Theoretical Perspectives

One major theoretical perspective on initiative and referendum voting is that citizens assess the merits of the case and vote in favor of a proposal when they expect that the benefits of adopting it outweigh the costs (e.g., Blais et al. 1996; Clarke, Kornberg and Stewart 2004; see also Alvarez and Kiewiet 2009).Footnote ³ However, if voters think costs exceed benefits or they are unsure, then they vote no. In this regard, uncertainty about the consequences of adopting a proposal to change the status quo is often used to explain a major stylized fact about such proposals, namely that they often are defeated (see LeDuc Reference LeDuc2003). The claim is that, when in doubt, voters discount possible benefits and say “no.” This latter hypothesis is consistent with widely cited research showing that in a wide variety of situations decision makers have asymmetric loss functions that prompt them to weigh costs more heavily than benefits (e.g., Kahneman, Slovic and Tversky Reference Kahneman, Slovic and Tversky1982; Thaler Reference Thaler1993; Thaler Reference Thaler1994; Kahneman Reference Kahneman2011).

A second major theoretical perspective on initiative and referendum voting involves the idea that in situations of high stakes and abundant uncertainty people rely on heuristic devices of various kinds. As pioneering work by Simon (Simon Reference Simon1955; Simon Reference Simon1957), decision making using heuristics has been associated with broader theories of bounded rationality. The latter may have served to color assessments of this approach by implying that the use of heuristics involves an inferior decision strategy when compared with the familiar optimizing models of rational choice theory. Indeed, Simon’s term “satisficing” implies something less than global optimization (Conlisk Reference Conlisk1996). Simon addressed these concerns by observing that the classical, optimizing model requires knowledge of all available alternatives, consequences and probabilities. It is a world without surprises (Gigerenzer and Gaissmeier Reference Gigerenzer, Hertwig and Pachur2011, 452; see also Gigerenzer, Hertwig and Pachur Reference Gigerenzer and Gaissmaier2011).

Simon’s analysis is consistent with that of Savage (Reference Savage1972), a major figure in the development of Bayesian decision theory. As noted by Gigerenzer and Gaissmeier, Savage differentiated between the perfect knowledge found in “small worlds” versus “large worlds,” where relevant information is either unknown, or must be estimated with uncertainty from small samples. Thus, and critically, “one can no longer assume that ‘rational’ models automatically provide the correct answer. Even small deviations from model conditions can lead to disaster when applied to the large world …” (Gigerenzer and Gaissmeier Reference Gigerenzer, Hertwig and Pachur2011, 453). In anticipating arguments below, it is reasonable to conclude that more knowledgeable and sophisticated people recognize the difference between “small-world” and “large-world” situations and act accordingly.

Gigerenzer et al. have been leading voices in making the case that simplifying heuristic devices may be superior to optimizing models of decision making and that, in fact, individuals rely heavily on heuristics when making a wide variety of decisions. Gigerenzer (Reference Gigerenzer2008) argues that: (a) optimization is often impossible because of computational intractability or is less robust, that is, subject to substantial measurement error; (b) individuals with greater cognitive capacity are not less likely to employ relatively simple heuristics when compared with those with lesser cognitive ability. More information and intensive computation is not always better. There are many situations where making good decisions requires ignoring some available information. This argument is consistent with Savage’s image of “large worlds,” which are fraught with uncertainty and where probabilistic reasoning is pervasive.

Political decisions confronting individuals in contemporary settings are ideally suited to Gigerenzer’s “fast and frugal” decision-making heuristics. Given the opaque nature of complex issues such as climate change and possible mitigating strategies and solutions, the uncertain nature of attendant political-economic trade-offs and serious difficulties involved in assessing the veracity of competing claims, a classical optimizing approach is likely not possible.

