THE POLITICS OF ECONOMIC EXPECTATIONS:

Transcription

1 INFORMATION SOURCES, BELIEF UPDATING, AND THE POLITICS OF ECONOMIC EXPECTATIONS: EVIDENCE FROM A DANISH SURVEY EXPERIMENT JAMES E. ALT DAVID D. LASSEN JOHN MARSHALL FIRST DRAFT: FEBRUARY 2014 When and in what way will different types of messages containing politically relevant information affect voter beliefs and ultimately political preferences? This paper varies the source and ideological content of messages that voters receive about aggregate unemployment prospects using two survey experiments in Denmark. We find that new information from political parties, but particularly the highly credible Danish Central Bank causes voters to change their beliefs about future unemployment prospects. Contrary to results in the US, belief updating is no greater among voters when the source is aligned with previously expressed political preferences. Information-induced changes in unemployment expectations support economic voting without affecting redistributive policy preferences, suggesting that economic information matters primarily for evaluating government competence. Those whose propensity to vote changes are not the poorly informed swing voters, but politically engaged and cognitively capable mainstream voters. We wish to thank Alberto Abadie, Charlotte Cavaille, Alex Fouirnaies, Torben Iversen and Horacio Larreguy for valuable advice, as well as participants at the Harvard Political Economy and Comparative Politics workshops. Department of Government, Harvard University, james alt@harvard.edu. Department of Economics, University of Copenhagen, david.dreyer.lassen@econ.ku.dk. Department of Government, Harvard University. jlmarsh@fas.harvard.edu. (Corresponding author.) 1

2 1 Introduction Possessing and processing politically relevant information is a central feature of how voters hold governments to account and express their preferences over policies. This is particularly true of information about macroeconomic variables that inform beliefs about a government s competence in office (Fearon 1999; Rogoff and Sibert 1988; Rogoff 1990) and an individual s preferences over taxation and welfare provision (e.g. Meltzer and Richard 1981; Moene and Wallerstein 2001; Romer 1975). Although there is some evidence that voters are well-informed about relevant macroeconomic variables (Duch and Stevenson 2008), most evidence suggests that voters lack basic information about their political or economic contexts (see Anderson 2007). Such limited information appears to be important in explaining political outcomes. Bartels (2008) argues that limited information explains why there is no majority of Americans against tax policies operating counter to their economic interests, while information about corruption (Ferraz and Finan 2008), information about economic performance (Bartels 2008; Healy and Lenz 2014), and transparent chains of accountability (Powell Jr. and Whitten 1993) have helped hold governments to account at the polls. These observations suggest that if voters become better informed, political outcomes better reflect voter preferences. However, informing voters is not a simple process: outside experimental contexts, information is rarely provided by independent sources and without an accompanying slant. Rather, economic and political information is typically communicated by actors with incentives to deceive or persuade its recipients. Recognizing that much of the information targeted at voters is biased (Goidel and Langley 1995; Nadeau et al. 1999), it would not be surprising if new information provision did not affect the beliefs of skeptical voters (Gentzkow and Shapiro 2006; Mullainathan and Shleifer 2005). This raises the key question of what types of messages containing political information will 2

3 change voter beliefs and political preferences. In order to illuminate this issue, we provide a conceptual framework based in the differing roles of a message s source and ideological content. This framework points to four main ways in which information may affect belief updating: the effect of the source of the message depends upon its objective and subjective credibility, while the ideological content of the message may update beliefs by reinforcing existing prior beliefs or persuading voters with different ideological predispositions. While these ideas synthesize previous psychological insights, our characterization is novel and critical for delineating the empirical tests that are the focus of this paper. We use survey experiments to test these relationships, specifically examining how the source and content of messages conveying politically relevant information about aggregate unemployment affect voter beliefs and political preferences. Our experiments are conducted in Denmark, an open economy where macroeconomic concerns have been highly salient in the aftermath of the financial crisis, and where left-right political divisions remain entrenched. Furthermore, the combination of a panel political survey and access to an extremely detailed administrative government data provides a unique opportunity to use detailed respondent histories to differentiate the mechanisms underpinning belief updating and its translation into political preferences. Our first main contribution examines how the source and content of political messages affect voter beliefs. We find that the objective credibility credibility that derives from the source itself of the information source matters: an unemployment projection from the Danish Central Bank, which is highly credible among citizens, causes voters to update their belief more than receiving information from government or opposition political parties. While information from political parties still causes voters to update their beliefs, there is no evidence of subjective credibility credibility that reflects the characteristics of individual voters such that voters update more in response to unemployment projection from the party they have recently voted for. This suggests that parties can credibly transmit economic information to all voters when providing simple facts without an ideological slant. 3

