## New Research In

### Physical Sciences

### Social Sciences

#### Featured Portals

#### Articles by Topic

### Biological Sciences

#### Featured Portals

#### Articles by Topic

- Agricultural Sciences
- Anthropology
- Applied Biological Sciences
- Biochemistry
- Biophysics and Computational Biology
- Cell Biology
- Developmental Biology
- Ecology
- Environmental Sciences
- Evolution
- Genetics
- Immunology and Inflammation
- Medical Sciences
- Microbiology
- Neuroscience
- Pharmacology
- Physiology
- Plant Biology
- Population Biology
- Psychological and Cognitive Sciences
- Sustainability Science
- Systems Biology

# Candidate preferences and expectations of election outcomes

Contributed by Charles F. Manski, January 17, 2012 (sent for review December 1, 2011)

## Abstract

Analysis of data from the American Life Panel shows that in the presidential election of 2008 and in multiple statewide elections in 2010, citizens exhibited large differences in their expectations of election outcomes. Expectations were strongly positively associated with candidate preferences, persons tending to believe that their preferred candidate is more likely to win the election. Committed supporters of opposing candidates regularly differed by 20–30% in their assessments of the likelihood that each candidate would win. These findings contribute evidence on the false consensus effect, the empirical regularity that own preferences tend to be positively associated with perceptions of social preferences. We used unique measures of preferences and perceptions that enabled respondents to express uncertainty flexibly. We studied a setting that would a priori seem inhospitable to false consensus—one where persons have little private information on social preferences but substantial common knowledge provided by media reports of election polls.

Before presidential and statewide elections, persons often have extensive common knowledge of the prospects of alternative candidates, but scant private knowledge. The common knowledge is provided by polls, which offer considerable public data about the voting intentions of the electorate. In settings with extensive common and scant private knowledge, one might predict that persons would tend to have similar expectations, or at least that expectations would not vary systematically. To the contrary, we have observed large differences in expectations of election outcomes, which are strongly positively associated with candidate preferences. We report the findings here.

To begin, we orient our work within the literature studying the association between persons’ own preferences and their perceptions of social preferences. We next describe the data we analyze from two cohorts of respondents to the American Life Panel (ALP). The first period of data collection concerns a national sample interviewed before the 2008 presidential election. The second period concerns multiple state-specific samples interviewed before the 2010 senatorial and gubernatorial elections. We describe several aspects in which these data improve on those previously available. We then present and discuss our findings.

## Research on Own Preferences and Perceptions of Social Preferences

Researchers who study attitudes sometimes ask persons to state their own preferences about specified subjects and to report their perceptions of social preferences. A common practice is to ask a person who is a member of group Y about subject X, about which there may be two points of view (say A and B). The wording varies across studies, but the questions go more or less as follow:

Q1. Regarding subject X, do you prefer option A or B?

Q2. What fraction of people in group Y do you think prefer each option?

Researchers have asked such questions about many subjects in varied settings. They have repeatedly found a positive association between own preferences and perceptions of social preferences. Fabrigar and Krosnick (1) traces this empirical regularity back to Wallen (2), whereas Ross et al. (3) begin their account with Katz and Allport (4). Ross et al. (3) appears to have first named the regularity “false consensus,” a term which has become widespread. Some psychologists refer to the phenomenon as social projection (5).

Psychologists seeking to explain the regularity commonly suppose that persons begin with knowledge of their own preferences and extrapolate from themselves to the group. This extrapolation is usually interpreted as a cognitive error. Thus, the adjective *false* is used as a descriptor in the term false consensus, and the word *bias* appears often in articles on the subject.

Not all psychologists agree that extrapolation from the self to the group is in error. Dawes and Mulford (6, 7) observe that the regularity can emerge from Bayesian updating in which a person begins with a prior distribution regarding social preferences, views himself as a randomly drawn member of the group, and uses observation of his own preferences to derive a posterior distribution.

Another possibility is that a person's perception of social preferences influences his private preferences, the person seeking to conform to perceived social norms. If causality runs in this direction, research attempting to explain false consensus needs to recognize the reflection problem (8)—that is, if social preferences affect private preferences, then private preferences must conversely affect social preferences, which are the algebraic aggregation of private preferences.

