See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228802065

The timing of voting decisions in presidential campaigns

Article · January 1999

CITATIONS READS
7 233

2 authors:

Janet Box-Steffensmeier David C Kimball

The Ohio State University University of Missouri - St. Louis
91 PUBLICATIONS   5,368 CITATIONS    80 PUBLICATIONS   1,694 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Voter Fraud Perceptions View project

All content following this page was uploaded by Janet Box-Steffensmeier on 28 May 2014.

The user has requested enhancement of the downloaded file.

 The Timing of Voting Decisions in Presidential Campaigns

Janet M. Box-Steffensmeier
Department of Political Science
Ohio State University
2140 Derby Hall
154 N. Oval Mall
Columbus, OH 43210-1373
Phone: 614-292-9642
Fax: 614-292-1146
E-mail: [email protected]

and

David Kimball
Department of Political Science
Southern Illinois University
Carbondale, IL 62901
Phone: 618-453-3191
Fax: 618-453-3163
E-mail: [email protected]

Prepared for presentation at the 1999 annual meeting of the Midwest Political Science
Association, Chicago. We thank Andrew Tomlinson for excellent research assistance and John
Kessel and John Brehm for helpful comments.
 Abstract

Voting analysts often make a distinction between "long-term" and "short-term" forces that
influence the voting decision in presidential elections (Campbell et al. 1960). Long-term forces
reflect information and considerations that are available to voter before the presidential campaign
starts, such as party identification, demographic attributes, and the record each candidate has
established previously in government. In contrast, short-term forces refer more specifically to the
campaign.
We posit that there is variation in the way voters integrate the long- and short-term forces
into a voting decision. Furthermore, the long-term forces are smaller in number and thus easier
for researchers to identify and measure. For example, much attention has been devoted to
conceptualization and measurement of party identification. However, short-term forces are
nearly infinite in number and are much harder to measure and link up to the voting decision.
This means that voting models should perform well when predicting the choices of voters who
are guided primarily by long-term forces. In contrast, voting models should not perform as well
for citizens who are strongly influenced by short-term forces. In statistical terms, there will be
heteroskedastic error variance in common vote models due to the differing influence of short-
and long-term forces.
We examine the variation among voters by using the standard NES question that asks
citizens how long before the election they made their voting decisions and test our expectations
using the heteroskedastic probit technique (Brehm and Alvarez 1995), which is like a standard
probit model except that there is a separate equation to model the error variance (the errors in
prediction). By using the timing of the vote decision to help model the error variance, we
produce unbiased estimates and improve our ability to explain voting behavior and the impact of
campaigns.
 Voting analysts often make a distinction between "long-term" and "short-term" forces that

influence the voting decision in presidential elections (Campbell et al. 1960). Long-term forces

reflect information and considerations that are available to the voter before the presidential

campaign starts. Long-term forces usually refer to political attitudes (such as party identification

and ideological preferences) and demographic attributes (such as race, religious affiliation, and

union membership) that are relatively stable over the long haul. Long-term forces may also

include the record each candidate has established previously in government, e.g., as a member of

Congress, as a governor, or as president for one seeking reelection. The period of governing has

been called the "long campaign" to denote its importance in subsequent elections

(Box-Steffensmeier and Franklin 1995). These long-term forces shape the voting decision and

often lead voters to develop a habitual pattern of voting for the same party every four years.

In contrast, short-term forces refer more specifically to the campaign and contemporary

events, and do not favor the same party every election. Traditionally, attitudes toward the

candidates and specific issue positions are considered short-term forces on the voting decision,

since the candidates and the salient issues change from election to election. Other short-term

factors include discussions with family members or co-workers during the campaign, exposure to

campaign advertising, or other contact with one of the campaigns.

We posit that there is variation in the way voters integrate the long- and short-term forces

into a voting decision. Specifically, some voters are influenced entirely by long-term forces,

while others are more influenced by short-term forces. For example, a strong partisan may have

his mind made up before the campaign begins, and he will likely interpret all the short-term

events of the campaign in a way that reinforces his partisan predispositions. In contrast, an

1
independent voter may base her vote more on the day-to-day events of the actual campaign.

