VOTER BEHAVIOR AND ATTITUDE FORMATION UNDER UNCERTAINTY

Voters face at least two kinds of uncertainty: candidate uncertainty and outcome uncertainty. Candidate uncertainty means that voters cannot be sure what policies a candidate will deliver once he or she takes office. Outcome uncertainty means that voters cannot be sure what the effects of a policy will be once it is enacted. These two forms of uncertainty are conceptually similar, but operate independently of each other. In particular, even if voters know with certainty what policies a candidate will enact once in office, outcome uncertainty may still be considerable because the effects are hard to predict.

Most work on uncertainty in mass elections focuses on candidate uncertainty.Footnote ¹ If voters have a dispreference for risk, the attractiveness of a candidate or policy proposal should decrease as uncertainty increases (Shepsle Reference Shepsle1972). While some studies show that perceived candidate uncertainty negatively affects both votes for and evaluations of candidates (Alvarez Reference Alvarez1997; Bartels Reference Bartels1986), more recent studies find that voters are hardly sensitive to uncertainty (Berinsky and Lewis Reference Berinsky and Lewis2007; Tomz and Van Houweling Reference Tomz and Van Houweling2009) or mitigate uncertainty by relying on heuristics such as candidate party affiliation (Milita et al. Reference Milita2017).Footnote ²

Research in political behavior has paid little attention to outcome uncertainty, but this has been examined in formal models of strategic communication, especially in legislative politics (see, for instance, Gilligan and Krehbiel Reference Gilligan and Krehbiel1987, Reference Gilligan and Krehbiel1989). In these models, legislators generally possess some private information about the mapping from policy to outcome and try to influence the beliefs of other legislators to affect their voting behavior. This literature often focuses on characterizing the behavior of senders under different institutional settings. For example, when are senders truthful (Austen-Smith Reference Austen-Smith1990) and when do they reveal or conceal private information (Austen-Smith and Riker Reference Austen-Smith and Riker1987)? How much information senders reveal, and their ability to influence receiver beliefs, have also been examined in theoretical work on outcome uncertainty in mass elections (for example, Minozzi Reference Minozzi2011; Roemer Reference Roemer1994).

These models are based on three assumptions about how persuasion works. First, legislators have preferences over outcomes, but vote on policy proposals (Austen-Smith and Riker Reference Austen-Smith and Riker1987). Secondly, persuasion works through influencing beliefs about the proposal outcomes and not changing outcome preferences (Austen-Smith Reference Austen-Smith1990). Thirdly, beliefs influence proposal preferences because they change the expected outcome or the uncertainty of the outcome (Gilligan and Krehbiel Reference Gilligan and Krehbiel1987). I argue that these modeling assumptions provide important insights into how outcome uncertainty structures persuasion not only between legislators, but also between politicians and voters.

Empirical work on outcome uncertainty suggests that this type of uncertainty matters, but these studies have only partially examined the core modeling assumptions from the theoretical literature. Jacobs and Matthews (Reference Jacobs and Matthews2012) find that as a policy's time horizon increases, both policy and outcome uncertainty increase and support for the policy decreases. However, they do not address whether political actors can themselves influence voter outcome beliefs by making predictions about the effects of policies. Jerit (Reference Jerit2009) exposes respondents to positive and negative predictions about the effect of a policy proposal and finds that respondents give more weight to the negative prediction, which is consistent with a status quo bias. However, neither of these studies manipulate expected outcomes separately from the uncertainty of outcomes, which is necessary to disentangle the two mechanisms. Nor do they fully address the relationship between preferences over policies, preferences over outcomes and outcome beliefs. In their seminal work, Tomz and Van Houweling (Reference Tomz and Van Houweling2009) manipulate both the position and uncertainty of the position of a candidate in the policy space and find that, on average, policy uncertainty does not repel voters. This has not been done in the context of outcome uncertainty. We do not know how outcome uncertainty, produced by conflicting information, affects public opinion.