Individuals with greater cognitive abilities may be at least as likely to use heuristics as those with lesser cognitive abilities. Although there is not a one-to-one correspondence between cognitive capacity and political knowledge and sophistication, there is very likely a strong linkage. Moreover, political choices may be especially amenable to such a linkage. When voters, or citizens in general, use heuristic strategies, they have to know where and when to look for useful cues. The more knowledgeable a person is, the more readily s/he can make a determination, which might be a “best guess,” as to where to seek, and how to use, relevant information. More knowledgeable individuals realize they do not and cannot know everything, and recognize the probabilistic and error-prone nature of the information they obtain. In contrast, citizens with little knowledge of the complexities of politics and policy are more likely to engage in indiscriminate and undisciplined information searches. The less knowledgeable, not knowing what they do not know, are likely to engage in efforts that are wasteful of scarce resources and conducive to problematic decisions.

The present paper focuses on citizen decision making in important political initiatives and referendums. Faced with making a choice on an initiative or referendum proposal, and lacking reliable information about the consequences of alternative courses of action or the cognitive resources to process that information effectively, voters employ cues provided by various groups and individuals. As noted above, Gigerenzer et al. describe these cues as “fast and frugal heuristics,” that is, cues that are used by voters who typically are unable to behave in accordance with the precepts of classic micro-economic rationality.Footnote ⁴ One such heuristic is provided by images of party leaders or candidates who are proponents or opponents of an initiative or referendum proposal. Voters use information about the proposition’s “friends and foes” together with information about these individuals to decide how to cast their ballots.Footnote ⁵

Elaborating this hypothesis, some scholars have conjectured that the effects of candidate and party leader images are heterogeneous, being stronger among people with lower levels of political knowledge (e.g., Mondak Reference Mondak1993; Gomez and Wilson Reference Gomez and Wilson2001; Bartle Reference Bartle2005; Gomez and Wilson Reference Gomez and Wilson2006; see also Converse Reference Converse1964; Zaller Reference Zaller1992; Delli Carpini and Keeter Reference Delli Carpini and Keeter1997). It is argued that, in contrast to uninformed voters who are attracted by slick, but vacuous, portrayals of leaders and candidates, knowledgeable individuals assess the pros and cons of initiative and referendum proposals. Discounting possibly misleading cues provided by leader and candidate images and relying more heavily on information about the policy alternatives on offer, knowledgeable voters concentrate on the merits of the case.

However, there is an alternative possibility. Consistent with findings that people often fail to behave in accordance with the strictures of micro-economic rationality when making political choices (e.g., Sniderman, Brody and Tetlock Reference Sniderman, Brody and Tetlock1991; Lupia and McCubbins Reference Lupia and McCubbins1998; Lupia, McCubbins and Popkin Reference Lupia, McCubbins and Popkin2000), it can be hypothesized that politically knowledgeable voters are actually more likely than less knowledgeable persons to utilize readily accessible cues such as those provided by leader and candidate images.Footnote ⁶ There are two reasons why this might be the case. First, as argued above, knowledgeable people may be more likely than less knowledgeable persons to recognize that they cannot possibly assemble and process the information necessary to assess the consequences of the adoption, or failure to adopt, particular resolutions in accordance with the canons of classic procedural rationality. Knowledgeable individuals react by relying on an alternative decision-making strategy.

There is yet another, simpler, possibility. According to this account, more knowledgeable voters have a higher probability of using leader and candidate heuristics because they are more likely to have information about where leaders and candidates stand on ballot propositions under consideration.Footnote ⁷ Even if all voters are equally likely to try to use heuristics, more knowledgeable people are most apt to be successful, even if they operate on a random “top of the head” basis (Zaller Reference Zaller1992; see also Converse Reference Converse1964). Using a random selection procedure, knowledgeable voters are more likely to find a relevant cue than are their less knowledgeable fellow citizens. Moreover, given the high salience of party leaders and candidates in most political campaigns, there is a substantial probability that a cue selected at random will involve some aspect of leader or candidate images. Below, we investigate these competing hypotheses about the interactive effects of candidate images and political knowledge in multivariate analyses of the determinants of voting on California’s Proposition 23.

