The crystallization of voter preferences during the 2008 presidential campaign.

Erikson, Robert S. ; Panagopoulos, Costas ; Wlenzien, Christopher 等

Even as the media dwell on--and speculate about the potential impact of--the many events that unfold daily during the course of presidential election campaigns, certain fundamental variables powerfully structure the election day vote. At the individual level, party identification is of great importance, increasingly so in recent years (Bartels 2000). (1) Other factors also matter at the individual level, including socioeconomic class and issue positions. Short-term forces such as economic conditions also shape the vote. On election day, voters tend to line up as political scientists predict they will.

To a growing number of scholars, the primary function of election campaigns is to deliver the so-called fundamentals (Andersen, Tilley, and Heath, 2005; Arceneaux 2005; Finkel 1993; Gelman and King 1993; Stevenson and Vavreck 2000; for a more nuanced view, see Vavreck 2009). Finkel (1993) shows that much of the change in presidential vote preference during the 1980 campaign was attributable to "activation" of political predispositions, where voters bring their preferences in line with their preexisting partisan and racial identities. Relatedly, Gelman and King's (1993) analysis of 1988 shows that the effects of party identification and various demographic variables on presidential vote preferences increased during the election year. Andersen, Tilley, and Heath (2005) show much the same pattern during the 1997-2001 and 2001-2005 general election cycles in the United Kingdom, where the effects of election day vote predictors increased as the campaign progressed; Arceneaux (2005) demonstrates a similar pattern across a set of nine European countries. (2) In Gelman and King's (1993) terminology, campaigns appear to "enlighten" voters about which candidates best represent their interests.

This model of campaign effects implies that voters' preferences come into focus during the course of the campaign. That is, preferences crystallize. Early on, voters are expected to be largely unformed. During this period, voters may simply be undecided. They also may express weak preferences. Some may "flirt" with candidates from the other party. As the campaign unfolds, according to the model, voters increasingly support the candidate of their preferred party. Voters' preferences also increasingly reflect their true "interests." This crystallization may be driven by a number of mechanisms. Following Gelman and King (1993), campaigns help voters learn which candidate best represents their interests, and this leads typically them back to their partisan attachments. (3) As Alvarez (1998) argues, uncertainty about candidates' issue positions is reduced as campaigns unfold. Alternatively, following Finkel (1993), the campaigns may activate voters' political predispositions, as they can learn that the candidates really are partisans, representing Democratic and Republican position and interests. Yet another possibility is that campaigns lead voters to perceive events using their partisan screens, prompting them to view candidates of their own party more favorably (and increase the intensity of their support) as the campaign unfolds. The latter mechanism is implied in The American Voter (Campbell et al. 1960) and in subsequent work showing the effects of partisanship on perceptions (Bartels 2002; Wlezien, Franklin, and Twiggs 1997).

The general pattern is clear: voters' preferences come into focus over the course of presidential election campaigns. The specifics, however, are less clear. What is the exact pattern of crystallization? That is, is it concentrated during the intense general election campaign after the conventions? Or does it play out over the longer campaign leading up to the conventions? These questions remain largely unanswered in the extant literature, and evidence is elusive. Our aim in this paper is to take a small step forward, focusing specifically on developments during the 2008 election campaign.

The 2008 presidential election was unusual in several respects. It was the first election in 56 years that did not include as a candidate either an incumbent president or vice president, and the first time the general election pitted two sitting U.S. senators against each other. Open contests in both parties attracted contenders who spanned the sociopolitical spectrum, including viable female (Hillary Clinton, Democrat), Mormon (Mitt Romney, Republican), and Latino (Bill Richardson, Democrat) candidates. Ultimately, Americans endorsed Democrat Barack Obama at the polls on election day, electing the nation's first African American president. Despite these peculiarities, however, the 2008 presidential cycle was typical, at least in the sense that the outcome was largely predictable, months in advance, based on "fundamental" elements such as the economy and presidential approval. With a souring economy, an unpopular GOP incumbent, George W. Bush, and unpopular wars in Afghanistan and Iraq, most analysts successfully forecasted the Democratic victory that ensued. Even still, presidential preferences evolved over the course of the campaign. In this article, we examine how their underlying structure changed.

On the Crystallization of Voter Preferences

If there is crystallization, as we have discussed, we will observe that preferences are increasingly structured over the course of the campaign. We remain agnostic about the hard question--what voters' interests should be. Instead, we want to see how the vote equation comes into focus over the yearlong campaign. From this point of view, crystallization implies that the variables that account for the vote on election day are less important at the campaign's outset but then become increasingly important predictors of vote choice over the course of the election year.

