What did the state polls suggest?
The selected archive estimates averaged Clinton leads of 5.3 points in Wisconsin, 3.2 in Michigan and 3.2 in Pennsylvania. Each average gives one equal-weight latest eligible estimate to a firm/race, with same-date ties averaged.
These are retrospective calculations from a stated archive window, not a claim about any particular website’s contemporaneous polling average. The archive’s interview midpoint dates are not publication dates or final fieldwork dates.
What did the votes show?
The FEC totals put Trump ahead in all three states: about 0.76 point in Wisconsin, 0.22 in Michigan and 0.72 in Pennsylvania. The selected mean margins therefore overstated Democrats by roughly 6.1, 3.4 and 3.9 points, respectively.
Clinton won the national popular vote while Trump won the presidency. This is why national popular-vote accuracy and Electoral College expectations must be evaluated separately.
Why did the state polls miss?
AAPOR's 2016 investigation found evidence that late-deciding voters and failure to adjust for educational differences contributed to state misses. Those general findings do not attribute a measured percentage of any specific retrospective aggregate's error to a single cause.
Education mattered because college graduates were more likely to participate in surveys and supported Clinton more strongly. Failing to correct the imbalance could make a survey look more Democratic than the electorate. Late-deciding voters also broke toward Trump in important states. These findings concern the evidence evaluated by AAPOR, not a precise decomposition of every poll’s miss.
Does this prove voters hid support for Trump?
No universal concealment explanation was established by the AAPOR 2016 evaluation. The documented evidence about education adjustment and late choice is stronger than an assumption that every respondent concealed a preference.
Errors in related states can move together. Averaging polls reduces some random noise, but a shared selection or turnout-model problem can survive the average. The three state outcomes are not independent demonstrations of three unrelated polling failures.
What should readers change?
Treat a small state lead as uncertain. Read the interviewing period, population, weighting disclosures and likely-voter assumptions. Distinguish a measurement of current preferences from a model assigning a probability of winning the Electoral College.
Do not carry the same correction into the next election automatically. A later cycle can have a different response pattern, candidate coalition, turnout environment or direction of error.
What to carry forward
A reasonably close national popular-vote estimate can coexist with consequential errors in the states that decide the presidency.