The empirical regularity has often been observed in studies of candidate preferences and expectations of election outcomes. Persons who support a given candidate tend to report that he will win, with potential explanations taking labels such as wishful thinking (people predict favorable outcomes according to their wishes) or bandwagon effects (individuals are more likely to support a candidate who is leading in the poll) (9). Bartels (10) cites work as early as Berelson et al. in 1954 (11), whereas Granberg and Brent (9) begin with Hayes in 1936 (12; see also refs. 13⇓⇓–16).

Interestingly, the question about social preferences has commonly been worded differently in election settings than elsewhere. If one were to apply the wording of Q1 and Q2 to voting, one might ask the following:

V1. Regarding upcoming election X, do you expect to vote for candidate A or B?

V2. What fraction of voters do you think will vote for each candidate?

In practice, V1 is asked more or less as above but the inquiry about social preferences goes as follows:

V2A. Who do you expect will win the election, candidate A or B?

Thus, persons are not asked to predict the fraction of the vote that each candidate will receive. They are asked to predict whether this fraction will exceed one-half.

## ALP Data

The ALP is a longitudinal survey of Americans age 18 and older, begun by the RAND Corporation in 2006 and administered over the internet.* We first describe the panelists recruited through August 2008, who form the basis for the sample interviewed during the 2008 presidential campaign, and those recruited through August 2010, who form the basis for the sample interviewed during the 2010 campaigns. We then explain ALP measurement of election-outcome expectations and candidate preferences in the two samples.

### 2008 and 2010 Samples.

From August through October 2008, the ALP administered seven biweekly surveys with questions about the 2008 presidential election. A survey administered in November after the election asked respondents if they voted and for whom. The interview response rate per wave varied between 68% and 77%. A total of 1,814 participants responded to at least one pre-election survey and to the postelection survey. Relative to the electorate, respondents were more often female (57%), non-Hispanic white (89%), middle aged (41% ages 50–64), and college educated (45% with 16+ y of schooling). Delavande and Manski (17) describe these surveys in detail.

In September and October 2010, the ALP administered three biweekly surveys with questions about the 2010 senatorial and gubernatorial elections to respondents living in Arizona, California, Florida, Georgia, Illinois, Maryland, New York, Ohio, Oregon, Pennsylvania, Utah, and Wisconsin. These states had elections for both governor and senator in 2010. A survey administered in November after the election asked respondents if they voted and for whom.

In 2010, only respondents who answered the first wave were eligible to participate in later waves. The interview response rate per wave ranged between 83% and 91%. A total of 1,143 panelists responded to the postelection wave, of whom 61% were female, 86% non-Hispanic white, 40% middle aged (ages 50–64), and 43% college educated.

Nonrepresentativeness of the electorate is a shortcoming of the ALP, but the ability to interview persons repeatedly is an advantage. Most polls are repeated cross-sections, drawing new samples each time they go into the field. Hence, one cannot compare the candidate preferences and expectations of election outcomes that persons state as an election draws nearer. The ALP enables this comparison, as well as the comparison of pre-election attitudes with actual voting behavior.

To enhance consistency between the 2008 and 2010 samples, our analysis focuses mainly on wave 44 of the 2008 data and the first wave of the 2010 data. These waves were administered in the middle of September of each year.

### Measuring Expectations of Election Outcomes.

Persons asked to predict election outcomes may be uncertain what will occur. The standard practice, embodied in question V2A, does not enable respondents to express uncertainty. To enable full expression of uncertainty, one could elicit joint subjective distributions of vote totals for all of the candidates. However, this would be a laborious task. Instead, the ALP asks respondents to state their subjective probabilities that each candidate will win. The 2008 wording was

Barack Obama is the Democratic candidate and John McCain is the Republican candidate. What do you think is the percent chance that each man, or someone else, will win the election?

Barack Obama will win ___ %.

John McCain will win ___ %.

Someone else will win ___ %.