The Climate of Opinion and Voting on Proposition 23

On 2 November 2010, the California electorate was invited to vote on a statewide initiative on climate change legislation. Proposed in the midst of the most serious economic downturn since the Great Depression of the 1930s with statewide joblessness in double digits, Proposition 23Footnote ⁸ would have suspended “implementation of air pollution control law (AB 32—The Global Warming Solutions Act) requiring major sources of emissions to report and reduce greenhouse gas emissions that cause global warning, until unemployment drops to 5.5 percent or less for full year.”Footnote ⁹ Although early polls suggested that the proposal might pass, it was defeated by a substantial 62–38 percent margin.

We first use survey data to describe factors relevant for understanding why Californians voted as they did on Proposition 23. The survey data were gathered in a representative national pre- and post-election internet survey with a California oversample conducted for the authors by YouGov/Polimetrix.Footnote ¹⁰ The California sample includes 1102 respondents.Footnote ¹¹ Figure 1 shows Californian’s opinions concerning the general importance of climate change.Footnote ¹² Responses vary widely: 22 percent rated climate change as “not important,” the lowest possible score, and 18 percent rated it as “very important,” the highest possible score. Other views were dispersed across the scale—43 percent rated climate change in the 0–4 range of importance and 50 percent gave scores in the 6–10 range. Similarly, when asked about trade-offs between environmental protection and economic growth, Californians were widely dispersed across an 11-point scale ranging from 0 (protect environment) to 10 (promote growth), with somewhat more (44 versus 40 percent) being on the pro-growth side of the scale (see Figure 2). Clearly, there was no consensus about whether environment or economy deserved top priority. In addition, it is noteworthy that a Cap-and-Trade principle was incorporated in the 2006 climate change legislation (AB 32) Proposition 23 would have suspended. Substantial numbers of the survey respondents voiced opposition to Cap and Trade—the idea was opposed 43 to 35 percent, with fully 37 percent expressing “strongly unfavorable” opinions and only 17 percent stating that they were “strongly favorable” (data not shown).Footnote ¹³

Fig. 1

Ratings of the importance of climate change

Fig. 2

Trade-off between environmental protection and economic growth

California senatorial candidates took clear and widely publicized stands on Proposition 23, with Republican, Carly Fiorina, endorsing the measure and Democrat, Barbara Boxer, voicing firm opposition. In the gubernatorial race, Democrat Jerry Brown was openly opposed, but Republican Meg Whitman tried to garner support from both proponents and opponents of the proposition, by seeking to straddle the issue. Although Whitman stated she would vote no on Proposition 23, she also said that she would suspend AB 32.Footnote ¹⁴ Figure 3 illustrates that the senatorial candidates’ images were powerfully correlated with voting.Footnote ¹⁵ Only 8 percent of those who strongly disliked Fiorina voted for the proposal, but among those who strongly liked her, 78 percent supported the proposal. The relationship between Boxer’s image and Proposition 23 voting is a mirror image—85 percent of those who strongly disliked her voted for the proposal, as compared with 10 percent of those who felt the opposite. Despite Whitman’s ambiguity, there are strong correlations between gubernatorial candidate imagesFootnote ¹⁶ and Proposition 23 voting. The percentage voting yes increased from 11 to 75 percent as attitudes toward her moved from very negative to very positive. Similarly, the percentage voting yes decreased from 82 percent among persons with highly negative images of Brown to only 16 percent among those with very positive images of him.

Fig. 3

Percentages voting yes on Proposition 23 by images of US senate candidates

In the next section, we specify two multivariate models of factors affecting Proposition 23 voting to determine if candidate images were influential when other theoretically interesting and statistically significant explanatory variables are controlled. The first model provides a baseline for assessing the effects of several predictor variables on the probability of voting for or against Proposition 23.Footnote ¹⁷ The second model specifies explicit interaction effects and is designed to test the two competing hypotheses articulated above concerning how candidate images interact with political knowledge to influence the choices voters make.

