The problem is the sampling is never truly random to begin with. As Morris wrote, the phone polling always picks up too many older people and other groups. So not adjusting would leave worse results. If you could do a truly random poll, that wouldn’t need adjusting, and you wouldn’t get margins of error.
Take 100 samples and it's 1%. Differentiate between Dems and Reps, or 2 factors, in the framework of those 100 samples, it doubles your level of error. look at three factors Dem/Rep/Ind, and you're up to 3.333% on a sample of 100.
You can increase sample size or reduce the number of factors recorded to change your confidence interval. Plus, if you do D/R/I and you get different levels of response, your confidence interval is the LEAST of the three different levels ~ usually the Ind which will be down there at 5% while the others are up at 2% or so.
Sample size used to be Call 100 Voters and now it's hardly ever less than Call 230 Voters ~
Br'r Morris is warning you that most of the polls you are looking at are too small to support the claims they make, and not all the elements reported on were found in sufficient numbers to give you a satisfactory confidence interval ~ meaning ~ NOT that they aren't statistically valid surveys but they aren't being reported consistently with the statistical methodology!
If the pollster knows that then he' lies. If he doesn't he's in the wrong business. If he's not a pollster with knowledge of what he's doing, but instead a reporter, he will lie as a matter of course because that's his profession.
Then there are the more expensive internal polls. That's where the astute politician has his own staff statistician familiar with these matters ~ because the pollster will still misrepresent what his polls say, and the news media, if they get aholt of it, will lie as well, so you need somebody around who will tell the truth ~ kinda'!
BTW when you do a presidential poll you go out and ask about Obamagube, Romney, Rep, Dem and Ind. You also ask for NO Response ~ that'll be down there on your list to reflect a call was made and you didn't get any response. That's actually 6 different factors in an array with 12 boxes or cells. The least number of responses of any kind sets the basis for the confidence interval ~ so if you want at most a 5% level of expected or tolerable error in your otherwise perfect random sample based survey you'd probably hope for no cell showing up with less than 21 responses!
That's probably why the pollsters never tell you about the surveys they took where they only got 15 Republicans to answer the phone, or where almost nobody happened to like either of the candidates ~