UNCERTAIN OUTCOMES IN A SPATIAL PREFERENCE MODEL

I provide a simple model of how voters form policy preferences under conditions of outcome uncertainty. The model builds on the assumption that parties campaign on policy proposals while voters care about policy outcomes. For example, voters do not care if the minimum wage is increased by $2 or $4. They care only about how the increase affects the living standard of minimum wage workers. To model this, we need to separate the policy space, Π, from the outcome space, Ω.Footnote ³ Although voters only have a preference over outcomes, indicated by the bliss point in the outcome space, O* ∈ Ω, they do not choose between outcomes. They choose between a policy proposal, P ∈ Π, and the status quo, SQ ∈ Π. While the effect of the status quo, O _SQ ∈ Ω, is known with certainty, voters do not know the policy outcome, O _p ∈ Ω, with certainty and must therefore assess the policy based on their beliefs about the outcome. Assuming that voters have quadratic utility and, consequently, are risk averse, a voter will prefer the policy over the status quo if:

(1)$$EU( P\vert m_p) = {-}[ O ^\ast {-}E( O_p\vert m_p) ] ^2-{\mathop{\rm var}} ( O_p\vert m_p) > -( O ^\ast {-}O_{SQ}) ^2 = EU( SQ) , \;$$

where E(⋅) is expectation and ${\mathop{\rm var}} ( {\cdot} ) $ is the variance of the belief about the policy outcome after receiving some prediction, m _p, about the policy outcome (Enelow and Hinich Reference Enelow and Hinich1981).Footnote ⁴ If the bliss point is fixed, the only way for political actors to influence public opinion is to change voters’ beliefs. As for models of candidate uncertainty (Alvarez Reference Alvarez1997; Hobolt Reference Hobolt2006), there are principally two ways that beliefs can affect attitudes. I illustrate these in Figure 1.

Figure 1. Illustrating two forms of persuasion

Note: the figure shows a policy space, Π, and an outcome space, Ω. P represents a policy proposal and SQ the status quo. O* represents the bliss point of the voter, O _SQ refers to the outcome of the status quo and O _p the outcome of the policy proposal. The outcome of the policy proposal is not perfectly known, which is represented by the probability distribution around O _p.

In all three panels in Figure 1, the policy proposal, P, and the status quo, SQ, are held constant. There is no uncertainty about what the policy or the status quo entail. The only thing that changes across the panels is the voter's belief about the policy outcome, O _p, represented by the probability distribution in the outcome space. The left panel is our benchmark case. The voter is certain that the policy proposal will produce outcomes closer to the bliss point, O*, compared to the status quo outcome, O _SQ; thus the voter prefers the policy proposal over the status quo.

The first way to change voter attitudes is by shifting the expected outcome of the policy proposal. This is captured by the first term of the expected utility of the policy in Equation 1 and is depicted in the center panel. The voter's belief has shifted compared to the left panel, and the voter is certain that the policy proposal will produce worse outcomes compared with the status quo. The voter now prefers the status quo over the policy proposal. From this, I derive the first hypothesis.

The second way to change voter attitudes is to change the variance of the beliefs – that is, the voter's perceived uncertainty about the outcomes. This is captured by the second term of the policy's expected utility in Equation 1 and is illustrated in the right panel. The expected outcome is identical to that in the left panel, but the uncertainty of the beliefs is much greater. There is a risk that the outcome of the policy proposal will be further from the bliss point compared to the status quo. In such a case, whether the voter prefers the policy over the status quo depends on its risk preferences. Assuming that voters are risk averse, increasing outcome uncertainty will decrease support for the policy compared with a safer option, and it follows that:

Separating policies and outcomes enables a natural understanding of voter persuasion in a spatial preference framework. I assume that outcome preferences are fixed, but that beliefs are subject to influence. Persuasion through changing beliefs enables voters to change their policy preferences while keeping their outcome preferences constant. Instead of changing voters' views of how the world should work, this form of persuasion changes the voters' views of how it does work. For example, to persuade voters to be less supportive of a minimum wage increase, voters do not have to be convinced that the purpose of the policy – helping low-income workers – is bad. It suffices to convince them that the policy may hurt low-income workers. This also implies that changing policy without changing beliefs may have little impact on public opinion. If no distinction is made between policy proposals and policy outcomes, persuasion implies shifting the voters' outcome preferences (Riker Reference Riker, Enelow and Hinich1990). Returning to the example, this means convincing voters that helping low-income workers is bad. This has implications for how outcome predictions affect voter preferences in the model. What the effect of a policy will be should not influence what a voter would ideally want the outcome to be. That is, outcome predictions should not affect what I call idealistic preferences. Consequently, I expect that:

Policy preferences, however, are a composite of outcome preferences and beliefs (Austen-Smith Reference Austen-Smith1990; Austen-Smith and Riker Reference Austen-Smith and Riker1987). Since policy preferences depend on outcome beliefs and outcome beliefs depend on predictions, I expect that:

From a Bayesian learning perspective, voters use predictions to update beliefs and, thereby, policy attitudes. But predictions about outcomes are not used to update idealistic preferences. In Appendix Section A, I provide a fuller interpretation of the model from a Bayesian learning perspective.

