日不到一个月，各项民调均预测民主党将大获全胜。根据选举预测专家内特·西尔弗整理的民调平均值显示，民主党在堪萨斯州的参议院候选人目前处于领先态势——该州在2024年大选中以16个百分点的差距支持特朗普；在阿拉斯加州、艾奥瓦州、俄亥俄州以及得克萨斯州，民主党候选人同样领先，而这些州当年特朗普的胜选优势均超过10个百分点。在众议院选举方面，民主党有望取得18年来最显著的多数席位优势。
Less than a month out from Election Day, the polls are predicting a Democratic rout. According to the election forecaster Nate Silver’s polling average, Democratic Senate candidates are leading in Kansas, which voted for Donald Trump by 16 percentage points in 2024, as well as Alaska, Iowa, Ohio, and Texas, all of which Trump won by more than 10 points. In the House, Democrats are poised to win their largest majority in 18 years.
以上便是民调给出的结论。这些结果可信吗？自2016年低估了特朗普的支持率后，民调机构便竭力试图重获公众信任，对调查方法进行了诸多或大或小的调整。或许这次他们终于做对了——至少他们自己是这么认为的。埃默森学院民调部门主任斯宾塞·金博尔对我说道：“我有信心，不仅我们机构，整个行业都掌握了可靠的技术与科学的方法论，能够准确反映并体现公众的意愿。”不过，该行业此前也曾信心满满，结果却事与愿违。
So the polls tell us. Can they be believed? Ever since underestimating Trump in 2016, pollsters have been busy trying to regain the public’s trust, tweaking their methodologies in ways both dramatic and subtle. Perhaps they’ve finally got it right. They seem to think so. “I’m confident that not just ourselves, but the industry, has good technique, good methodology, and that they are reflecting and representing public opinion,” Spencer Kimball, the director of polling at Emerson College, told me. But the industry has been confident before, only to be proved wrong.
2016年之后，民调机构意识到自己未能充分重视“教育程度导致的政治分化”这一现象——即受过大学教育的选民与未受过高等教育的选民在政治倾向上存在显著差异。此后，整个行业试图在调查中纳入更多无大学学历的选民；若实在难以实现，则会对这类选民的反馈数据加以加权处理，以匹配其在选民群体中的实际占比。2018年中期选举时，民调结果几乎准确无误，这似乎印证了这些调整措施卓有成效。
After 2016, pollsters concluded that they had failed to pay enough attention to educational polarization, the somewhat-new phenomenon of college-educated voters having political preferences that were sharply different from those of voters without a degree. Going forward, the industry would try to include more non-college-educated voters in surveys or, if that proved impossible, would weight those voters’ responses to align with their share of the electorate. In 2018, the pollsters nailed the midterms with little error, suggesting that their adaptations had paid off.
然而2020年的情况却截然不同。尽管民调准确预测了多数选民更青睐乔·拜登而非特朗普，但其对特朗普实际支持率的估算值却比2016年时更低，尤其是在各州层面。最典型的例子便是威斯康星州：民调曾预测拜登能以最多10个百分点的优势胜出，最终结果却是他的领先幅度不足1个百分点。
Then came 2020. Although the polls correctly gauged that most people preferred Joe Biden to Trump, they underestimated Trump’s support by even more than they did in 2016, especially at the state level. The most memorable example was Wisconsin, where polls had Biden winning by up to 10 percentage points. He wound up winning by less than one.
这些结果让民调机构更加困惑。既然控制教育水平这一变量也无济于事，那还有什么办法呢？他们推测，特朗普的支持者或许与其他选民截然不同：相较于其他人口特征相似的选民，他们更不愿意与民调人员交谈。这种现象被称为“党派性无回应偏差”，它有助于解释为何特朗普执政时期的中期选举民调结果通常比总统选举民调更为准确：中期选举的选民群体以政治参与度较高的群体为主，这类人群对各类机构更信任，也更愿意参与民调。
These results left pollsters even more perplexed. If controlling for education didn’t work, what would? Perhaps, they figured, Trump supporters were just different from other voters: less likely to talk to pollsters than other demographically similar voters were. This phenomenon, known as partisan-nonresponse bias, would help explain why midterm polls in the Trump era have generally been more accurate than presidential-election polls: The midterm electorate skews toward politically engaged voters who are more trusting of institutions and more likely to have answered polls.