Multivariate Models of Voting on Proposition 23

Consonant with the theoretical perspectives on initiative and referendum voting discussed above, the two models incorporate variables tapping beliefs about the merits of the case, that is, beliefs about whether a proposal, if passed, would produce what the voter considers to be a desirable outcome from a cost-benefit perspective. Specifically, we hypothesize that people with pro-environmental views would vote no, as would those who evaluate the economy positively. In contrast, people less strongly concerned about the environment and those judging the economy negatively would be likely to vote yes. We measure environmental attitudes using a factor score variable that summarizes information concerning attitudes toward Cap-and-Trade legislation, willingness to pay higher taxes to protect the environment, the importance ascribed to climate change and self-placement on a proposed trade-off between environmental protection and economic growth.Footnote ¹⁸ A second factor score variable summarizes information regarding national and personal economic evaluations over retrospective and prospective time horizons.Footnote ¹⁹

The second category of explanatory variables contains heuristics voters might use as cues for making political choices. As discussed above, analysts often have argued that partisanship, liberal–conservative ideological positions and candidate images are cues that people use when deciding how to cast their ballots. Given arguments advanced by competing groups during the Proposition 23 campaign, we anticipate that Democratic identifiers and self-identified liberals would be more likely to vote against the proposition, whereas Republican partisans and conservatives would be more likely to vote in favor of it.Footnote ²⁰ Regarding candidate images, noting that Republican Senate candidate Carly Fiorina strongly endorsed Proposition 23 during the campaign and her Democratic rival, Barbara Boxer, strongly opposed it, we hypothesize that people with a relatively favorable image of Fiorina as compared with Boxer would tend to vote in favor, and those favoring Boxer over Fiorina would tend to vote against it. The situation for gubernatorial candidates is less clear-cut because of Whitman’s equivocation, rejecting Proposition 23, whereas endorsing suspension of AB 32. However, we expect individuals with relatively favorable images of her would likely vote yes, and those with relatively favorable images of Jerry Brown likely would vote no.Footnote ²¹

There are other possibly relevant heuristics as well. One concerns voters’ reactions to President Barack Obama. Although Obama was not on the ballot in 2010 mid-term congressional elections, his widely publicized endorsement of Cap-and-Trade legislation and his advocacy of measures to promote environmental friendly solutions to America’s energy and economic problems had clearly identified him as favoring green policy alternatives. Accordingly, images of Obama were a potentially influential heuristic available as voters sought cues about what to do about Proposition 23. People with favorable images of the PresidentFootnote ²² should have been inclined to vote no and those with unfavorable images should have been disposed to vote yes.

We also consider the effects of the campaigns mounted by the political parties and their candidates. When parties contact voters, they attempt to mobilize support not only for themselves but also for their positions on major policy initiatives such as Proposition 23. Accordingly, people contacted by Republicans should be more likely to vote yes, whereas those contacted by Democrats would be more likely to vote no.Footnote ²³ In addition, risk acceptance–risk aversion may be relevant, with the latter group voting no to avoid threats that might arise if the status quo was overturned.Footnote ²⁴

Finally, socio-demographic characteristics may be relevant to both the cost-benefit and heuristic perspectives. For example, by proposing to delay implementation of AB 32 until unemployment was below 5.5 percent for four consecutive quarters, Proposition 23 invited voters to think in terms of a trade-off involving taking action to combat climate change and the need to jump start California’s flagging economy. In this regard, unemployed people might be swayed by their personal adversity in the labor market to support the proposition. Other economically vulnerable groups such as racial and ethnic minorities, women and young people and those with lower incomes and educational levels,Footnote ²⁵ might react similarly.