For this exercise, assume that the actual election day vote equation consists of party identification plus a vector of other unspecified variables. Now, imagine that we have surveys of voters on each day t of the campaign. In these surveys, voters are asked about which presidential candidate they prefer, their family income, and various other things. We can represent intended vote choice (V) for individuals i over the election cycle as follows:

[V.sub.i,t] = [[alpha].sub.i] + [[beta].sub.1] [[Party ID].sub.i,t] + [[beta].sub.1] [SIGMA] [Z.sub.j,i,t] + [[epsilon].sub.i,t], (1)

where Party ID designates an individual's party identification and Z is a set of other (j) variables that tap each individual's characteristics and interests; t equals the number of days into the campaign.

If campaigns cause voters to learn, we will observe two things. First, the coefficients ([beta]) will increase as the number of days into the campaign (t) increases. Second, the fit of the equations will likewise increase as t increases. The amount of increase indicates the amount of crystallization. In the extreme, all of it occurs during the election year. That is, vote choice will show no structure at the beginning of the year. In Equation 1, the [beta]s will equal 0 and vote choice will be random. This is unlikely, and we expect some initial crystallization at the beginning of the year. But how much? How much occurs thereafter? We also want to know the pattern of crystallization. Does it occur linearly over time? Or is it more pronounced during some periods than others? What explains the variation?

As we have already discussed, the literature offers only very general answers to these questions. Plus, the evidence that we do have is from a handful of elections more than 20 years ago. We simply do not know the specific pattern of crystallization and whether it is the same for different election years. In this article, we are interested in seeing what happened during the 2008 election campaign. Some features of the campaign--an incumbent was not running and one of the candidates was a relative newcomer to the national stage--worked in favor of finding a great deal of preference crystallization during the election year. In other words, there was a lot for voters to learn. Other factors--partisan sorting and polarization and the stretching of the primary campaign period--may have accelerated the process so that voter preferences became more partisan earlier than usual in the campaign cycle.

Data and Methodology

To test for crystallization, ideally, we would have panel data over the full election year. We then could assess the evolution of individuals' own preferences over time. Such data were unavailable for this project, making it necessary to rely on repeated cross-sections. With these data, we cannot actually observe individuals' preferences evolve during the election year; we can, however, observe aggregate patterns and, more importantly, examine moving cross-sections of individual-level data to observe how the vote choice equation evolves. This way, it is possible to analyze how the relative impact of the various factors that explain preferences (or intended vote choice) changes over the course of the campaign. That is, we can assess whether and how the coefficients in Equation 1 change.

The demographic variables on the right-hand side of the equation are virtually constant over the campaign. The electorate has a near-constant demographic composition over the course of a campaign. And people do not change their demographic characteristics as a function of vote choice. This resistance to change is less clear in the case of subjective variables, such as party identification, and individuals may increasingly bring their partisanship in line with their vote choice. This is crystallization, but it is not of the kind imagined by proponents of activation or enlightenment discussed earlier. While there is little denying that party identification has an endogenous component--for example, as voters sort themselves in accordance with their ideologies (e.g., Levendusky 2009)--the fact is that party identification is highly stable at the individual level (see, e.g., Erikson, MacKuen, and Stimson 1998; Gerber and Green 2002).

For our analyses, we rely on individual-level survey data collected by the Pew Research Center for People and the Press over the course of the 2008 campaign for president. Pew surveyed Americans at 13 regular intervals between February 24 and November 1, 2008. In each of these "trial heat" polls, Pew asked respondents about their presidential preferences in an Obama-McCain matchup. Pew's final, national pre-election presidential preference estimates were among the more accurate in 2008 (Panagopoulos 2009b). Restricting the analyses to data gathered by the same survey organization alleviates concerns about "house effects" that may compromise the comparability of poll findings from multiple sources (Erikson and Wlezien 1999). (4) All data we report here are weighted and restricted to registered voters. Because of the special appeal of Obama to racial minorities, we focus our analysis on white non-Hispanics.

Key Findings

Let us begin with an overall picture of preference dynamics over the duration of the campaign. For this exercise, we focus on the "percent for Obama" among those who indicated they would vote for either Obama or McCain, ignoring minor candidates and undecideds. Figure 1 displays the numbers separately for non-Hispanic white voters, minorities, and all respondents. Here we can see that the level of support differed fairly consistently across groups but that it tended to vary together over time--the average correlation between the three series is 0.61. Overall support for Obama was steady over much of the campaign, drifting a few points up here and down there, before the sharp climb in mid-September 2008. This presumably reflects, at least partly, growing voter dissatisfaction with Republicans as the economic downturn intensified during this period.

