Where are we to place responsibility for the conduct of our government? When we go to the polls, who can we hold accountable for the successes and failures of national policies? The president? The House? The Senate? The unelected Supreme Court? Or, given our federal system, the states, where governments are, in their complexity, a microcosm of the national government?
Even for those who spend their lives studying politics, these can be extremely difficult questions to answer.
Robert Dahl (Reference Dahl2002, 115)Voters are not merely ignorant; their beliefs about policy-relevant subjects are often systematically biased. Voters systematically overestimate the federal budget share of foreign aid and welfare, and underestimate the share of Social Security and health (Kaiser Family Foundation and Harvard University 1995).
Less-informed voters favor systematically different policies than otherwise identical more-informed voters (Althaus Reference Althaus2003). The public's beliefs about economics, the causes of cancer, and toxicology systematically diverge from the beliefs of experts (Caplan Reference Caplan2007; Kraus, Malmfors, and Slovic Reference Kraus, Malmfors and Slovic1992; Lichter and Rothman Reference Lichter and Rothman1999). Taken together, the evidence raises a troubling question: If politicians cater to the policy preferences of the median voter, will inefficient and counter-productive policies win by popular demand?
The strongest potential answer is that citizens vote for results, not policies. The retrospective voting literature argues that politicians win popularity by delivering prosperity, peace, safe streets, and educated students—not by pandering to the public's beliefs about the best means to achieve these ends (Fiorina Reference Fiorina1981; Gasper and Reeves Reference Gasper and Reeves2011; Lewis-Beck and Stegmaier Reference Lewis-Beck and Stegmaier2000). One simple heuristic—reward success, punish failure—seems to allow voters who have little, zero, or even negative knowledge about policy to extract socially beneficial behavior from their leaders.
Unfortunately for democracy, this heuristic is not as foolproof as it seems. To reward success and punish failure, voters need to know which government actors—if any—influence the various outcomes that voters care about and how great that influence is (Arceneaux Reference Arceneaux2006; Cutler Reference Cutler2004, Reference Cutler2008; Delli Carpini and Keeter Reference Delli Carpini and Keeter1996; Lewis-Beck Reference Lewis-Beck1997; Leyden and Borrelli Reference Leyden and Borrelli1995; Rudolph and Grant Reference Rudolph and Grant2002; Somin, Reference Sominforthcoming, ch. 4). As Achen and Bartels (Reference Achen and Bartels2004, 6) put it:
If jobs have been lost in a recession, something is wrong, but is that the president's fault? If it is not, then voting on the basis of economic results may be no more rational than killing the pharaoh when the Nile does not flood…
Well-functioning democracy does not require “whodunit” knowledge to be universal. If well-informed voters know the right people to reward and punish, and the rest of the electorate votes randomly, politicians still have clear incentives to deliver good results. The greatest danger to democracy comes from systematically biased beliefs about political influence (Caplan Reference Caplan2007; Kahneman Reference Kahneman2011). Just as the market for automobile repair works poorly if the average customer blames his grocer for engine trouble, local elections work poorly if the average voter blames the president for the quality of public schools.
One of the main ways that scholars have tested for the presence of systematic bias is to see whether average beliefs of laypeople and experts diverge (Caplan Reference Caplan2007; Kraus, Malmfors, and Slovic Reference Kraus, Malmfors and Slovic1992; Lichter and Rothman Reference Lichter and Rothman1999). We extend this approach to questions of political influence.Footnote 1 To test the American public's beliefs about political influence for systematic bias, we designed and administered a survey to two distinct groups: (1) a nationally representative sample of Americans, and (2) members of the American Political Science Association (APSA) who specialize in American politics.
Systematically biased attributional beliefs turn out to be common: 14 out of 16 survey questions exhibit statistically significant biases. Compared to experts on American politics, the public greatly overestimates the influence of state and local governments on the economy, the president, and Congress on the quality of public education, the Federal Reserve on the budget, Congress on the course of the Iraq war, and the Supreme Court on crime rates. The public also moderately underestimates the influence of the Federal Reserve on the economy, state and local governments on public education, and the president and Congress on the budget. Perhaps noncognitive factors explain observed belief gaps. But controlling for demographics and various measures of self-serving and ideological bias does little to alter our results.