To summarize, the model of voting on Proposition 23 (without interaction terms) is:

(1) $$\eqalignno{ {\rm Pr}({\rm Vote}=1)= \,& \Phi (\beta _{0} {\plus}B_{{\rm 1}} {\rm ENVIR}{\plus}B_{{\rm 2}} {\rm ECON}{\plus}B_{{\rm 3}} {\rm LIBCON} \cr & {\plus}B_{{\rm 4}} {\rm SCAND}{\plus}B_{{\rm 5}} {\rm GCAND}{\plus}B_{{\rm 6}} {\rm OBAMA} \cr & {\plus}B_{{{\rm 7}{\minus}{\rm 9}}} {\rm PID}{\plus}B_{{{\rm 1}0{\minus}{\rm 11}}} {\rm CAMP}{\plus}B_{{{\rm 12}{\minus}{\rm 14}}} {\rm RISK} \cr & {\plus}B_{{{\rm 15}}} {\rm UNEMP}{\plus}B_{{{\rm 16}{\minus}{\rm 23}}} {\rm DEMOS}). $$

In this model, the dependent variable is Pr(Vote=1), the probability that a voter will cast a yes ballot in favor of Proposition 23. The predictor variables are: ENVIR=environmental attitudes; ECON=economic evaluations; LIBCON=liberal–conservative beliefs; SCAND=relative senatorial candidate images; GCAND=relative gubernatorial candidate images; OBAMA=Obama image; PID=party identification; CAMP=contact by Democratic or Republican parties; RISK=risk orientations; UNEMP=employment status; DEMOS=socio-demographic characteristics (age, education, gender, income, race/ethnicity); B ₁₋₂₃=coefficients to be estimated. Since voting on Proposition 23 is measured as a dichotomy (vote no=0, vote yes=1), a binomial probit link function (Long and Freese Reference Long and Freese2014) is specified.Footnote ²⁶

The model’s overall fit is satisfactory—the McKelvey R ² is 0.72, and fully 87.7 percent of the voters are correctly classified (Model A, Table 1).Footnote ²⁷ Regarding specific predictors, several variables operate as hypothesized. Pro-environmental attitudes and liberal ideological beliefs have the hypothesized negative effects on the probability of voting yes, whereas negative evaluations of economic conditions encourage a yes vote. Senatorial candidate images also behave as anticipated, with people more favorably disposed toward Carly Fiorina than Barbara Boxer being more likely to vote yes, and those favorable toward Boxer rather than Fiorina being prone to vote no.Footnote ²⁸ However, relative dispositions toward the gubernatorial candidates (Brown and Whitman) are not statistically significant. Voters’ image of President Obama and their party identification also fail to exert significant effects.

Table 1

Binomial Probit Models of “Yes” Voting on Proposition 23 Initiative

AIC=Akaike information criterion; —=variable not included in model.

*p≤0.10; **p≤0.05; ***p≤0.01; ****p≤0.001; one-tailed test for all predictors except socio-demographics.

As for other predictors, party contacting works as advertised—net of other considerations, people contacted by the Republicans are more likely to vote yes, whereas those contacted by the Democrats are more likely to vote no. Risk orientations also behave as expected, with risk averse people being more apt to endorse the status quo by voting no. Moreover, although employment status, level of formal education and annual family income are not statistically significant, young people, women, Hispanics and those in the “other” ethnic category are more likely to vote yes than are older people, men and whites.

