The Art of Statistics: Learning from Data

by David Spiegelhalter

My personal, rather sceptical heuristic is that any quoted margin of error in a poll should be doubled to allow for systematic errors made in the polling.