Prevalence and Moderators of the Candidate Name-Order Effect: Evidence from All Statewide General Elections in California

Transcription

1 Prevalence and Moderators of the Candidate Name-Order Effect: Evidence from All Statewide General Elections in California Josh Pasek, University of Michigan, 5413 North Quad, 105. S. State St., Ann Arbor, MI 48103, , (corresponding author) Daniel Schneider, Eberswalder Str 35, Berlin, Germany, , Jon A. Krosnick, Stanford University, 450 Serra Mall, Room 432, Stanford, CA 94305, , Alexander Tahk, University of Wisconsin, Madison, 1050 Bascom Mall, 305 North Hall, Madison, WI 53706, , Eyal Ophir, Rockmelt, Inc., 800 California St. Suite 300, Mountain View, CA 94041, , Claire Milligan, Claire Milligan Design, 870 Market Street, Suite 755, San Francisco, CA 94102, , Running Header: Candidate Name Order Word Count: 5436 Words; 2 Figures; 4 Tables 1

2 Author Note JOSH PASEK is an assistant professor of communication studies and faculty associate at the Center for Political Studies in the Institute for Social Research at the University of Michigan, Ann Arbor, MI, USA. DANIEL SCHNEIDER is an independent researcher, Berlin, Germany. JON A. KROSNICK is University Fellow at Resources for the Future and Frederick O. Glover Professor of Humanities and Social Sciences and professor of communication, political science and psychology at Stanford University, Stanford, CA, USA. ALEX TAHK is assistant professor of political science at the University of Wisconsin Madison, Madison, WI, USA. EYAL OPHIR is an independent researcher with Rockmelt Inc., Mountain View, CA, USA. CLAIRE MILLIGAN is an independent researcher with Claire Milligan Design, San Francisco, CA. The authors thank the members of the Political Psychology Research Group at Stanford University, particularly Dan Blocksom, Ellen Cerf, Eric Chen, Alice Cheung, Lance Cidre, Janelle Cornwall, Felicia Cote, Allison Dencker, Caroline Galindo, Lucinda Gibbs, Joshua Harder, Julie Lein, Rachel Linn, Leslie Mercado, Lilly Wolfson, La Rena Woods, and Steve Zhang, for their work in obtaining the data and building the datasets analyzed here. Please direct all correspondence to Josh Pasek, Assistant Professor, Department of Communication Studies, University of Michigan, 105 South State Street, 5413 North Quad, Ann Arbor, MI 48109, or Jon Krosnick, 432 McClatchy Hall, Stanford University, Stanford, CA 94305, 2

3 Prevalence and Moderators of the Candidate Name-Order Effect: Evidence from All Statewide General Elections in California Abstract Although some past studies suggest that candidates may receive more votes when their names are listed first among their competitors than when listed later, two recent studies challenged this conclusion with regard to major party candidates running in statewide races and raised questions about the impact of analytic methods on the conclusions of name order research. This study tests a series of hypotheses about the conditions under which name order effects are most likely using the largest dataset to date: a set of quasi-randomized natural experiments involving 402 candidates running in 76 statewide California elections. Regardless of the analytic method used, a small primacy effect appeared consistently that could have a substantive impact on some contests. This effect was larger in races for lower visibility offices, in years with higher turnout, and in races that were not close. All this is consistent with the claim that name order effects occur among voters who have little or no information about the candidates or among voters who feel ambivalence about the candidates. 3

4 Prevalence and Moderators of the Candidate Name-Order Effect: Evidence from All Statewide General Elections in California Elections are moments for citizens of a democracy to send messages to their elected officials about what they would like government to do in the coming years. For this process of communication to work effectively, it is important that the exchange of information be distorted as little as possible by the ways in which votes are collected. But for more than a century, political scientists have worried that the specifics of voting methodology might alter outcomes (Allen 1906; Beard 1909; Herrnson et al. 2008). In Florida in 2000, studies of the so-called butterfly ballot suggest that its layout of candidate names and punch card holes might have caused some people to vote for a candidate different than the one they intended to support. (Brady, Herron, Mebane, Sekhon, Shotts, and Wand 2001; Wand, Shotts, Sekhon, Mebane, Herron, and Brady 2001). Other investigations have explored the impact of ballot design on unrecorded votes (Kimball and Kropf 2005), and still other work has examined the possibility that choices of voting equipment might have differentially affected electoral participation by various racial groups (Tomz and van Houweling 2003). In this paper, we examine another ballot design factor that might influence election outcomes: the order of candidate names on the ballot. Research on this topic dates back 60 years. Most studies reported evidence of primacy effects, whereby candidates received more votes when listed first than when listed later (for a review, see Miller and Krosnick 1998; see also Brockington 2003; Krosnick, Miller, and Tichy 2004; Meredith and Salant 2013). However, two recent papers found few primacy effects in voting for major party candidates in California general elections and suggested that the use of suboptimal analytic methods led past studies to 4

