Pollster.com

Articles and Analysis

 

Frank Newport on Likely Voters

Topics: 2006 , Likely Voters , The 2006 Race

Editor's Note:  This post inaugurates a new feature on Pollster, our "Guest Pollster's Corner."  We hope this new forum will provide opportunity for professional pollsters of all stripes -- media and campaign; Democrat, Republican and non-partisan -- to occasionally share their own thoughts on the art and science of political polling. We are honored to receive our first contribution from Frank Newport, the Editor and Chief of the granddaddy of them all, the Gallup Poll.

The average turnout level among the voting age population in midterm elections is typically well below 50%, significantly lower than in presidential election years. This means by definition that the actual group of voters who turn out and vote on Election Day is a relatively small sub-set of the large poll of all eligible voters. If there is no difference in the voting intentions between these two groups, then reports of pre-election generic ballot results based on registered voters are all that is needed. If the pool of those who have the highest probability of voting is significantly different from those who are less likely to vote, however, then the effort to identify likely voters in pre-election polls becomes critical to accurately predicting and understanding the outcome.

Gallup's past history of polling indicates there is a high probability of a significant difference in the voting intentions of the large pool of registered voters and the smaller subset of likely voters in lower turnout midterm elections.

In 1994, Gallup's final generic ballot showed a dead heat between the Republicans and Democrats among all registered voters, but a 7-point lead for the Republicans among likely voters. According to estimates of the national-two party vote for that election, the Republicans had a nearly 7-point advantage in all votes cast for Congress that year (52.4%-45.5%). (In the penultimate Gallup poll in late October 1994, the likely voter "gap" showed a 10-point Republican advantage, while the registered voter gap in the same poll showed a 3 point Democratic lead, representing a 13-point difference in the gap between the two groups.)

In 2002, Gallup's final generic ballot among registered voters -- in the poll conducted Oct 31-Nov 3, 2002 -- showed a 5 point- Democratic edge, 49%-44%. Among likely voters it was 51% to 45% Republican, for a difference in the gap between registered and likely voters of 11 points. The final national House vote in 2002 was 50.5% for the Republicans vs. 45.9% for the Democrats, a 5 point Republican advantage).

In both of these years, the distinction between the vote intentions of all registered voters and likely voters was significant. The likely voter estimate was more predictive of the real world outcome.

Gallup's first use of the likely voter model in 2006 -- in the USA Today/Gallup Poll conducted Sept 16-17 -- provided an early suggestion that the standard pattern of turnout by party will continue this midterm cycle. Among the pool of all registered voters in the sample, Democrats led Republicans by a 9-point gap, 51% to 42%. Among the pool of those identified as likely voters, the ballot was tied at 48% to 48%.

This can change during the course of the election between now and Nov. 7. Likely voter estimates are more volatile than estimates based on larger samples of registered voters or all national adults. The gap between registered voter and likely voter estimates often fluctuates in September and October, particularly in response to the high-intensity campaigning likely to occur over the next month. Still, the mid-September Gallup results suggest that the historical turnout advantage Republicans have enjoyed in mid-term elections appears to be operative again this year -- at least as of this point.

Frank Newport
Editor in Chief, The Gallup Poll

 

Comments
Alan Abramowitz:

Frank Newport uses evidence from two midterm elections in which the national political climate strongly favored Republicans, 1994 and 2002, to argue that a substantial pro-GOP gap between likely and registered voters is entirely plausible. Before accepting this claim, however, I would like to see evidence from other midterm elections, especially those in which the national political climate favored Democrats--elections such as 1974 and 1982. One would expect a Republican turnout advantage in an election in which the national political climate favors the GOP but not in an election in which it favors Democrats. That is what makes Gallup's recent results puzzling.

____________________

Frank,

I'd like to know how you control for state effects in a midterm election. Turnout is correlated with the competitive races, and without hot contests in California and New York, I expect national turnout to be lower than say, 1994. However, in the states with competitive Senate or Governor's races, I expect turnout will be higher than the national average.

My favorite example is Hawaii: the state has one of the lowest turnout rates in presidential elections, but ranks in the top five over the past two midterm elections. Simply, Hawiian presidential elections are uncompetitive, but they've had a couple of close governor races in the midterms.

Do you use a national probability sample or do you tweak for state effects?

____________________

Michael Cobb:

I had started writing my question when I looked up and realized Alan already asked it; why does Gallup pick out two midterms to tout the LV model as superior when many midterms exist? What happened in all midterms with the LV model compared to the RV model? Also, I was under the impression (maybe for presidential elections?) that Gallup's LV model was inferior to its RV model. Can you comment on this? And if I'm right about the LV model not being any stronger in presidential elections, can you explain why it would only work better in midterms?

thanks....

____________________

Gary Kilbride:

Nice sample. Two.

A pollster provided a sample of two.

The editor in chief of the granddaddy of all polls cherrypicked a sample of two.

What is this, Monty Python?

No, that would be funny. This is pathetic.

Alan Abramowitz and Michael Cobb already made the same point but I'm shocked it hasn't been shouted a dozen or more times. I work in sports stats and I guarantee if I provided examples from '94 and '02 as representative while ignoring everything in between or previously, I would be ridiculed prior to being fired.

Where is '98? Is that AWOL like the '02 exit polls, or doesn't it fit snugly with our teeter totter agenda? This would be bad enough, a flimsy sample of two, if it were only the most recent examples, but when you skip one it smells and not a little bit.

