Princeton Election Consortium

Innovations in democracy since 2004

Outcome: Biden 306 EV (D+1.2% from toss-up), Senate 50 D (D+1.0%)
Nov 3 polls: Biden 342 EV (D+5.3%), Senate 50-55 D (D+3.9%), House control D+4.6%
Moneyball states: President AZ NE-2 NV, Senate MT ME AK, Legislatures KS TX NC

What’s The Real Source Of Inaccuracy In Alaska?

September 25th, 2014, 5:30pm by Sam Wang

On Tuesday, I suggested that control of the Senate could come down to as few as four key races in Colorado, Iowa, Louisiana, and Arkansas. There’s a fifth state where voters are exceptionally powerful: Alaska, because of the closeness of the race and its small population. But even though get-out-the-vote efforts in Alaska are certain to be valuable, activists will have to do their work without a clear picture of which candidate is ahead, Democrat Mark Begich or Republican Dan Sullivan.

As you can see, the Meta-Margin moved toward Republicans today. This is caused by two recent polls in Alaska. This race has had only seven surveys since Sullivan clinched the nomination in August. I think we may continue to see fluctuation in Alaska surveys in the coming weeks. Today I want to point out some special characteristics of Alaska that complicate any exercise in poll aggregation: its small population and extreme mobility in and out of the state. The bottom line is the possibility of polling error like what happened in the Nevada 2010 Senate race…though in the opposite direction.

Fewer than 500,000 people are registered to vote in Alaska as of this year. Senate polls for the general election and the Republican primary election were based on approximately 15,000 responses. It has been estimated that fewer than 10 percent of people who are called in a political opinion survey actually pick up the phone and complete the survey. Most pollsters predominantly use telephones to reach respondents, so over this campaign season they have dialed up at least 150,000 people – nearly 1 in 3 Alaska voters, which is an amazingly high fraction. As Edward Freeland, director of Princeton’s Survey Research Center, put it to me, “you can crank up polling machinery that works so well in most cases, but in Alaska you start reaching limits that don’t occur elsewhere.”

However, I think a source of even worse trouble for polling state is the fact that many of Alaska’s voters are relatively recent arrivals. Sullivan himself moved to Alaska in 1997 and has only lived in the state for eight of the last seventeen years, mainly as a government official. (Begich has highlighted the difference by running a campaign called “True Alaska.”) But Sullivan is hardly alone. Out of a total of approximately 730,000 residents in Alaska, over 340,000 arrived between 2000 and 2011. Over 320,000 people left the state during that same period.

In other words, in any given year, approximately 10% of Alaska’s population has just arrived or is about to move away. This is reminiscent of Nevada, which has very similar statistics. Nevada is also the site of a spectacularly wrong polling error. In fall of 2010, surveys unanimously showed Republican Sharron Angle favored to unseat Democratic Senator (and Majority Leader) Harry Reid, by a median of four percentage points. Every data-based prognosticator, including me, expected Angle to win. Yet Reid won the election by 5 points, outperforming polls by 9 points. I regard this as the worst miss of Senate polling in the last three election cycles.

What can the Nevada case tell us about possible polling errors in Alaska? In Nevada, underpolled populations probably included several groups that tilt Democratic, including cell-phone-only users and Hispanics, as well as itinerant workers in Nevada-specific industries. Similarly, there are groups of Alaskans, given its unusual geography and demographics, who are likely to be underrepresented in polls because they are mobile: Alaska has substantial populations of seasonal and short-term workers such as oil-industry workers and fishermen. 49 percent of Alaskans use cell phones exclusively or predominantly. Across the U.S., nearly 10 percent of cell phone users bear an area code that does not match their current home state, and in Alaska the percentage is likely to be considerably higher. These voters would not be identifiable as Alaskans by telephone-based surveys, whether landline or cell phone-based.

A recent analysis by FiveThirtyEight has suggested that in Alaska, Republicans tend to outperform opinion polls on Election Day, and offered as the explanation the fact that partisan polling can distort polling averages. However, I don’t think that explanation entirely holds water, for the simple reason that it does not account for why Alaska would be different from other states. Partisan polls are everywhere.

Based on the extreme mobility of Alaska residents, I suggest a problem of potentially far larger impact. Like Nevada, telephone-based surveys in both states simply fail to capture a sample that is fully representative of voters. In other words, I suggest that ideologically diverse states with lots of recent arrivals are subject to inaccuracy.

Pollsters could get around the mobile-voter problem by reaching out through the Internet; Alaska has one of the highest rates of Internet access in the United States. Of polls conducted since the GOP primary in August, two organizations used Internet-based sampling, YouGov (entirely online, Sullivan +6%) and Public Policy Polling (20% online respondents, Sullivan +2%). Recent polls using telephone-only sampling have been done by Hays Research (Begich +5%), Harstad (Begich +5%), and Rasmussen (Sullivan +5%). The average difference is that phone-only surveys are 5.7% better for Begich. This is a little hard to interpret because there is also the partisan-pollster issue. However, it is nearly the same size as the 5-point bias that FiveThirtyEight highlighted. Internet-based sampling is a growing technology in polling, and has attracted some controversy. But if we believe that use of the Internet helps reach otherwise-inaccessible voters, Sullivan is in better standing than a simple median would suggest.