5 yield illusory evidence (e.g., Alvarez, Sinclair, and Hasen 2006; Ho and Imai 2008). This paper reports a new, comprehensive analysis of name order effects in California elections to assess the source of these discrepant results. We begin below by providing a psychological rationale for expecting name order effects and deriving a series of hypotheses about when name order effects might be largest. Then we review the analytic methods used in past studies and propose reasons why some might have led to misleading conclusions. Leveraging natural quasi-experiments in California, we assessed how outcomes of every California election that rotated name order varied depending on name order. Furthermore, we explored how assumptions made in conducting analyses, analytic techniques, and control variables employed can impact research conclusions. We conclude by spelling out some of the implications of our findings for theories of voting and for the conduct of elections. A Rationale for Name Order Effects Voters are often asked to choose among candidates when doing so on substantive grounds is understandably difficult. Some races receive little or no coverage from the news media, so learning about the candidates poses a substantial challenge for even the most dedicated voters (Schaffner, Streb, and Wright 2001; cf. Squire and Smith 1988). And in highly-publicized campaigns for very visible offices, competing claims by major party candidates could paralyze some well-informed individuals with feelings of deep ambivalence (Keele and Wolak 2008). Thus, voters may sometimes find themselves in voting booths lacking a preference between candidates in a race, due either to insufficient information or to ambivalence. 1 In such situations, voters might lean toward selecting the first listed candidate for a 1 Although voters should perhaps abstain in such situations (Feddersen and Pesendorfer, 1996), many people may not be aware that abstention is permitted or may worry that their votes will not be counted if they do so. Indeed, answers to frequently asked questions issued by some state elections officials assure voters that they may abstain without invalidating their entire ballots, thereby suggesting that citizens are sometimes uncertain about this possibility (see e.g., Minnesota Office of the Secretary of State, n.d.; Oklahoma State Election Board, n.d.). 5

6 number of reasons: (1) in some states, the first-listed candidate is the incumbent (see Dencker, Ackner, Cerf, Tichy, and Krosnick 2010), so voters who know this might use this cue in order to avoid changing horses, (2) voters might rely on the up is good or first is best heuristic (see e.g., Meier and Robinson 2004) 2, (3) voters who have some information about the candidates but have not made a choice between them might think about them with a confirmatory bias (Klayman and Ha 1987) coupled by a fatigue-induced premature stopping rule (a form of satisficing), meaning that initial thinking will yield mostly reasons to vote for first-listed candidates (Krosnick 1991; Simon 1955), or (4) citizens may attempt to vote completely randomly in some races if they mistakenly think that they need to vote in all races on a ballot in order for any of their votes to count, but when attempting to choose randomly, people unintentionally lean toward first-listed options (cf., Ayton, Hunt, and Wright 1989; Yeager and Bardo 1983). All of this logic anticipates primacy effects in voting, which would parallel primacy effects in many other arenas of study, including responsiveness to persuasive arguments (e.g., Haugtvedt and Wegener 1994; Lund 1925), the rendering of personality judgments (e.g., Anderson 1965), recollections of events (e.g., Glanzer and Cunitz 1966), and responses to closed-ended survey questions (e.g., Holbrook, Krosnick, Moore, and Torangeau 2007; Krosnick and Alwin 1987; for a review, see Krosnick and Presser 2010). Contest-Level Moderators of Name Order Effects 3 High versus low visibility offices. Voters seem likely to learn more from the media about candidates running for high visibility offices, such as President, Governor, and Senator, than about candidates running for lower visibility offices, such as Insurance Commissioner or 2 Though there is some controversy around the cause of apparent embodiment effects (see e.g. Casasanto 2009; Crawford 2009). 3 We do not discuss various potential moderators of name order effects that we do not investigate here, including the extent of undervoting, the number of candidates competing in a race, and voter educational attainment. 6

7 Secretary of State (Kahn 1991), to discuss the former more often (Mondak 1995), and to be contacted more often by candidates running in the form more often (cf., Boyd 1989; Campbell 1960; Smith 2001). Therefore, name order might be more likely to influence the outcomes of races for lower visibility offices. Turnout. People who participate in low-turnout elections are those chronically interested in politics (cf. Boyd 1989; Smith 2001), whereas high-turnout elections attract momentary participation from people who are normally not particularly engaged (Converse 1962). If the latter are most likely to manifest name order effects, due either to lack of information or ambivalence, higher turnout elections would be expected to manifest especially strong name order effects, particularly for low-visibility candidates. Absentee voting. When making electoral choices, voters casting mail-in ballots before election day can seek information about candidates between encountering a name on the ballot and making a choice. For contests involving low visibility offices, this means that absentee voters may be able to inform themselves about the candidates and make substantive choices, thus perhaps decreasing their susceptibility to name order effects. When similarly-prepared voters confront those same contests in a voting booth where last minute learning is not so easy name order effects might be stronger. Hence, in elections with more absentee voting, candidate name order might be less influential. Close contests. When pre-election polls indicate that one candidate is likely to beat the other(s) by a large margin, voters may have little incentive to carefully consider and cast thoughtful votes, because they will make little difference. In contrast, when pre-election polls suggest that a race outcome will be close, voters might think that they have more potential to determine the victor and may therefore be more motivated to think carefully when making their 7

