Polls and elections: preelection poll accuracy in the 2008 general elections.

Panagopoulos, Costas

Voters in the 2008 presidential election decidedly endorsed Democrat Barack Obama over his Republican opponent John McCain in the first-ever contest between two sitting U.S. senators. Amid growing economic turmoil and frustration with the war in Iraq and with two-term incumbent Republican president George W. Bush, Obama bested McCain at the polls by nearly 10 million votes. The historic election attracted 61.6% of the eligible electorate to the polls on election day--the highest turnout since 1968--and produced the first African American president of the United States (McDonald 2009). Obama ultimately captured 52.9% of the popular vote, while McCain received 45.7%. The Democratic nominee carried 28 states plus the District of Columbia and prevailed in Nebraska's Second Congressional District to garner 365 electoral votes; McCain earned 173 electoral votes from victories in 22 states. (1)

The 2008 election was the first election in 56 years that did not include an incumbent president or vice president. Both parties held spirited primaries that attracted a broad range of candidates, but the Democratic primary race that pitted former first lady and U.S. senator Hillary Clinton, the early front-runner, against Barack Obama was especially vivacious. The 17-month contest remained close throughout, and Obama did not secure enough delegates to clinch the nomination until June 3, 2008.

Despite noteworthy preelection polling setbacks during the primaries (most pollsters predicted an Obama victory in New Hampshire, for example, but Clinton won) (Liss 2008), the final national preelection polls in the general election unanimously projected a Democratic victory. Estimates of Obama's margin over McCain were among the most precise on record. In this essay, I analyze national and subnational preelection polls conducted in the 2008 cycle to present an in-depth evaluation of poll accuracy.

Preelection Poll Volume and Poll Aggregation in 2008

The number of preelection polls conducted during a typical presidential election cycle has grown substantially over the past few decades. Michael Traugott reports that the "explosion" in polls started in the 1980s and that the occurrence of the standard trial heat items in polls increased about 900% between 1984 and 2000 (2005, 644). The overall volume of preelection polling conducted during the 2008 cycle remained high. Based on national samples reported on Pollster.com, 221 non-overlapping presidential trial heats were conducted between Labor Day (September 1) and election day 2008, substantially more than in 2004 (although fewer than in the 2000 cycle). (2) Several polling organizations conducted daily tracking polls in 2008; Gallup and Rasmussen tracked preferences daily for at least six months leading up to the November election. In addition, DailyKos, George Washington University Battleground, Hotline, and Zogby provided daily tracking estimates during the fall campaign, and two additional organizations, IBD/TIPP and ABC News/Washington Post, reported daily tracking poll data toward the end of the campaign. A total of 975 trial heat items in statewide polls fielded between Labor Day and election day probed presidential preferences in 2008, while 394 trial heats reported statewide preferences for U.S. Senate candidates and 110 for gubernatorial candidates over the same duration.

One of the main developments in terms of preelection polling in 2008 was the emergence of online poll aggregators. Web sites such as FiveThirtyEight.com, Pollster.com, the Princeton Election Consortium, and RealClearPolitics.com collected, analyzed, and graphed an unprecedented amount of state and national polling data and helped voters interpret and assess campaign dynamics by providing singular, statistical measures that captured all of the available information daily. Poll aggregators used different mathematical models to contextualize polling information and to refine the predictive capabilities of available polling data (Harmanci 2008). Frequent visits to poll aggregator Web sites became a popular way to monitor preference dynamics in the 2008 cycle; in late-October 2008, Nate Silver, founder of FiveThirtyEight.com, reported that the Web site was receiving more than 600,000 hits per day (Harmanci 2008). RealClearPolitics.com, the most visited site in 2008, reportedly received more than 140 million page views in September alone (Becker 2008).

National Preelection Polls in 2008

National estimates of presidential preferences attracted perhaps the most attention consistently over the course of the 2008 campaign. Here, I assess how well these polls projected the eventual election outcome. Table 1 presents data on 20 final preelection estimates of the 2008 presidential vote based on national samples. Overall, pollsters fared quite well in the 2008 general election cycle. Obama's actual margin over McCain in the popular vote was 7.2 percentage points. All of the final poll estimates were within the range of [+ or -] 4 percentage points of the actual electoral result, and all showed Obama in the lead.