We are interested here in examining the evolving structure of voter preferences--the factors that influenced voter preferences over the campaign and how, if at all, their impact changed over the course of the 2008 campaign. For our analysis, we use probit regression to predict support for Obama (coded 1 throughout) over McCain (coded 0), excluding respondents who preferred other candidates. We remind readers that the analysis is limited to non-Hispanic white voters. The independent variables include a set of salient demographic characteristics that Pew ascertained from respondents in all 13 Pew polls--age, gender, education, income, religious affiliation, and residency in a Southern state or rural area. For each of the 13 surveys, we model non-Hispanic white vote choice as a function of these demographics. For a second set of equations, we model the non-Hispanic white vote as a function of these demographics plus party identification and ideology. (5)

[FIGURE 1 OMITTED]

We adopt the following coding scheme for the key variables discussed earlier. We code partisanship as a set of two dummy variables for Democratic and Republican identification. All independents, including "leaners" (e.g., independents who admit to leaning toward one of the parties), make up the base category. We code ideology as a set of two dummy variables for liberal and conservative identification (combining "very liberal" with "liberal" and "very conservative" with "conservative") responses; "moderate" responses are the excluded category. On coding, we note the following for other variables: Education is categorized as follows: "low education" respondents claim less than a high school education; "high education" respondents consist of those with a college degree. Those in the middle range on education, including high school graduates and those with some college education, make up the base category. Income is categorized as follows: low = $0-30,000 per year, high = $75,000+; the middle categories make up the base. Age is coded as 18-29-year-olds and respondents older than 65 years old, with ages in between as the base category. Religious affiliations are coded as a series of dummy variables for Catholic and Jewish, with others (mainly Protestants) as the base category. "Born again" respondents responded affirmatively to the question of whether they indeed consider themselves born again. Seculars are those who answered "no religion/atheists" to the survey item on religious affiliation. "Rural" is coded dichotomously based on the population density variable provided by Pew.

Figures 2 and 3 display the estimated impact (coefficients are indicated by circles, with vertical bars representing the associated standard errors) for each trait separately for each survey. Figure 2 presents coefficients over time for the equation including only demographics, while Figure 3 displays effects when partisanship and ideology are added. (6) The horizontal lines represent the pattern of effects over time for each characteristic.

The results in Figure 2 suggest that the estimated effects of most demographic characteristics remained stable over the duration of the campaign, although we detect some movement for Catholics and Evangelicals toward McCain and Jews toward Obama. These shifts are clearest in the first set of equations, in which demographics are entered alone (Figure 2). They largely disappear when party and ideological identification are added in the second set of equations. (7) This is clear in Figure 3. Here, we can also observe growth in the impact of party identification, noting that the impact of the party coefficients represents the effect of Democratic and Republican identification relative to self-identification as an independent. (8) Predictably, the "Democrat" coefficient grows increasingly positive and the "Republican" coefficient increasingly negative. That is, Democratic identifiers were increasingly likely to support Obama as the campaign progressed, and Republicans were increasingly likely to support McCain, compared to independents. While both major parties' identifiers initially may have held reservations about their parties' standard-bearers in the election, their vote preferences eventually evolved to coincide with their partisan identities.

Next we consider the overall distributions of predicted presidential preferences. Figure 4 presents histograms of the predicted values (preferences) yielded by two separate probit regressions estimated for each survey. In the left-most column, we present histograms of fitted values predicted by probit regressions including only demographic variables; the right-hand column shows estimates produced by adding partisanship and ideology to demographic attributes are presented. (9) We also report the sample sizes (N), variance of the predicted values, and the pseudo [R.sup.2] for each probit accordingly. The evidence reinforces the findings we reported earlier but offers a more nuanced portrait.

[FIGURE 2 OMITTED]

[FIGURE 3 OMITTED]

[FIGURE 4 OMITTED]

The predictions yielded by the probit equations containing only demographic variables reveal only a modest amount of crystallization at any point in the campaign. That is, predicted vote preferences based solely on demographics tend to be clustered in the middle of the distribution. This is especially true early in the election year, during the spring nomination season. We observe some spreading out through the summer, though the tendency is modest. Based on demographics alone, preferences look to be substantially unstructured even close to election day.