Voters who are more influenced by short-term forces in a campaign may be more uncertain or

conflicted about their ultimate decision, which could produce an unequal error variance in

statistical models of the vote choice. We think it is important for voting models to account for

uncertainty and conflicting attitudes, and we aim to do so by estimating a heteroskedastic probit

model.

In the next section of the paper we elaborate on the ideas of the long- and short-campaign

forces and their effects on the error variance of a model of presidential voting decisions. Next

we discuss the usefulness of the heteroskedastic probit model to address concerns about unequal

variance across observations. In the third section of the paper we present some data on the

timing of voter decisions in the 1988 presidential elections. There is a discussion of the results

of the heteroskedastic probit model of voting decisions in the fourth section and the role of

timing, knowledge, and attitude consistency in the error variance part of the model is highlighted.

Finally, we conclude with the import of our study.

Long- and Short-term Factors

Long-term forces are smaller in number and easier for researchers to identify and

measure. For example, much attention has been devoted to conceptualization and measurement

of party identification. However, short-term forces (which can include various characteristics of

the candidates, a discussion with a neighbor, or a particular television ad) are nearly infinite in

number and are much harder to measure and link up to the voting decision. This means that

voting models should perform well when predicting the choices of voters who decide early and

2
are guided primarily by long-term forces. In contrast, voting models should not perform as well

for citizens who are strongly influenced by short-term forces, because it is difficult to account for

all possible short-term forces and the events of the fall campaign may heighten attitude conflict

in the mind of the voter.

Citizens who decide which candidate to vote for prior to the start of the general election

campaign base their vote solely on long-term forces; they decide before any of the short-term

campaign events occur. By comparison, people who decide shortly before election day are likely

to be more influenced by something that took place during the campaign.1 This timing argument

is the basis of the claim that “the general election campaign affected voters’ choices” (Abramson,

Aldrich, and Rohde 1995, 64).

The timing of the vote decision may also be related to the level of conflict between the

various considerations that go into that decision. Those who wait longer to decide tend to be

ambivalent between competing candidates because some considerations favor one party while

other considerations favor another. In one of the earliest uses of regression analysis to examine

voting behavior, the authors of The American Voter (1960) used six "partisan attitude"

consistency measures to predict presidential vote choice. Campbell, Converse, Miller, and

1
We recognize that the date of decision is subject to recall problems. Campbell et al.
(1960) respond by saying that regardless, the responses do “undoubtedly give a sense of the
location in time of individual choices” (1960, 78). One possibility is that the recalled date of
decision is primed by the pre-election interview. However, we find that the correlation between
pre-election interview date and the time of decision reported in the post-election wave is not
statistically different from zero (r = -0.04, p = 0.13). Another possible bias might be a related to
the date of the post-election interview as suggested by Wright’s (1993) bandwagon hypothesis.
However, there is convincing evidence against the bandwagon hypothesis in Gronke (1992) and
Box-Steffensmeier, Jacobson, and Grant (1999). Thus we do not think the hypothesis is
applicable to the question of when voters made their decisions.

3
Stokes (1960) found that “the proportion deciding late always is greater as the degree of attitude

conflict is greater” (1960, 82). The idea of attitude inconsistency manifesting in late voting

decisions can be traced to Lazarsfeld, Berelson, and Gaudet (1944). Lazarsfeld et al. (1944)

devote a chapter to examining “time of final decisions,” from which we posit that as political

knowledge increases, voters are less likely to be ambivalent.

The explained variance for Campbell et al.’s (1960) regression equation is very large for

people who made their voting decisions before the campaign ever started, but it declines

dramatically when the sample is limited to those who decided during the campaign. Similarly,

when analyzing the vote for Nixon, Humphrey, and Wallace and dividing the electorate by the

timing of their vote decision, Kessel (1973) can explain roughly two-thirds of the variance for

respondents who decide early and one-third for late deciders.

