Valence and Party Support: The Measurement of Mood in Issue Competence

Transcription

1 Valence and Party Support: The Measurement of Mood in Issue Competence Jane Green University of Manchester Will Jennings University of Manchester This paper presents a newly compiled aggregate measure of issue competence for the Republicans and Democrats based on 2,450 survey items on issue competence over six decades. Using quarterly time series data between 1980 and 2008 we test three theoretical propositions about the issue valence of U.S. parties: citizens ratings of party competence on valence issues tend to move in common over time, consistent with the existence of a prevailing mood in public opinion, controlling for partisan bias; issue valence is driven by presidential performance when the party is in power, but by other valence indicators when the party is in opposition; and, the effect of valence exerts a substantive influence upon electoral support. The paper therefore reveals that existing performance measures of party choice are useful indicators of competence, but issue valence provides a valuable contribution to explaining party choice. The implications are significant for the analysis of electoral choice in U.S. politics and for our understanding of dynamics of public opinion and party competition. Paper prepared for the Elections, Public Opinion and Parties Annual Conference 9-11 September 2011, the University of Exeter 1

2 Presidential and party reputations on policy have long been watch-words for electoral success. Issue-by-issue policy evaluations are an important component of the narrative of political campaigns, and scholars highlight the effects of these ratings upon party and candidate strategies (Budge and Farlie 1983; Budge 1987; 2001; Druckman 2004; Holian 2004; Jacobs and Shapiro 1994; Petrocik 1996; Petrocik et al 2003; Riker 1993; Simon 2002). Recent scholarship also points to the significance of competence or handling upon vote choice for electorates in Canada (Fournier et al. 2003) and Europe (Clarke et al. 2004; 2009; Green-Pedersen 2007). These studies analyse the electoral effects of voters perceptions of party competence, trust, delivery or handling ratings according to the idea, first put forward by Stokes (1963; 1992), that on issues which do not divide the parties such as economic growth, management of services, reducing crime voters should base their vote choice on which candidate or party will be most likely to deliver, or handle that issue competently. There has been little research however, which undertakes analysis of similar predictors in the United States. Previous studies have highlighted the effects of presidential issue handling upon vote choice (Miller and Wattenberg 1985; Cover 1986) and the association between a party s ratings on its owned issues and its vote share (Petrocik 1996) but U.S, issue voting is more commonly studied via the lens of Downsian (1957) spatial voting, whereby issue voting is specified as agreement with a candidate s policy position rather than with its competence (e.g. Aldrich et al. 1989; Carmines and Stimson 1980; Enelow and Hinich 1984; Kedar 2009; Meguid 2008). Valence is also described as an inherent candidate attribute (Abney et al. forthcoming; Hayes 2005) and a factor used to explain party positions (Groseclose, 2001; Schofield, 2003; Adams et al. 2005; Somer-Topcu 2007). Performance voting, on the 2

3 other hand, is defined according to handling of the economy, in line with Fiorina s (1981) retrospective punishment model, inspiring a vast literature on economic voting (see Lewis-Beck and Stegmaier 2000; Duch and Stevenson 2008, for reviews). There is a significant gap in our understanding of how public ratings of party issue competence vary over time, how voters formulate evaluations on competence, and whether and how party competence ratings exert an electoral cost or benefit in U.S. elections. This paper makes a contribution to the literature on performance and valence voting by presenting a new aggregate measure of party valence ratings over six decades, from 1944 to 2008, and by analysing these data, at a quarterly level, between 1980 and 2008 over 15 U.S. congressional electoral cycles, to reveal the dynamic properties of party valence, the way these dynamic evaluations are shaped, and their impacts upon electoral support. The paper estimates alternative valence models of the generic congressional ballot, to reveal a distinctive contribution of party valence to the understanding of U.S. electoral choice, and reveals that party valence is an important factor in the calculus of U.S. electorates. The implications of our findings are important for scholars of public opinion, party competition and electoral behavior. With respect to public opinion, our analyses reveal that public ratings of parties on policy exhibit common variance, such that party valence can be conceptualised and analysed as aggregate-level mood in public opinion. While mood in evaluations of handling of valence issues is shaped by voters preferences towards a party, it is also found to precede shifts in macro-partisanship, and is shaped by exogenous political events, presidential approval and economic evaluations. With respect to party competition, our findings suggest that scholars may turn to general evaluations of valence, rather than issue-specific evaluations on issues, to understand the competition strategies of political parties. Parties develop general 3