Several measures have been proposed in the literature to assess preelection poll accuracy. Two commonly used measures were developed by Frederick Mosteller and colleagues (1949). Mosteller's Measure 3 is the average absolute difference between the poll estimate for each of the leading candidates in the final estimate, while Mosteller's Measure 5 is the absolute value of the difference between the margin separating the two leading candidates in the poll and the difference in their margins in the actual vote (Traugott 2001, 2005).

An alternative method for assessing poll accuracy was developed by Elizabeth Martin, Michael Traugott, and Courtney Kennedy (2005). The new measure of predictive accuracy (A) is based on the natural logarithm of the odds ratio of the outcome in a poll and the actual election outcome (see Martin, Traugott, and Kennedy 2005 for a complete description). Among several advantages associated with this measure is the ability to compare accuracy across elections and polling firms and to detect the direction of bias because a signed statistic is produced (not an absolute value). A positive sign indicates a pro-Republican bias, while a negative sign indicates a pro-Democratic bias (Traugott 2005). (3)

Table 1 presents the values for the three measures for each of the final national 2008 polls that I evaluate. The average value for the Mosteller Measure 3 is 1.53 for the 20 polls included in the analysis, while the average value for the Mosteller Measure 5 is 1.55. Table 2 helps situate poll accuracy in 2008 in historical context by presenting summaries of Mosteller's Measures 3 and 5 for elections since 1956 (see Traugott 2005). The evidence reveals that the 2008 polls overall performed better than average based on the Mosteller Measure 3 (the average error for the 1956-2004 period was 1.9). Using the Mosteller Measure 5, preelection polls in 2008 were more accurate, on average, than in any presidential election cycle since 1956.

Although there is no comparable time series of values for the measure of predictive accuracy developed by Martin, Traugott, and Kennedy (2005), the authors computed the average values of the statistic for 1948, 1996, 2000, and 2004. They report that the average value of A for the final preelection polls conducted in 1996 was -0.0838, suggesting a slight Democratic bias that overestimated Bill Clinton's margin over Bob Dole. In 2000, polls overestimated support for George W. Bush, and the average value for A was 0.0630. For the 2004 election, the average value of A was -0.024, suggesting that Bush's electoral margin of victory was slightly underestimated (see Traugott 2005). Using the 20 polls that I analyze, the average value of A in 2008 was -0.013, indicating that polls reflected a slight Democratic bias that overestimated Obama's margin over McCain. Assuming a tied election, our estimate of bias implies that polls favored Obama by about one-third of a percentage point on average. (4) However, the standard error associated with the mean value for A that I report for the full sample of national polls is 0.009, indicating that the bias overall is not statistically significant; thus, the final national polls as a whole were not significantly biased in 2008. Moreover, based on this measure of poll accuracy, the average value of A was smaller in 2008 than in any of the preceding cycles for which estimates of the measure are available, suggesting that polls were more precise in 2008 than in 1996, 2000, and 2004. In fact, the gain in precision over the 2004 cycle was considerable; the average value of A in 2008 was about 50% smaller than in 2004.

Statewide Preelection Polls in 2008

As I noted earlier, state-level preelection polling was widespread across the nation during the 2008 cycle, with pollsters assessing statewide preferences for presidential as well as U.S. Senate and gubernatorial candidates. Following the elections, the National Council on Public Polls (NCPP) compiled and analyzed a compendium of 507 final state-level preelection polls conducted after October 15, 2008 (see NCPP 2008 for details and a complete list of polls included). The NCPP reported that most state polls (53.3%) were conducted by telephone using live interviewers, while 28.2% were conducted using Interactive Voice Response (IVR), 18.3% were Internet polls, and one poll was conducted by mail (NCPP 2008).

I use the complete set of polls included in the NCPP report to assess predictive accuracy in statewide polls in 2008. Using each poll as a single (unweighted) observation, I present the frequency distribution of A in Figure 1. I note that the polls include estimates of support for presidential as well as other statewide candidates (U.S. Senate and governor). In the absence of overall bias, I would expect the distribution to be centered on zero. The mean value of A in the complete sample of polls is -0.002, suggesting a negligible bias favoring Democratic candidates in statewide polls overall, but the bias is not statistically significant (standard error = .005). Assuming all races were perfectly tied, this would translate into a difference of merely 0.05 percentage point (see footnote 2).