As the parameters in Model A are estimated using a nonlinear (probit) link function, the magnitude of the effect of any predictor variable on the probability of voting yes depends upon the values of other predictors (Long and Freese Reference Long and Freese2014). To provide intuition about the substantive importance of statistically significant predictors, we construct a scenario where all predictors are set at their means in the case of continuous variables, and at 0 in the case of dichotomous variables. Then, we vary each significant predictor in turn from its lowest to its highest value and compute the difference in the probability of voting yes.Footnote ²⁹ The results indicate that environmental attitudes have a predictably impressive impact on voting on Proposition 23. As these attitudes move from highly negative (anti-environment) to highly positive (pro-environment), the probability of voting yes falls by fully 0.81 points on the 0–1 probability scale. Liberal–conservative beliefs also have impressive effects, with strong liberals being 0.41 points less likely to vote yes than strong conservatives. Comparative senatorial candidate images and economic evaluations also had sizable effects. As voters’ images of the senatorial candidates move from strongly pro-Boxer to strongly pro-Fiorina or economic evaluations move from highly positive to highly negative, the probability of voting yes increases by 0.27 points. Other predictors are less powerful.Footnote ³⁰

Viewed generally, these results are consonant with the hypotheses that perceived costs and benefits and various heuristics both exert sizable effects on initiative and referendum voting. Given the subject of Proposition 23, attitudes toward the environment are expected to have—and do have—strong effects. The fact that the Proposition 23 vote was conducted in a context of economic hardship and explicitly pitted employment growth against environmental protection gives economic evaluations a cost-benefit tenor as well. Although the effects of economic evaluations are not as large as those associated with environmental attitudes, economic judgments have a substantial impact. Several heuristics matter too—the two strongest involve general liberal–conservative beliefs and voters’ images of the senatorial candidates taking opposing stands on Proposition 23. In the next section, we investigate competing hypotheses about interactions between candidate images and political knowledge.

Candidate Image Heuristics and Political Knowledge

The preceding analysis documents that the images of senatorial candidates Carly Fiorina and Barbara Boxer had substantial effects on voting on Proposition 23. As discussed above, similar results have been reported in earlier research, with some analysts conjecturing that the effects of candidate images are stronger among less politically knowledgeable voters. However, we hypothesize that the effects of candidate heuristics may be stronger among more knowledgeable individuals. Such voters are arguably more likely to recognize their decision-making limitations, to have images of the candidates and to know candidates’ positions on an initiative or referendum proposal under consideration.Footnote ³¹

We test these competing hypotheses by re-specifying the Proposition 23 vote model to include a political knowledge variable plus a multiplicative interaction term involving this variable and the Fiorina–Boxer comparative candidate image variable.Footnote ³² If the traditional hypothesis is correct, then this interaction variable will have a negative effect on Proposition 23 voting. However, if the alternative hypothesis is correct, the interaction variable will have a positive effect.

Testing these rival interaction effect hypotheses is not straightforward. In recent years, analysts have become increasingly aware of the need for care when drawing inferences about interaction effects in multivariate models.Footnote ³³ To date, this literature has focused heavily on interpreting interaction effects in linear regression models. However, Norton et al. (Ai and Norton Reference Ai and Norton2003; Norton, Wang and Ai Reference Norton, Wang and Ai2004; see also Brambor, Clark and Golder Reference Brambor, Clark and Golder2006; Berry, DeMeritt and Esarey Reference Berry, DeMeritt and Esarey2010) caution that additional complexities arise when one specifies interaction effects in nonlinear models such as the binomial probit model of interest here. Consider an interaction effect in a binomial probit model involving two predictor variables:

(2) $$E(y\,\mid\,X1,X2,X)=\Phi (B_1X1{\plus}B_2X2{\plus}B_{12}X1X2{\plus}XB)=\Phi (u),$$

where E(y|X1,X2,X)=conditional mean of dependent variable Y, $\Phi $ the cumulative normal distribution, X1 and X2 the predictor variables interacted as the product X1×X2 and XB the effects of other predictor variables in the model.Footnote ³⁴

Following the analogy with what is done with linear models, one might be tempted to interpret the marginal effect for X1X2 as the derivative of Φ(u) with respect to the interaction effect (XIX2), that is,