4 reputations and track records for performance, management and competence. Moreover, our new measure of valence makes possible analysis of the effects of valence on party positions, on party issue emphasis and on the determinants of party strategies via the drivers of party valence. Finally, the paper reveals that valence can help explain the aggregate shifts in vote choice in U.S. electoral cycles, both as a mediating variable for other performance determinants of the vote and as a direct explanatory variable on party choice. Performance and Valence Voting Parties and candidates should prime voters to cast their votes on the basis of their most advantageous issues, via weighting of these issues in the electoral calculation (Budge and Farlie 1983; Budge 1987; Budge et al. 2001; Iyengar and Kinder 1987; Jacobs and Shapiro 1994; Enelow and Hinich 1981; Vavreck 2009). Parties have different utilities which vary by party and which vary by policy issue (see Petrocik 1996; Petrocik et al. 1996; Simon 2002; Sigelman and Buell 2004; Damore 2005; Sides 2006; Belangér and Meguid 2008; Bellucci 2006; Green-Pedersen 2007). The received wisdom implies, therefore, that voters judge parties and candidates on different issues in different ways, according to whether those issues and utilities are positive or negative for the party and the voter. The characteristics of an advantageous issue may relate to spatial proximity (Downs 1957; Enelow and Hinich 1984; Kedar 2009; Meguid 2008; Vavreck 2009), or to competence, trust, delivery and commitment (Budge and Farlie 1983; Budge 1987). The latter are particularly salient when valence issues dominate the agenda (Stokes 1963; Green and Hobolt 2008; Clarke et al. 2004; 2009). Another facet of competence relates to ownership (Petrocik 1996; Petrocik et al. 2003), referring to associations of parties with issues, generated via long-term commitments and 4

5 reputations. Petrocik (1996) found a strong association between ratings on a party s owned issues and that party s electoral support. Therefore, the weighting of issues to the vote depends on whether spatial, valence or owned issues dominate the agenda. These theories overlap with work on the effects of the economy and retrospective voting (Beck, 1987; Duch and Stevenson 2008; Erikson 1989; Fiorina 1981; Hibbs 1987; Lewis-Beck and Stegmaier 2000; Powell and Whitten 1993). If a candidate has a positive track record on the economy and the economy is performing well, an incentive exists to prime the issue in campaigns (Vavreck 2009), as other issues offer strategic competitive advantages. Further research has analysed the issue of foreign policy and its effects on presidential approval (Aldrich et al. 1989), the electoral consequences of presidential handling of war (Krosnick and Brannon 1993), the effects of crime levels and economic mismanagement and issue handling upon trust in government (Chanley et al. 2000; Citrin and Green 1986) and presidential issue handling and the vote (Miller and Wattenberg 1985; Cover 1986). Despite these literatures, research on the effects of valence issues and party competence is remarkably sparse. One possible reason for this is the lack of survey data available to researchers that provides regular and continuous observations. The National Election Studies offer little by way of continuous ratings of candidates and parties on performance handling of issues. Since 1956, the NES has asked just 45 questions on issue valence of political parties across 12 different questions, with some of those items surveyed on just one occasion and no question asked on more than 14 occasions. Another reason is the expectation that questions on competence reveal little more than partisan preference, rather than informed and varied ratings of parties on different policy areas. Furthermore, the U.S. system of government means that party performance or competence must be decoupled from the performance of presidents. 5

