What does margin of error mean in a poll?

Margin of error quantifies uncertainty from sampling: because a poll interviews a few hundred or thousand people rather than everyone, the result is an estimate. A margin of plus or minus 3 points at 95 percent confidence means that if you repeated the poll many times, about 95 percent of the intervals constructed this way would contain the true population value. The crucial and widely botched implication is for comparing candidates, since a 48 to 45 lead with a 3-point margin involves two separate estimates each with their own uncertainty, and the margin on the difference between them is larger than the margin on either one. Such a race is statistically indistinguishable from a tie.

What the margin of error does not cover is often the bigger problem. It measures sampling error only, and says nothing about non-response bias when certain groups are harder to reach, question wording effects, the accuracy of the likely-voter model, house effects from methodology choices, or people who change their minds after answering. Polling misses in recent elections were largely failures of these non-sampling factors rather than of sample size. Two practical reading habits: look at averages of many polls rather than any single one, since idiosyncratic errors partially cancel, and treat crosstabs, the subgroup breakdowns, with real suspicion, because a subgroup of 90 respondents carries a margin roughly three times the full sample's and is frequently reported as though it were solid.