8 candidate choices (cf., Cox and Munger 1989; Larcinese 2007). Furthermore, the news media may be more likely to devote attention to the candidates participating in close races, because of the suspense involved in the contest, thus educating voters about such races more. These processes would be expected to reduce the size of name order effects in close contests. Partisan versus non-partisan contests. When ballots display the party affiliations of the candidates in a race, this cue may allow voters who have no information about the competitors but do have partisan loyalties to cast a vote substantively (Popkin 1991; Schaffner, Streb, and Wright 2001). But when no party affiliations are listed on the ballot, low information voters lack this cue and are perhaps therefore more likely to be influenced by name order. Prior Research on the Effect of Candidate Name Order In line with these expectations, most past studies of candidate name order have uncovered evidence of primacy effects. Miller and Krosnick (1998) found that the vast majority of 24 studies done before 1998 pointed to a first-position advantage, but most were methodologically flawed. More recent studies have overcome the problems apparent in early research by focusing on natural experiments in real elections. Four post-1998 studies explored order effects in non- American election. Three of these studies documented a sizable influence of name order (Faas and Schoen 2006; Geys and Heyndels 2003; King and Leigh 2009), and one found no discernable effects (Orr 2002). 4 Two other studies examining only primaries or special elections found relatively large benefits for candidates listed closer to the top of the ballot (Ho and Imai 2006; Koppel and Steen 2004). Among six studies of general elections in the U.S., four found evidence of primacy effects for all types of candidates (Brockington 2003; Krosnick, Miller, and 4 A variety of studies since 1998 have also explored the influence of ballot order in cases that were not natural experiments (e.g., Curtice and Marsh 2008; Bochel and Denver 2007; Lutz 2010). These studies generally documented primacy effects, but the lack of randomization or rotation presents a major confound limiting informativeness. 8

9 Tichy 2004; Meredith and Salant 2013; Miller and Krosnick 1998). Of the remaining two studies, one found primacy effects for candidates unaffiliated with major political parties (Ho and Imai 2008), but both found little evidence of order effects for major party candidates (Alvarez, Sinclair, and Hasen 2006; Ho and Imai 2008). The two recent studies that did not produce evidence of name order effects for major party candidates were similar in various ways. Alvarez et al. (2006) and Ho and Imai (2008) analyzed data from statewide contests in California, with most data coming from the 1998 elections. Both papers tested for the presence of a name order effect in each race individually. Each study also introduced analytic methods that had not been used previously: Alvarez et al. (2006) used seemingly unrelated regression (cf. Zellner 1962), and Ho and Imai (2006, 2008) implemented a non-parametric approach using permutation tests. The Current Study In this paper, we explored whether research conclusions would be different using a wider array of elections and conducting analyses with more statistical power. In addition, we examined whether research conclusions differ depending on the analytic method used, and we tested the hypotheses offered above about the moderators of name order effects. To do so, we used data from natural quasi-experiments wherein the orderings of 402 statewide candidates names were rotated across California s 80 assembly districts during elections between 1976 and Methods Data We analyzed data on all statewide general elections in California run between 1976 and 2006 in which candidate name order was rotated across assembly districts. Election returns were 9

10 obtained from the California State Archives and the Secretary of State. The dataset we built included complete records for each of 402 candidates who ran in 76 contests in all 80 Assembly Districts (ADs) (Table 1 shows the contests examined and the number of candidates running in each). California s Name Order Rotation Method To determine the ordering of candidates names in statewide races on California ballots, since 1976, the Secretary of State has drawn the letters of the alphabet randomly, one at a time, dubbing the resulting order of letters as a random alphabet constructed for each election year. Using this random alphabet, the candidates were ordered by their last names in the CA 1 st Assembly District (1 st AD), located in the northwest corner of the state. The first name listed for each contest on the ballot in the 1 st AD was shifted to the last position in that contest on the ballot in the 2 nd AD, and all remaining candidate names were shifted up one position. Repeating this procedure, candidate name orders were assigned successively for the 3 rd through 80 th ADs. In some respects, this rotation scheme acts like randomization. Each candidate is equally likely to be listed first, second, or in any other position in any given district. Hence, in aggregate, we can compare a particular candidate s performance in districts where he/she was listed first to that candidate s performance in districts where he/she was listed later. However, because the state has only 80 ADs, and most contests involved many candidates competing for the same office, name order could have been confounded with district attributes. One obvious potential confound is political party loyalty. Figure 1 shows a histogram of the distribution of the proportion of major party registrants (Democrats and Republicans) who were Democrats in each AD at the time of each election between 1976 and Needless to 5 The distribution in Figure 1 shows the proportion of all AD-candidate combinations in the dataset and hence is weighted by the number of contests in each year analyzed, as well as the number of candidates per contest in those 10