Our original contribution is twofold. First, to the best of our knowledge, no other study uses a large, representative lay-expert comparison to test whether voters have systematically biased beliefs about political influence.Footnote 2 Second, our full array of outcomes (macroeconomic performance, budget, education, crime, and the war in Iraq) and actors (president and Congress, Supreme Court, Federal Reserve, and state and local government) is the largest and most comprehensive to date, with information about both vertical and horizontal clarity of responsibility (Anderson Reference Anderson2006; Arceneaux Reference Arceneaux2006; 1995; Cutler Reference Cutler2008).
Our results do not imply that the American public's beliefs about political influence are biased in every conceivable respect. Voters' attributional judgments sometimes respond in rational ways to divided government (Lewis-Beck Reference Lewis-Beck1997; Leyden and Borrelli Reference Leyden and Borrelli1995; Powell and Whitten Reference Powell and Whitten1993; Rudolph Reference Rudolph2003a) and federalism (Anderson Reference Anderson2006; Arceneaux Reference Arceneaux2006; Cutler Reference Cutler2004). Nevertheless, the American public's beliefs about political influence are biased in some important respects, raising serious questions about the ability of retrospective voting to counter other cognitive shortcomings in the democratic process.
DATA
We administered our Perceptions of Political Influence on Policy Outcomes Survey in two phases—one for laypeople, the other for experts. In phase one, conducted February 13–18, 2008, Zogby International included our questions on an omnibus telephone survey of adults nationwide. The targets were randomly drawn from telephone compact discs of nationally listed samples, with selection probabilities proportional to population size within area codes and exchanges. Zogby achieved a typical response rate of 14.6%, collecting a total of 1,215 responses.
On March 17, 2008, we began phase two of our survey. We mailed our political influence questions—plus Zogby's demographic and control questions—to a subset of members of APSA.Footnote 3 All APSA members with US addresses who specialize in American politics were included in our sample. To qualify as “specialists in American politics,” APSA members had to list at least one of the following fields of interest: federalism/intergovernmental relations, law and courts, legislative studies, public policy, representation/electoral systems, presidency research, or state politics and policy. This approach yielded 2,894 names, approximately 90% of which had US mailing addresses. We continued to accept responses until July 29, 2008. By that date we had 673 responses from APSA members, with a response rate of 26%.
Table 1 lists the public's and political scientists' mean responses to our main questions. Lower numbers indicate more perceived influence. Table 2 lists both groups' mean responses to Zogby's demographic and control questions. As expected, political scientists are markedly more educated, affluent, male, Democratic, and liberal than the general public.
Perceptions of Political Influence: Summary Statistics
1 = “very influential” 2 = “somewhat influential” 3 = “not very influential” 4 = “not at all influential”
Demographic/Control Variables: Summary Statistics
BENCHMARK RESULTS
In standard rational choice models of belief formation, additional information reduces the variance of beliefs without changing their mean. One implication is that the public and experts will have the same average beliefs. As long as experts are correct on average, we can test the public's political influence beliefs for systematic bias simply by checking whether American politics specialists in APSA systematically disagree (Caplan Reference Caplan2007). In principle, admittedly, belief gaps could indicate bias in either group—or both. Before we consider the main challenges to political scientists' credibility, though, we estimate some benchmark results. We use ordered logits to measure the lay-expert belief gap for all of the beliefs in table 1. Table 3 displays the estimated coefficients and z-stats when the political scientist dummy is our sole independent variable.Footnote 4
Benchmark Results – Ordered Logits on PoliSci
* Significant at the 5% level;
** significant at the 1% level;
*** significant at the .1% level
The initial case for systematic bias is strong. Differences between political scientists and the general public are statistically significant in 15 out of 16 questions; the one exception is the president's influence on the war in Iraq. The absolute value of the z-stat exceeds 4 in 14 out of 16 questions. The average absolute value of the lay-expert gap is .36 on our 4-point scale. This substantial difference—much larger than the gap between laypeople with median education levels and those with graduate degrees—is highly unlikely to result from chance or mere noise in the data.
The most obvious difference between political scientists and the public is that the public thinks that politicians have more influence over outcomes. Eleven out of the 15 statistically significant belief gaps are positive, indicating that political scientists ascribe less influence to politicians than the public does. For example, the public thinks that all of the actors mentioned in our survey—the president, Congress, the Supreme Court, and state and local governments—have more influence over crime rates than political scientists will admit.