Voters are not experts, and need to rely on external information to form beliefs. Therefore, they should be more susceptible to persuasion through changing outcome beliefs compared to changing idealistic preferences. However, forming outcome beliefs is a complex task and voters will face conflicting messages from politically motivated actors. Political actors have an incentive to misrepresent reality for their own political gain. But they may also have an incentive to be sincere, if the policy outcomes are dire enough (Minozzi Reference Minozzi2011). Since voters cannot be sure what motivates politicians, communication becomes ineffective. In the model, I implicitly assume that voters update their beliefs when presented with predictions. Yet, it may be that voters discount predictions from politicians as cheap talk. The view of voters as rational and cognizant has been the subject of a great deal of scholarship. Numerous studies find that voters are ignorant, misinformed and persuadable (for example, Achen and Bartels Reference Achen and Bartels2016; Carpini and Keeter Reference Carpini and Keeter1996; Lenz Reference Lenz2013), yet other studies find that voters are sophisticated enough to distinguish cheap talk from costly signals (Alt, Marshall, and Lassen Reference Alt, Marshall and Lassen2016; Fernandez-Vazquez Reference Fernandez-Vazquez2019). Thus, we do not know whether voters are competent enough to identify and discount cheap talk signals. It is therefore of substantial interest to examine how voters update their beliefs when the self-interest of the sender varies. Even if voters update their attitudes strongly when their beliefs change, it is ineffective for political actors to attempt to shape public opinion by changing voter beliefs if voters discount information from political actors.

EXPERIMENTAL DESIGN

In this section, I present a survey experiment testing the hypotheses derived from the theoretical model. In the experiment, the respondents are presented with predictions of outcomes of three real policy proposals from the US political debate.Footnote ⁵ The treatments consist of manipulating the message, m _p, in Equation 1, to influence respondents’ posterior beliefs. The predictions vary in the predicted outcome, the uncertainty about the predicted outcome and sender credibility. I examine how the predictions influence individual outcome beliefs, idealistic preferences and support for the policy proposals.

Treatments

In the experiment, all respondents are presented with a vignette that briefly describes the purpose of the policy and explains that there is disagreement over its expected effects. Importantly, the vignette emphasizes that not implementing the reform will leave the outcome variable unchanged. This is all the information the control group receives about the reform. The treatment groups are presented with either one or two quantitative predictions from either experts or partisan actors regarding the effects of the policy. For the treatment groups, a graphic complements the vignette and summarizes the information in the prediction. It shows both the predictions and who their senders are. A treatment example is available in Appendix Section C.

The treatments are constructed as follows. First, the respondents are randomly assigned to the expert sender or partisan sender treatment. To avoid confounding, the same sender is assigned for all three reforms. Secondly, the respondents are randomly assigned to receive either one or two predictions. Thirdly, the respondents are randomly assigned to a specific value for the prediction(s) received. The second and third steps are performed once for each policy.

Compared with qualitative predictions stating that the reform will, for example, ‘strongly increase unemployment’, using numerical predictions gives precise meaning to the effects of the reform. This allows me to precisely change both the predicted outcome and the uncertainty of the proposals. However, when presenting respondents with quantitative or numerical information, it is important to provide them with a benchmark that anchors the relative magnitude of the quantity (Ansolabehere, Meredith and Snowberg Reference Ansolabehere, Meredith and Snowberg2012). Therefore, the graphic shows the development of the outcome variable over the past 10 years and the predicted effect of the proposed policy in the future. The purpose of illustrating the effect of a policy as an effect over time is to emphasize the causal nature of the prediction. As our intuition about causality is strongly associated with change over time, this should make it easier for respondents to understand the predictions. In Table 1, I provide six examples of the treatment graphs. Below I discuss the treatments in greater detail. I provide the full treatment texts used in the experiment in Appendix Section D.

Table 1. Examples of treatment graphs

Note: the table shows examples of treatment graphs for the single prediction treatment, the expert uncertainty treatment – that is, the multiple predictions treatment and the partisan uncertainty treatment.

Predicted outcomes

The first treatment numerically varies the predicted outcome to examine whether respondents change their support and beliefs as the predicted outcomes change. Comparing the responses among respondents who get two different predicted outcomes is equivalent to comparing the left panel with the center panel in Figure 1.

