order for any scientific poll ( that uses small sample sizes - at least 1500 ) to be accurate, the sample has to be TRULY RANDOM. This is where these polls fail. If they are just polling landlines then it is not truly random of the population as a whole. The poll results should say "Of those polled with landlines, the results are such and such ". Even if cell phones and landlines are polled, it is still not truly random. Many people don't answer calls on cellphones that they do not know who the caller is. Even if it is answered , many people hang up ( with the IRS targeting conservatives, I can see where a lot of conservatives would not want to answer a poll ). So any of these poll results should start out with " Of those who agreed to be polled". Once again, not a truly random sample.
If they do a truly random poll, and discover that the results don’t match what they have decided in advance is to be the R/D/I mix (e.g. 2012 turnout numbers), they have to throw away responses to achieve that mix.
Which responses are thrown away and how do they decide? Are they more inclined to toss “anomalous” responses, such as Democrats supporting Republicans?