2024년 선거의 투표 정확도 이해하기

### 여론조사의 이해: 선거 예측의 이면에 숨겨진 복잡성
* 📊 최근 여론조사에 따르면 카말라 해리스와 도널드 트럼프가 박빙의 승부를 벌이고 있는 것으로 나타났는데, **10월 25일 타임스 시에나 여론조사에 따르면 해리스가 48%, 트럼프가 48%의 지지율**을 기록했습니다.
* 📉 그러나 이러한 수치는 전체 인구가 아닌 표본 조사에서 나온 것이며 응답률은 **1~2% 정도로 낮을 수** 있으므로 여론조사 기관은 일반적인 표본 규모인 **1,000명을 모으기 위해 최대 10만 명에게 연락**을 취해야 합니다.
* 🔍 여론조사 수치에는 일반적으로 **3% 내외의 오차 범위**가 있으므로 실제 지지도가 보고된 것보다 높거나 낮을 수 있다는 점에 유의해야 합니다.
* 📈 여론조사 기관에서는 가중치라는 방법을 사용하여 일반 인구를 더 잘 반영하도록 표본을 조정하여 **연령, 인종, 학력 등의 인구 통계가 정확하게 대표**되도록 합니다.
* ⚖ 과거 선거에서 대학 교육을 받지 않은 트럼프 지지자와 같은 특정 그룹이 여론조사에서 과소 대표되는 경우가 많았기 때문에 **가중치 적용은 특히 중요**합니다.
* 🗳 이러한 조정에도 불구하고 일반적으로 투표 연령의 미국인 중 **약 3분의 2만이 투표에 참여**하기 때문에 실제 투표율을 예측하는 것은 여전히 어려운 일입니다.
* 🔮 궁극적으로 여론조사는 **여론에 대한 통찰력을 제공**하고 접전 주를 강조할 수 있지만, 주로 경합이 치열하다는 것을 나타내며 **실제 결과는 개표가 완료된 후에야 알 수 있다**는 점입니다.
* 💬 이 주제에 대한 정확한 저널리즘을 지원하려면 **Vox의 멤버십 프로그램에 가입**하여 **심층 기사 및 전문가 분석**에 액세스하는 것이 좋습니다.

여론조사에 따르면 카말라 해리스와 도널드 트럼프가 박빙의 승부를 벌이고 있지만 표본 추출 방식에 따라 정확도가 제한됩니다.

[NEWSCASTER] “A new poll has Kamala Harris and Donald Trump locked in a dead heat.” [NEWSCASTER] “Latest Wall Street Journal poll...” [NEWSCASTER] “New Washington Post poll out this morning.” [NEWSCASTER] “A new exclusive poll.” New polls have been released what feels like every day this election cycle. [NEWSCASTER] “Trump ahead by two.” [NEWSCASTER] “Harris up by four.” [NEWSCASTER] “Deadlocked race.” [NEWSCASTER] “A tied race.” And you might be staring at ones like these, hoping it'll answer the question of who will win this presidential election. But there's a lot more to these numbers than meets the eye. And there's a few reasons they'll disappoint you if you rely on them to predict the future. [NEWSCASTER] “It is still anybody's race.” These numbers are from the Times Siena poll on October 25th,

which suggests that 48% of American voters planned to vote for Kamala Harris, while 48% planned to vote for Donald Trump. But of course, pollsters didn't survey every American. They surveyed a sample of them. Today, pollsters find their sample of people to survey through a mix of phone calls, text messages, online surveys, or by mail. And if you're thinking, who even answers those kinds of things? The answer is... not a lot of people. The response rate for polling can be as low as 1 to 2%, which means pollsters have to reach out to as many as 100,000 people to pull 1,000 people, which is a typical sample size for a national poll, or 50,000 people to reach the roughly 500 people typical for a state poll. Which sounds small, but that amount of people can still be accurate at measuring the opinion of a population if you read the polls with a few caveats in mind.

The first thing to know about these numbers is there is a margin of error, which means this number is best thought of as the center point in a range of possible outcomes. On average for a sample size of 1,000, any number that is 3% different in either direction could still technically be accurate. And that's simply because any time you depend on a sample to represent a larger population, there's a lack of precision. A general rule of thumb is that the bigger the sample size, the smaller the margin of error. But there's only so much pollsters can do. Polling more people is a lot more costly and time-consuming, which matters because big news events can alter people's opinions. And ultimately, increasing the sample size can have diminishing returns on accuracy. [COURTNEY KENNEDY] “And it really starts to level off around 1,000... 1,500...2,000 interviews in terms of if you keep putting in more effort,

are you going to be more accurate? The answer becomes no pretty quickly. And part of it is because as a pollster, you're either doing a phone poll or you're doing an online poll. And if you keep interviewing thousands and thousands of more people through that mechanism, the next 10,000 are going to be just like the 2,000 people you already spoke with. This margin of error doesn't account for other errors in the polling process. Things that are harder to measure, like excluding certain demographic groups, low response rates among certain groups, and people misunderstanding the survey questions or misreporting their opinions. Which means this 3% margin is actually a low estimate. [KENNEDY] The rule of thumb is that you should actually double it. Before a poll is released, pollsters have to do some work to make sure the sample size looks just like the entire population. Based on sources like the US Census and voting files. It's called weighting.