6 Presidential approval offers a readily available performance assessment of the governing capability in power, alongside economic performance and, where available, various questions on presidential handling of foreign affairs, the economy, and war. We contend that citizen s competence ratings of political parties make up an important dynamic of mass opinion, shaped by presidential performance in power, by events and by other political and economic judgements across the electoral cycle, representing meaningful shifts in public evaluations of parties on their competence across the policy agenda. A Theory of Public Opinion on Valence Prior to September 11, 2001, President George W. Bush had a positive rating on his handling of the issue of foreign affairs of 54% (July 2001, Gallup). In a subsequent poll, in October, that rating had increased to 81% (also Gallup). However, Bush s ratings did not increase on the issue of foreign affairs alone. His ratings also increased on his handling of social security (41% to 54%), on the economy (54% to 72%), on the federal budget (46% to 59%), and even on the environment (50% to 54%). This example serves to illustrate a point: when voters are provided with information about presidential handling on a salient issue or event, that evaluation, positive or negative, will extend to ratings across the policy agenda. Publics lack information about politics and yet still make coherent political judgements, via the use of heuristics and short-cuts (Popkin 1991; Sniderman et al. 1991; Zaller 1992). Research in social psychology reveals the way citizens form their opinions about politics via the use of heuristics (Tversky and Kahneman 1982). Heuristics represent the short-cuts or cues voters use to overcome cognitive burdens of information processing. Zaller and Feldman (1992, p. 609) note, the making of these judgements requires an aggregation of one s feelings across frequently diverse 6

7 concerns. Heuristics may be particularly useful for evaluating performance. According to Campbell et al. (1960; also Converse 1964) the vast majority of voters rely on general judgments of which party will be dependable or capable, and do not form independent evaluations of candidates on individual policies. Performance on one issue is like to provide a heuristic to form competence evaluations on others. As other issues rise in salience, voters are provided with high profile cues to signal the competence of the government. The economy is another classic example. Economic shocks may be relatively short-lived, moving issue handling judgments in the shortterm. Other shocks may introduce considerable persistence and long-term memory into citizens ratings of issue competence and offer voters reliable information about competence on other issues. The effect of shocks on competence evaluations can persist indefinitely or otherwise decay over time, with political judgements often exhibiting at least some long-term persistence (Durr 1993; Box-Steffensmeier and Smith 1996; 1998; Wlezien 2000; Clarke and Lebo 2003). The implications of these expectations are that public evaluations on competence should move in common. If a party, via its president, performs well on a salient policy issue, we would expect performance judgements to improve on other issues. If a large scale event results in damage to trust and to performance ratings, damage should extend across the issue agenda. These expectations are different to the concept of the ownership of issues (Petrocik 1996; Petrocik et al. 2003), but not in conflict with those ideas. A party s reputation may be higher on some issues and lower on others, but there will still be common variation in the relative fluctuations across different levels of public ratings of competence. Public opinion on valence will be related to movements in partisan preference, whether endogenous or indicative of running tally evaluations (Fiorina 1981), and to 7

8 other existing indicators of performance; presidential approval and the economy. Valence will also be shaped by party reputations on issues, and party and candidate rhetoric and commitment. Public opinion on issue competence can be led through strategic action of elites, where issue evolution leads to partisan realignment or change (e.g. Carmines and Stimson 1989). However, variation in public opinion on valence should be responsive to shocks and changing performance judgements, and insofar as publics use performance information to make judgements about parties on competence, these variations in valence should provide additional information about public evaluations of political elites, with electoral consequences. We suggest that parties have incentives to perform competently on policy in office, and to nurture and promote reputations for competence in opposition, because voters evaluate political parties on competence, and deliver a valence-based verdict at the ballot box. Valence: Empirical Expectations In order to operationalise the above theoretical arguments, a number of empirical expectations can be derived. The first expectation is that we should observe common variation in issue competence ratings of U.S. political parties, irrespective of the issue domain in question. This common movement should be manifested in the loading of issues onto a single dimension of citizens evaluations on issue competence, indicative of an underlying factor of public opinion. The time series properties of this latent dimension can provide substantive interpretation of the nature of persistence in public ratings of parties on competence. It is argued that the aggregation of individual-level behaviour can yield long-term persistence in time series if that behaviour is heterogeneous and if autoregressive processes are present (see Granger 1980; Granger and Joyeaux 1980; and Box-Steffensmeier and Smith 1996). 1 Such an inertial element is crucial for the 8