Similar questions were asked in 2010. For example, respondents in California were asked

What do you think is the percent chance that each of the candidates for governor will win the election?

Jerry Brown (Democrat) ___ %

Meg Whitman (Republican) ___ %

Someone else ___ %

In both years, the ordering of Democratic and Republican candidates was randomized, and the “someone else” option always appeared last. Respondents whose answers did not sum to 100 received an error message and were invited to change their answers. The item response rates were 93.9% in 2008 wave 44 and 97.2% in 2010 wave 1.

These questions apply ideas about probabilistic measurement of expectations that have been implemented in numerous settings over the past 20 y (see refs. 18⇓–20 for a review of the literature). Delavande and Manski (17) have previously analyzed the responses to 2008 ALP questions asking respondents to probabilistically predict their own voting behavior in the presidential election. The present paper studies the responses to the questions seeking expectations of election outcomes.

### Measuring Candidate Preferences.

We measure candidate preference by the subjective probabilities of candidate choice that ALP respondents express.^{†} In fall 2008, the ALP asked respondents to probabilistically predict their own voting behavior in the presidential election. The first question sought the subjective probability that a person placed on voting. The second elicited her probability of voting for each candidate, conditional on voting. The wording was as follows:

If you do vote in the presidential election, what do you think is the percent chance that you will vote for

John McCain (Republican) ___ %

Barack Obama (Democrat) ___ %

Someone else ___ %

Analogous questions were asked in fall 2010. For example, respondents in California were asked

If you do vote in this year's election for governor, what do you think is the percent chance that you will vote for

Jerry Brown (Democrat) ___ %

Meg Whitman (Republican) ___ %

Someone else ___ %

The ordering of the Democratic and Republican candidates was randomized in all cases. The item response rates were 99.4% in 2008 wave 44 and 96.1% in 2010 wave 1.

Delavande and Manski (17) analyzed the 2008 ALP data and found that stated probabilities of candidate choice are good predictors of actual voting behavior. In theories of rational voting, responses to the probability questions are interpretable as measuring strength of candidate preference.^{‡}

## Association of Candidate Preferences and Election Expectations

We next studied the empirical association of candidate preferences and expectations of election outcomes. We found that the ALP data strongly and persistently exhibited the empirical regularity predicted by the false consensus hypothesis. We observed substantial positive association of expectations and preferences in the 2008 presidential election and in all of the 2010 statewide elections. Disaggregation into demographic groups shows that the association occurs among both men and women; whites and blacks; and less- and more-educated respondents. Indeed, the magnitude of the association is remarkably stable across elections and groups.

To study the association between preferences and expectations, we compare the mean expectations of persons who state different preferences. Empirical analysis of expectations conditional on preferences has a clear causal interpretation under the hypothesis that persons extrapolate from their own preferences to the community. However, we do not insist on this interpretation. As mentioned earlier, it may be that causality is bidirectional, with social preferences affecting private preferences and vice versa.

For consistency across elections, we always measure a person's preference by her stated percent chance of voting Democratic. We measure her election expectation by the percent chance she gives to the Democrat winning. Stated percent chances of voting Republican or that the Republican would win were typically 100 minus the values given for the Democrat. The ALP respondents expressed very little support for third-party candidates in these elections, and they gave such candidates essentially zero chance of winning.

Let *e* denote an election and *w* denote an ALP wave. Let *x* denote the candidate preference that a person states in wave *w* regarding election *e*, and let *y* denote her corresponding expectation for the election outcome. Let *z* denote covariates, such as years of schooling or sex. We examine how the conditional mean *E*(*y* | *x*, *e*, *w*, *z*) varies with *x*. That is, we study the regression of *y* on *x* among persons with covariates *z* interviewed in wave *w* about election *e*.

So as not to constrain the shape of a regression, it is preferable to estimate it nonparametrically and present findings graphically. However, it would take considerable space to display the regressions estimated for many values of (*e*, *w*, *z*). Hence, we first present nonparametric findings that do not disaggregate persons into groups. We then present more succinct logistic regressions that disaggregate by group.

We first view the ALP as a cross-sectional survey. We then exploit the longitudinal structure of the survey. Here we examine the temporal covariation of election expectations with candidate preferences across different waves.