Kessel (1968a), theorizing about different types of citizens at different stages of the long

and short campaign, provides insight into the different degrees of success in modeling the

explained variance. “The first persons to decide are traditionalists, persons whose attitudes

strongly predispose them to cast Democratic or Republican votes and who do so almost as a

matter of course. Next come the activists, persons who are themselves involved in politics and

who tend to acquire information about the candidates and issues rather early” (1968a, 283). We

contend that both traditionalists and activists are likely to be heavily influenced by party and

ideology. Certainly activists will acquire information and use it, for example, to decide which

campaigns to volunteer to work for, however, they are not deciding between a Democratic and

Republican candidate. Rather they are deciding between the Republican candidate running for

U.S. Senator and the Republican candidate running for U.S. Representative. “During the

4
campaign proper ‘policy-oriented’ citizens pay close attention to what the candidates and parties

are saying, and tend to make issue-based decisions. Finally, just before the election, weakly

motivated voters finally have their attention caught by the mounting hoopla of the campaign.

The inclination of these ‘peripheral citizens’ are hard to predict because of their lack of real

involvement, and they are rather easily moved from one camp to another” (1968a, 283).

In other work Kessel (1968b) eloquently states that, “there might be different orders of

calculus depending on the complexity of the voter’s choice. A first order decision, the normal

partisan decision, is the simplest and the most frequent. Little information or psychological

effort is required to vote for the candidate of the party with which one identifies. A second order

decision calculus involves a more elaborate attitudinal structure. Here one decides on the basis

of attitudinal proximity in a more inclusive system. The insufficiencies of these explanations,

specifically in choices concerning a weak third party candidate and in late decisions, suggest the

existence of some third order decision. It may be that this third order decision involves

mechanisms we have not yet included in our explanatory model, but we should also be mindful

of the possibility that random chance becomes more significant as decisions become more

difficult” (1968b, 111).

All of this work points to the statistical conclusion that there will be heteroskedastic error

variance in common vote models due to the differing influence of short- and long-term forces (as

indicated by the reported time of the vote decision), as well is variation in attitude conflict and

uncertainty about the available choices. Variation in attitude conflict also means that errors in

prediction will be smaller for voters with no apparent conflict and larger for conflicted voters.

The picture that Kessel (1968a, 1968b) paints strongly suggests an error variance that is cone

5
shaped over the course of the campaign. That is, the errors grow over time as the decision

process extends through the campaign season. This problem of heteroskedastic errors or unequal

variance across observations can be addressed with a heteroskedastic probit model

The Heteroskedastic Probit Model

There are methodological and substantive reasons for using a heteroskedastic probit

model to estimate vote choice. First, there is a major statistical problem with using standard vote

choice models, such as logit and probit, when heteroskedasticity is present. Namely, the

maximum likelihood estimates are inconsistent (Greene 1993). However, the inconsistency can

be addressed by directly taking into account the heterogeneity of the process under study

(Alvarez and Brehm 1995).

The heteroskedastic probit model can accommodate our expectations about the error

variance, i.e., that it increases over the course of the campaign based on the differences between

early and late deciders. The heteroskedastic probit model is like a standard probit model except

that there is a separate equation to model the error variance or the errors in prediction. The log-

likelihood function for the heteroskedastic probit is:

6
Alvarez and Brehm (1995) point out that the difference between the likelihood for the

heteroskedastic probit and a regular probit “is the inclusion of the variance model in the

denominator” for the former (1995, 1062). Thus, in addition to a statistical model with variables

predicting vote choice, the heteroskedastic probit model includes a set of covariates to account

for variation in the error variance across observations.

Based on the discussion above, we expect the error variance in a model of voting to be a

function of (1) the short campaign, (2) voter uncertainty about the candidates and parties, and (3)

conflicting attitudes about the major-party alternatives. First, we expect that the error variance

increases for late deciders, those voters most susceptible to the short campaign. Second, voters

with more political knowledge should be more certain the qualities of the competing candidates

and thus the error variance should decrease as political knowledge increases. Similarly, we

expect a smaller error variance for respondents who can see a difference between what the major

parties stand for. Third, the error variance is expected to decrease as the absolute value of the

difference between the number of pro-Republican and pro-Democrat responses grows. That is,

if the respondent can provide favorable responses about both parties, there is more likely to be

conflicting attitudes about which party’s candidate to support. This expectation is based on the

idea that attitude conflict leads to greater variation (and less predictability) in vote choices.