Still, the pattern is more complex than “political scientists see more randomness in politics than the public” or “the public scapegoats leaders for outcomes beyond their control.” For three of our five outcome variables, experts single out political actors with influence that the average layperson overlooks. On the economy, political scientists single out the Federal Reserve Board. On public schools, political scientists single out state and local governments. On the budget, political scientists single out Congress and the president. If the consensus of political scientists is correct, the public does not merely blame leaders too much. It also shows some crucial actors undue leniency.
EXPERT BIAS?
Large, systematic disagreements between the general public and political experts provide prima facie evidence of systematic public bias. But the prima facie case can be rebutted. Political scientists differ sharply from the broader public on several noncognitive dimensions. They are disproportionately affluent white men, and much more liberal and Democratic than the general public.
Fortunately, our data set is rich enough to test both of these doubts about the experts' credibility. Suppose political scientists' distinctive views stem entirely from self-serving bias (Dahl and Ransom Reference Dahl and Ransom1999). Controlling for income, sex, race, and other measures of self-interest should then drive the coefficients on the political science dummy variable to zero. Similarly, if political scientists' distinctive views stem entirely from their politics, then the estimated effect of training in political science should vanish after controlling for party identification and ideology.
Self-Serving Bias
We re-estimate all of the ordered logits in table 3 with controls for race, gender, age, age squared, income, job security, and expected income growth. Table 4 shows (a) the revised coefficients on the political scientist dummy, (b) the revised z-stats, and (c) the expected beliefs of the laypeople and experts after setting all of the control variables equal to their median values for lay respondents.
Controlling for Self-Serving Bias
Ordered Logits on Race Dummies, Male, Age, Age2, Income, Job Security, Expected Income Growth, and PoliSci (Comparisons set variables other than PoliSci equal to medians for general public).
* Significant at the 5% level;
** significant at the 1% level;
*** significant at the .1% level
The results offer virtually no support to the self-serving bias hypothesis. After adding all of these controls, the political scientist variable becomes statistically significant in all 16 equations. The z-stat exceeds 4 in all but three cases. The average magnitude of the predicted belief gaps is .35, compared to .36 in the raw data. While political scientists are indeed economically and demographically unusual, these potentially self-serving differences have no apparent effect on their attributional beliefs.
Ideological Bias
Political scientists are decidedly more Democratic and liberal than the broader population. Earlier research suggests that these political variables sway political scientists' beliefs in two ways. First, because our survey was run during the final troubled year of the Bush administration, with both houses of Congress under Democratic control, the evidence on partisan bias suggests that political scientists would exaggerate the influence of the president relative to other branches of government. (Bartels Reference Bartels2002; Marsh and Tilley Reference Marsh and Tilley2009; Rudolph Reference Rudolph2006, Reference Rudolph2003a, Reference Rudolph2003b) Second, as Rudolph (Reference Rudolph2003b, 701–2) predicts and broadly confirms, liberals tend to give government actors more credit and blame for economic outcomes. Liberals' belief in governments' centrality arguably generalizes to noneconomic outcome variables as well. Conservatives might hold, for example, that good schools and safe streets depend primarily on family values rather than government policy.
To test for ideological bias, we re-estimate all of the ordered logits in table 3 with controls for party and ideology. Table 5 shows (a) the revised coefficients on the political scientist dummy, (b) the revised z-stats, and (c) the expected beliefs of the public and experts after setting all of the control variables equal to their median values (party = independent, ideology = moderate) for the lay respondents.
Controlling for Ideological Bias
Ordered Logits on Party, Ideology, and PoliSci (Comparisons set variables other than PoliSci equal to medians for general public).
* Significant at the 5% level;
** significant at the 1% level;
*** significant at the .1% level
At best, the data provide sporadic support for the ideological bias hypothesis. Conservatives think the Supreme Court has more influence over crime rates and assign marginally more budgetary influence to Congress and the president. For the Federal Reserve Board's influence on the budget, and the president's influence on the Iraq war, party and ideology actually push in opposite directions. But none of these effects are large. After controlling for ideological bias, the coefficient on the political scientist dummy remains statistically significant in 14 out of 16 equations. The z-stat exceeds 4 in all but three cases. The average magnitude of the lay-expert belief gap does not budge from its benchmark level of .36.