11 say, the ADs differed a great deal from one another in terms of their party leanings. Some districts were overwhelmingly Democratic and appear on the right side of Figure 1. Others were majority Republican and appear on the left side of Figure 1. Although others were more evenly balanced, few were 50-50, in the middle of Figure 1. Such variation would not be a threat to internal validity in a natural experiment in which there were a large number of districts. But because California has only 80, and most statewide contests run since 1976 involved between four and eight competing candidates, relatively few districts were assigned each unique candidate name order; wide variation in partisan leanings across districts might introduce error variance in significance testing that could obscure name order effects. 6 We set out to overcome the limitations created by the small number of ADs in two ways. First, we controlled for the partisan leaning of each AD in our analyses. Specifically, we controlled for the proportion of major party registrants in the district who were registered as Democrats. 7 Second, we aggregated across candidates and races to produce a large number of data points for our summary analysis. That is, rather than looking at name order effects for one candidate at a time, we simultaneously analyzed many candidates and conducted a single, integrated analysis using various statistical techniques. Analytical Strategies To assess whether analytical methods altered the conclusions researchers would reach, we assessed the influence of name order using a number of different methods. Within-contest analytical strategies we examined include: (1) permutation analyses, (2) three types of Ordinary years. 6 When 5 candidates were competing in a race, one candidate s name was listed first in the first district as well as the sixth, eleventh, sixteenth, twenty first, and so on. If those districts happened to include a larger percentage of Republicans voters than were present in other districts, this would lead to a confounding of name order with Republican voting tendencies. 7 Party Registration = Democratic Party registrants / ( Democratic Party registrants + Republican Party registrants). Registration statistics were obtained from the California Secretary of State and were calculated separately for each assembly district in each election year. 11

12 Least Squares (OLS) regression, and (3) Seemingly Unrelated Regression (SUR) to produce estimates of the performance of each candidate in each ballot position and to determine the error for any given candidate estimate of vote share. All within-contest strategies were conducted twice, once without controlling for party registration in the assembly district in the relevant year, and once with this control. We also compared the performance of each candidate when listed first vs. last on the ballot, and we compared the performance of each candidate when listed first vs. all later positions. 8 Across-candidate analytical strategies included (1) permutation analyses, (2) OLS regressions, (3) multilevel modeling of order effects, and (4) Monte Carlo simulations. Using each analytical strategy, we gauged the impact of controlling for the partisan leaning of each district (see Online Appendix A for a description of the analytic methods). 9 We also evaluated the influence of moderators of name order effects by estimating the parameters of a series of regression equations. Because some of the proposed moderators were correlated with one another, we tested each moderator individually before combining testing all the moderator hypotheses simultaneously. All moderator analyses were conducted as multilevel models with candidate effects clustered at the contest level. Results Overall Name Order Effects No controls. When not controlling for party registration, 82% of candidates performed better when listed first than when listed later (see Table 2, row 1, column 7), and 78% of 8 Both metrics provided equivalent results for two-candidate contests. 9 As in much past research (cf. Miller and Krosnick, 1998; Ho and Imai, 2008), candidate was the unit of analysis. The outcomes of multiple candidates in the same race are not independent of each other, such that one candidate doing better in the first position necessarily causes other candidates to do less well in later positions. Seemingly unrelated regression is explicitly designed to account for this non-independence and yielded the same results as other methods did. Multilevel model tests of moderators of the name order effect accounted for this possibility by clustering the results for candidates in the same race. 12

13 candidates performed better when listed first than when listed last (Table 2, row 3, column 7). On average, across all contests, candidates received nearly half a percentage point of additional votes compared to when they were listed either in the average of all later positions (.41 percentage points; Table 2, row 1, column 1) or last (.45 percentage points; Table 2, row 3, column 1). Effect sizes varied widely across contests, with individual candidates performing as much as 10 points better or 7 points worse when listed first as compared to later (Table 2, rows 1-4, columns 5 and 6). 10 Figure 2a shows the distribution of candidate name order effects for all candidates examined, comparing first position with the average of all later positions. The differences between the first and later positions in terms of candidate performance were statistically significant in all analyses; the same conclusion was supported by results yielded by all estimation methods (p s<.0001 for sign tests, bootstrap analyses, and t tests; see Online Appendix B for all tests). Differences in means and linear regressions also yielded significant name order effects when using Monte Carlo simulations, and SURs produced marginally significant order effects when statistical significance was assessed this way (Table 2, column 9, rows 1-4). A first position advantage, therefore, appeared to be endemic across California contests; 70.4% of the 402 candidates received their largest vote share when listed first (p<.001 sign test). Average performance for candidates in each position is shown in Online Appendix C. Controlling for party registration. Controlling for party registration did not alter the average apparent influence of candidate name order, but doing so did drastically reduce the variance of first position advantages. Candidates received about half a percentage point more votes when listed first than when listed either later or last (Table 2, column 1, rows 5-8). 10 The wide variations can be attributed both to the small number of assembly districts examined in each contest (n=80) and to the non-random distribution of voters predispositions within those districts. 13