Table 2 gives an overview of the three policy proposals, the outcome variables and the levels of the predicted outcomes. The treatments include widely different predictions about the effects of the reforms. Some of the predicted effects closely correspond to the estimated elasticities from the policy evaluation literature, whereas others do not. The predicted effects include both large and small changes to the outcome variables. In Table 1, the first row shows one example per policy proposal of the predicted outcome treatment graphics.

Table 2. Reforms, outcomes and treatment levels

Note: the treatment levels for the Trans-Pacific Partnership reform are percent change from the baseline, but the information presented to respondents is in millions of manufacturing jobs.

Outcome uncertainty

The second treatment category is outcome uncertainty. I manipulate this by presenting the respondents with two predictions simultaneously, varying the distance between them. Similar to Tomz and Van Houweling (Reference Tomz and Van Houweling2009), who manipulate policy uncertainty by presenting a range of possible policy positions taken by a candidate, this exposes the respondent to a range of possible outcomes. In contrast to previous studies, this allows me to manipulate both the extensive and intensive margins of uncertainty. By keeping the predicted outcome constant and changing the spread between the predictions, I can disentangle the effect of increasing uncertainty from shifting the expected outcome.

To ensure comparability with the single prediction treatments, the two predictions in the uncertainty treatment are distributed symmetrically around the predicted outcome treatment levels. This is equivalent to identifying the treatment effect by comparing the left and right panels in Figure 1. To preserve the correspondence to the estimated effects from the policy evaluation literature, I bound the magnitude of uncertainty in the treatments. I put a maximum bound on the uncertainty to an interval of 4 percentage points (or percent for the Trans-Pacific Partnership reform). In the second row of Table 1, I provide one example per proposal of the uncertainty treatment when experts are senders.

I do not provide the respondents with explicit information about the predicted outcome for the uncertainty treatments. That is, I do not inform them that the two predictions are centered on some predicted outcome. Providing respondents with the predicted outcome risks attenuating the effect of uncertainty. When asked about their outcome beliefs, respondents might simply defer to the predicted outcome and ignore the uncertainty about it. Not providing this information allows me to better examine whether a treatment effect of uncertainty is driven by the perceived likelihood or the valuation of good versus bad outcomes. For example, suppose that uncertainty decreases support for a proposal. This could be driven by respondents perceiving bad outcomes as more likely, consistent with negativity bias, shifting their expected outcome toward the pessimistic prediction. This could also be driven by respondents being more concerned with avoiding negative outcomes than seeking positive outcomes, as suggested by the risk preference mechanism. If I do not detect any evidence of negativity bias, but find that uncertainty decreases support, this suggests that the effect is channeled through the risk preference mechanism.

Sender

To examine whether respondents can tell self-interested from credible actors, I manipulate the sender of the predictions. The senders are either non-partisan experts, or representatives from the Democratic or Republican parties. This addresses the question of whether politically motivated actors are able to influence voters’ beliefs and thereby shape public opinion.

I impose some restrictions on the predictions when the senders are partisans. For the minimum wage and corporate tax proposals, the partisan representatives always send cheap talk signals in the uncertainty treatment. Democrats always provide a more optimistic prediction about the effect of the minimum wage on the unemployment rate, and Republicans always provide a more optimistic prediction about the effect of the corporate tax rate on the employment rate. For the Trans-Pacific Partnership proposal, I randomly assign which representative provides the optimistic prediction. This enables me to test the extent to which partisan respondents rely on party cues when forming beliefs and attitudes. The last row of Table 1 shows one example of the treatment graphs per policy when partisans send conflicting predictions.

Main Specification

To test the theory, I estimate the following pre-registered specification, for individual i and reform j,

(2)$$\eqalign{& y_{i, j} = \beta _0 + \beta _1\,predictionCenter_{i, j} + \beta _2\,predictionSpread_{i, j} + \beta _3\,partisanSender_{i, j} \cr & \quad \quad + \beta _4\,partisanSender_{i, j}\cdot predictionCenter_{i, j} \cr & \qquad + \beta _5\,partisanSender_{i, j}\cdot predictionSpread_{i, j} + \varepsilon _{i, j}} $$

where Prediction Center is the predicted outcome around which the predictions are symmetrically distributed, Prediction Spread is half the absolute distance between the predictions, and Partisan Sender is a dummy indicating whether the senders are partisans or not. Prediction Center ranges from −4 to 4 and Prediction Spread from 0 to 2. The outcome variable, y, is either support for the reform, outcome beliefs or idealistic preferences. To avoid the monotonicity problem of heterogeneous priors, only treated individuals are included in the main specification. All models are estimated using ordinary least squares (OLS) with robust standard errors.