Predictions including partisanship and ideology show much more structure throughout the campaign. Early in the election year, the pattern is asymmetrical. At that point in time, the probit regressions were better at locating McCain supporters (compared with Obama supporters). Perhaps this reflected the fact that, entering the campaign, Obama was much less well known nationwide than McCain, and that this changed through the nomination season as he emerged as the front-runner. Regardless of the details, the polarization of preferences is clear. It also was largely complete by summer, well before the national conventions, and remained so for the duration of the post-Labor Day general election campaign.

The McKelvey-Zavoina pseudo [R.sup.2] values reported in Figure 4 show a similar pattern. (10) The amount of explained variance in the models that include only demographic characteristics is substantially lower than that in models expanded to include partisanship and ideology. Pseudo-[R.sup.2] values for the demographics-only model range from a low of 0.10 in the March 22 and April 28 surveys to a high of 0.29 on October 13, while the amount of explained variance in the full model ranges from a low of 0.51 in the March 22 survey to a high of 0.76 in the September 29 survey. On average, the explanatory power of the full models is more than three times greater than the demographics-only models over the 13 surveys we examine, 0.63 to 0.19, respectively. Adding partisan and ideological identification adds appreciably to the explained variance, as expected. Overall, we observe real increases in both sets of models' predictive capacities over time. While growth in the pseudo-[R.sup.2] is slightly larger with the full model in terms of amount of variance explained, the more modest pseudo-[R.sup.2] for the lesser, demography-only model roughly doubles in size. These findings support our contention of meaningful crystallization of preferences during the 2008 presidential campaign.

Discussion and Conclusions

Researchers demonstrate that although big shifts in voter preferences over the course of presidential campaigns are rare, small movements in preferences are common (Wlezien and Erikson 2002; Panagopoulos 2009a). Even though vote preferences may not change dramatically during campaigns, the relative weight of factors that explain presidential preferences may change as a campaign evolves and as voters learn about the candidates and the available options. In this paper, we assessed the crystallization of voter preferences during the 2008 presidential race. As expected, we found that the structure of preferences did evolve over time and that most of the evolution was attributable to the increasing impact of partisan identity. We also demonstrated that the evolution was largely complete during the summer, before the fall general election campaign even began, a key finding that speaks to the temporal dimensions of contemporary presidential campaigns.

From a practical perspective, early crystallization suggests that the ability of candidates to seek crossover support may be maximized well in advance of election day, with persuading voters becoming a more daunting challenge toward the end of the campaign. Ironically, this is precisely when campaign activity and mass media attention are most intense. Thus, these findings have implications not only for our scholarly understanding of evolving electoral choice, but also for the campaign practices of candidates and parties themselves. Of course, the extent to which the findings generalize to other presidential campaign cycles remains an open question. We speculate that it is the campaign that induces the crystallization process, and that the characteristics of campaigns (e.g., their length) and the political context (e.g., partisan sorting and polarization) may have sped up the process. More research is required to understand this process over multiple election cycles and different electoral settings. What we know for sure is that the absence of an incumbent and the presence of a less well-known candidate did not hinder or delay the evolution of preferences during the 2008 campaign.

Appendix: Probit Estimation Details

Demographics-Only Models

Variable              2/24           322            4/28

Education (low)    0.12 (0.26)   -0.33 (0.26)   -0.08 (0.29)
Education (high)   0.06 (0.12)   -0.19 (0.13)   -0.24 (0.12)
Income (low)       0.04 (0.16)    0.29 (0.15)    0.16 (0.15)
Income (high)     -0.11 (0.12)    0.13 (0.12)   -0.09 (0.12)
Age (18-29)        0.01 (0.25)    0.26 (0.18)   -0.01 (0.19)
Age (65+)         -0.10 (0.13)    0.01 (0.13)   -0.27 (0.12)
Female             0.24 (0.11)    0.19 (0.11)    0.32 (0.10)
South             -0.23 (0.11)   -0.43 (0.12)   -0.22 (0.12)
Catholic           0.21 (0.13)    0.19 (0.13)    0.15 (0.13)
Jew                1.18 (0.34)    0.11 (0.39)    0.79 (0.29)
Evangelical       -0.76 (0.37)   -0.35 (0.37)    0.14 (0.33)
Secular            0.65 (0.16)    0.48 (0.17)    0.56 (0.16)
Rural             -0.15 (0.11)   -0.12 (0.11)   -0.05 (0.11)
(Intercept)       -0.28 (0.15)   -0.20 (0.15)   -0.22 (0.14)