14 Candidates whose vote returns suggest that they performed less well when listed first than when listed later appear to have lost many fewer votes than was suggested by the analysis without controls: slightly more than a two-percentage point deficit, as compared to more than seven percentage points before controls. Candidates who apparently benefitted the most by being listed first also appear to have gained fewer percentage points in the analysis including controls: four instead of nine (Table 2, columns 5-6, rows 5-8). Figure 2b illustrates the much smaller variance after controlling for party registration. About 90% of candidates enjoyed an advantage when listed first according to all metrics (Table 2, column 7, rows 5-8). After controlling for party registration, a candidate s performance when listed first was significantly better than that when listed later, according to all of our statistical tests across all of the analytical strategies used (p s < ). Controlling for party registration thus revealed that primacy effects were present in the vast majority of California elections. Major Versus Non-major Party Candidates Major party candidates manifested larger and more variant primacy effects than did nonmajor party candidates. On average, major party candidates gained about eight-tenths of a percentage point when listed first, as compared to any other position (Table 3, column 1, rows 1-4). Non-major party candidates, in contrast, garnered about a quarter of a percentage point of additional vote share when listed first as opposed to either last or later (Table 3, column 1, rows 5-8). This distinction is made particularly stark in the histograms in Figures 2a and 2b, where major party candidates are shown in black, and non-major party candidates are shown in white. Although effect sizes for major and non-major party candidates differed in our analyses (b=.56, p<.001; Table 4, column 1, row 2), name order effects were clearly present for both groups. The intercept in Table 4, column 1 shows that the projected benefit for first-listed non- 14

15 major party candidates was.27 percentage points (p<.001, row 9). The projected advantage for major party candidates can be seen by adding this difference to the coefficient for major party candidates shown in Table 4, row 1, yielding a first-position benefit of.83 percentage points. All tests for all models tested indicated that both major and non-major party candidates performed significantly better when listed first than when listed in other positions (Table 3). Larger effect sizes for major party candidates made it more likely that name order would indeed influence election results (see Online Appendix D). Studying Candidates Individually Consistent with past studies of California voting, name order effects for candidates tested separately were rarely significantly different from zero. Caterpillar plots showing the name order effect and margin of error estimates for individual candidates illustrate the large errors hampering assessments of name order on a candidate-by-candidate basis (see Figure E1 in Online Appendix E). 11 Without controlling for party registration, confidence intervals for many major party candidates spanned ten or more percentage points (column 1, Figure E1). To detect significant name order effects in such circumstances, the primacy effect would need to be at least 5 percentage points. Moderators of the Primacy Effect High-visibility offices. As expected, voters were less influenced by the order of candidates names when voting in contests for high visibility offices. In races for low visibility offices, candidates averaged.56 percentage points of additional vote share when listed first as compared to when listed later (p<.05; see Table 4, column 2, row 9), compared to only.33 percentage points for high visibility offices (p<.05), significantly smaller than for low visibility 11 Figure E1 in Online Appendix E shows the effects and margin of errors when comparing first position with the mean of all other positions. Because of the smaller sample size when comparing first and last positions, standard errors are even larger for those comparisons. 15

16 offices (difference: b=-.23, p<.001; Table 4, column 2, row 2). Turnout. As expected, among candidates running for low visibility offices, name order effects were greater in years with higher turnout. In 2002 the lowest turnout year examined when only 36% of eligible citizens voted, our model (based on Table 4, column 3) predicts a primacy effect of.38 percentage points for candidates running for low visibility offices. In contrast, in 1984 the highest turnout year examined where 59% of eligible voters cast ballots, the same model anticipates a primacy effect of.99 percentage points in an otherwise identical contest (difference: p<.001). 12 There was no significant difference between expected name order effect sizes for high visibility offices in the highest and lowest turnout years (range:.29 [lowest turnout year] to.35 [highest turnout year], n.s.). Absentee voting. As expected, the advantage of being listed first in races for lowvisibility offices varied depending on the proportion of individuals who cast ballots by mail. In 1978 races for low-visibility offices, when only 4% of California ballots were cast absentee, our model (based on Table 4, column 4) predicted that being listed first garnered.70 percentage points of additional vote share. In contrast, in 2006 races for low visibility offices, when 42% of California ballots were cast absentee, the predicted first position advantage dropped to.38 percentage points (a significant difference at the p<.01 level). 13 Among candidates running for high-visibility offices, the rate of absentee voting was not related to the size of name order effects (p=n.s.) The predicted values were generated using the parameter estimates in Table 4, column 3 (-.59[intercept] +.36[turnout]*2.68[turnout coefficient] =.38[estimated first position benefit]; -.59[intercept] +.59[turnout]*2.68[turnout coefficient] =.99[estimated first position benefit]). 13 The predicted values were generated using the parameter estimates in Table 4, column 4 (.74[intercept] +.04[proportion absentee]*-.86[absentee coefficient] =.70[estimated first position benefit]; 74[intercept] +.42[proportion absentee]*-.86[absentee coefficient] =.38[estimated first position benefit]). 14 It is not possible to separate absentee votes from in-person votes in California election statistics, so we could not explore the presence of name order effects separately among voters who voted absentee vs. those who voted in person on election day. Therefore, we explored the impact of absentee voting by comparing races run in years with 16