Finally, a decrease is expected for respondents that say they see a difference between the parties.

We examine the variation among voters by using National Election Study (NES) data, in

particular the standard NES question that asks citizens how long before the election they made

their presidential voting decisions. The exact question wording is: “How long before the election

did you decide that you were going to vote the way you did?” The question follows the series of

7
questions: "Did R[espondent] Vote for President in 1988 election?" then "Who did R vote for in

1988 election?" and "Was R's preference Strong or Not Strong?"

The Timing of Voter Decisions in Presidential Elections

Table 1 presents information about when citizens made their voting decisions in

presidential election campaigns from 1948 to 1996. The timing of the vote decision is broken

into three categories: before the end of the conventions, from the end of the conventions to three

weeks before the election, and less than three weeks before the election. We see a trend of more

people deciding late in the campaign in more recent elections, which is consistent with on work

by Flanigan and Zingale (1994). Such a trend increases the importance of the campaign. In

addition to looking for patterns over time, we see that there are election characteristics that affect

timing. For example, if the election is a landslide more voters appear to decide early. Similarly,

there are more early deciders in races with an incumbent, a candidate with whom voters are

already familiar.

[Table 1 About Here]

Because of the shifting nature of some presidential campaigns, there often is a

relationship between the timing of the voting decision and the candidate chosen (Abramson,

Aldrich and Rohde 1995; Campbell 1999; Stein 1998). Table 2 provides some evidence of

shifting voter alliances over time from the 1988 NES. Early deciders in 1988 slightly preferred

George Bush to Michael Dukakis. Those who decided during the middle portion of the general

election campaign (after the conventions but before the final two weeks) favored Bush even more

heavily. This period coincided with some of the Bush campaign’s hardest-hitting television ads

8
(e.g., Willie Horton, Boston Harbor, Dukakis riding a tank). Finally, those who decided at the

end of the campaign actually leaned toward Dukakis, who at this time point was the likely loser.

Thus, we have some tentative evidence that the reported timing of the voting decision may

capture the effect of the campaign.

[Table 2 About Here]

We now turn to testing our expectations about the effect of the timing of voting decisions on the

error variance by using the heteroskedastic probit technique.

Results for the Heteroskedastic Probit Model of Presidential Voting Decisions

In this section, we estimate a heteroskedastic probit model of presidential vote choice in

1988. Our dependent variable is a dichotomous indicator of the candidate chosen (with Bush

coded as 1, Dukakis as 0). Our voting model includes a standard set of predictors, including

party identification, ideology, demographic characteristics, and evaluations of the economy,

President Reagan, and the two candidates. These variables are described in more detail in the

appendix.

A second component of the model includes a set of variables to account for the size of the

error variance. These explanatory variables include the time of the vote decision, a standard

political knowledge scale, a measure of perceptions of party differences, and the net number of

pro-Democratic versus pro-Republican responses in open-ended questions about the parties and

candidates. Specifically, timing is used to test the expectation of increasing error variance over

the course of the campaign. We also expect that greater political knowledge, and a perception of

party differences, is associated with smaller error variances in the vote choice model. Finally, the

9
more one-sided the open-ended responses about the two parties (indicating minimal attitude

conflict), the smaller the error variance. By using the timing of the vote decision and other

factors to help model the error variance, we produce unbiased estimates and improve our ability

to explain voting behavior and the impact of campaigns.

Table 3 presents the results for the heteroskedastic probit model of vote choice between

Bush and Dukakis in the 1988 presidential election.2 First, the goodness-of-fit test examines the

null hypothesis that the coefficients of the model are equal to zero. We see the heteroskedastic

probit model is statistically significantly better than the null model: P 2 = 975.29, p<.001.3 More

interesting, the heteroskedasticity test examines the zero restrictions on the error variance

coefficients imposed by the probit model compared to the heteroskedastic probit model. The

probit model assumes no heteroskedasticity; the heteroskedastic probit model allows the

researcher to model the variance. There is a statistically significant improvement: P 2 = 67.79,

p<.001 (see Alvarez and Brehm 1995 for more details on the test for heteroskedasticity). So we

prefer the heteroskedastic probit model, specifically its treatment of the variance.