### Nonparametric Regression Analysis.

#### 2008 presidential election.

Fig. 1 displays the estimated regression of presidential election expectations on candidate preferences among the 1,488 respondents to wave 44. This and other nonparametric estimates were computed using the kernel smoothing method with Gaussian kernel and bandwidth determined by Silverman's rule of thumb. The solid curve gives the estimate, and the dashed band shows a 95% bootstrap confidence interval.

Fig. 1 shows a rising S-shaped curve. Persons who stated a 0%, 25%, 50%, 75%, or 100% chance of voting for Obama on average were estimated to give Obama a 40%, 43%, 51%, 61%, and 65% chance of winning, respectively. The regression estimate barely rises as *x* increases from 0 to 20, rises gradually as *x* increases from 20 to 80, and then rises slightly as *x* increases from 80 to 100. The confidence interval is tight, indicating that the estimate is statistically precise across the domain of *x*.

An alternative perspective is provided by grouping respondents into response categories instead of using the kernel method to generate a smooth curve. Table S2, section A, corroborates the main features of Fig. 1 and adds further information. Large numbers of respondents expressed extreme percent-chance values of voting for Obama. Of the 512 respondents who stated a 0% chance of voting for Obama, their 25th-, 50th-, and 75th-percentile responses to the election expectations question were 30%, 45%, and 50%, respectively. Of the 489 who stated a 100% chance of voting for Obama, their 25th-, 50th-, and 75th-percentile responses to the election expectations question were 50%, 60%, and 80%, respectively. Thus, committed McCain and Obama supporters reported quite different election expectations. Both groups believed their own candidate was more likely to win.

Table S2, section B, provides an analogous summary of wave 51, fielded at the end of October. Of the 620 respondents who then stated a 0% chance of voting for Obama, their 25th-, 50th-, and 75th-percentile responses to the election expectations question were 48%, 50%, and 55%, respectively. Of the 615 who stated a 100% chance of voting for Obama, their 25th-, 50th-, and 75th-percentile responses to the election expectations question were 60%, 80%, and 95%, respectively.

Thus, respondents as a whole tended to see better prospects for election of Obama at the end of October than in mid-September. This movement in expectations was ordinally consistent with the polls, which showed greater support for Obama near the end of the campaign than earlier. Nevertheless, the regularity predicted by the false consensus hypothesis is manifest in both waves. Indeed, further analysis shows that the regularity occurs in all seven waves conducted before the election.

#### 2010 elections for governor and senator.

Fig. S1 *A* and *B* display the estimated regressions of California gubernatorial and senatorial election expectations on candidate preferences in 2010 wave 1. Fig. S1 *A* and *B* show kernel estimates of *E*(*y* | *x*, *e* = 2010 California governor, *w* = 1) and *E*(*y* | *x*, *e* = 2010 California senator, *w* = 1).

The sample of California respondents was smaller than the national sample in the 2008 presidential election, with 248 persons providing data on the election for governor and 250 on the election for senator. However, the California sample was large enough to yield reasonably precise estimates. The confidence intervals are fairly tight.

As with Fig. 1, Fig. S1 *A* and *B* shows rising S-shaped curves. In the election for governor, persons who stated a 0%, 25%, 50%, 75%, or 100% chance of voting for Jerry Brown on average were estimated to give Brown a 37%, 40%, 47%, 54%, and 59% chance of winning, respectively. In the election for senator, persons who stated a 0%, 25%, 50%, 75%, or 100% chance of voting for Barbara Boxer on average were estimated to give Boxer a 44%, 48%, 55%, 62%, and 65% chance of winning, respectively.

We have analyzed the data on elections in states other than California and have repeatedly observed the empirical regularity, albeit with less precision, because the sample sizes are smaller. Rather than show a multitude of figures, Fig. S1*C* aggregates the 2010 gubernatorial elections and provides a kernel estimate of *E*(y | *x*, *e* = 2010 governor, *w* = 1) based on the 1,215 respondents in the 12 states sampled. Again we find a rising S-shaped curve. Persons who stated a 0%, 25%, 50%, 75%, or 100% chance of voting Democratic on average were estimated to give this candidate a 32%, 37%, 48%, 59%, and 63% chance of winning, respectively. Thus, the regularity predicted by the false consensus hypothesis appears again, and it occurs as well in the data on senatorial elections. See also Table S2, section C.