The measure of predictive accuracy (A) developed by Martin, Traugott, and Kennedy (2005) permits us to compare accuracy across a range of poll characteristics. Table 3 presents mean levels of (A) and the corresponding standard errors for statewide polls grouped by a variety of characteristics, including election type, survey mode, sample type, interviewing period, and sponsor. I begin by comparing the poll performance of 11 individual polling organizations that conducted at least 10 statewide polls and that polled in multiple (more than three) states. (5) The evidence presented in Table 3 for each organization reveals that slight biases in a Democratic direction can be detected for statewide polls conducted by four organizations, while pro-GOP bias appears for the remaining seven polling entities. The standard errors associated with these estimates indicate that the biases are statistically insignificant across polling organizations, with one exception: polls conducted by Research 2000 in collaboration with the Democrati-cleaning blog DailyKos exhibit a significant bias in favor of Democratic candidates.

[FIGURE 1 OMITTED]

Our sample includes polls conducted by partisan as well as nonpartisan organizations. Mean values of A for statewide preelection polls conducted by Democratic and Republican organizations and for all nonpartisan polling firms combined are presented in Table 3. I detect hints of bias among partisan polls in 2008 in favor of the respective party's candidates, but only Republican polls are significantly biased toward GOP candidates in statewide polls. The pro-Democratic bias among the 59 Democratic organizations included in our sample (the mean value for A is -0.030) implies an average bias of 0.75 percentage point in favor of Democratic contenders in tied races, but the bias is not statistically significant (standard error = .019). By contrast, among the 18 polls that I analyze conducted by Republican polling organizations, the mean value for A is 0.042. The pro-Republican bias that I detect in polls conducted by Republican organizations is statistically significant (standard error = .016). In perfectly tied races, our estimate implies that Republican polls overestimated support for GOP candidates by 1.05 percentage points on average, relative to election day results. While this advantage is admittedly quite modest, biases of this magnitude can be important in close races. On the whole, nonpartisan polls reflect virtually no bias in 2008; the mean value for A is -0.0003, suggesting that nonpartisan polls overstated support for Democratic candidates by about 0.01 percentage point on average (given a tie), but this bias is, as expected, statistically insignificant (standard error = .006).

Using A, I can investigate bias in subsamples of polls by the level of electoral contests. In statewide presidential polls (N = 327), the mean value of A is -0.004, suggesting a slight but statistically insignificant (standard error = .006) bias favoring Obama. Similarly, statewide polls in U.S. Senate races (N = 145) reveal a slight but statistically insignificant pro-Democratic bias; the mean value of A is a mere -0.001 (standard error = .012). Statewide gubernatorial polls (N = 35) appear to reflect a modest bias in favor of Republican candidates--the mean value for A is 0.011--but the bias is also statistically insignificant (standard error = .032). I conclude from these results that overall bias was minimal in the final statewide polls conducted in 2008, whether examining statewide presidential, U.S. Senate, or gubernatorial preelection polls.

I turn next to examining overall bias by poll mode. The key results are presented in Table 3. Mean values of A across modes suggest that polls conducted via the Internet reflect the highest degree of bias. Among the 93 Internet polls, the mean value for A is 0.039, suggesting a statistically significant bias in the Republican direction (standard error = .014). IVR polls (N = 143) reveal the lowest degree of overall bias (0.018), but these polls appear to have significantly favored Republican candidates as well (standard error = .008). By contrast, the mean value of A is -0.027 (standard error = .008) for statewide polls conducted by telephone (N = 270), implying a significant bias in a Democratic direction.

There has been considerable debate in recent election cycles about the range of procedures employed by polling organizations in their estimations of likely voters in preelection polls (Erikson, Panagopoulos, and Wlezien 2004; Martin, Traugott, and Kennedy 2005). Table 3 displays mean levels of predictive accuracy (A) for three survey sample types that were included in the overall pool of 2008 statewide polls: samples of likely voters, registered voters, and adult populations. The initial evidence that I present suggests that likely voter samples were indeed more accurate than samples of registered voters or adults in 2008. The directions of the biases suggested by the estimates indicate pro-GOP bias only for registered voter samples, but the biases are statistically insignificant across sample types.