(3) $${{\partial \Phi (u)} \over {\partial (X1X2)}}=B12\Phi '(u).$$

However, the interaction effect in this nonlinear model is actually the cross-derivative with respect to X1 and X2. Thus, if X1 and X2 are continuous variables, as is the case in the probit specification of the Proposition 23 voting model, the interaction effect is:

(4) $$\eqalignno{ & {{\partial ^{2} \Phi (u)} \over {\partial X1\partial X2}}=B_{12}\Phi ^{'} (u){\plus}(B_1{\plus}B_{12}X2)(B_2{\plus}B_{12}X1)\Phi ^{{''}} (u). {\rm } $$

Ai and Norton (Reference Ai and Norton2003, 124) note four important implications: (i) if B ₁₂ is 0, the interaction effect is not necessarily 0; (ii) the statistical significance of the interaction effect is not a simple t-test on B ₁₂ but rather varies across cases (here Proposition 23 voters); (iii) the interaction effect is conditional on the full set of predictor variables in the model; and (iv) the sign of the interaction effect is not necessarily the sign on B ₁₂, but rather can vary depending on the values of other covariates in the model. Here, we employ these ideas to study interaction effects between candidate image heuristics and political knowledge in a binomial probit model of voting on Proposition 23. As noted, the model of interest (Model B) is identical to Model A except that we add a political knowledge variable and a multiplicative term for relative senatorial candidate image×level of political knowledge.Footnote ³⁵

Model B’s parameter estimates show the values of the coefficients for most of the predictor variables that appeared in Model A are largely unchanged (see Table 1).Footnote ³⁶ The smaller Akaike information criterion (AIC) value for Model B in comparison with Model A indicates that Model B has a superior fit.Footnote ³⁷ In addition, the coefficient for the comparative senatorial candidate image×political knowledge multiplicative term in Model B is positive (0.033) and statistically significant (p<0.001). However, there is considerable variation in the nature of this interaction effect across the electorate, with the effect being positive for 525 voters and negative for 327 voters. Tests of statistical significance (p≤0.05) for these effects are displayed in Figure 4 which plots the Z-scores for the interaction effects against the probability of voting yes.Footnote ³⁸ Given competing hypotheses that the comparative leader heuristic×political knowledge interaction effects could be either positive or negative, these significance tests are two-tailed (critical values of Z=+/−1.96, p=0.05). The shaded portions of Figure 4 show which voters have statistically significant positive interaction effects and which voters have significant negative interaction effects.

Fig. 4

Relative senatorial candidate images×political knowledge interaction effects in Proposition 23 vote model

The distribution of Z-scores shows that a large number (270) of cases are positive and statistically significant. In contrast, only a small number (seven) are negative and statistically significant. Thus, fully 97.5 percent of the statistically significant interaction effects are positive and only 2.5 percent are negative. These two groups constitute 31.3 and 0.8 percent, respectively, of the total sample of Proposition 23 voters (N=862). The remaining Z-scores (for 67.9 percent of the sample) are not statistically significant. The relatively large number of significant positive interaction effects is consistent with the hypothesis that, compared with their less knowledgeable fellow citizens, politically knowledgeable people tend to place more, not less, emphasis on candidate heuristics when deciding how to cast their ballots.

The concave shape of the plot of interaction effect Z-scores against the probability of voting yes in Figure 4 also is informative. This pattern illustrates that candidate image×political knowledge interaction effects tend to be statistically insignificant when the overall probabilities of voting yes are either very small or very large. In contrast, the size of the interaction effect tends to increase as the likelihood of voting yes or no approaches 0.5.Footnote ³⁹ Regressing the size of the interaction effect on the probability of voting yes folded at 0.5 shows that this relationship is very strong—the (folded) probability can explain fully 86 percent of the variance (Figure 5). Substantively, this relationship is consistent with the idea that as various forces influencing a voter’s decision on Proposition 23 become more evenly balanced, there is heightened uncertainty about what a voter will do. In this context of ceteris paribus uncertainty, the interaction between candidate image heuristics and political knowledge exerts its greatest force.