4 [We are still waiting on results regarding message content. All results pertain to the effects of providing information about unemployment projections, assigned to political and independent sources.] How are changes in unemployment expectations manifested politically? Using instrumental variable techniques, we find that the provision of unemployment information increases economic voting. In particular, a percentage point decrease in unemployment expectations increases the probability of supporting the leading parties on economic policy in Denmark s coalition government by 3.5 percentage points. These findings could instead reflect changes in voter policy preferences, rather than beliefs regarding the competence of the government. However, this interpretation receives limited support as unemployment expectations do not affect the most likely arenas for altered preferences, attitudes toward redistributive or unemployment insurance policies. The results suggest that economic voting is an importance feature of political behavior even in Denmark s very open economy, but is not ubiquitous. Rather, given assigning responsibility for policy outcomes is especially challenging in Denmark s complex political system, only a subset of voters appear to be willing and capable of voting economically. In particular, we find that economic voting is neither concentrated among swing voters nor ideologues. Rather, we find economic voting is only observed among a reasonably well-educated subset of politically-engaged voters, and especially among those who already believe the economy is improving. These results have important implications for understanding democratic accountability, in addition to potentially explaining why parties tend to target target their messages at the more politically-engaged portion of the electorate (Adams and Ezrow 2009; Gilens 2005). This paper is structured as follows. Section 2 distinguishes the source and content of political messages, and considers how economic information might affect political preferences. Section 3 details our experiments designed to parse out these effects. Section 4 examines how beliefs change, before Section 5 maps these beliefs to vote intention and welfare policy preferences. Section 6 concludes. 4

5 2 Theoretical motivation This section first provides a conceptual framework distinguishing the potential roles for the source and content of politically-relevant messages. Focusing on economic information, we consider two prominent hypotheses economic voting and changing policy preferences as to how providing economic information could affect political preferences. These distinctions and outcomes motivate our experimental analysis. 2.1 What types of new information matter? Despite long-running attention to economic voting and growing interest in political information, it remains unclear what types of new information will change the beliefs and behavior of voters. Many researchers treat information as an unbiased resource helping voters to make the right decision (e.g. Feddersen and Pesendorfer 1996), or assume that voters start from a common prior (e.g. Rogoff and Sibert 1988; Rogoff 1990). In experimental work, information is frequently provided without a source and consequently relies upon the experimenter s credibility. However, in the real world, most politically-relevant information is conveyed by agents with distinct and often wellunderstood ideological biases and incentives to distort perceptions of the true state of the world (e.g. Gentzkow and Shapiro 2006; Mullainathan and Shleifer 2005; Zaller 1999). Accordingly, voters must evaluate the information they receive in terms of both the credibility of the source and the persuasiveness of the content of the message. We distinguish two forms of source credibility that could affect belief updating after receiving new information. Objective credibility reflects the beliefs about the source that depend upon the institutional characteristics of the source itself (Ansolabehere, Meredith and Snowberg 2012; Zaller 1999). For example, central banks are often highly credible because they have few political incentives to deceive voters and often successfully establish a reputation for sending accurate messages. Conversely, political parties and certain news providers have widely-understood biases. 5

6 H1. (Objective credibility) Fixing message content, belief updating is greater when the source is objectively credible. To be precise, greater belief updating entails larger shifts in the mean of an individual s belief distribution. Whether an objectively credible source affects beliefs more than receiving a message going against the expected bias of a less objectively credible source is an empirical question we address below. Subjective credibility instead depends upon the characteristics of the receiver of the message. Political psychologists have argued that respondents differ in their accuracy goals and directional goals, where the former types seek to make decisions based on the most accurate information, while the latter only seek information that confirms their prior beliefs (Taber and Lodge 2006). In particular, a large literature in the US has suggested that knowing the position of a political party on an issue strongly conditions a voter s beliefs and preferences (see Boudreau and MacKenzie 2014; Bullock 2011; Malhotra and Kuo 2008; Healy and Malhotra 2013). Similarly, Mullainathan and Shleifer (2005) show that with heterogeneity in priors over politically divisive issues, newspapers separate in their reporting of the news and cater to a segmented market where the credibility of information in the eyes of the consumer varies considerably. Gentzkow and Shapiro (2006) show that in segmented markets, media bias will exist and be expected by voters. These considerations imply, H2. (Subjective credibility) Fixing message content, belief updating is greater when the source is subjectively credible for a given voter. Both objective and subjective credibility could simultaneously cause a voter to update their beliefs, although our empirical design can separate their relative magnitudes. 1 A second key facet of politically relevant messages is the extent to which their content taps into compelling political or ideological arguments. Here, ideology defines fundamental beliefs 1 Our approach can be clearly shown in a Bayesian updating framework. Specifically, we write 6