Variable              5/24           629            7/27

Education (low)    0.64 (0.26)   -0.01 (0.22)    0.39 (0.19)
Education (high)   0.11 (0.14)   -0.12 (0.08)    0.03 (0.09)
Income (low)      -0.10 (0.16)    0.14 (0.10)    0.18 (0.11)
Income (high)     -0.05 (0.13)   -0.11 (0.08)   -0.38 (0.09)
Age (18-29)        0.14 (0.23)    0.05 (0.10)   -0.14 (0.13)
Age (65+)         -0.05 (0.13)   -0.08 (0.09)    0.12 (0.09)
Female             0.04 (0.12)    0.20 (0.07)    0.26 (0.08)
South             -0.54 (0.13)   -0.07 (0.07)   -0.23 (0.08)
Catholic           0.23 (0.14)   -0.02 (0.09)    0.13 (0.10)
Jew                0.41 (0.40)    0.24 (0.24)    0.15 (0.20)
Evangelical       -0.69 (0.51)   -0.57 (0.09)   -0.74 (0.10)
Secular            0.70 (0.17)    0.64 (0.12)    0.47 (0.12)
Rural             -0.21 (0.12)   -0.20 (0.07)    0.06 (0.08)
(Intercept)       -0.08 (0.16)   -0.01 (0.10)   -0.07 (0.11)

Variable              8/10           9/14           9/29

Education (low)    0.21 (0.13)    0.48 (0.16)   -0.60 (0.28)
Education (high)   0.14 (0.06)    0.01 (0.06)   -0.23 (0.09)
Income (low)       0.12 (0.08)    0.08 (0.08)   -0.17 (0.12)
Income (high)     -0.10 (0.06)   -0.17 (0.06)   -0.18 (0.09)
Age (18-29)        0.21 (0.09)    0.33 (0.09)   -0.17 (0.14)
Age (65+)          0.01 (0.07)   -0.04 (0.07)   -0.33 (0.10)
Female             0.28 (0.06)    0.32 (0.05)    0.31 (0.08)
South             -0.27 (0.06)   -0.31 (0.06)   -0.40 (0.09)
Catholic           0.05 (0.07)   -0.09 (0.07)   -0.44 (0.11)
Jew                0.43 (0.27)    0.95 (0.25)    0.83 (0.44)
Evangelical       -0.52 (0.07)   -0.72 (0.07)   -0.81 (0.10)
Secular            0.47 (0.09)    0.50 (0.09)    0.44 (0.12)
Rural             -0.08 (0.06)   -0.09 (0.05)    0.03 (0.08)
(Intercept)       -0.18 (0.08)   -0.07 (0.07)    0.35 (0.12)

Variable             10/12          10/19          10/26

Education (low)    0.08 (0.27)   -0.11 (0.21)    0.26 (0.35)
Education (high)   0.16 (0.11)    0.05 (0.08)    0.08 (0.11)
Income (low)       0.12 (0.16)    0.46 (0.11)    0.29 (0.17)
Income (high)     -0.31 (0.12)   -0.14 (0.08)   -0.13 (0.13)
Age (18-29)        0.12 (0.18)    0.21 (0.13)    0.30 (0.17)
Age (65+)          0.07 (0.14)   -0.11 (0.09)   -0.02 (0.15)
Female             0.42 (0.11)    0.21 (0.08)    0.26 (0.11)
South             -0.34 (0.12)   -0.34 (0.08)   -0.39 (0.12)
Catholic           0.10 (0.14)   -0.15 (0.10)   -0.04 (0.14)
Jew                0.60 (0.33)    0.78 (0.23)    1.19 (0.40)
Evangelical       -0.97 (0.15)   -0.85 (0.10)   -0.76 (0.15)
Secular            0.33 (0.17)    0.23 (0.13)    0.60 (0.18)
Rural             -0.14 (0.11)   -0.16 (0.08)   -0.24 (0.12)
(Intercept)        0.03 (0.16)    0.23 (0.11)    0.12 (0.16)

Variable              11/1

Education (low)    0.23 (0.20)
Education (high)   0.00 (0.08)
Income (low)       0.46 (0.10)
Income (high)     -0.06 (0.08)
Age (18-29)        0.07 (0.12)
Age (65+)         -0.10 (0.09)
Female             0.18 (0.07)
South             -0.25 (0.08)
Catholic          -0.10 (0.09)
Jew                0.74 (0.24)
Evangelical       -0.71 (0.09)
Secular            0.53 (0.12)
Rural             -0.19 (0.07)
(Intercept)        0.10 (0.10)