9 formation of partisan identifications (Fiorina 1981, p. 102) and we would similarly expect that autoregressive tendencies of individuals ratings of parties on competence, and their exposure to exogenous political forces, would generate fractional dynamics, where shocks to parties competence ratings cumulate for extended periods, exhibiting long-memoried, but not infinite, persistence. In particular, we expect heterogeneity in the formation of competence evaluations because of individual-level variation in the strength of partisan attachments, education, political commitment and knowledge (e.g. Converse 1964; Zaller 1992), all of which would lead shocks for some individuals to persist for long periods of time, while dissipating for others at a much faster rate. This inertial element of public evaluations of competence is therefore of substantive importance. Furthermore, if issue competence variation represents meaningful information about ratings of parties on competence, variance in issue competence ratings should not be accounted for by fluctuations in partisanship. If this were the case, issue competence ratings would simply be endogenous to partisan preference, and scholars would be correct to be sceptical about these estimates of parties on issues. The first steps of the empirical analysis in this paper therefore examine the dynamic properties of issue competence ratings, or valence, and the time series interrelationships between valence and macro-partisanship (see MacKuen et al. 1989). The second expectation is that valence should be influenced by existing measures tapping a party s performance. Available measures are performance ratings of a president belonging to a party (presidential approval), mediated by periods when a party holds the presidency, economic evaluations, events, and party identification, allowing for party identification to be construed as a dynamic running tally, rather than social identity (Fiorina 1981; MacKuen et al. 1989; Green et al. 2002). We expect economic evaluations to influence issue valence judgements, but economic 9

10 evaluations should be influential irrespective of party incumbency. An upturn in the economy may be as good for the incumbent as it is bad for the opposition, and a downturn would extend a negative effect on the party in office and a positive effect for the challenger. 2 Furthermore, we expect significant political or economic events to have lasting effects upon party valence. Voters recognise policy decisions for which a party is responsible for, and punish a party accordingly (Duch and Stevenson, 2010). Major political events should therefore have significant effects on competence, especially where these denote signals of incompetence, loss of trust, or of strong political leadership. The third and final expectation is that valence will have unique explanatory power in models of generic congressional ballot. If valence is a meaningful construct of shared underlying dynamics in public evaluations of performance, its behavioural consequences should be significant. We therefore expect valence to mediate the effects of presidential approval, of economic evaluations, party identification and events, but to also account for a unique proportion of variance in vote choice over consecutive U.S. congressional election cycles. Data and Measures Scholars analyzing macro-politics in the US have used measures of party identification data (see MacKuen et al 1989; Bartels 2000), leader ratings (Erikson and Wlezien 1994; 1996; Wlezien and Erikson 2004), trust in government (Keele 2007), economic expectations (MacKuen et al. 1992), in addition to other available measures of macro-economic performance. Using similar approaches, and similar aggregate-level analysis, we can provide an exploration of valence on policy issues. The macro-level consequences of the theory of public opinion and valence give rise to the concept of a mood in competence evaluations. The work of Tversky 10

11 and Kahneman (1982) on framing suggests that individuals judgements are often reducible to a single, dominant consideration. If voters ratings of parties move in tandem, those ratings will also be indicative of an underlying valence dimension of public opinion. This can be shown by the degree to which issue competence ratings are reducible to a single dimension of public opinion, or mood (Stimson 1991; 2002; 2004). In the concept of mood, changes in individual attitudes drive trends in macrolevel opinion whether or not these shifts in public opinion are heterogeneous or homogeneous (e.g. Converse 1990, p. 382; Erikson et al. 2002, pp. 5, ; Stimson 2002, p. 268; Enns and Kellstedt 2008, p. 450). If change occurs at random, those variations are cancelled out through sampling and measurement over time. Noise, response instability and individual errors are therefore removed through aggregation (see Page and Shapiro 1992; Erikson et al. 2002). Furthermore, the effect of short-term shocks or contamination between issues can accumulate and persist in macro-level processes, where heterogeneous micro-processes mean that disturbances can persist in the behaviour of some individuals and decay for others (Granger 1980; Box-Steffensmeier and Smith 1996; Wlezien 2000). Even if the issue competence evaluations of a small number of individuals are subject to long-term memory, as shocks cumulate, this produces macro-level persistence. The concept of mood in valence therefore represents a combination of change and persistence in the competence evaluations of parties, expressed at a macro-level of aggregation where persistence indicates the extent to which people do and do not discount the past in their assessments of the political parties. Using data collated from a large range of opinion polls, and data available from election surveys, we gathered 2,512 individual survey items on policy competence, handling, trust or performance of the Republican and/or Democratic 11