### Logistic Regression Analysis.

We now disaggregate respondents in various ways, stratifying by sex, race, education, or state. The recurring rising S-shape of the regressions suggests use of logistic regression to fit the data. Suppose that among persons with covariates *z* interviewed in wave *w* about election *e*, *y* is related to *x* as follows:

Here, α_{ewz}, β_{ewz} are parameters common to the triple *e*, *w*, *z*. We estimate the parameters by nonlinear least squares. The resulting estimates are consistent if the regression truly is logistic. If not, we obtain a consistent estimate of the best predictor of *y* in the logistic family of predictor functions.

#### 2008 presidential election.

Table S3 presents findings for the 2008 election, giving the parameter estimates and estimated SEs. Here and below, we use one covariate value as a base case. When we discuss the estimates for other covariates, we add the base-case estimates to the estimates specific to those covariate values. In performing the logistic regressions, we divided the percent-chance data on candidate preference by 100. Thus, the reader should be aware that *x* is measured on a 0–1 probability scale rather than a 0–100% chance scale.

As previously, we focus on wave 44 administered in mid-September 2008. Dividing the sample by sex, the estimates for the parameters α_{ewz}, β_{ewz} are −0.369, 0.969 for males and −0.506, 1.159 for females. Dividing by race, the estimates are −0.678, 1.520 for nonwhites and −0.427, 1.024 for whites. Dividing by education, the estimates are −0.509, 1.571 for persons with 12 or fewer years of schooling, −0.568, 1.390 for 13–15 y, and 0.105, 0.750 for 16+ y of school. The estimates are generally precise. Here and elsewhere, the parameter estimates are typically similar across groups.

We are mainly interested in the slope parameter β_{ewz}, which measures the strength of the association between candidate preferences and election expectations. The findings suggest some detailed differences in β_{ewz} across groups, the slope parameter being larger for females, nonwhites, and persons with less schooling than for males, whites, and persons with more schooling. However, the main lesson is that β_{ewz} is positive and has similar magnitude whether one divides the sample by sex, race, or schooling. Thus, the regularity predicted by the false consensus effect appears regardless of how we stratify the sample.

The main lesson holds up when we stratify in other ways. For example, dividing the sample into age groups shows essentially no difference in the slope parameters of younger, middle-aged, and older persons.

#### 2010 elections for governor and senator.

Tables S4 and S5 present findings for the 2010 elections. As earlier, we focus on wave 1 administered in mid-September 2010. For the sake of sample size, we aggregate elections across states when dividing the sample by group. We later will discuss state-by-state results.

Table S4 gives the findings for the gubernatorial races. Dividing the sample by sex, the estimates for α_{ewz}, β_{ewz} are −0.705, 1.220 for males and −0.793, 1.439 for females. Dividing by race, the estimates are −0.653, 1.403 for nonwhites and −0.765, 1.393 for whites. Dividing by education, the estimates are −0.882, 1.600 for 12 or fewer years of schooling, −0.791, 1.472 for 13–15 y, and −0.665, 1.178 for 16+ y.

Table S4 also gives the findings for the senatorial races. Dividing by sex, the estimates are −0.780, 1.435 for males and −0.772, 1.447 for females. Dividing by race, the estimates are −0.702, 1.378 for nonwhites and −0.782, 1.445 for whites. Dividing the sample by education, the estimates are −0.907, 1.640 for 12 or fewer years of schooling, −0.746, 1.480 for 13–15 y, and −0.754, 1.369 for 16+ y.

These findings are similar to those obtained in 2008. Indeed, the differences in parameter values across groups are even smaller here. The slope parameter is positive in every case and remarkably stable in magnitude.

