http://fivethirtyeight.com/features/harrys-guide-to-2016-election-polls/?ex_cid=538twitter

"Margin of error and sample size matter less than who’s in the sample. Good polling costs a lot of money, so many times the best polls have a smaller sample size (the more people you call, the costlier the survey). That raises the statistical margin of error, but the margin of error for a sample of 400 is less than double that for a sample size of 1,000. What you don’t want is coverage error, in which you’re polling people who won’t even vote or ignoring people who will."