Between now and November you will consume hundreds of polls, and the discourse around them will cycle through the same two poles: blind faith in the averages, and blanket declarations that polling is dead. Both are wrong. Polling is an instrument with well-understood failure modes — and once you know them, you can extract real signal from noisy data.
Failure mode one: nonresponse bias
The foundational modern problem. Response rates have collapsed to the low single digits, and the people who still answer are not a random slice of America — they are more engaged, more educated, more institutionally trusting. If the voters a pollster cannot reach differ politically from the ones it can, no sample size fixes it. This is the leading suspect in the 2016 and 2020 state-level misses, when low-trust, non-college voters — disproportionately Trump supporters — were systematically underrepresented.
Failure mode two: weighting and its discontents
Pollsters correct their skewed samples by weighting — adjusting respondents to match the electorate’s known demographics. The catastrophic 2016 lesson was that many state polls failed to weight by education, at the precise moment education became the sharpest divide in white America (that realignment here). Weighting has improved, but it can only correct for categories you measure; if the divide shifts to something unweighted — social trust, media diet — the error returns wearing a new coat.
Failure mode three: the likely-voter guess
A midterm poll is not measuring opinion; it is measuring the opinion of people who will actually vote — a population that does not exist yet. Every likely-voter screen is a forecast wearing a lab coat, and in low-turnout midterms the screen is the poll: model an electorate two points too old or too young and you have manufactured a two-point error before fielding a single call.
Failure mode four: herding
Watch the final three weeks, when outlier results quietly vanish and every pollster converges on the same comfortable number. That is herding — suppressing or massaging results that deviate from the average, because being wrong alone is career death while being wrong together is Tuesday. A suspiciously tight cluster of polls is a sign of less information, not more.
1. Consume averages, never single polls — but check whether the average is a real spread or a herd. 2. Assume a normal error of a few points in either direction, and remember polling error is correlated: if one Senate poll is off, its neighbors are off the same way. 3. Weight real votes — specials and primary turnout — alongside surveys (how). 4. Distrust precision theater: a race polling 48-46 is a coin flip, full stop. The professionals’ secret is not better polls. It is appropriate humility, applied consistently.