[FIGURE 2 OMITTED]

Longitudinal analysis can also be used to analyze the dynamics of poll accuracy, relative to electoral outcomes, and to gauge whether the predictive capacity of preelection polls improves as election day approaches. Scholarly evidence about the relationship between poll timing and accuracy is mixed; while some studies find that accuracy improves over the course of a campaign (Crespi 1988), others find no significant impact of poll timing on accuracy (Lau 1994; Martin, Traugott, and Kennedy 2005). Figure 2 presents lowess-smoothed levels of overall bias and accuracy in the 2008 statewide polls by the number of days until election day. The solid line presents the smoothed pattern of the absolute value of A over this period and suggests that preelection poll accuracy improved steadily within the final three weeks of the election cycle (the absolute value of A trended toward zero). Accuracy improved most dramatically during the final few days prior to the election. The dashed line in Figure 2 plots lowess-smoothed levels of mean predictive accuracy (A) over the same duration. The pattern suggests that statewide preelection polls initially reflected a pro-Democratic bias (two to three weeks prior to election day), but this bias eroded steadily; within the final few days of the cycle, statewide polls overall increasingly reflected a pro-Republican bias. These initial patterns suggests overall poll accuracy and bias can change over the course of a campaign, particularly during the final campaign period that I examine; substantively, however, these changes may not account for very much. I investigate this matter more rigorously next using multivariate techniques.

To explain overall levels of poll accuracy and bias more rigorously, I conduct a series of multivariate regression analyses. One advantage of A as a measure of predictive accuracy is that it (or variants of A as discussed later) is amenable to explanation using multivariate techniques that utilize various poll attributes as explanatory variables. The results of two such estimations are presented in Table 4. In both regressions, I include controls (fixed effects) for "house" (polling organization) and state effects. In Model 1, the dependent variable is the absolute value of A; as such, higher values represent less accurate poll estimates, relative to election outcomes. The results of the regression analysis reveal that statewide preelection polls for U.S. Senate candidates and gubernatorial candidates were significantly less accurate in 2008, compared to statewide presidential polls (the excluded category); all else being equal, levels of the absolute value of A were higher on average relative to presidential polls. Controlling for other factors, I also find that statewide preelection polls conducted via the Internet in 2008 were significantly less accurate than polls conducted by telephone (the excluded category), while the overall accuracy of mail and IVR polls did not differ significantly from phone polls. The results also suggest that registered voter samples were, in fact, more accurate than likely voters samples (the excluded category), while samples of adults were less accurate than likely voters samples, all else being equal. Once controls for other factors are incorporated, I find no evidence that polls conducted by nonpartisan organization were significantly more accurate overall compared with partisan polls. All else being equal, the analysis reveals poll timing did not significantly impact overall accuracy in statewide polls conducted in 2008.

Model 2 presents the results of a probit regression analysis in which the dependent variable is coded 1 if the preelection poll reflected a pro-Republican bias (A > 0) and 0 if the poll reflected a pro-Democratic bias (A < 0). Using the same poll attributes as in Model 1 to explain whether polls reflected a pro-Republican bias overall, I find that both U.S. Senate and gubernatorial polls were, all else being equal, significantly more likely to be biased in a Republican direction in 2008, compared to presidential polls (the excluded category). Controlling for other poll characteristics, Internet and IVR polls were significantly less likely than polls conducted by phone (the excluded category) to exhibit bias favoring Republican contenders. With controls in place for other factors, samples of registered voters were more likely than likely voter samples to exhibit bias in favor of GOP contenders, all else being equal. Nonpartisan polls overall were not significantly biased, relative to polls conducted by partisan organizations. Similarly, poll timing did not exert a significant impact on overall bias, all else being equal.