此后，民调机构决定采用手头最“直接”的方法：“回忆投票法”。既然特朗普的支持者存在系统性差异，或许最好的办法就是直接询问：“上次选举您是否把票投给了特朗普？”随后再根据这一结果对样本进行加权处理，使其中的特朗普支持者比例与上一次总统选举时的实际比例保持一致。
Moving forward, pollsters decided to reach for the bluntest object they had access to: “recalled vote.”If Trump supporters were systematically different, perhaps the best solution was just to ask, “Did you vote for Trump last time?” and then weight the sample so that it had the same Trump-voting percentage as the previous presidential election.
2022年中期选举时，民调结果再次准确无误。然而到了2024年，“回忆投票法”却未能奏效：民调对特朗普支持率的预测值比实际值低了约3%，这一情况与2016年时如出一辙。至少在总统选举年份，这种无回应偏差实在难以纠正。即便是在那些自称曾投票给特朗普的受访者中，民调机构似乎也只能接触到那些最有可能转而投票给民主党人的人群。
When 2022 rolled around, the pollsters once again nailed the midterm. But the recalled-vote tweak didn’t solve the problem in 2024, when polls underestimated Trump by about 3 percent, just as they had in 2016. In presidential years, at least, nonresponse bias had proved to be extremely difficult to correct for. Even among people who said they had voted for Trump, pollsters could seem to find only the ones who were most likely to vote for a Democrat.
“从某种程度上说，受访者需要配合，”共和党民调专家帕特里克·鲁菲尼告诉我，“你们通过电话联系到的这些人，其实并不是目标人群。”特拉法加集团的共和党民调专家罗伯特·卡哈利则对我说：“我发现，共和党人往往更担心：我的名字会不会被列在某个名单上？他们想知道，你们收集这些信息到底是为了什么？”为了应对这一棘手问题，民调专家们正采用越来越复杂的解决方案。
“At some level, the respondents need to cooperate,” Patrick Ruffini, a Republican pollster, told me. “The individuals that you’re getting on the phone are the wrong people.”Robert Cahaly, the Republican pollster behind the Trafalgar Group, told me, “I’ve found that Republicans tend to be a little more concerned about: Is my name going to be on a list? They want to know, Who are you getting this information for?”To deal with such an intractable problem, pollsters are resorting to ever more complex solutions.
YouGov的首席科学家道格拉斯·里弗斯告诉我，他的团队现在正在开展选区级别的建模，不仅要确保样本中包含足够多声称自己投票给特朗普的人，还要确保其中足够多的人既声称投票给了特朗普，又居住在大体上也支持特朗普的社区。《纽约时报》和锡耶纳学院为它们联合开展的民调发明了两种新技术：“支持度评分”和“能量平衡”，听起来像是某种疗愈水晶，但实际上是统计权重。
Douglas Rivers, the chief scientist at YouGov, told me that his team is now doing precinct-level modeling, making sure it has not just enough people who say they voted for Trump but also enough people who say they voted for Trump and live in neighborhoods that for the most part voted for Trump. The New York Times and Siena College have invented two new techniques for their jointly conducted polls, “support score” and “energy balancing,” which sound like healing crystals but are in fact statistical weights.
这些权重被用来根据多种因素调整调查数据，例如受访者的名字听起来有多“共和党”，他们开什么车，以及选民中有多少人在人口特征上与受访者完全一致（也就是说，不仅是同为拉丁裔，还要是同样年轻、拉丁裔、不可知论的保守派，并且居住在郊区）。
They’re used to adjust survey data based on such factors as how Republican a respondent’s name sounds, what car they drive, and how many people in the electorate are demographically identical to them (as in, not just fellow Latinos but fellow young, Latino, agnostic conservatives who live in the suburbs).
即使他们成功校正了无回应偏差，民意调查员也必须精准构建投票率模型。要准确预测谁会投票给谁，首先必须弄清楚到底谁会去投票。投票率模型可能对最终结果产生巨大影响。例如，YouGov 在 9 月进行的一项民调显示，民主党在美国众议院普选中领先 6 个百分点，这足以让他们以微弱优势重新夺回众议院的控制权。
Even if they manage to correct for nonresponse bias, pollsters have to nail the turnout model. Accurately predicting who’s voting for whom requires figuring out who’s voting at all. Turnout models can swing the results massively. For example, a September YouGov poll had Democrats leading by six percentage points in the U.S. House popular vote, enough for them to narrowly retake the lower chamber.