Demographics, Partisanship, and Ideology Models

Variable              2/24           3/22           4/28

Democrat           0.76 (0.15)    1.16 (0.16)    0.60 (0.14)
Republican        -1.28 (0.18)   -0.87 (0.16)   -1.43 (0.17)
Liberal            0.36 (0.29)    0.69 (0.25)    1.39 (0.25)
Conservative       0.82 (0.19)    0.51 (0.19)    0.87 (0.18)
Education (low)    0.11 (0.29)   -0.49 (0.26)   -0.33 (0.34)
Education (high)   0.07 (0.15)   -0.01 (0.15)   -0.32 (0.13)
Income (low)      -0.09 (0.22)    0.26 (0.17)    0.06 (0.17)
Income (high)     -0.07 (0.14)    0.23 (0.14)    0.04 (0.15)
Age (18-29)       -0.02 (0.32)    0.43 (0.21)    0.21 (0.23)
Age (65+)         -0.18 (0.16)   -0.04 (0.15)   -0.29 (0.14)
Female             0.08 (0.12)    0.07 (0.12)    0.28 (0.13)
South             -0.14 (0.14)   -0.22 (0.14)   -0.22 (0.14)
Catholic          -0.11 (0.17)    0.23 (0.15)    0.08 (0.15)
Jew                0.56 (0.31)   -0.55 (0.32)    0.42 (0.27)
Evangelical       -0.79 (0.38)    0.02 (0.47)    0.65 (0.43)
Secular            0.35 (0.20)    0.06 (0.21)    0.20 (0.20)
Rural             -0.19 (0.14)   -0.10 (0.13)    0.01 (0.13)
(Intercept)       -0.09 (0.18)   -0.50 (0.19)   -0.15 (0.19)

Variable              5/24           6/29           7/27

Democrat           0.85 (0.15)    1.01 (0.10)    1.00 (0.11)
Republican        -1.23 (0.19)   -1.07 (0.10)   -1.36 (0.14)
Liberal            1.00 (0.39)    1.51 (0.24)    1.05 (0.22)
Conservative       0.74 (0.20)    1.03 (0.14)    0.53 (0.15)
Education (low)    0.53 (0.29)    0.10 (0.23)    0.74 (0.15)
Education (high)   0.19 (0.19)   -0.23 (0.09)    0.31 (0.11)
Income (low)       0.05 (0.18)    0.13 (0.12)    0.14 (0.13)
Income (high)     -0.01 (0.16)    0.18 (0.10)   -0.08 (0.11)
Age (18-29)        0.05 (0.26)    0.03 (0.12)    0.08 (0.18)
Age (65+)         -0.18 (0.14)    0.17 (0.11)    0.04 (0.10)
Female            -0.03 (0.14)    0.01 ((.09)    0.12 (0.09)
South             -0.57 (0.16)   -0.02 (0.09)   -0.33 (0.10)
Catholic           0.08 (0.17)   -0.05 (0.12)   -0.09 (0.13)
Jew               -0.36 (0.37)   -0.58 (0.20)   -0.25 (0.24)
Evangelical       -0.52 (0.62)   -0.25 (0.12)   -0.54 (0.13)
Secular            0.21 (0.20)    0.24 (0.14)    0.07 (0.14)
Rural             -0.23 (0.14)   -0.20 (0.09)    0.07 (0.10)
(Intercept)       -0.02 (0.20)   -0.23 (0.15)   -0.19 (0.14)

Variable              8/70           9/14           9/29

Democrat           1.14 (0.08)    1.35 (0.08)    1.80 (0.14)
Republican        -1.23 (0.08)   -1.40 (0.10)   -1.26 (0.14)
Liberal            0.43 (0.22)    1.03 (0.26)    1.86 (0.34)
Conservative       0.82 (0.11)    0.76 (0.11)    1.19 (0.23)
Education (low)    0.07 (0.18)    0.45 (0.25)   -0.32 (0.29)
Education (high)   0.18 (0.08)    0.03 (0.08)   -0.14 (0.12)
Income (low)      -0.05 (0.10)    0.09 (0.10)   -0.14 (0.18)
Income (high)      0.02 (0.08)   -0.08 (0.08)   -0.07 (1.11)
Age (18-29)        0.38 (0.11)    0.32 (0.12)   -0.10 (0.19)
Age (65+)          0.15 (0.08)   -0.05 (0.09)   -0.54 (0.13)
Female             0.04 (0.07)    0.19 (0.08)   -0.02 (0.11)
South             -0.30 (0.07)   -0.34 (0.08)   -0.38 (0.11)
Catholic           0.04 (0.09)   -0.06 (0.10)   -0.43 (0.14)
Jew                0.28 (0.37)    0.56 (0.20)    0.16 (0.53)
Evangelical       -0.29 (0.09)   -0.25 (0.09)   -0.68 (0.12)
Secular            0.24 (0.11)    0.31 (0.13)    0.13 (0.21)
Rural             -0.05 (0.07)   -0.08 (0.07)    0.08 (0.12)
(Intercept)       -0.25 (0.11)   -0.29 (0.10)    0.15 (0.16)