For one, some groups are more likely to be reached by pollsters and respond to them. [KENNEDY] “You have, usually too many college graduates, often too many women, and maybe too many white adults. If you look at 2016 and 2020, Trump supporters were more difficult to reach in surveys. After data collection, there's ways in which your sample now departs from being random and representative. Now, weighting is how we pollster fix that.” So if each individual in a poll starts off with a value of one, weighting is when people who are underrepresented compared to the larger population get a value higher than one. While groups who are overrepresented get a smaller value. So if our imaginary sample here was made up of 60% women and 40% men, pollsters would make the women count less, and the men count more to match the fact that about 52% of US voters are women, while 48% are men.

They repeat this for factors like age, race, and geographic location, too, until each individual is weighted in a way that makes the sample a good stand-in for the larger population it represents. Recently, pollsters started consistently weighting for education, too. In the 2016 election, some state polls underestimated Donald Trump's support among non-college-educated people, an oversight that resulted in an election outcome few saw coming. Weighting respondents based on how their education level matches the general population is intended to help solve for that. In the 2024 election, pollsters are increasingly weighting for things like how people voted in past elections and party affiliation, too, to avoid any other ways it could be over- or underrepresenting a candidate's supporters. But the thing is, weighting a sample to match the general US adult population is pretty simple. Through sources like the US census, pollsters know how many people are in each age group, or gender, or racial group, or whether someone went to college.

But for a poll to be a precise prediction of an election outcome, pollsters are attempting to weight against an unknown population: the people who will actually vote in the future, on election day. There's no guarantee that anyone surveyed ahead of an election will actually vote. Some polls make educated guesses, counting people more who say they are likely to vote, or who are registered to vote, or who have voted in the past. But at most, only two-thirds of Americans of voting age turn out to vote. And pollsters can only make informed guesses on who those people will be. Polls are good for a lot of things – at capturing where the population stands on big issues. Or figuring out which states are really going to be decisive in the election outcome. But when the polls on voting choice are as close as they are in this election, the main thing they tell us is – simply that the race is close. And the only way we'll know who will win is to pay less attention to the polling news cycle,

and wait until the votes are counted. Before you go, I want to tell you about Vox's new membership program. We put this video together as quickly as we could when we realized there was a lot of curiosity, confusion, and obsession over the election polls. This kind of journalism takes a lot of work: interviewing experts, writing, editing, and animating. Our membership program is the best way to support this kind of work – both here on our YouTube channel but also on Vox's news site. If you join, you'll get access to more articles from Vox's newsroom and other member-only benefits. I hope you'll consider becoming a member by going to the link on the screen or the one in the video description below.

What polls can actually tell us

Vox helps you understand our world.

Vox video is Bridgett Henwood, Cath Spangler, Adam Freelander, Joey Sendaydiego, Coleman Lowndes, Christophe Haubursin, Kim Mas, Laura Bult, Ed Vega, Rajaa Elidrissi, and with research support from Melissa Hirsch. 

Become a member today: https://www.vox.com/support-now

Check out our FAQ on permissions and licensing of our videos: https://www.voxmedia.com/pages/licensing

✉️ Write us: voxvideo@vox.com


새로운 여론 조사에 따르면 카말라 해리스와 도널드 트럼프가 박빙의 승부를 벌이고 있지만, 이러한 수치는 오해를 불러일으킬 수 있습니다.

여론조사 기관은 여론을 측정하기 위해 응답률이 낮은 소수의 표본에 의존하는 경우가 많습니다.

표본이 적더라도 이러한 여론조사는 신중하게 해석하면 인구의 의견을 정확하게 측정할 수 있습니다.

여론조사 수치에는 오차 범위가 있으며, 표본 규모가 1,000명일 경우 어느 방향이든 3%의 차이가 있어도 잠재적으로 정확할 수 있습니다.

표본 크기를 2,000명 이상으로 늘리면 정확도에 대한 수익률이 감소합니다.

오차 범위는 다른 여론조사 과정의 오류를 고려하지 않기 때문에 낮은 추정치이며, 여론조사 기관에서는 표본 크기가 전체 모집단을 대표하도록 가중치를 사용합니다.

여론조사 기관은 종종 대표성 있는 표본을 확보하는 데 어려움을 겪으며 특정 그룹이 과대 대표되는 경우가 있습니다.

가중치는 여론조사에서 개인의 값을 조정하는 데 사용되며, 과소 대표되는 그룹에는 더 높은 값을, 과대 대표되는 그룹에는 더 낮은 값을 부여합니다.

최근 여론조사 기관에서는 더 많은 인구를 더 잘 대표하고 후보자에 대한 지지를 과소평가하지 않기 위해 교육, 과거 투표 행동, 소속 정당에 따라 가중치를 부여하기 시작했습니다.

여론조사를 통한 선거 결과 예측의 어려움

여론조사업체들은 미래의 유권자 수를 알 수 없기 때문에 선거 결과를 정확하게 예측하는 데 어려움을 겪고 있습니다.

투표 가능성 및 과거 투표 행동과 같은 요소를 바탕으로 추측을 하지만, 일반적으로 유권자 중 3분의 2만이 투표에 참여합니다.

여론조사는 여론을 파악하고 결정적인 상황을 파악하는 데 유용하지만, 투표 선택이 박빙일 때는 주로 박빙의 승부를 나타내며 실제 결과는 개표 후에야 드러납니다.