7 about the economic world; in our empirical application, the relevant issue is whether voters tend to believe expansionary fiscal policy can be effective at stimulating the economy. For simplicity, we divide voters into three types: right-wing ideologues, left-wing ideologues, and non-ideologues. We further define a political message as congruent when its ideological argument matches the voter s type. Because the appeal of different ideological arguments will vary with the prior beliefs of different types of voters, we expect the content of messages to operate on a more specific subset of like-minded and persuadable voters. Like-minded voters share the ideological argument of the message, and therefore new information should have a reinforcing effect. H3. (Reinforcement) When provided with new information containing a political argument that is congruent with their prior ideology (holding the source fixed), voters update their beliefs in accordance with the argument more than when faced with new information and an incongruent (especially opposing) argument. Rather than provide new information that fits with a voter s prior belief structure, new information can also persuade voters. This is likely to be particularly prevalent among non-ideological voters. H4. (Persuasion) Persuasion occurs if non-congruent voters provided new information with a political argument (holding the source fixed) update their beliefs in accordance with the individual i s conditional posterior belief about future unemployment level, U, as: P(U = u X i,z i ) = P(U = u X i ) P(Z i U = u,x i ), P(Z i X i ) where Z i is an information shock received by i, and X i captures i s relevant pre-information history or type (e.g. their ideology). i s prior belief, P(U = u X i ), depends upon their pre-information characteristics X i. The likelihood P(Z i U = u,x i ) represents i s interpretation of the informativeness of the signal they received: P(Z i U = u,x i )/P(Z i X i ) = 1 or P(Z i U = u,x i ) = P(Z i X i ) captures i not believing that receiving signal Z i is related to the likelihood that the state of the world is U = u. H1 hypothesizes that P(Z i U = u,x i ), or more simply P(Z i U = u) because individual characteristics play a weak role in objective credibility, is large. H2 instead hypothesizes that updating depends upon the characteristics X i of the receiver. 7

8 argument. This effect should be larger for non-ideologues than voters with opposing ideologies. It is important to distinguish persuasion by political argument from the source of the argument, which is the claim in H The political role of economic information We focus on two important ways in which economic information, namely aggregate unemployment prospects, could affect political preferences: economic voting and changing welfare policy preferences Economic voting The idea that governments may be rewarded or sanctioned by voters on the basis of their economic performance is well-established (see Anderson 2007; Lewis-Beck and Stegmaier 2000). The logic underlying this argument is that voters impose sanctions on the basis of economic outcomes to deter re-election seeking politicians from choosing suboptimal policies (Barro 1973; Ferejohn 1986), or looking forward use the available information to select the most competent candidate (Fearon 1999; Rogoff and Sibert 1988; Rogoff 1990). 2 Both backward and forward-looking information can help to evaluate office-holders. The empirical evidence assessing whether economic success translates into higher likelihoods of an incumbent being re-elected has been mixed (Anderson 2007), and has struggled to provide compelling evidence of a causal relationship (Healy and Malhotra 2013). Nevertheless, to the extent that economic performance is a key election issue and is deemed to possess the capacity to affect the economy (Duch and Stevenson 2010), information about macroeconomic performance is expected to increase economic voting: 2 That motives underpinning this approach could be either sociotropic or self-interested. As Ansolabehere, Meredith and Snowberg (2012) have shown, parsing out these effects is challenging. 8

9 H5. (Strong economic voting) If an individual s unemployment expectations decrease, the likelihood that they vote for a party in government increases. Economic voting models somewhat heroically assume that voters obtain and process sufficient information about policy choices or at least their (expected) outcomes to attribute responsibility and evaluate incumbent performance. These assumptions are now receiving greater scrutiny, primarily in the US context (see Anderson 2007; Healy and Malhotra 2013). Research has shown that voters have often lacked even the minimal information required to vote according to economic performance (e.g. Campbell et al. 1960) or suffer partisan biases in attribution (Fiorina 1981; Rudolph 2003a, 2003b, 2006; Malhotra and Kuo 2008; Tilley and Hobolt 2011), while informed voters have lacked the motivation or cognitive capacity to translate information into responsibility designation (Bartels 1996; Delli Carpini and Keeter 1996; Zaller 1992). These problems can be multiplied in institutional contexts characterized by multiple loci of decision-making power, where even the most willing economic voter may struggle to assign responsibility for economic performance (Duch and Stevenson 2008; Nadeau, Niemi and Yoshinaka 2002; Powell Jr. and Whitten 1993). Combining these behavioral insights, the economic voting hypothesis can be re-stated with greater conditionality: H6. (Weak economic voting) If an individual s unemployment expectations decrease, the likelihood that they vote for a party in government (in any given institutional context) increases only if the individual has the cognitive capacity to assign government responsibility to economic performance. In this light, economic voting is not the inevitable by-product of providing economic information at least for some voters. In particular, our focus on the source and content of the messages received by voters provides a foundation for evaluating the role of subjective factors in belief updating, and thus a natural basis for understanding what types of information could push voters toward the normatively appealing action of voting according to government economic performance (Anderson 9