17 Margin of victory. As expected, name order effects were smaller when the margin of victory between leading candidates was smaller. In nail-biter contests, such as the 2002 California Controller s race that was decided by.23 percentage points, name order had a small but significant effect (b=.36 percentage points, p<.001; based on Table 4, column 5). Our model predicts considerably larger name order effects in blowouts. In races such as the 1986 Treasurer s contest, where Democrat Jesse Unruh cruised to reelection with 74 percent of the popular vote, the expected first position advantage approaches one percentage point (p<.001 difference). Non-partisan contests. Finally, as expected, name order effects appear to have been larger in races in which candidate party affiliations were not listed on the ballot. When partisan cues were available, candidates gained.47 percentage points of additional vote when listed first as opposed to later (p<.001; Table 4, column 6, row 9). But among the Superintendent contests, for which party affiliations were not listed, the primacy effect averaged 3.58 percentage points, more than 7 times larger (p<.001 difference; Table 4, column 6). Combined model. When controlling for all moderators simultaneously, the moderating effects of being a major party candidate, running in elections with higher turnout, running in close contests, and running in non-partisan contests remained significant (Table 4, column 7). Among non-major party candidates running for high visibility offices in low turnout elections and with the smallest observed margin of victory (the conditions under which we expect the smallest name order effects), candidates were predicted to perform, on average,.22 percentage points better when listed first than when listed later. This advantage was not significantly different from zero. many vs. fewer absentee votes. California voters can ask to vote absentee without giving a reason, and all absentee voting is done via mailed ballots.. 17

18 Discussion Name order effects appear to have been ubiquitous in California statewide elections and were more pronounced when voters were either uninformed or ambivalent. Multiple analytic approaches yielded equivalent evidence of frequent primacy effects. Consistent with past work, the average effect size was small in the statewide races we examined. Nonetheless, as Online Appendix D shows, simulations suggest that if name order had not been rotated in these races, the outcomes of 9% of the races could have been different due to the primacy effect if one particular name order had been used throughout the state instead of rotating. Effects of the magnitude observed here were far larger than the margin of victory in some especially visible recent elections. For example, George W. Bush (listed first) defeated Al Gore (listed second) in Florida in 2000 by less than.009 percentage points to win the presidency. 15 Al Franken (listed third) bested Norm Coleman (listed second) in 2006 by less than.007 percentage points, claiming the Democrats 60 th vote in the Senate and a supermajority. Both elections resulted in considerable changes in American policy based on margins of victory one-tenth the size of the effect of name order observed in the current study for similar races. Had either Gore or Coleman been listed first, life in America might well have been different. Moderators Major party candidates were particularly prone to the influence of name order, perhaps because more voters were ambivalent about these candidates due to the extensive publicity that they received. Name order had more impact on the outcomes of candidates running for less visible offices during high-turnout years, presumably because voters were less informed about 15 Although the butterfly ballot has been implicated as a cause of Bush s victory (Brady et al. 2001; Niemi and Herrnson 2003; Wand et al. 2001), our results suggest that a change in name order would have changed the victor as well. 18

19 such candidates, and average voter information was presumably lower in such elections. 16 In past studies, statistically significant name order effects appeared more often for minor party and independent candidates than for major party candidates (cf., Ho and Imai 2008; King and Leigh 2009; Miller and Krosnick 1998). This appears to be an artifact of greater variance in the proportion of votes cast for major party candidates. Minor party and independent candidates receive small proportions of votes, so the variance of their vote shares is small, allowing for easier detection of significant name order effects. In contrast, major party candidates receive large numbers of votes, increasing the variance of estimates and requiring greater statistical power in attempts to detect effects. Aggregating across many contests, we documented that name order effects were larger for major party candidates than for others. In the present study, name order effects were smaller in contests that were relatively close, perhaps because these races were more publicized and inspired more voter effort to learn about the candidates and choose between them carefully. Name order effects were also more likely to appear in non-partisan contests where especially helpful cues were not on the ballot, a finding that resonates with similar patterns reported by Miller and Krosnick (1998) and Ho and Imai (2008). 17 Each of these findings echoes what we would expect to see if some voters who were relatively uninformed selected the first-listed name. More absentee voting was not associated with weaker name order effects. Our hypothesis in this regard was based on the assumption that voters who cast absentee, mail-in ballots have opportunities to seek out needed information. Therefore, perhaps absentee voters did not acquire more information when choosing candidates than did people who voted on 16 Because lack of information about a race may cause both name order effects and undervoting, we correlated the proportion of undervotes in each contest with the apparent order effect for the candidates in the contest. The correlation was.51, indicating that the two circumstances often co-occurred. Undervoting was not treated as a covariate in the regressions we conducted, because it was viewed as an outcome rather than a predictor. 17 Although heuristic cues such as party identification are often discussed in political science as a type of low information rationality (Popkin 1991), they are nonetheless informational (cf. Lau and Redlawsk 2001). 19