Variable             10/12          10/19          10/26

Democrat           1.47 (0.20)    1.20 (0.13)    0.95 (0.18)
Republican        -1.54 (0.17)   -1.66 (0.12)   -1.55 (0.17)
Liberal            0.54 (0.27)    0.68 (0.34)    1.20 (0.29)
Conservative       0.98 (0.27)    0.84 (0.18)    0.75 (0.27)
Education (low)   -0.25 (0.36)   -0.30 (0.21)    0.55 (0.44)
Education (high)   0.36 (0.16)    0.11 (0.11)    0.22 (0.15)
Income (low)      -0.02 (0.20)    0.44 (0.14)    0.19 (0.22)
Income (high)     -0.33 (0.17)    0.02 (0.12)    0.03 (0.16)
Age (18-29)        0.06 (0.24)    0.16 (0.16)    0.48 (0.25)
Age (65+)         -0.08 (0.18)   -0.25 (0.11)    0.23 (0.17)
Female             0.23 (0.15)   -0.05 (0.10)    0.13 (0.14)
South             -0.40 (0.15)   -0.38 (0.11)   -0.49 (0.16)
Catholic           0.04 (0.19)   -0.11 (0.13)   -0.10 (0.18)
Jew               -0.43 (0.44)    0.28 (0.36)    0.62 (0.35)
Evangelical       -0.56 (0.18)   -0.50 (0.13)   -0.56 (0.19)
Secular            0.25 (0.23)   -0.02 (0.16)    0.17 (0.25)
Rural             -0.16 (0.14)   -0.13 (0.11)   -0.35 (0.15)
(Intercept)        0.07 (0.22)    0.30 (0.15)    0.27 (0.18)

Variable              11/1

Democrat           1.34 (0.11)
Republican        -1.25 (0.12)
Liberal            1.43 (0.35)
Conservative       1.05 (0.18)
Education (low)    0.01 (0.26)
Education (high)   0.03 (0.10)
Income (low)       0.36 (0.14)
Income (high)      0.11 (0.10)
Age (18-29)        0.15 (0.16)
Age (65+)         -0.15 (0-12)
Female             0.02 (0.10)
South             -0.22 (0.10)
Catholic          -0.04 (0.12)
Jew                0.25 (0.26)
Evangelical       -0.35 (0.12)
Secular            0.25 (0.17)
Rural             -0.17 (0.10)
(Intercept)       -0.07 (0.14)

Notes: Tables report coefficients with standard errors in parentheses.
Probit regressions. Samples restricted to registered, non-Hispanic,
white voters. Data weighted. Dependent variable coded 1 if respondent
expressed preference for Obama, and 0 for McCain. See Figure 4 for
fit statistics and sample sizes.

AUTHORS' NOTE: We are grateful to Yair Ghitza for invaluable assistance with the data and analysis. We also thank Jeff Cohen for his helpful direction and comments.

References

Alvarez, R. Michael. 1998. Information and Elections. Ann Arbor: University of Michigan Press.

Andersen, Robert, James Tilley, and Anthony Heath. 2005. "Political Knowledge and Enlightened Preferences: Party Choice through the Electoral Cycle." British Journal of Political Science 35 (April): 285-302.

Arceneaux, Kevin. 2005. "Do Campaigns Help Voters Learn: A Cross-National Analysis." British Journal of Political Science 36 (January): 159-73.

Bartels, Larry. 2000. "Partisanship and Voting Behavior, 1952-1996." American Journal of Political Science 44 (January): 35-50.

--. 2002. "Beyond the Running Tally: Partisan Bias in Political Perceptions." Political Behavior 24 (June): 117-50.

Campbell, Angus, Philip E. Converse, Warren E. Miller, and Donald E. Stokes. 1960. The American Voter. New York: Wiley.

Dalton, Russell J., and Martin Wattenberg. 2000. Parties without Partisans: Political Change in Advanced Industrial Democracies. New York: Oxford University Press.

Erikson, Robert S., Michael B. MacKuen, and James A. Stimson. 1998. "Macropartisanship." American Political Science Review 83 (December): 1125-42.