10 2007) Updating welfare policy preferences In contrast to economic voting models arguing that new information helps to identify the best candidate, new information could also change voter beliefs about what the best policy is. Standard political economy models assume that voters form their preferences on the basis of complete information about the effects of policies. However, since voters are often poorly informed about the implications of different policies, they can support policies that may not be in their best interests (e.g. Bartels 1996, 2008). Information about aggregate unemployment could affect an individual s calculus by altering perceptions of macroeconomic aggregates, altering perceptions of an individual s own risks, or through altruistic preferences. Here, we assess whether the changes in beliefs about aggregate unemployment cause voters to update their policy preference, focusing on how preferences over welfare policy might change. In terms of preferences over redistribution, information about lower aggregate unemployment has two conflicting effects at least the individuals below the mean income for whom a positive tax rate is preferred. The Romer-Meltzer-Richard (Meltzer and Richard 1981; Romer 1975) model predicts that individuals perceiving an increase in taxable wages (due to decreased unemployment relative to their priors) should, holding their perception of their own income constant, increase their support for redistribution as the amount of income available for redistribution grows. Conversely, to the extent that lower aggregate unemployment decreases an individual s own expectations of being unemployed (and thus their expected income), the model predicts a decline in support for redistribution, holding the average income in the economy constant. Similarly, altruistic individuals whose own utility increases with the employment and wages of others would similarly oppose redistribution as unemployment prospects improve. We show these implications formally in the Appendix. 10

11 A specific policy that could be affected is unemployment insurance. Social insurance models, where replacement rates are received only once an individual becomes unemployed, make similar predictions (Iversen and Soskice 2001; Moene and Wallerstein 2001, 2003). If voters do not update their own employment expectations, but see that the rest of the economy is not doing better than expected, then their preference for unemployment insurance decreases in anticipation of paying more for a resource they will not use. Altruistic voters and individuals who update their own expectations about the risk of unemployment experience the opposite effect. These results easily follow from Moene and Wallerstein s (2003) simplified models. These logics imply the following predictions: H7a. (Nationally-oriented policy preference change) If an individual s unemployment expectations decrease, and primarily updates their beliefs about the country s economic prospects, then unemployment expectations should be negatively correlated with preferences for redistribution and social insurance. H7b. (Own-oriented policy preference change) If an individual s unemployment expectations decrease, and primarily updates their beliefs about their own economic prospects (or the prospects of others that are included in their utility functions), then unemployment expectations should be positively correlated with preferences for redistribution and social insurance. 3 Research design 3.1 Danish political context Danish political divisions have traditionally revolved around left-right divisions over economic policy, and governments have oscillated between center-left and center-right coalitions. In 2011, Social Democrat Helle Thorning-Schmidt became Denmark s first female Prime Minister, having 11

12 led the left bloc containing the Social Democratic, Social Liberal and Socialist People s parties to victory over a center-right coalition led by the Liberals that had held office since Dissatisfaction with the government s economic performance was a major issue in the 2011 election. 3 Having sustained very low levels of aggregate unemployment throughout the 2000s, the financial crisis hit Denmark s open economy badly. In early 2008 unemployment hit new lows of nearly 3%, but had increased to around 8% by the 2011 election. 4 The budget deficit also ballooned, leaving Denmark with hard fiscal choices. The center-right s austerity policies were widely blamed for the failure to produce a strong economic recovery. 5 Since the 2011 election, the Danish economy improved. In January 2013, gross unemployment had officially fallen to 7.4%. 6 Importantly for our study, the Danish Central Bank (DCB) expected this rate to fall to just below 7% by January 2014 (which turned out to be exactly right). However, within-coalition tensions between the economically liberal Social Liberals and the socialist Socialist People s parties increased. The Social Liberals only joined the coalition after agreeing a significant conservative welfare reform with the center-right before the election, and these differences culminated in the Socialist People s Party leaving the coalition in January 2014 over unpopular plans to privatize the country s state-owned energy company. Economic policy has remained particularly contentious: just before our January 2013 survey began, Parliament passed a law reducing the maximum length of unemployment benefits from four to two years, only for this to be repealed in lieu of agreement in June 2013 to instead reduce benefits in the final two years to 60% of their initial level. 3 See this Economist article. 4 Unemployment data from Eurostat here. Although Eurostat computes unemployment using surveys to ensure cross-national comparability, the Danish government uses administrative records to calculate gross unemployment (which is very similar, and what is used in this paper). 5 Even though the financial crisis itself was not the fault of Denmark s government at the time, governments can still held responsible for exogenous shocks (see Duch and Stevenson 2008). 6 Gross unemployment is the official unemployment figure used by the government, and is calculated using administrative register data. Gross unemployment differs from net unemployment in that participants in active labor market programs are included in the unemployment rate. 12