The multivariate analyses described here are useful in that they reveal the impact of a range of poll attributes on overall accuracy and the direction of bias while simultaneously controlling for other poll characteristics. In several instances, the results may cause analysts to reconsider, update, or confirm initial conclusions about the impact of various factors on accuracy and bias in preelection polls. For example, the interview period does not appear to influence overall levels of accuracy or bias. Moreover, the multivariate results suggest that preference estimates based on registered voter samples may actually be more accurate than likely voter samples, even as they are prone to pro-Republican bias. Curiously, although the value of multivariate approaches along these lines to explain poll bias and accuracy more rigorously is trumpeted in the literature (Martin, Traugott, and Kennedy 2005), such analyses are infrequently advanced. (6)

Discussion

Unprecedented interest in the 2008 election campaign and in preelection poll estimates of preference dynamics stimulated widespread awareness of persistent and emerging challenges associated with preelection polling and led to a national dialogue about poll quality and polling methodology. Concerns about polling procedures, including the estimation of likely voters, the presence of a so-called Bradley effect, Internet-based or IVR polls, and a variety of other considerations fueled tremendous speculation about eventual poll performance in 2008. The results that I present in this report indicate that, despite these apprehensions, polls across the board performed quite well in 2008. That said, concerns about poll methodologies should not be wholly and readily dismissed. While the evidence suggests that improvements in accuracy and declining bias in preelection polls as a whole, compared to previous election cycles, sources of inaccuracy and bias can also be detected using 2008 polls. Pollsters are wise to devote attention to monitoring and to investigating these sources further in the pursuit of still greater improvements in accuracy and bias.

References

Becker, Bernie. 2008. "Political Polling Sites Are in a Race of Their Own." New York Times, October 27, p. A14.

Crespi, Irving. 1988. Pre-Election Polling: Sources of Accuracy and Error. New York: Russell Sage Foundation.

DeSart, Jay, and Thomas Holbrook. 2003. "Campaigns, Polls and the States: Assessing the Accuracy of Statewide Presidential Trial-Heat Polls." Political Research Quarterly 56 (December): 431-39.

Erikson, Robert, Costas Panagopoulos, and Christopher Wlezien. 2004. "Likely (and Unlikely) Voters and the Assessment of Poll Dynamics." Public Opinion Quarterly 68 (Winter): 588-601.

Harmanci, Reyhan. 2008. "Poll Analysis Sites Put New Spin on Statistics." San Francisco Chronicle, October 18.

Lau, Richard. 1994. "An Analysis of 'Trial-Heat' Polls During the 1992 Presidential Election." Public Opinion Quarterly 58 (Spring): 2-20.

Liss, Sharon. 2008. "Good Night for Clinton But Not So Good for Pollsters, Pundits." Fox News, January 9. http://www.foxnews.com/politics/elections/2008/01/09/ good-night-for-clinton-butnot-so-good-for-pollsters-pundits/[accessed August 17, 2009].

Martin, Elizabeth, Michael Traugott, and Courtney Kennedy. 2005. "A Review and Proposal for a New Measure of Poll Accuracy." Public Opinion Quarterly 69 (Fall): 342-69.

McDonald, Michael P. 2009. "The Return of the Voter: Voter Turnout in the 2008 Presidential Election." The Forum 6 (4). http://www.bepress.com/forum/v016/iss4/art4 [accessed August 17, 2009].

Mosteller, Frederick, Herbert Hyman, Philip McCarthy, Eli Marks, and David Truman. 1949. The Pre-Election Polls of 1948: Report to the Committee on Analysis of Pre-Election Polls and Forecasts. New York: Social Science Research Council.

National Council on Public Polls (NCCP). 2008. "Analysis of Final Presidential Pre-Election Polls, 2008." http://www.ncpp.org/files/NCPP_2008_analysis_of_election_polls_121808 %20pdf_0.pdf [accessed August 17, 2009].

Traugott, Michael. 2005. "The Accuracy of the National Preelection Polls in the 2004 Presidential Election." Public Opinion Quarterly 69 (5): 642-54.

--. 2001. "Assessing Poll Performance in the 2000 Campaign." Public Opinion Quarterly 65 (Fall): 389-419.

COSTAS PANAGOPOULOS

Fordham University

(1.) Nebraska and Maine are the only two states do not adopt the traditional winner-take-all method of electoral vote allocation and permit electoral votes to be split based on candidate vote shares in congressional districts.