但这一结果仅基于已投票人群。而在 YouGov 认定为“可能投票者”的受访者中，民主党以 16 个百分点的优势领先，这一结果很可能意味着民主党将在众议院获得自 20 世纪 70 年代以来的最大多数席位。就连负责 YouGov 美国民调业务的里弗斯（Rivers）也承认，这一结果令人难以置信。他在最近一篇博客文章中写道，这一结果“源于一系列都将结果推向民主党有利的假设，而这些假设很可能应该被适当打折看待。”
But that was amonged voters. Among respondents whom YouGov deemed to be likely voters, the poll had Democrats winning by 16 points, an outcome that would likely usher in the largest Democratic majority in the House since the 1970s. Even Rivers, who runs YouGov’s American polling outfit, admits that that result is hard to believe. In a recent blog post, he wrote that it “comes from a mix of assumptions that all push the result in Democrats’ favor and that should probably be discounted.” (A newer YouGov poll published yesterday suggests a 10-to-12-point lead among likely voters, depending on the definition of likely.)
（YouGov 昨天发布的一项更新民调显示，在“可能投票者”中，民主党领先幅度为 10 至 12 个百分点，具体取决于“可能投票者”的定义。）投票率模型在很大程度上依赖于有依据的推测。在美国总人口中，黑人、白人或亚裔成年人的数量是可以查明的，因此民意调查员可以在这一层面确保样本具有代表性。但要预测这些群体中任何一组的实际投票率，始终需要依赖判断。
Turnout models involve a fair amount of informed guessing. Within the general population, the number of Black or white or Asian adults in America is knowable, and pollsters can thus make their samples representative in that sense. But predicting how any of those groups will turn out is always a judgment call.
如果某些群体的投票率最终远低于或远高于民意调查员的预测，就可能导致最终结果的误差增大。
If turnout among certain groups winds up being much lower or much higher than the pollsters predict, that can increase the final error.
民主党人以及民调行业都有理由对2026年抱有乐观预期：那将又是一场特朗普并未参与竞选的中期选举，那些投票积极性最低的选民预计也会选择不投票。这类选举的结果恰恰是民调机构近期能够准确预测的。不过，特朗普执政以来的选举样本数量实在有限——仅有两次中期选举和三次总统选举。选民行为有时会出现难以预测的变动。
Democrats—and the polling industry—have a reason to be optimistic about 2026: It’s another midterm in which Trump is not on the ballot and the least engaged voters can be expected to stay home. That’s the kind of election that pollsters have recently gotten right. Then again, the sample size of Trump-era elections—two midterms and three presidential elections—is tiny. Voter behavior sometimes changes in unpredictable ways.
而且，特朗普几乎以各种方式让自己“参与”到了选举之中：亲自挑选大部分候选人、在全国各地开展巡回竞选、动用纳税人的资金来投放政治广告，还强迫本党举办以他为核心的“中期选举大会”。倘若出现类似2016年或2024年的系统性民调误差，民主党将无法掌控参议院。如果民调在所有州的预测值都比实际高出整整3个百分点，那么民主党仅能勉强赢得阿拉斯加州、得克萨斯州、密歇根州和俄亥俄州，却无法拿下堪萨斯州、艾奥瓦州和缅因州，还会失去明尼苏达州。
And to whatever degree possible, Trump has managed to put himself on the ballot: hand-selecting most of the nominees, barnstorming the country, using taxpayer money to pay for political ads, and forcing his party to host a Trump-centric “midterm convention.”A systematic polling error on the level of 2016 or 2024 would mean that Democrats fall short of Senate control. If the polls overestimate their chances by exactly three points in every state, then Democrats will barely win Alaska, Texas, Michigan, and Ohio, but they won’t pick up Kansas or Iowa or Maine, and they’ll lose Minnesota.
届时参议院将呈现50对50的平局局面，副总统万斯将拥有决定投票结果的权力，民主党掌控参议院的梦想也将化为泡影。如果误差程度达到2020年时的水平，对民主党而言则更为糟糕：他们最终连一个参议院席位都无法增加。虽然民主党仍会重新掌控众议院，但优势会极其微弱。
The Senate would then be 50–50, Vice President Vance would be the tiebreaker, and dreams of a Democratic Senate would stay dreams. A 2020-level miss would be even worse for Democrats: They’d gain no Senate seats at all on net. Democrats would still take back the House, but much more narrowly.
当然，以上都只是假设情况。现实中，无论出现何种民调误差，各州的情况也会各不相同。倘若我们能提前准确预知这些误差的具体表现，也就根本不需要进行民调了。
Of course, these are hypotheticals. In reality, whatever polling errors materialize will vary from state to state. And if we could say in advance exactly how they’d vary, we wouldn’t need polls in the first place.