A final point of interest: Do political scientists' distinctive views reflect their high level of education, their training in politics, or some mixture of the two? In other words, to what extent do laypeople with graduate educations “think like political scientists”? To answer this question, we re-estimate all of the ordered logits in table 3 with controls for self-serving bias, ideological bias, and educational attainment. Table 6 shows the results.
Controlling for Self-Serving Bias, Ideological Bias, and Education
Note: Ordered Logits on Race Dummies, Male, Age, Age2, Income, Job Security, Expected Income Growth, Party, Ideology, Education, and PoliSci (Comparisons set variables other than Education and PoliSci equal to medians for general public).
* Significant at the 5% level;
** significant at the 1% level;
*** significant at the .1% level
Training in political science has a much larger effect than educational attainment. Even after controlling for education, the coefficient on the political scientist dummy remains statistically significant in 14 out of 16 equations. The belief gap between political scientists and the public reflects roughly 90% training in political science, and only 10% education per se. The difference between political scientists and the public is some eight to ten times greater than that between people with median education levels and those who have graduate degrees.
DISCUSSION: THE EFFECTS OF BIAS
Theory
Retrospective voting is the last, best safety net for democratic efficiency. As long as ignorant and irrational voters know enough to properly reward success and punish failure, democracy can still work well. Unfortunately, retrospective voting requires a largely undefended assumption: Voters' beliefs about political influence are unbiased. How precisely do systematically biased beliefs about political influence impede democratic performance? Consider these three basic cases:
Case 1: Underestimating influence. Retrospective voters who underestimate political actors' influence will be too willing to vote against incumbents when conditions are good and too reluctant to vote against incumbents when conditions are bad. This, in turn, weakens politicians' incentives to excel and encourages political shirking (Albouy Reference Albouy2011). If voters falsely attribute the fruit of your efforts to luck, why struggle to deliver the goods? If voters falsely attribute your errors and misdeeds to outside failures, why bother with caution and probity?
Case 2: Overestimating influence. The dangers of overestimating politicians' influence on outcomes are less obvious, but no less real. Retrospective voters who overestimate political actors' influence over outcomes will be too eager to vote against incumbents when conditions are bad and too willing to vote for incumbents when conditions are good.
It is tempting to object that, “The stronger politicians' incentives are, the better.” But in a noisy world, incentives can easily be too strong (Gibbons Reference Gibbons2005). Suppose voters overestimate the effect of the president on the quality of public schools and vote accordingly: If the public schools don't measure up, voters fire the incumbent in the next election. This amplifies presidents' incentive to improve public schools. But if the president has little influence in this area, voters will frequently fire executives who did well overall given their constraints. The greatest drawback of overestimation of political influence, though, may simply be needless disruption every time the polity replaces one scapegoat with another.
Even if voters and experts agree that the president has more influence over the Iraq war than over education, the size of the gap matters. Suppose the Iraq war goes fairly well, but schools perform very badly. The more severely swing voters overestimate the president's influence on education relative to war, the more likely they are to reason, “Yes, the president has slightly more influence on the war than education. But the abject failure on education outweighs the modest success with the war”—and vote against reelection.
Overestimation is particularly dangerous when there is a cap on the penalty for failure. In most democracies, an incumbent's worst-case scenario is merely losing office. As a result, an incumbent with slightly subpar performance has a clear incentive to take big risks to make the cut: Heads he wins, tails he suffers the same fate he would have met if he played it safe. In the extreme case, politicians fearing electoral defeat might instigate “diversionary” wars or other crises (Smith Reference Smith1996). If the war or crisis results in success, the imperilled leader might stave off electoral defeat. If it ends in failure, the leader is not much worse off than before, because the incumbent was likely to lose power anyway.
Case 3: Misallocating influence. The effects of systematically biased beliefs about political influence become more complex if voters misallocate influence—that is, reward and punish one branch of government for the successes and failures of another. In this situation, standard models of team production (Dixit Reference Dixit2002) suggest that retrospective voting will perversely encourage bad performance.
Suppose voters underestimate the president's influence on the Iraq war, and overestimate Congress's influence on the same outcome.Footnote 5 The president might actually have an electoral motive to prolong the war. Even if the president and Congress belong to the same party, the president might deliberately underperform to enhance his bargaining position: If you don't cooperate with me across the board, you're more likely to lose your job than I am.