20 election day (cf. Stein 1998). Relations to Past Studies of California Elections The appearance of primacy effects for major party candidates in statewide California elections is at variance with the conclusions reached by Alvarez et al. (2006) and by Ho and Imai (2008). Aggregation across contests was necessary to determine that these effects were reliable; controlling for party registration made this detection task even more effective. Indeed, without controlling for party registration and aggregating the effects across numerous contests, we would have had little hope of detecting anything short of a huge influence of name order. Instead, we found an almost universal benefit (more than 85% of contests using all metrics), but one that could only be identified by aggregating across contests, with or without controls. Beyond aggregation across races, there is an additional likely reason for the discrepancies between our results and those reported by Alvarez et al. (2006) and Ho and Imai (2008). The datasets analyzed by those investigators were prepared by the California Secretary of State s office and distributed by U.C. Data ( Those reports contain errors, as explained in Online Appendix F. Correcting those errors in the dataset constructed for our study may also explain some discrepancies between our findings and those reported by these other investigators. Finally, both Alvarez et al. (2006) and Ho and Imai (2008) chose a particularly problematic set of contests with which to assess name order effects. The years selected for their analyses 1998 and 2000 were based on a single redistricting in California and, in both years, seven candidates competed in the majority of the contests (see Table 1). Assignment of name order to AD in 7-candidate contests happened to confound order with partisan variation across ADs (cf. Ho and Imai 2008). The most Democratic and least Democratic set of ADs in 7-20

21 candidate contests based on the 1990s redistricting averaged a difference in party registration of 9.08 percentage points. This was considerably larger than the 5.74 percentage point average range across all contests during the period of our study. Distortion of conclusions caused by this set of confounds can be avoided by aggregating across contests, using contests with varying numbers of candidates, controlling for party registration, and including many elections over multiple redistricting cycles (all implemented herein). These techniques dilute the potential influence of any single source of systematic error and thereby improve the validity of the estimates. The choice of analytic method used to assess the influence of name order had at most a minor influence on the relations we observed between name ordering and vote shares (see Online Appendix B). Standard errors for the within-contest evaluations tended to be slightly larger for the SURs than for other models. And the Monte Carlo simulations produced estimates with larger variances than did models generalizing from maximum likelihood estimates at the candidate level. Nonetheless, all of the models told the same story, contradicting claims to the contrary. Limitations One limitation of the current study is the fact that it does not provide direct evidence of the psychological processes that caused primacy effects. To determine how and why voter thinking leads to primacy effects will require other approaches, such as experiments, that test the individual-level mechanisms more directly. Almost certainly, the average size of the primacy effect reported here would have been different had we analyzed a different set of races, a different mix of candidate types, or elections in other states or countries. For example, given the explanations we proposed for the 21

22 moderators effects, elections in countries with compulsory voting might manifest larger name order effects, because more individuals with less political information may go to the polls. In fact, the moderating factors identified in this paper lead us to expect findings similar to those identified by Miller and Krosnick (1998) for the contests they examined. In contrast with the current study, Miller and Krosnick (1998) explored a large number of non-partisan contests and contests at the county level. Only two statewide general election contests with partisan candidates were included in their analyses a presidential contest with 8 candidates and a senatorial contest with 3 candidates. In those contests, the first-listed candidate received an additional.44 percentage points of vote share on average, almost exactly what we would predict using the parameter estimates from the current paper s model. Including county-level partisan offices, the average effect size increased to 1.25 percentage points in their study, a result that nicely extends what we find in the move from the higher to lower visibility statewide contests in the current study. Implications for Electoral Procedures Whether our findings merit changes in electoral procedures is a matter of perspective. The electoral impact of name order is generally rather small: half a percentage point on average for partisan statewide general elections and likely no larger than a few percentage points on average in other general elections. But because most states use a single name order for candidates on all ballots, biases due to name order can (and apparently do) influence outcomes. In a political environment where hundreds of millions of dollars are spent to induce marginal shifts among voters, even small biases can be highly substantive. Correctly assessing voter intent appears to be a priority enormous resources are regularly expended to address a handful of annual cases of voter fraud (Lipton and Urbina 2007) and to manually recount ballots when an 22

23 election proves too-close-to-call. In this context, randomizing or rotating candidate names across ballots seems to be a good investment to promote fairness and minimize the distortion of election outcomes. 23

24 References Allen, Philip L Ballot Laws and Their Workings. Political Science Quarterly 21: Alvarez, R. Michael, Betsy Sinclair, and Richard L. Hasen How Much is Enough? The Ballot Order Effect and the Use of Social Science Research in Election Law Disputes. Election Law Journal: Rules, Politics, and Policy 5: Anderson, Norman H Primacy Effects in Personality Impression Formation Using a Generalized Order Effect Paradigm. Journal of Personality and Social Psychology 2:1-9. Ayton, Peter, Anne J. Hunt, and George Wright Psychological Conceptions of Randomness. Behavioral Decision Making 2: Beard, Charles A The Ballot's Burden. Political Science Quarterly 24: Bochel, Hugh, and David Denver. (2007). A Quiet Revolution: STV and the Scottish Council Elections of Scottish Affairs 61:1-17. Boyd, Richard W The Effects of Primaries and Statewide Races on Voter Turnout. Journal of Politics, 51: Brady, Henry E. Michael C. Herron, Walter R. Mebane Jr., Jasjeet S. Sekhon, Kenneth W. Shotts, and Jonathan Wand Law and Data : The Butterfly Ballot Episode. PS: Political Science and Politics 34: Brockington, David A Low Information Theory of Ballot Position Effect. Political Behavior 25:1-27. Campbell, Angus Surge and Decline: A Study of Electoral Change. Public Opinion 24