(2.) I am grateful to Mark Blumenthal from Pollster.com for providing 2008 data. See Martin, Traugott, and Kennedy (2005) for comparable data for previous cycles. For 2008, we include only non-overlapping poll releases; for tracking polls that reported three-day, rolling averages daily, we include only every third release.

(3.) Polls reflect no bias when A equals zero.

(4.) See Martin, Traugott, and Kennedy (2005, 11 n. 11) for a discussion and formula used to convert the parameter A to a percentage point difference.

(5.) This is consistent with the approach adopted in Martin, Traugott, and Kennedy (2005, 362).

(6.) See Lau (1994) and DeSart and Holbrook (2003) for exceptions analyzing presidential trial heat data.

Costas Panagopoulos is an assistant professor of political science and director of the Center for Electoral Politics and Democracy at Fordham University.

TABLE 1
Final National Preelection Poll Results and Poll
Accuracy Estimates, 2008

 Field
Firm Period McCain Obama Barr

Election Result 45.7 52.9
Democracy Corps (D) 10/30-11/2 44 51 2
FOX News/Opinion Dynamics 11/1-11/2 43 50 --
CNN/Opinion Research 10/30-11/1 4G 53 --
Ipsos/McClatchy 10/30-11/1 4G 53 --
American Research Group 11/1-11/3 45 53
IBD/TIPP 11/1-11/3 44 52 --
Harris Interactive 10/30-11/3 44 52 --
YouGov/Polimetrix 10/18-11/1 45 51 --
Pew 10/29-11/1 46 52 1
Rasmussen 11/1-11/3 46 52 --
NBC News/Wall Street
 Journal 11/1-11/2 43 51 --
GWU (Lake/Tarrance) 11/2-11/3 44 49 --
ABC News/Washington Post 10/30-11/2 44 53 --
Diageo/Hotline 10/31-11/2 45 50 --
DailyKos.com (D)/
 Research 2000 11/1-11/3 46 51 1
Marist College 11/3 43 52 --
CBS News 10/31-11/2 42 51 --
Gallup 10/31-11/2 44 55 --
Reuters/ C-SPAN/Zogby 10/31-11/3 43 54 --
CBS News/New York Times 10/25-10/29 41 52 --
Average

 Predictive
 Obama Accuracy
Firm Nader Other Lead (A)

Election Result 7.2
Democracy Corps (D) 1 1 7 -0.001
FOX News/Opinion Dynamics -- 2 7 -0.005
CNN/Opinion Research -- 7 0.005
Ipsos/McClatchy -- 1 7 0.005
American Research Group 8 -0.017
IBD/TIPP -- 4 8 -0.021
Harris Interactive 1 2 8 -0.021
YouGov/Polimetrix -- 2 6 0.021
Pew 1 6 0.024
Rasmussen -- 6 0.024
NBC News/Wall Street
 Journal -- 2 8 -0.024
GWU (Lake/Tarrance) -- 5 0.039
ABC News/Washington Post -- 2 9 -0.040
Diageo/Hotline -- 5 0.041
DailyKos.com (D)/
 Research 2000 1 0 5 0.043
Marist College -- 3 9 -0.044
CBS News -- 9 -0.048
Gallup -- 11 -0.077
Reuters/ C-SPAN/Zogby -- 11 -0.081
CBS News/New York Times -- 11 -0.091
Average -0.013

 Mosteller Mosteller
Firm Measure 3 Measure 5

Election Result
Democracy Corps (D) 1.80 0.20
FOX News/Opinion Dynamics 2.80 0.20
CNN/Opinion Research 0.20 0.20
Ipsos/McClatchy 0.20 0.20
American Research Group 0.40 0.80
IBD/TIPP 1.30 0.80
Harris Interactive 1.30 0.80
YouGov/Polimetrix 1.30 1.20
Pew 0.60 1.20
Rasmussen 0.60 1.20
NBC News/Wall Street
 Journal 2.30 0.80
GWU (Lake/Tarrance) 2.80 2.20
ABC News/Washington Post 0.90 1.80
Diageo/Hotline 1.80 2.20
DailyKos.com (D)/
 Research 2000 1.10 2.20
Marist College 1.80 1.80
CBS News 2.80 1.80
Gallup 1.90 3.80
Reuters/ C-SPAN/Zogby 1.90 3.80
CBS News/New York Times 2.80 3.80
Average 1.53 1.55

NOTE: To be consistent with previous years' analyses of poll
accuracy, I include poll estimates produced within the final
week of the election. The analysis excludes polls which
completed interviewing prior to October 29, 2008.