Finally, we use state as the covariate, presenting separate estimates for each gubernatorial and senatorial race (Table S5). Here, too, the estimates are stable in magnitude, with typically small variation across states. In the races for governor, the estimated slope parameter ranges from 1.016 (California) to 1.662 (Oregon). In the races for senator, the estimated slope parameter ranges from 0.763 (Utah) to 1.954 (Florida).

### Temporal Covariation of Expectations and Preferences.

The above analysis treated each ALP wave as a separate cross-section. The ALP is a longitudinal survey, with the same respondents interviewed in multiple waves; this offers the opportunity to analyze the within-respondent temporal covariation of candidate preferences and election expectations.

Consider someone interviewed in adjacent waves, say waves *w* − 1 and *w*. Let *x _{w}* −

*x*

_{w}_{− 1}denote the change between waves in this person's stated percent chance of voting Democratic. Let

*y*−

_{w}*y*

_{w}_{− 1}denote the corresponding change in the percent chance she gives to the Democrat winning. The regression

*E*(

*y*− y

_{w}_{w − 1}|

*x*−

_{w}*x*

_{w}_{− 1},

*e*) summarizes how, on average, election expectations temporally vary with candidate preferences.

Fig. 2 presents a kernel estimate for the 2008 election. The sample of 7,631 observations used to produce this estimate pools the data for the seven waves before the election. Thus, a respondent who participated in all waves contributes six observations. Fig. S2 *A* and *B* present analogous estimates for the 2010 elections. Here the samples pool the data for the three waves before the election, so a respondent who participated in all waves contributes two observations. The resulting sample sizes are 1,917 and 1,932.

Each estimate shows a mildly increasing function within its range of precise estimation. Thus, temporal changes in candidate preference are associated with ordinally consistent movements in election expectations. The confidence intervals show that the estimate is statistically precise when *x _{w}* −

*x*

_{w}_{− 1}is not too large in absolute value (roughly in the range −40 to +40). Precision decreases when

*x*−

_{w}*x*

_{w}_{− 1}takes more extreme values; the reason is the relative abundance of observations. It is common for respondents to change their stated preference moderately from one wave to the next, but it is rare for them to report changes >40%.

To summarize the estimates, we computed corresponding least-squares fits of *y _{w}* −

*y*

_{w}_{− 1}to

*x*−

_{w}*x*

_{w}_{− 1}(Table S6). The three sets of slope parameters, which are precisely estimated, are highly similar in magnitude—0.258, 0.191, and 0.265, respectively. Thus, a temporal increase of 1% in a respondent's stated chance of voting Democratic tends to be accompanied by a temporal change of one-fifth to one-quarter of a percent in the chance she places on the Democrat winning the election.

## Discussion

We have found in multiple elections, one presidential and many statewide, that the regression of election expectations on candidate preferences is a rising S-shaped curve. At the extremes, committed supporters of opposing candidates (those with 0% or 100% chance of voting Democratic) regularly differ by 20–30% in their assessments of the likelihood that the Democrat will win. These findings are common to males and females, nonwhites and whites, and to persons with different levels of schooling. We have also found that, when persons change their candidate preferences over time, their election expectations tend to move in the same direction.

The regularity predicted by the false consensus hypothesis has been reported in many settings. However, previous studies have not measured the uncertainty of expectations regarding social preferences. Moreover, studies of false consensus have often concerned topics about which there exists little or no public information about social preferences. In such circumstances, it is reasonable to conjecture that persons extrapolate their own preferences to the community.

In contrast, considerable common knowledge of social preferences in presidential and statewide elections is available from the polls conducted in the months before the election. Given the availability of poll data, and that persons ordinarily have little private information about prospective election outcomes, one might anticipate that ALP respondents would not vary much in their expectations for election outcomes, or at least not vary systematically. However, respondents expressed large differences in their expectations of election outcomes, which were strongly positively associated with their own candidate preferences. It thus appears that Americans, despite having access to the same publicly available information, nevertheless inhabit disparate perceptual worlds.

We see two major open questions that warrant future investigation. First, what explains the false consensus effect in expectations of election outcomes? Although psychologists have long conjectured various generic explanations of false consensus, no consensus has emerged. Moreover, a generic explanation may not fit elections and other settings where persons have substantial common but scant private knowledge.