With divided government and party loyalty, the danger is even greater. A Republican president could improve his party's chances of regaining Congress in the next election simply by dragging out the war, safe in the knowledge that Congress will shoulder most of the blame. The precise effects of blame-shifting are model-specific. For example, it could be Congress trying to shift blame to the president, rather than vice versa. But extremely dysfunctional equilibria are plainly possible.
Empirics
Our data suggest that all three cases are empirically relevant. But Case 2—overestimation—predominates. In our data, voters exaggerate politicians' influence, so retrospective voters typically overreward politicians for success and overpunish them for failure. This does not mean that reelection rates are too low. The implication, rather, is that reelection rates are too high when outcomes are good and too low when outcomes are bad.
Still, there are important exceptions: Voters underestimate the influence of the Federal Reserve on the economy, of state and local government on the quality of public schools, and of both the president and Congress on the budget. In these areas, we should expect retrospective voters to underreward success and underpunish failure. If American politics specialists know what they are talking about, these are areas where voters should accept fewer excuses and demand more results.
Finally, there are at least three outcomes—the economy, public schools, and the budget—where voters seem to misallocate influence—to overestimate the role of some actors, while underestimating the role of others. On the economy, the public overestimates the role of the president, Congress, and especially state and local governments, while underestimating the role of the Federal Reserve. The case of the economy is especially important given its pivotal role as a major issue—often the major issue—in most modern elections. State and local government may often be scapegoats for the central bank's mistakes (Hansen Reference Hansen1999).
For public schools, similarly, the public overestimates the influence of Congress and the president while underestimating the role of state and local government. The expected result is that state and local governments will shift blame for schools' shortcomings onto the federal government. Finally, on the budget, our data indicate that voters sharply overestimate the role of the Federal Reserve and underestimate the influence of Congress and the president. When retrospective voters are dissatisfied with the budget, an unelected body apparently siphons off some of the blame from the politicians who actually control the outcome.
In a few situations overestimation of influence in one area might alleviate the impact of underestimation in another, and vice versa. But such fortuitously beneficial ignorance is likely rare (Somin, forthcoming, ch. 2). In most cases, systematic public misunderstanding of policy influence tends to undermine the overall effectiveness of retrospective voting.
CONCLUSION
The prima facie evidence of voter bias is strong. Political scientists and the public systematically disagree on 15 out of 16 questions. Their belief gaps are usually large in magnitude and highly statistically significant. After adding controls for self-serving bias, ideological bias, and education to the list of control variables, more than 90% of the raw belief gap between political scientists and the public remains.
These findings shed light on two broader topics. First, they undermine the view that systematically biased beliefs about policy can be safely ignored. Retrospective voting may partially mitigate the effect of popular misconceptions about economics, toxicology, and other subjects. But retrospective voting is a flawed filter. Second, our findings show that retrospective voting actually adds new contaminants to the democratic process. Systematically biased beliefs about political influence make some politicians' incentives overly weak and others' excessively strong.
The obvious direction for future research is to explore the robustness of our findings using other samples and other benchmarks of voter competence. But perhaps more importantly, our findings highlight the need for political models with realistic assumptions about human cognition (e.g., Kuran and Sunstein Reference Kuran and Sunstein1999). If presidents know that voters will at least partially blame Congress for their errors, how does this change presidential behavior? If members of Congress expect to be the president's scapegoats, how will they respond? Can both branches profit by creating an unelected agency to deflect the blame for bad outcomes? The best response to unrealistic formal models is not to abandon models but to rebuild them on empirically sound assumptions.
ACKNOWLEDGMENTS
We thank Zogby International for including our questions on their survey, Martha Grabowski and Matt Loveland of Le Moyne College's Contemporary Catholic Trends for assistance, George Mason University's Department of Economics for financial support, Kevin Arceneaux, Michael Delli Carpini, Tyler Cowen, Robin Hanson, Andrew Healy, Craig Heinicke, Garett Jones, Scott Keeter, Suzanne Mettler, John Nye, Alex Tabarrok, Midwest Political Science Association panel discussants Erin Cassese and David Doherty, and seminar participants at George Mason University for useful comments, and Zachary Gochenour and Diana Weinert for excellent research assistance.