13 3.2 Experimental design To examine the hypotheses derived above, we embedded survey experiments in the 2013 and 2014 waves of the Danish Panel Study of Income and Asset Expectations (Kreiner, Lassen and Leth- Peterson 2013), an annual panel survey of around 6,000 broadly nationally representative Danes conducted every January/February. 7 The panel, which has been conducted since 2010 and suffers from limited attrition, asks wide-ranging question about the respondent s financial position as well as their political preferences. Moreover, Statistics Denmark have linked the data to the Central Person Registry, an extraordinarily rich administrative dataset used by the government to compile wide-ranging information about all Danes. This enables us to construct very detailed respondent histories. The central goals of the experiments are to evaluate under what conditions the provision of economic information affects individual beliefs and political preferences. We designed our treatments to differentiate the effects of political sources and political content, and use heterogeneous effects to further isolate hypotheses. As shown in Figure 1, our first experiment provides factual content in an apolitical manner by varies the source of the information. This permits tests of H1, H2, and H5-H7. Our second experiment is more ambitious, distinguishing the source of political content and also separating out factual from political content. The former aspect tests H2 with political content by comparing political to non-political sources of the same information, and H3 by comparing whether respondent differentially update their beliefs according to the ideological nature of the political content. The later aspect incorporates variation in content, seeking to isolate the effect of political arguments from purely factual information about a policy. This distinction separates the persuasion effects inherent in H4 from information effects. Experiment 2 also examine political 7 The first wave in 2009 randomly chose around 6,000 respondents from the national register. Annual attrition has been around 20-30%. The sample has been replenished with randomly chosen respondents from the register. 13

14 Source Content Factual Political Independent Exp. 1 and 2 Exp. 2 Political Exp. 1 Exp. 2 Figure 1: Experimental framework for differentiating the source and content of economic information outcomes to test H5-H Treatments Experiment 1. In January 2013, we focused on source credibility by varying the source of unemployment forecasts containing no overt political content, as well as the forecast itself. After being asked what they believed the current unemployment rate is, respondents were randomly assigned to one of eight treatment conditions with around 700 members each. The control group received no information, while six treated groups were read the following statement: DCB/Government/Opposition 7/5% treatment: Assume that that the [DCB/government/Liberals] estimates unemployment in 2013 to be [almost 7%/around 5%]. (Survey treatments and questions translated from Danish.) Respondents were therefore informed that the DCB, the government or main opposition party project that unemployment of the next year will be almost 7% or around 5%. The true DCB projection for gross unemployment was almost 7%. However, because only the DCB has publicly stated this, ethical considerations required that our other primes begin with assume that.... In order to examine the extent to which such wording weakens the treatment, our final treatment group was truthfully told The DCB estimates unemployment in 2013 to be almost 7%. We compare this treatment to the analogous assume version, and ultimately find no statistical difference in the distribution of unemployment expectations. 14

15 These sources vary considerably in their credibility among voters of all political stripes. Unlike some other central banks, the DCB is highly regarded by voters, and is not regarded as having a right-wing agenda or being an instrument of government. Asking respondents how much trust they place in each source, the DCB is attributed far greater credibility: 67% of respondents trusted or greatly trusted the DCB, while only 17% and 27% trusted or greatly trusted the government and Liberals respectively. 8 These numbers are broadly in line with mass surveys conducted by Statistics Denmark: in 2011, they found that while 82% trusted the DCB, only 59% trust Parliament. 9 Eurobarometer data indicates that trust in Denmark s political parties is very similar to the EU mean (European Commission 2011). [Note: experiment 2 is currently being run!] Experiment 2. The January 2014 treatments were designed to examine how political content changes the updating process. Rather than provide specific unemployment projections, these treatments focused on arguments regarding the unemployment implications of the government s new policy of increasing government consumption by 0.7% per year on average until This issue has divided the leftist coalition government and the Liberals. The government has argued that the policy will support private sector growth by increasing productivity in the economy. Conforming to traditional left-right divisions, the Liberals have argued that government consumption should be frozen in order to best stimulate the economy by reduced crowding out. We presented both arguments to treated respondents. All treated respondents were told of the policy, but differed in whether they received the government s argument, opposition s argument, or no argument. Just being informed about the policy itself distinguishes the ideological message from information about the policy itself. To separate the effects of the political source from political content of the message, the same government and opposition arguments were also attributed to some economists in both cases. As in the first experiment, respondents were read one of the 8 Only the control group responses were used because this question followed the treatment, and thus including the full sample could bias our estimates. 9 See report summary here. 15

16 following five statements after being asked what they believed the current unemployment rate is: Policy information: The government plans to increase public consumption by about 0.7% per year until Policy information + Government argument: The government plans to increase public consumption by about 0.7% a year until The government argues that increasing productivity in the Danish economy will make it more attractive for Danish firms to hire people and thereby lower unemployment. Policy information + Left-wing argument: The government plans to increase public consumption by about 0.7% a year until Some economists argue that increasing productivity in the Danish economy will make it more attractive for Danish firms to hire people and thereby lower unemployment. Policy information + Opposition argument: The government plans to increase public consumption by about 0.7% a year until Venstre argue that increasing productivity in the Danish economy is not sufficient to lower unemployment, while government consumption may happen at the expense of the private sector employment. Venstre instead argue that zero growth in public consumption will lower unemployment by increasing private sector activity. Policy information + Right-wing argument: The government plans to increase public consumption by about 0.7% a year until Some economists argue that increasing productivity in the Danish economy is not sufficient to lower unemployment, while government consumption may happen at the expense of the private sector employment. These economists instead argue that zero growth in public consumption will lower unemployment by increasing private sector activity. These messages, which unlike many other survey experiments actually reflect the true positions of the parties (see Boudreau and MacKenzie 2014), are political in content in that they both provide 16