[Table 3 About Here]

The vote choice model for the 1988 election is presented in the upper part of Table 3. We

see an effect of political attitudes and the party record. Demographic attributes did not have a

statistically significant effect once the error variance is modeled. Party identification with the

Republican party has a positive and statistically significant effect on voting for Bush.

2
Limdep 7.0 is used for the analysis. Data, code, and output are available from the
authors.
3
Not surprising, the goodness-of-fit test of the probit model versus a null model also
shows that the probit model is statistically significantly better.

10
Respondents who were more favorable to Ronald Reagan as indicated by the Reagan feeling

thermometer were also more likely to vote for Bush. The more favorable traits that the

respondent identified for Bush, the more likely to vote for Bush. Similarly, the more favorable

traits identified for Dukakis, the less likely the respondent would vote for Bush. Economic

approval, race, gender, union membership, and ideology are not statistically significantly

different from zero.

The second part of the heteroskedastic probit model “estimates the error variance in the

binary vote choice part of the model” (Alvarez and Brehm 1995, 1065). Considering the timing

variable first, we see confirmation that the error variance increases for late deciders. The

decisions of voters affected by the short campaign are harder to explain (there is a larger error

variance) because the factors of the short campaign are larger in number and harder to identify

and measure. The short campaign factors include, for example, the campaign styles and

characteristics of the candidates, a radio talk show about the campaign issues, or a conversation

with a neighbor.

The variance model also shows that the more knowledgeable the voter, the less

ambivalent he or she is about choosing a candidate. This result contrasts with work on policy

choices where there is conflict between core beliefs. That is, Alvarez and Brehm (1995) find that

more information does not make a respondent less ambivalent on abortion policy. Finally, the

error variance decreased if the voter could give more positive comments about one of the two

parties, i.e., the number of pro-Republican and pro-Democrat answers were counted, then the

absolute difference between the two numbers was taken. The more evenly balanced the

comments about the parties were, the more likely that there are conflicting attitudes about which

11
party’s candidate to support, resulting in increased error variance. Finally, the error variance in

our model appears to be unrelated to whether voters see any differences between the two parties.

Conclusion

Based on our knowledge about the long- and short-campaign and the variation in the way

that voters integrate the factors that affect voting decisions during the stages of a presidential

election, we expect that the error variance will increase for voters deciding later in the course of

the election cycle. A heteroskedastic probit model confirmed these expectations about the error

variance in a model of the 1988 presidential elections.

Voters vary in the weights attached to long- and short-term factors. Long-term factors,

such as political attitudes, demographic attributes, and the party record, reflect the information

and considerations that are available to the voter before the presidential campaign even begins.

Short-term factors focus on the day-to-day campaign events. Our evidence of the effect of the

timing variable provides support for these ideas as well as for the argument that the campaign has

an effect, although not the sort of effect typically contemplated by scholars. We argue that the

campaign has a subtle effect on the variation (or uncertainty) associated with voting decisions. If

the campaign did not have an effect, we would not see the increase in variance according to the

timing of the decision in our model. The timing of the vote decision is also related to the level of

attitude consistency. Those who wait longer to decide may also be ambivalent between

competing candidates because some considerations favor one party while other considerations

favor another. Finally, political information reduced the error variance in our model.

Clearly, not all voters are the same in the way they integrate a variety of factors into a

12
voting decision. It is important for voting behavior research to incorporate differences in the

extent to which voters are influenced by campaign events, as well as differences in voter

knowledge and attitude conflict. As a result, some voters have more uncertainty associated with

their decisions and a wider range of potential choices than other voters, and these expectations

can be included in a model of vote choice.