Second, what are the implications for voting behavior? Expectations of election outcomes loom large in theories of instrumental voting, which suppose that a person votes if the expected benefit exceeds the expected cost. The expected benefit depends on a person's candidate preferences and her expectation that her vote will affect the outcome. Thus, the empirical regularity reported here may be consequential for voting behavior, the specific consequences depending on the explanation of the regularity. Theories hypothesizing that persons extrapolate from their own preferences to the community may have different voting implications than ones hypothesizing a bidirectional process in which perceptions of social preferences affect private preferences and vice versa.

## Acknowledgments

The authors thank Charles Bellemare and Baruch Fischhoff for helpful comments on this manuscript. Support for this work was provided by National Institute of Aging Grant 5R01AG020717-05 (to A.D.) and National Institute of Aging Grant P01-AG026571-06 (to C.F.M.).

## Footnotes

- ↵
^{1}To whom correspondence should be addressed. E-mail: cfmanski{at}northwestern.edu.

Author contributions: A.D. and C.F.M. designed research; A.D. and C.F.M. performed research; A.D. and C.F.M. analyzed data; and A.D. and C.F.M. wrote the paper.

The authors declare no conflict of interest.

Data deposition: The American Life Panel data are publicly available on the RAND Web site at https://mmicdata.rand.org/alp/index.php/Main_Page.

↵*See https://mmicdata.rand.org/alp/index.php/Main_Page. Panel members stem from multiple sources; until September 2009, most were recruited from outgoing participants in the Reuters/Michigan Surveys of Consumers, who were asked if they would be willing to participate in internet surveys. The ALP recruited from those who gave any response except “no, certainly not.” Eighty percent of the ALP panelists who responded to at least one of the 2008 election surveys were recruited in this manner, whereas 20% were a snowball sample of persons generated from the Reuters/Michigan recruits. Beginning in September 2009, new respondents were recruited after participating in the National Survey Project (NSP), a collaboration of Stanford University and Abt SRBI. Following participation in the NSP, respondents to that survey were invited to join the ALP. Beginning in 2010, the ALP also recruited through random mail and telephone solicitations. Among respondents who participated in at least one of the 2010 election surveys, 61% stem from the Reuters/Michigan survey, 23% from the snowball sample, 15% from the NSP, and 1% from the new mailing/telephone recruitment.

↵

^{†}A second measure of candidate preference, a thermometer rating, was obtained from respondents in 2010. Comparison of this measure with the subjective probabilities of candidate choice reveals a strong empirical association, corroborating the validity of the subjective probability measure (*SI Text*and Table S1).↵

^{‡}Suppose that there are two candidates, A and B. Consider a voter named j. Suppose that when j casts her ballot, she places utilities UjA and UjB on the candidates and votes for the one giving the higher utility. Suppose that before the election, person j is polled and asked to state the percent chance that she will vote for candidate A, should she vote. She may be uncertain how she will vote because she is aware that she may learn more about the candidates as the campaign progresses. A rational respondent will interpret the question as seeking her current probability that UjA will exceed UjB when she evaluates these quantities on election day. She may judge that, the stronger her current preference for A, the more likely it is that she will still prefer A on election day. If so, the probability she places on voting for A measures the strength of her current preference for A. C.F.M. (21, 22) develops this reasoning.This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1200861109/-/DCSupplemental.

## References

- ↵
- ↵
- ↵
- ↵
- Katz D,
- Allport F

- ↵
- Robbins JM,
- Krueger JI

- ↵
- Dawes R

- ↵
- Dawes R,
- Mulford M

- ↵
- ↵
- ↵
- ↵
- Berelson B,
- Lazarsfeld P,
- McPhee W

- ↵
- ↵
- ↵
- ↵
- ↵
- Nir L

- ↵
- Delavande A,
- Manski C

- ↵
- ↵
- ↵
- ↵
- ↵

## Citation Manager Formats

## Sign up for Article Alerts

## Article Classifications

- Social Sciences
- Economic Sciences