TABLE 2
Average Errors in Presidential Polls, 1948-2008

 Morteller Morteller
 Measure 3 Measure 5
 # of # of Average Average
Year Polls Candidate Error (%) Error (%)

2008 20 2 1.5 1.5
2004 19 2 1.7 2.1
2000 19 3 1.7 3.5
1996 9 3 1.7 3.6
1992 6 3 2.2 2.7
1988 5 2 1.5 2.8
1984 6 2 2.4 4.4
1980 4 3 3.0 6.1
1976 3 3 1.5 2.0
1972 3 2 2.0 2.6
1968 2 3 1.3 2.5
1964 2 2 2.7 5.3
1960 1 2 1.0 1.9
1956 1 2 1.8 3.5
Yearly
 Average
1956-2008 1.9 3.2

NOTE: Data for the 1956-2004 period were obtained from
Traugott (2005, 649). The 2008 update was compiled by
author.

TABLE 3
Mean Predictive Accuracy (A) by Poll Characteristics,
2008 Statewide Polls

 Mean
 Predictive
Poll Characteristics Number Accuracy Standard
(Type/Sponsor) of Pools (A) Error

Presidential 327 -.004 .006
U.S. Senate 145 -.001 .120
Governor 35 .011 .032
Democratic 59 -.030 .019
Republican 18 .042 .016
Nonpartisan 430 -.0003 .006
Internet 93 .039 .014
Phone 270 -.027 .008
IVR 143 .018 .008
Likely voters 377 -.007 .006
Registered voters 123 .015 .011
Adults 7 -.041 .065

Sponsor *
AP-GfK 14 -.055 .041
ARG 16 -.030 .027
Allstate/National
 Journal/FD 11 -.011 .019
DailyKos (D)/
 Research 2000 28 -.085 .034
Mason-Dixon 25 .034 .019
Public Policy Polling (D) 30 .018 .015
Rasmussen 44 .016 .014
Research 2000 16 .044 .039
Strategic Vision (R) 15 .031 .016
SurveyUSA 49 .017 .014
YouGov/Polimterix 84 .028 .014

* Consistent with Martin, Traugott, and Kennedy (2005), only
polling organizations that conducted at least 10 statewide
polls and that polled in multiple (three or more) states in
2008 are included in the analysis.

TABLE 4
The Impact of Poll Attributes on Bias and Accuracy in
statewide Preelection Polls, 2008

 Model 2:
 Model 1: Pro-Republican
Independent variables Accuracy Bias
(Poll characteristics)
U.S. Senate .031 *** .448 **
 (.008) .183
Governor .079 *** 1.129 ***
 (.015) (.355)
Internet .114 *** -4.943 ***
 (.035) (1.850)
IVR -.010 -8.266 ***
 (.058) (.962)
Mail -.046 --
 (.075)
Registered voters -.291 *** 6.433 ***
 (.101) (1.435)
Adults .157 ** --
 (.075)
Nonpartisan Polling .012 -1.247
 organization (.041) (1.178)
Days to election .000 -.054
 (.002) (.041)
Constant .127 ** -6.533 ***
 (.056) (2.183)
N 507 384
[R.sup.2]/Pseudo [R.sup.2] .53 .28
Log-likelihood -- -191.787

Notes: Model 1: OLS; dependent variable is the absolute
value of A. Model 2: Probit; dependent variable = 1 if A >
0, and 0 if A < U. Fixed effects for polling organization
("house") and state are included. Observations with
covariate patterns that predict outcome perfectly are
excluded from the model, resulting in the smaller number of
cases.

*** Signifies statistical significance at the p < .01 level,

** p < .05, two-tailed. Standard errors in parentheses.