13
 Appen dix

Dependent variable:

Presidential vote choice: a dichotomous variable indicating whether the respondent voted for
Bush (=1) or Dukakis (=0).

Independent variables:

Reagan feeling thermometer: range = [0,100]

Economic Approval: approve of Reagan’s handling of the economy: 1 = Disapprove Strongly, 2

= Disapprove Not Strongly, 3 = Approve Not Strongly, and 4 = Approve Strongly.

Party Identification: a six category variable coded such that higher values of the variables are
more Republican. So 0=strong Democrat, 1=weak Democrat, 2=Independent that leans
Democrat, 3=Independent, 4=Independent that leans Republican, 5=weak Republican, and
6=strong Republican.

Bush and Dukakis traits: answers to questions about positive traits, such as "How well does
trustworthy describe _______ ?" were recoded into dichotomous variables (the trait fits or it
does not) and then summed for each candidate. There is a standard battery of such items on the
NES. The traits in the questions are: intelligent, compassionate, moral, inspiring, provides strong
leadership, decent, really cares about people like you, knowledgeable, and honest.

Ideology: R’s self-reported ideology. The scale is from 0 to 9, with 0 being that the respondent
never said that any of the traits fit the candidate.

Race: dichotomous variable with non-white = 1.

Gender: dichotomous variable with female =1.

Union member: dichotomous where someone in the family belongs to a labor union = 1.

Ideology: self-reported ideology, ranges from 1 to 7 with higher values being more conservative.
1 = Extremely Liberal, 2 = Liberal, 3 = Slightly Liberal, 4 = Moderate or Middle of Road, 5 =
Slightly Conservative, 6 = Conservative, 7 = Extremely Conservative

Time of vote decision: three categories of before the end of the conventions, from the end of the
conventions to three weeks before the election, and less than three weeks before the election.

14
Political knowledge: The number of correct answers to political knowledge questions were
summed, i.e., what office does Jim Wright hold? Who has a majority in the House of
Representatives? A total of nine questions were used, so the variable ranges from 0 to 9.

Partisanship of answers: the number of pro-Republican and pro-Democrat answers were

counted, then the absolute difference between the two numbers was taken (range from 0 to 85).
Trait questions about Reagan were included but any questions that were used in creating other
variables, such as the Bush traits variable, were excluded. Subtracting the scores is simpler and
more intuitive than multiplying the scores. The weakness is that someone with two zeroes gets
the same score as someone with 25 pro-Republican and 25 pro-Democratic responses, while it
can be argued that the latter person would be more ambivalent. The multiplicative variable also
suffers from the problem of different combinations giving the same score, but the problem is not
as severe, e.g., 4x4 is 16 but so is 8x2. Standardizing by dividing by the total number of
responses (or by the number of responses in the largest category) helps alleviate that problem.
We ran the partisanship variable both ways and found similar results so we present the simpler
subtracted scores.

Differences between parties: a dichotomous variable, 1= the respondent says that they see
differences between the Republican and Democratic parties.

15
 References

Abramson, Paul R., John H. Aldrich, and David W. Rohde. 1995. Change and Continuity in the
1992 Elections. Washington, D.C.: CQ Press.

Alvarez, R. Michael, and John Brehm. 1995. “American Ambivalence Towards Abortion
Policy: Development of a Heteroskedastic Probit Model of Competing Values.”
American Journal of Political Science 39(4): 1055-1082.

Box-Steffensmeier, Janet M., and Charles H. Franklin. 1995. “The Long Campaign: Senate
Elections in 1992.” In Democracy’s Feast: Elections in America, Herbert F. Weisberg,
ed. Chatham, NJ: Chatham House Publishers.

Box-Steffensmeier, Janet M., Gary Jacobson, and J. Tobin Grant. 1999. “Question Wording and
the House Vote Choice: Some Experimental Evidence.” Ohio State University.
Typescript.

Campbell, James E. 1999 “The Competitive Effect of Presidential Campaigns.” Presented at the
Annual Meeting of the Midwest Political Science Association, Chicago, April 15-17,
1999.

