Margin Of Error: Breaking down the polls

Margin Of Error: Breaking down the polls

Is Hagan ahead?

Posted February 28, 2014

— Election polls are inherently interesting to political junkies, but when they take place many months before the election – or years away – they have important limitations and need to be interpreted cautiously.

Such is the case with a recent survey about the North Carolina U.S. Senate race.

American Insights, a new polling firm in Raleigh with Republican ties, conducted the survey. AI surveyed 611 registered voters Feb. 11-15, using calls to landlines and cellphones as well as via the Internet. In the poll, respondents said that, if Democratic Sen. Kay Hagan was to face Republican Thom Tillis in November, they preferred Hagan over Tillis 38 percent to 35 percent. Hagan was also preferred by nearly identical margins over all other potential Republican nominees.

After AI released its data, headlines quickly announced Hagan was in the lead. Yet, this interpretation is misleading, and compared with earlier polls, it is confusing.

Just last month, Rasmussen reported that Hagan trailed Tillis by 7 points – 47 percent to 40 percent – while Public Policy Polling said that Hagan trailed the GOP field by just a few points. If you reach back to news stories from the summer of 2013, you can see Hagan "widening her lead" over her GOP contenders.

The real problem is that none of these headlines is accurate.

There is no Republican nominee yet – the primary is May 6 – and voters are not paying close attention to the race.

In the AI poll, for example, 27 percent did not express a preference for either Hagan or Tillis. The primary reason is because voters are unfamiliar with anyone but Hagan. A recent Public Policy Poll found that most North Carolinians had no feelings about Tillis.

Polls are volatile and non-predictive in these circumstances. Unfortunately, most polling firms tend to manufacture the illusion of settled opinions. They do this by:

  • Not providing respondents with an explicit answer option of "undecided”
  • Asking respondents who initially say they are undecided to please choose someone
  • Allocating a portion of undecided respondents to each candidate based on some decision rule

That is why the Rasmussen poll in January reported just 10 percent of respondents were undecided. How then did AI find 27 percent are undecided in the race? The answer is, to AI’s credit, the firm did not push people to answer a question in which they had no real opinion.

Of course, some people do know how they will vote in November, but since the lead for any candidate is always within the margin of error, none of these polls can actually be used to distinguish which candidate is ahead.

For the AI poll, this error is plus or minus 4 percentage points, meaning each candidate’s true level of support could be anywhere between 4 points more or less than was actually measured in that poll. So Hagan’s support is anywhere between 34 and 42 percent, and Tillis’ is between 31 and 39 percent. Since the possible range of each candidate’s level of support overlaps with each other –Tillis could be "up" 38 to 36 percent – it is impossible to distinguish whether any candidate is truly ahead. This is what pollsters commonly call, "a dead heat."

A final observation: It would be a mistake to rely on a single poll to judge a race. In the case of AI, it lacks a track record and uses a combination of Internet, cellphone and landline modes of administering surveys, which is a promising but largely uncharted approach.

Instead, as more polling on the race accumulates, we can construct an average of the polls to more accurately evaluate it. This technique is what made Nate Silver a national name. For reputable websites that use aggregate polling data to track the Senate race, try Real Clear Politics or Huffington Post's Pollster. The Pollster website has the advantage of being user interactive.


Please with your account to comment on this story. You also will need a Facebook account to comment.

Oldest First
View all