25 Quarterly 24: Casasanto, Daniel Embodiment of Abstract Concepts: Good and Bad in Right- and Left- Handers. Journal of Experimental Psychology 138: Converse, Phillip. E Information Flow and the Stability of Partisan Attitudes. Public Opinion Quarterly 26: Cox, Gary W. and Michael Munger Closeness, Expenditures, and Turnout in the 1982 U.S. House Elections. American Political Science Review 83: Crawford, L. Elizabeth, Conceptual Metaphors of Affect. Emotion Review 1: Curtice, John, and Michael Marsh How Did They Vote? Voters Use of the STV Ballot Paper in the 2007 Scottish Local Elections. Representation 44: Dencker, Allison, Nitzan Ackner, Ellen Cerf, Michael P. Tichy, and Jon A. Krosnick Procedures for Ordering Candidate Names on General Election Ballots. [Unpublished Report]. Stanford, CA. Faas, Thorsten, and Harald Schoen. (2006). The Importance of Being First: Effects of Candidates List Positions in the 2003 Bavarian State Election. Electoral Studies 25, Feddersen, Timothy J., and Wolfgang Pesendorfer The Swing Voter s Curse. American Economic Review 86: Geys, Benny, and Bruno Heyndels Influence of Cognitive Sophistication on Ballot Layout Effects. Acta Politica 38: Glanzer, Murray, and Anita R. Cunitz Two Storage Mechanisms in Free Recall. Journal of Verbal Learning and Verbal Behavior 5: Haugtvedt, Curtis P., and Duane T. Wegener Message Order Effects in Persuasion: An Attitude Strength Perspective. Journal of Consumer Research 21:

26 Herrnson, Paul S. Richard G. Niemi, Michael J. Hanmer, Benjamin B. Bederson, Frederick C. Conrad, and Michael W. Traugott Voting Technology. Washington, D. C.: Brookings Institution Press. Ho, Daniel E. and Kosuke Imai Randomization Inference With Natural Experiments: An Analysis of Ballot Effects in the 2003 California Recall Election. Journal of the American Statistical Association 101: Ho, Daniel E. and Kosuke Imai Estimating Causal Effects of Ballot Order from a Randomized Natural Experiment: The California Alphabet Lottery, Public Opinion Quarterly 72: Holbrook, Allyson L., Jon A. Krosnick, David Moore, and Roger Torangeau Response Order Effects in Dichotomous Categorical Questions Presented Orally: The Impact of Question and Response Attributes. Public Opinion Quarterly 71: Kahn, Kim F Senate elections in the news: Examining campaign coverage. Legislative Studies Quarterly 16: Keele, Luke, and Jennifer Wolak Contextual Sources of Ambivalence. Political Psychology 29: Kimball, David C. and Martha Kropf Ballot Design and Unrecorded Votes on Paper- Based Ballots. Public Opinion Quarterly 69: King, Amy, and Andrew Leigh Are Ballot Order Effects Heterogeneous? Social Science Quarterly 90: Klayman, Joshua, and Young-Won Ha Confirmation, Disconfirmation, and Information in Hypothesis Testing. Psychological Review 94: Koppel, Jonathan G. S. and Jennifer A. Steen The Effects of Ballot Position on Election 26

27 Outcomes. Journal of Politics 66: Krosnick, Jon A Response Strategies for Coping with the Cognitive Demands of Attitude Measures in Surveys. Applied Cognitive Psychology 5: Krosnick, Jon A. and Duane F. Alwin An Evaluation of a Cognitive Theory of Response- Order Effects in Survey Measurement. Public Opinion Quarterly 51: Krosnick, Jon A., Joanne M. Miller, and Michael P. Tichy "An unrecognized need for ballot reform: Effects of candidate name order. In Rethinking the Vote: The Politics and Prospects of American National Election Reform, eds. Ann N. Crigler, Marion R. Just and Edward J. McCaffery. New York: Oxford University Press, Krosnick, Jon A. and Stanley Presser Questionnaire Design. In Handbook of Survey Research (Second Edition) eds. James D. Wright and Peter V. Marsden. West Yorkshire, England: Emerald Group. Larcinese, Valentino The Instrumental Voter Goes to the News-Agent: Demand for Information, Marginality and the Media. Journal of Theoretical Politics 19: Lau, Richard R. and David P. Redlawsk Advantages and Disadvantages of Cognitive Heuristics in Political Decision Making. American Journal of Political Science 45: Lipton, Eric, and Ian Urbina. In 5-Year Effort, Scant Evidence of Voter Fraud. New York Times (New York). Apr. 12, Lund, Frederick H The Psychology of Belief. IV. The Law of Primacy in Persuasion. Journal of Abnormal and Social Psychology 20: Lutz, Georg First Come, First Served: The Effect of Ballot Position on Electoral Success in Open Ballot PR Elections. Representation 46: