Reading and Trusting Poll Results

Random sampling gives everyone in a population an equal chance of being picked, but it cannot force the chosen people to pick up the phone. Nonresponse bias creeps in when certain kinds of voters are consistently harder to reach or less willing to talk to a pollster, so the people who do respond stop looking like the population as a whole. In 2016, many pollsters underestimated support for Donald Trump partly because non-college-educated voters, who leaned heavily toward him, were underrepresented among respondents even in properly randomized samples.

Pollsters try to fix this with weighting: they compare their sample's demographics, education, race, age, region, to known census figures, then adjust the influence of underrepresented groups upward and overrepresented groups downward. Weighting helps, but it is a correction built on assumptions about who is missing and why, and if the pollster weights by the wrong variable, education instead of ideology, for instance, the fix falls short. This is also why the same underlying population produces different numbers depending on who gets surveyed: a poll of all adults includes people who never vote, a poll of registered voters narrows that, and a poll of likely voters, filtered by a screen like past voting history or stated enthusiasm, usually comes closest to an actual electoral outcome but depends heavily on how accurately the pollster predicts turnout.

Because every individual poll carries its own sampling error and its own weighting choices, one poll's number is a single noisy estimate, not a verdict. Averaging many polls together, which sites and forecasters do with polling averages, smooths out the noise from any one house's methodology and gives a steadier read on where opinion actually stands.