Campbell, Angus, Philip E. Converse, Warren E. Miller, and Donald E. Stokes. 1960. The
American Voter. New York: Wiley; reprinted 1980 by University of Chicago Press.

Flanigan, William H., and Nancy H. Zingale. 1994. Political Behavior of the American
Electorate, 8th ed. Washington, D.C.: CQ Press.

Greene, William H. 1993. Econometric Analysis. 2nd Edition. New York: MacMillan
Publishing Company.

Gronke, Paul. 1992. “Overreporting the Vote in the 1988 Senate Election Study: A Response to
Wright.” Legislative Studies Quarterly 17: 113-29.

Kessel, John H. 1968a. The Goldwater Coalition: Republican Strategies in 1964. The Bobbs-
Merrill Company, Inc.

Kessel, John H. 1968b. “Pennsylvania.” In Explaining the Vote: Presidential Choices in the
nation and the States, 1968, David M. Kovenock, James W. Prothro, and Associates, eds.
Institute for Research in Social Science. University of North Carolina at Chapel Hill.

16
Lazarsfeld, Paul F., Bernard Berelson, and Hazel Gaudet. 1944. The People’s Choice: How the
Voter Makes up his Mind in a Presidential Campaign. New York: Columbia University
Press.

Stein, Robert M. 1998. “Early Voting.” Public Opinion Quarterly 62: 57-69.

Wright, Gerald C. 1993. "Errors in Measuring Vote Choice in the National Election Studies."
American Journal of Political Science 37:291-316.

17
 Table 1: Time of Vote Decision from 1948 to 1996

1948 1952 1956 1960 1964 1968 1972 1976 1980 1984 1988 1992 1996
Decided by end of 72.0 67.7 78.5 62.8 69.2 60.4 62.5 54.1 58.6 69.5 60.4 53.8 65.5
conventions
Decided between 15.4 20.9 11.7 25.1 19.7 18.6 21.8 21.8 15.2 16.9 22.2 21.9 16.1
conventions and last
two weeks of
campaign
Decided during last 12.6 11.4 9.8 12.1 11.1 21.0 13.7 23.8 25.8 13.3 17.3 24.3 18.1
two weeks of
campaign
Other* 2.1 0.3 0.4 0.3 0.1 0.3
N 382 1202 1230 877 1248 1016 1487 1332 988 1395 1200 1657 1131

* Includes answers coded as “other” as well as answers in 1972 coded as “when Eagleton was dropped from the Democratic ticket,”
“when Kissinger announced that peace was at hand,” and “when Alabama Governor George Wallace was shot.”

18
 Table 2: Time of Vote Decision by Vote Choice, 1988

Timing of Voting Decision

Early: Middle:
Vote for President Before conventions are over After conventions Late: Last two weeks
Bush 52.9% 57.4% 46.4%
Dukakis 47.1% 42.6% 53.6%
Number of cases 719 256 194

N=1169
Chi-square, 2 df = 5.39, p = .067

19
 Table 3: Heteroskedastic Probit Results of the Presidential Vote in 1988

Vote Model Coefficient Standard Error |Z| p |Z|$ z

Reagan feeling thermometer 0.010 0.006 1.69 0.09

Economic approval 0.053 0.078 0.68 0.50

Party Id 0.142 0.082 1.73 0.08

Bush traits 0.142 0.080 1.77 0.08

Dukakis traits -0.138 0.069 2.00 0.05

Race -0.570 0.385 1.48 0.14

Gender 0.034 0.119 0.29 0.77

Union member -0.038 0.140 0.27 0.79

Ideology 0.087 0.087 1.01 0.32

Constant -1.536 0.953 1.61 0.11

Variance Model

Time of vote decision 0.375 0.125 3.01 0.01

Political knowledge -0.202 0.052 3.91 0.01

Partisanship of answers -0.027 0.015 1.81 0.07

(pro-Republican vs. pro-Democratic)

See a differences between parties 0.287 0.211 1.36 0.17

N=970
Goodness of Fit: 2 df=13 = 975.29, p<.001
P

Heteroskedasticity Test: 2 df=4 = 67.79, p<.001

P

20

View publication stats