Sorry for the tone but it's apropos. This is a second term midterm of a GOP president. Let's see those examples, from '74 and '86. Instead, we are handed the always-obvious GOP landslide from '94 and another one that broke hard toward the Republicans over the final week, with Bush's approval rating in the low 60s.

As a Democrat I am concerned that our anti-Bush, no clear message strategy will not encourage turnout and is overly dependent on a depressed GOP turnout, which I don't think will play out. So I agree with Frank Newport's basic conclusion that the national congressional vote will be much closer than the polls are indicating. But come on, I'm sure Mark Blumenthal would have allowed and encouraged a more thorough and relevant report.

This one takes me back to that advertisement in 2004, evaluating Gallup. I won't repeat the title but I'm sure many remember it. Yeah, Gallup got that one correct but that was when it correctly identified the '04 dynamic, that party ID had turned toward the Republicans and was dead even. Now, instead of isolating similar examples in which the party ID change and voter intensity level favored Democrats, or the out party, or anything similar to what we have in '06, Gallup prefers to identify the only two examples that off the top of my head I would not pick to evaluate '06.

____________________

Frank Newport

I will be happy to see if I can answer some of the comments which have been posted.


I chose the 1994 and 2002 elections as examples for no reason other than the fact that they demonstrated how significant a difference there has been at times in past midterm elections between the registered voter estimate and the likely voter estimate. I certainly wasn?t saying that this type of gap exists in every election. And I was careful to point out that it may not persist in this election. As I noted, it is not uncommon to find variations from poll to poll in the likely voter pool estimate compared to the registered voter estimate.

I think the point to be made is this: In any midterm election, one has to be very careful to assess the impact of turnout. There are enough examples historically to suggest that a difference between registered voter estimates and likely voter estimates should be a not-to-be overlooked possibility. Election observers should be cautious, as a result, in looking only at registered voter results.

Here are the data from some other years.

In 1982, Gallup?s October 15-18 poll showed an 18 point Democratic lead among registered voters, [54% to 36%] and a 10 point lead among likely voters [55% to 45%]. The final national two-party popular vote was 56% Democratic, 44% Republican.

In 1998 a Gallup poll conducted Sept. 11-12 had a 13 point Democratic lead among registered voters [52% to 39%] and a 1 point lead among likely voters [47% to 46%]. That difference changed as the race progressed. A poll conducted in late September (Sept. 23-24) had a 9 point among registered voters [50% to 41%] and a 6 point Democratic lead among likely voters [51% to 45%]. The final poll conducted Nov 2 had a 7 point Democratic lead among registered voters [48% to 41%] a tie among likely voters [47% to 47%]. The final national two-party popular vote was 50-50.

In terms of the question about differential turnout at the state level, that should be reflected in the national data. In other words, if this year there is higher than the national average turnout in California because of the governor?s race, higher than the average turnout in Virginia because of the Senate race, and lower than the average turnout in Arizona because of a lack of a highly competitive state race, then the assessment should produce a higher proportion of likely voters in California and Virginia, and a lower proportion of likely voters in Arizona, etc.

Finally, our historic analysis suggests that in almost every national midterm and presidential election, the estimates produced by the likely voter model have been as accurate as or more accurate than the estimates produced by looking at all registered voters. For example, Gallup?s final registered voter estimate in the 2004 presidential election was Kerry 48, Bush 46. Our final likely voter estimate was Bush 49, Kerry 47. (The final national vote was Bush 51, Kerry 48.) If the evidence didn?t suggest that the likely voter model produced more accurate estimates than the much simpler procedure of just using registered voters on average, we would take the easy way out and drop the extensive questioning and calculations necessary for producing likely voter estimates. To this point in time, that decision would not be warranted by the data

____________________

Alan Abramowitz:

The evidence concerning the performance of Gallup's likely vs. registered voter samples is actually more ambiguous than Frank Newport acknowledges. By selectively citing evidence from various midterm and presidential polls, Newport makes it appear that the LV estimate is almost always superior to the RV estimate. In fact, during the 2004 presidential campaign, the RV estimate was actually closer to the national popular vote in four of the last five Gallup polls, sometimes by a substantial margin. In the October 14-16 poll, the LV sample showed an 8 point Bush lead while the RV sample showed a 3 point Bush lead. In the October 22-24 poll, the LV sample showed a 5 point Bush lead while the RV sample showed a 2 point Bush lead. If we average the results from Gallup's last three polls, the RV estimate (Bush +1) was substantially closer to the results of the election than the LV estimate (Bush +5). The evidence from the 2004 presidential election does not indicate that Gallup's LV sample is more accurate than its RV sample when applied several weeks before an election.

____________________

E evidente che il luogo e stato fatto dalla persona che realmente conosce il mestiere!

____________________



Post a comment




Please be patient while your comment posts - sometimes it takes a minute or two. To check your comment, please wait 60 seconds and click your browser's refresh button. Note that comments with three or more hyperlinks will be held for approval.

MAP - US, AL, AK, AZ, AR, CA, CO, CT, DE, FL, GA, HI, ID, IL, IN, IA, KS, KY, LA, ME, MD, MA, MI, MN, MS, MO, MT, NE, NV, NH, NJ, NM, NY, NC, ND, OH, OK, OR, PA, RI, SC, SD, TN, TX, UT, VT, VA, WA, WV, WI, WY, PR