17 politically-charged arguments to claim that their preferred policy would effectively reduce unemployment. Since the Liberals are out of government and therefore cannot choose policy their message cannot state what will happen but instead provides a partisan logic for believing that the government s policy may not actually reduce unemployment Outcome variables We consider two types of outcome: unemployment forecasts and political preferences. To capture unemployment expectations we asked respondents What is your best estimate of the unemployment rate in 2013? We would like your best guess, even if you are not completely sure. 10 This question was asked immediately after respondents received their treatment. The 20 respondents in 2013 who answered that the unemployment rate would exceed 50% were removed. 11 Political preferences are measured by voting intentions and self-reported policy preferences. First, we code indicator variables for intending to vote for the various individual parties, as well those in the governing coalition (Social Democrats, Social Liberals and Socialist People s parties) and right-wing parties. 12 Because turnout in Denmark typically exceeds 85%, 13 and 72% of respondents ultimately voted for the party they intended to vote for eight months after the 2011 survey, vote intention represents good approximation for what would happen if an election were held immediately. Second, policy preferences are captured by support for redistribution and support for unemployment benefits benefits. The five-point redistribution responses range from every man for himself (1) to the government should help the poor a lot (5), after being asked whether the 10 This variable can be thought of as i s posterior belief P(U = u X i,z i ) assigned greatest probability or the mean of their distribution E[U Z i,x i ], while we estimate the average prior P(U = u X i ) over i using the control group s mean unemployment expectation. 11 These individuals were very evenly spread across treatment conditions, with between 2 and 4 omitted respondents in each group. 12 This question was asked 18 questions after the treatment was administered. 13 See Institute for Democracy and Electoral Assistance. 17

18 government should be responsible for maintaining the living standards of the poor. 14 The unemployment insurance policy question informed voters that The economic crisis has meant that many people have lost their job before offering respondents a 3-point less-same-more response as to whether the government should support the unemployed Identification and estimation Treatment status is well balanced across pre-treatment covariates. Tables 5 and 6 in the Appendix confirm balance across variables frequently included in observational studies regressing political preferences on a set of covariates. Because the survey is administered without breaks where interference could occur, violations of the stable unit treatment value assumption are implausible. Given random assignment and non-interference, our empirical analysis can straight-forwardly identify the causal effects of the treatments. To estimate the average treatment effect on the treated for each information treatment in the vector Z i on unemployment expectations U exp i, we estimate the following equation using OLS: U exp i = Z i α + λu now i + ε i, (1) where the respondent s belief about the current unemployment rate, U now i, is included to enhance efficiency in some specifications. 16 Robust standard errors are reported throughout. Interaction terms are added to allow for heterogeneous responses to treatments, and thus aid characterization of which types of individual the treatments affect. To identify our ultimate quantity of interest, the causal effects of unemployment expectations on political preferences, we use our information treatments as instruments for unemployment ex- 14 This question was 19 questions after the treatment. 15 This question was asked immediately after unemployment expectations were elicited, and thus is most likely to find an effect. 16 Unsurprisingly, the results are highly robust to the inclusion of control variables. 18

19 pectations. This approach overcomes the obvious concern that economic expectations may be correlated with omitted variables that also affect political preferences. Taking equation (1) as the first stage, we thus estimate the local average causal response (LACR) (Angrist and Imbens 1995), averaging the causal effects for compliers (individuals for whom our randomly-assigned information treatments induced respondents to change their unemployment expectations) across different unemployment expectation levels. 17 Accordingly, we estimate the following structural equation using 2SLS: Y i = τu exp i + δu now i + ξ i, (2) where Y i is vote intention, support for redistribution, or support for extending unemployment insurance. Consistent estimation of the LACR requires two further assumptions: monotonicity and an exclusion restriction (Imbens and Angrist 1994; Angrist, Imbens and Rubin 1996). Monotonicity entails that each individual would update their unemployment expectations in the same direction upon receipt of the treatment. Although it is hard to imagine when prominent public sources would induce voters to update their beliefs against the information provided, Figures 5 and 6 (below) show that the information treatments appear to violate monotonicity by causing a small number of respondents with especially low prior unemployment expectations to converge upon the rate proclaimed by the treatment at least for the 7% treatments. Fortunately, the monotonicity assumption can be weakened in ways consistent with our data. 18 In general, IV estimation recovers a very similar quantity of interest to the LACR when few 17 The local average causal response here is the linearized causal of effect of unemployment expectations, weighted such that areas where the density function of complier responses is greatest. 18 One sufficient assumption is that there are more compliers than defiers for any combination of potential outcomes (Assumption 2.4 and equation 2.2 in de Chaisemartin 2013), while a weaker condition requires only that some subset of compliers has the same size and marginal distribution over potential outcomes as defiers, or that there are more compliers and defiers at each potential outcome (de Chaisemartin 2013). 19