Erikson, Robert S., and Christopher Wlezien. 1999. "Presidential Polls as a Time Series: The Case of 1996." Public Opinion Quarterly 63 (Summer): 163-77.

Finkel, Steven E 1993. "Reexamining the 'Minimal Effects' Model in Recent Presidential Elections." Journal of Politics 55 (February): 1-21.

Gerber, Alan, and Donald Green. 2002. Partisan Hearts and Minds: Political Parties and the Social Identities of Voters. New Haven, CT: Yale University Press.

Gelman, Andrew, and Gary King. 1993. "Why Are American Presidential Election Polls So Variable When Votes Are So Predictable?" British Journal of Political Science 23 (July): 409-51.

Holbrook, Thomas M. 1996. Do Campaigns Matter? Thousand Oaks, CA: Sage.

Holmberg, Soren. 2007. "Partisanship Reconsidered." In Oxford Handbook of Political Behavior, eds. Russell J. Dalton and Hans-Dieter Klingemann. Oxford: Oxford University Press, 557-70.

Levendusky, Matthew. 2009. The Partisan Sort: How Liberals Became Democrats and Conservatives Became Republicans. Chicago: University of Chicago Press.

Mair, Peter, and Ingrid van Biezen. 2001. "Party Membership in Twenty European Democracies, 1980-2000." Party Politics 7 (January): 5-21.

McKelvey, Richard D., and William Zavoina. 1975. "A Statistical Model for the Analysis of Ordinal Level Dependent Variables." Journal of Mathematical Sociology 4: 103-20.

Panagopoulos, Cosras. 2009a. "Campaign Dynamics in Battleground and Nonbattleground States." Public Opinion Quarterly 73 (April): 119-30.

--. 2009b. "Polls and Elections: Preelection Poll Accuracy in the 2008 General Elections." Presidential Studies Quarterly 39 (December): 897-908.

Stevenson, Randolph, and Lynn Vavreck. 2000. "Does Campaign Length Matter? Testing for Cross-National Effects." British Journal of Political Science 30 (April): 217-35.

Vavreck, Lynn. 2009. The Message Matters: The Economy and Presidential Campaigns. Princeton, NJ: Princeton University Press.

Wlezien, Christopher, Mark N. Franklin, and Daniel Twiggs. 1997. "Economic Perceptions and Vote Choice: Disentangling the Endogeneity." Political Behavior 19 (March): 7-17.

Wlezien, Christopher, and Robert S. Erikson. 2002. "The Timeline of Presidential Election Campaigns." Journal of Politics 64 (November): 969-93.

ROBERT ERIKSON

Columbia University

COSTAS PANAGOPOULOS

Fordham University

CHRISTOPHER WLEZIEN

Temple University

(1.) The opposite may be true in other countries, where the level and effect of party identification have been on the decline (Dalton and Wattenberg 2000; Holmberg 2007; Mair and van Biezen 2001).

(2.) Some research shows that this is more likely the longer the campaign (Stevenson and Vavreck 2000). Other research suggests that big campaign events such as conventions and debates serve the purposes of learning, helping voters bring their vote choices in line with the fundamentals (Holbrook 1996).

(3.) Interest presumably is at least part of the reason people hold the attachments in the first place.

(4.) The survey dates that we report are for the final day the survey was in the field. The complete survey periods for the surveys we analyze are 2/20-2/24, 3/19-3/22, 4/23-4/28, 5/21-5/25, 6/18-6/29, 7/23-7/27, 7/31-8/11, 9/9-9/14, 9/27-9/29, 1019-10/13, 10/16-10/19, 10/23-10126, and 10/29-11/1. For additional information about the polls conducted by Pew in 2008, see http://www.people-press.org.

(5.) We note that measures of other items (e.g., issue preferences, candidate attributes) were not consistently asked by Pew and therefore were unavailable for inclusion as additional variables.

(6.) See Appendix 1 for complete estimation results for Equations 1 and 2.

(7.) Evidently the modest increase in religion effects on the vote operated through party and ideological identification.

(8.) We remind readers that we classify "leaners" with independents in our analyses.

(9.) Estimates range from 0 (far left; support for McCain) to 1 (far right; support for Obama) accordingly.

(10.) McKelvey and Zavoina (1975) derived their pseudo [R.sup.2] explicitly to show the ratio of explained to unexplained variance in the latent variable.

Robert S. Erikson is a professor of political science at Columbia University.

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

Christopher Wlezien is a professor of political science and faculty affiliate in the Institute for Public Affairs at Temple University.