Fig. 5

Candidate image–political knowledge interaction effects by folded probability of yes vote on Proposition 23

Conclusion: Heuristics, Heterogeneity and Political Choice

One important dimension of the multifaceted debate on climate change has focused on the economic consequences of measures designed to mitigate the threat of global warming. Some political jurisdictions, including California, have taken action to curtail greenhouse gases, using the Cap-and-Trade concept for policy guidance. Opponents have argued that these policies will have seriously deleterious economic effects. When Californians went to the polls in November 2010, Proposition 23 gave them a voice in this debate—if they wished, they could vote to suspend AB 32, a climate change bill passed by the state legislature in 2006. Given California’s economic difficulties, some observers thought that voters might conclude that the bill was an unaffordable jobs killer. However, Proposition 23 was defeated by a sizable margin. Hard times notwithstanding, a majority of voters were not ready to abandon AB 32.

As hypothesized in previous studies, the analyses presented above testify that cost-benefit judgments and various heuristics significantly influenced how Californians cast their ballots. People who were impressed by the dangers posed by climate change strongly tended to vote no and those who were unimpressed strongly tended to vote yes. Economic evaluations were relevant too and, again, the effects were strong and straightforward. People who judged economic conditions negatively were more likely to support Proposition 23 than were those who were more sanguine about the economy’s current condition and future prospects. General liberal–conservative beliefs and risk aversion mattered as well—people identifying themselves as liberals and those who were highly risk averse were more likely to vote no than were self-identified conservatives and risk acceptant individuals. In addition, although images of gubernatorial candidates and President Obama were not influential, voters’ reactions to the senatorial candidates, Carly Fiorina and Barbara Boxer, had sizable effects. Fiorina enthusiastically endorsed Proposition 23 and people who were more favorably impressed by her than Boxer tended to vote yes. Contrariwise, Boxer opposed the proposition and people who were more favorably disposed toward her than her opponent were likely to vote no.

The analyses also shed light on how candidate heuristics influence political choice among voters with varying levels of political knowledge. Contradicting the hypothesis that more knowledgeable people accord less weight to candidate images, multivariate analyses indicate that the effects of senatorial candidate images tended to be greater among more knowledgeable voters. These interaction effects increased in strength as the overall set of forces prompting a yes versus a no vote became more evenly balanced. The latter finding suggests that candidate image cues are apt to be emphasized by knowledgeable voters when they encounter what is otherwise a difficult decision.

Historically, most studies of voting in initiatives, referendums and general elections have assumed homogeneous electorates. Despite findings stretching back to Converse’s (Reference Converse1964) much cited research indicating that political belief systems in mass publics vary significantly in content and structure, the assumption that forces driving political choice have the same weight for all voters remains widespread.Footnote ⁴⁰ However, the assumption is not ubiquitous—in recent years it has been challenged by researchers who hypothesize that the use and impact of heuristics varies across electorates. Consistent with this heterogeneity conjecture, present analyses indicate that knowledgeable voters tend to make more—not less—use of heuristics when faced with potentially important decisions such as the one posed by a ballot initiative to suspend climate change legislation in a period of economic distress. In addition, the statistical evidence suggests that the impact of candidate heuristics is greatest among voters who are most uncertain about how to cast their ballot. The interactive effects of candidate image heuristics and political knowledge warrant additional inquiry by scholars interested in the psychology of political choice in direct democracy and general election settings.

More generally, the present study illustrates the potential theoretical and practical significance of the “fast and frugal” heuristic theory developed by Gigerenzer and others. Applied here to studying the determinants of voting on an important ballot initiative, these theoretical ideas may be very useful for understanding decision making not only in direct democracy contexts, but also to choice behavior in the larger set of political arenas where information costs are substantial and people are subjected to an array of competing and conflicting messages.