20 subjects are defiers, or if defiers and compliers have reasonably similar distributions of potential outcomes (de Chaisemartin 2013: 7) and is identical under constant causal effects (see Angrist, Imbens and Rubin 1996). 19 In this application, there are few defiers, while we provide empirical evidence indicating that the effect is similar when we restrict the sample to treatments with almost no defiers. The exclusion restriction, which requires that the instrument only affects Y i through U exp i, is usually more problematic in empirical studies. Although such violations are unlikely, perhaps the most plausible violation arises where information treatments prime respondents to think more carefully about government performance and policies (beyond the effect of changing beliefs about unemployment expectations), inducing bias if such thinking systematically affects support for the government. We assess this possibility by looking at whether belief in the importance of political information for either private economic decisions or as part of the respondent s job differs across treatments groups (or comparing the control to all treated respondents), and find no difference. 4 Effects of information on economic expectations This section shows that the information treatments substantially change unemployment expectations, principally through objective credibility rather than subjective credibility, and especially among the least informed voters whose prior considerably exceeds the DCB s projection of around 7%. We first examine the distribution of the data, before proceeding to regression analyses quantifying average effects and identifying which types of individuals update their beliefs most. 19 Rather than recover the local average treatment effect, the Wald estimator (with no covariates) recovers the average treatment effect for a smaller group of compliers (precisely those not canceled out by the defiers. 20

21 4.1 Results Figures 2 and 3 plot the distribution of unemployment expectation responses by treatment condition. It is clear to see that the assume wording barely affects the distribution of DCB 7% projection responses. In fact, although stating the claim as fact slightly increases the density at its peak, tests comparing the mean and variance of the distributions cannot reject the null hypothesis of identical sample moments. This suggests that the statement wording is not biasing the results. Henceforth we pool the DCB 7% treatment groups. The leftward shift in density associated with the treatments demonstrates that all types of information sources significantly reduce average unemployment expectations. That we observe a reduction reflects systematic pessimism in a population where the average member of the control group expected an unemployment rate of 9.0%. Furthermore, despite its optimism relative to the true DCB claim, the 5% treatments dragged expectations below those receiving the 7% treatments. In all cases, the information treatments significantly reduced the variance of the distributions, providing further evidence that the treatments induce a response from voters. 20 Consistent with objective credibility (H1) and the differences in institutional credibility noted above, receiving information from political parties caused voters to update their beliefs less than information from the DCB did. Furthermore, distributional tests show that the DCB treatments induced substantially more similar responses from voters (i.e. smaller standard deviation). Nevertheless, partisan sources still substantially reduced unemployment expectations. Surprisingly, given that the Liberals have a strong incentive to criticize the government, the Liberal projections of lower unemployment than the prior of the average Dane did not cause voters to update more than receiving the expected optimistic message from the government. Statistical tests clearly indicate that the political party treatments cannot be differentiated at either unemployment 20 Although these apparent changes in beliefs could in part reflect anchoring biases (Tversky and Kahneman 1974), it is hard to see how such explanations could explain the changes in political preferences we document below. 21

22 Density Control DCB 7% Assume DCB 7% Assume DCB 5% Unemployment forecast 2013 Figure 2: Unemployment expectations by DCB treatments 22

23 Density Control Assume govt. 7% Assume govt. 5% Assume opp. 7% Assume opp. 5% Unemployment forecast 2013 Figure 3: Unemployment expectations by political party treatments 23

24 projection. In sum, it is clear that although sources differ in their objective credibility, respondents are willing to update their beliefs in response to all sources. Table 1 confirms these observations by estimating equation (1). The results reiterate the large changes in unemployment expectations: receiving the 7% treatment reduces unemployment expectations by around 1 percentage point, while the 5% treatment subtracts a further 0.5 percentage points. Supporting the importance of differences in objective credibility, the p-values associated with t-tests comparing the DCB coefficients to the party coefficients suggest that the credibility differences border statistical significance for the 7% and 5% treatments. The inclusion of current unemployment beliefs as a control substantially increases estimate efficiency in specification (2), and is thus retained for the subsequent analysis where instrument strength is particularly important. Rather than objective credibility driving our results, it could be the case that the treatments increase confidence in the source themselves. However, we find no evidence that the treatments increase confidence in either political party, and only marginal increases in trust of the DCB. 4.2 Heterogeneous effects: which individuals update their beliefs? The potential impact of different messages depends on how voters update their beliefs. Consequently, it is critical to understand which voters update most in order to be able to predict how new information will affect political preferences. We move beyond Kendall, Nannicini and Trebbi (2013) to explore this key question by using heterogeneous effects as a means of identifying the set of voters for whom additional information is important. We first test for subjective credibility (H2) by assessing how the source of our message providing unemployment projections affects voters with different political preferences. Table 2 shows that there is no strong evidence for differential updating: respondents who voted for a government (right) party at the last election did not update their beliefs significantly more when provided with information from the government (opposition). 21 In combination with the finding that voters 21 As a robustness check we also defined left and right-wing supporters as respondents who 24