The PEC calculation is polls-only without adjustments, and I can’t really change the calculation on an ad hoc basis – it defeats the purpose of having a clean snapshot. At the moment, since the median shows a small Begich lead, the effect of assuming that Sullivan is ahead in Alaska would be to shift today’s Senate Meta-Margin from D+0.6% to D+0.2%, basically a perfect toss-up. If another pro-Sullivan poll shows up, the Meta-Margin will fully reflect a Sullivan lead, and this difference will disappear.

Alaska voters have exceptional power, on a per-vote basis, to influence Senate control. Over in The Power Of Your Vote, at right, I have calculated that in terms of influencing the overall probability of Senate control by either party, individual Alaska voters have a tremendous amount of power. Alaskans might want to ready themselves for a new type of incoming demographic: Republican and Democratic campaign workers.

Tags: 2014 Election · Senate

13 Comments so far ↓

  • Randall

    Hold Iowa, NH, NC, and Colorado. Steal Kansas. Independents caucus with Dems.

    Is that 50/50, Sam? Please correct me if I am wrong :)

    Making Alaska (or Georgia or Arkansas or Louisiana) icing on the cake?

    • Sam Wang

      Your counting is correct, but Iowa and Colorado are extremely close. Bake the cake first.

    • Fritz Pettyjohn

      Alaska’s itinerants include a lot of military who maintain Alaska residence when they leave the state. They continue to get Permanent Fund dividends that way. They vote in order to bolster their claim to residence. Republicans are pro-military, and pro-permanent fund, so 90% of these votes go R.

  • Insidious Pall

    Most importantly, the trend for poll to poll results over time is toward Repubs in AK. For example, Sullivan’s totals have increased significantly across PPP, Razz, and CBS News, et al polls over the last month; each having 3 polls. If AK goes red, that is roughly R-50 with CO and IA very nearly up for grabs. Of course others are close but they seem to lean, albeit very slightly, one way or the other. I still think there is a good chance Senator Orman caucus red in this scenario, even if Dems take CO and IA. Mine is admittedly a neophyte’s view, but I cannot envision Orman standing on semantics after inferring to deeply red Kansas voters he will caucus with whomever reaches a majority. So Sam’s discussion about the transient nature of the Alaskan electorate may turn out to be the wild card here. I well recall he called the Kansas mess days before any of the others could even find it on the map.

  • atothec

    Assuming Orman goes w/ Dems I think it will come down to IA, CO, and AR.

    Folks have to remember Pryor’s father was a rep, a governor, AND a Senator over a 30yr period and is still active. Pryor Sr. won in a Reagan year no less and Mark eked out a slim win in 2010. Given that this is not a wave year and some 40+ yrs of state connections, I think Pryor has a real shot.

    Alaska I do think is toast.

    Just need Iowa to come to its senses.

  • Brian G

    I have a question for you? How can all these different pollsters likely voter models affect their outcomes? I’m wondering if some of these pollsters have to restrictive likely voter models.

  • JayBoy2k

    I have been lurking for a month and did read the “basics” of the model — mostly use medium on all polls. As I look at the header, it says R51 D&I49 and meta-margin is R+.01. So what are the other factors that contribute to a probability of 70%D&I having 50 of more seats on Election day?

  • Pat

    Today the meta margin is R+0.1% and there is only 47% chance that Dems hold the Senate if the election were held today (‘snapshot’).
    If future evolution is a random walk from today’s conditions, I am puzzled by why it would drift back to a probability of 70% Dem control on election day. I had assumed the random walk was symmetrical with only the ‘diffusion’ coefficient setting the extent of the fluctuations based on past history.
    So you are actually not only considering the extent of past *fluctuations* to get your random walk, but also the past values themselves? (meaning, since past values were rather favorable to Dems, the situation is more likely than not to drift back into Dem territory?)

    • Sam Wang

      Yes. The same approach that stabilized it at 70% when the Meta-Margin approach D+2% is doing the same now when we’re at R+0.1%.

      Eventually the present becomes more predictive of where we are in the range from June until now…that launches next Tuesday, at Election minus five weeks.

  • Olav Grinde

    The simple fact is that favorable polls are a very cheap way to influence the “Who is winning” narrative. Call me paranoid if you wish, but I expect an overabundance of right-leaning polls between now and election day.

    • Matthew

      I wouldn’t call you paranoid just Rovian. Republicans dismissed the polls as biased because they didn’t like the results. If you similarly dismiss the polls because you dislike the results you are making the same error the Republicans did in 2012. Fact is the polling companies do a very good job and are very accurate. And they have an incentive to be accurate, money. If people can’t trust the polls less people will buy the polling data and the polling companies will lose customers and thus their profits will suffer.

  • Joseph

    I am increasingly convinced that the recent “militarism” of President Obama is pulling down his numbers, especially from his liberal flank. The problem is that even the most surgical of air strikes will have collateral damage that will be exposed in the media. (Note that, although I categorize myself as a “liberal”, I am also a realist, and understand that the President isn’t undertaking these steps lightly.) Unfortunately, this new militarism may have serious connotations for the present election cycle. If I am right, then the polls may be picking this up, which would explain the sudden swing towards the Republican side in this ultra-close election.

Leave a Comment