25 Table 1: Effect of information treatments on unemployment expectations (%) (1) (2) DCB 7% treatment (combined) -1.12*** -0.93*** (0.20) (0.10) DCB 5% treatment -1.66*** -1.51*** (0.23) (0.13) Government 7% treatment -0.85*** -0.79*** (0.21) (0.12) Government 5% treatment -1.22*** -1.36*** (0.23) (0.13) Opposition 7% treatment -0.92*** -0.76*** (0.22) (0.12) Opposition 5% treatment -1.34*** -1.34*** (0.24) (0.14) Test: DCB 7% = Government 7% p = 0.03 p = 0.16 Test: DCB 7% = Opposition 7% p = 0.16 p = 0.06 Test: Government 7% = Opposition 7% p = 0.65 p = 0.73 Test: DCB 5% = Government 5% p = 0.02 p = 0.24 Test: DCB 5% = Opposition 5% p = 0.10 p = 0.22 Test: Government 7% = Opposition 7% p = 0.57 p = 0.90 Control for current unemployment estimate! Observations 5,705 5,705 R Notes: all specifications estimated using OLS; robust standard errors in parentheses; p < 0.1, p < 0.05, p <

26 Table 2: Heterogeneous effects of information treatments on unemployment expectations (%), by conditional marginal effect DCB 7% DCB 5% Govt. 7% Govt. 5% Opp. 7% Opp. 5% Linear effect *** ** *** ** *** ** (2.054) (2.606) (2.09) (2.857) (2.353) (2.503) News every day 0.462* (0.253) (0.308) (0.304) (0.309) (0.271) (0.33) Improving economic prospects 0.665*** 0.565** 0.827*** 0.647** 0.576** 0.527** (0.208) (0.25) (0.235) (0.264) (0.238) (0.259) Expected income (log) 0.546*** 0.345* 0.436*** ** 0.328* (0.157) (0.199) (0.155) (0.222) (0.179) (0.194) Medium education 0.663** 0.768** 0.797** 0.629* 0.642* 0.732* (0.308) (0.374) (0.406) (0.381) (0.348) (0.425) High education 0.948*** 0.967** * 0.913** 1.377*** (0.362) (0.429) (0.466) (0.445) (0.397) (0.513) Woman *** * *** *** *** *** (0.216) (0.266) (0.253) (0.269) (0.256) (0.285) Voted left at last election (0.215) (0.259) (0.249) (0.266) (0.247) (0.276) Notes: all coefficients are estimated from a single OLS equation interacting all six treatments conditions with the variables on the left hand side of the table (see Appendix for their definitions), and controlling for current unemployment expectations; the sample size is 5,465; robust standard errors in parentheses; p < 0.1, p < 0.05, p <

27 do not update more when the opposition says that the government is performing well, this result suggests that the main factor determining differential updating when the content of a message is non-ideological resides in the objective credibility of a source rather than its political identity (i.e. government or opposition) which matters for subjective credibility. 22 While the extent to which a voter updates their beliefs does not depend on their previous political allegiances, we do find support for intuitive explanations for heterogeneous updating. Table 2 shows that respondents with higher levels of education, higher expected income, greater economic prospects for the Danish economy relative to the previous year, and men update least. 23 Including the interaction of the respondents current unemployment rate estimate with the information treatment substantially causes only the interaction with national economic prospects perceptions to remain statistically significant. This clearly indicates, in conjunction with correlations within our control group, that female, low education and poorer citizens update their beliefs most. 5 Effects on political preferences The preceding analysis has shown that information about aggregate unemployment projections affects voter beliefs about the economy s prospects. However, does this translate into affecting political preferences? Supporting H6, this section finds that exogenously changing expectations causes informed and cognitively able voters to change their vote intentions in accordance with economic voting motivations, but does not affect their policy opinions (as H7a and H7b hypothesized). These results imply that aggregate unemployment expectations are principally used to evaluate the competence of the government. intended to vote for the same left or right party in the 2011 and 2012 surveys. We again find to evidence of subjective credibility. 22 Furthermore, looking across all interaction terms, the respondent s own political position does not affect their updating. 23 At the cost of 1,000 observations, we also included the respondent s subjective probability of being without a job in the forthcoming year, and found no interaction effect. 27