12 Parties, spanning a period of six decades. Sufficient data was gathered for the period between 1980 and 2010 to enable the reliable estimation of Valence at quarterly intervals, for use in conjunction with quarterly estimates of other aggregate level variables (e.g. macro-partisanship, presidential approval, economic evaluations), outlined below, facilitating its analysis, over 15 congressional elections. Stimson s (1991; Stimson et al. 1995) dyad ratios algorithm is used to extract a global dimension of competence evaluations. Stimson s dyad ratios algorithm offers a solution to the problem of missing data in circumstances where 85% of all possible values are missing and where the distribution of data is irregular (Stimson 2004, p. 10), ideally suited to issue competence data which has not been collated continuously. The dyad ratios algorithm first calculates valence as an average of the available observations of positive evaluations of a given party s competence in handling an issue (for example, the percentage of respondents rating a party as best to run government efficiently or the percentage of respondents rating a party as doing quite well or very well in handling the economy), at time period t, through recursive estimation of dyad ratios (the dyad ratio is estimated for each pair of cases in which there is an observation of competence evaluation at two points in time), weighted for the degree of common variance of each evaluation with valence itself. 3 In order to ensure our measure is reliable as an indicator of valence, we estimated four preliminary series; one containing all policy issue items irrespective of topic; one containing all policy issues but removing economic ratings, in case valence could be simply a proxy for economic evaluations; one containing issues traditionally associated with strongly partisan positions in U.S. politics (healthcare and social security reform, immigration, morality and ethics, and environmental issues), and one series containing typical valence issues (the economy, foreign affairs, crime, security, 12

13 inflation, governing responsibility and education). Although we recognise that there is a valence component to spatial issues (and most likely a spatial component to valence issues), we eliminate positional issues because we suspect they are most prone to partisan bias, and may be very closely associated in many periods, and because we seek to estimate a measure of issue valence which is distinct and conceptually valid. These indices were then subjected to scrutiny. The analysis employs existing aggregate level performance proxies as covariates; macro-partisanship, presidential approval, economic conditions and evaluations, and key events specific to each party. The dependent variable in the first set of multivariate models is the measure of valence, and in the second set of models is the two-party share of the generic congressional ballot (vote intention). The measure of macro-partisanship used in analysis is the series generated by Stimson and colleagues (Mackuen et al. 1989; Erickson et al. 2002), being equal to the percent Democratic of party identifiers, with the estimated series compiled from more than 4,000 opinion polls between 1952 and The measure of presidential approval was compiled from the data collection on presidential performance held at the Roper Center for Public Opinion Research, consisting of 3,261 polls over the period between 1944 and The measure of egocentric economic evaluations (expectations of individuals future economic situation) is compiled from historical data series made available by the Thompson Reuters/University of Michigan Surveys of Consumers. Its survey asks respondents Now looking ahead--do you think that a year from now you (and your family living there) will be better off financially, or worse off, or just about the same as now?" The survey has been running since The Michigan Index of Consumer Sentiment (ICS), used in supplementary analyses, is a composite measure of five 13

14 different items from the survey (which include current personal finances, expected personal finances, expected business conditions, national unemployment expectations and current consumption conditions -- Generally speaking, do you think now is a good or bad time for people to buy major household items?. The Index of Consumer Sentiment dates to We code political event dummies using variables for the 1980 to 1991 period taken from Erikson et al. (2002, p. 52), in addition to coding other major events that occurred between 1992 and These include the federal government shutdown in December 1995 and January 1996, the abandonment of the Clinton health care plan in 1994, the terror attacks on the World Trade Centre in September 2001 and the start of intervention in Afghanistan, Hurricane Katrina, and Lehman s filing for bankruptcy in September Finally, our measure of vote intention is the generic congressional ballot, the survey question asked in a number of forms across a wide range of polling houses about how respondents intend to vote in their district in the upcoming congressional elections. The traditional Gallup formulation of the generic ballot question asked respondents... If the elections for Congress were being held today, which party s candidate would you vote for in your congressional district -- the Democratic Party s candidate or the Republican Party s candidate?. This question dates back to 1942, and we have 1,997 generic ballot polls from Wlezien and Erikson (2002; Erikson and Wlezien 2010), supplemented with data from the Gallup Brain database and trend data on generic ballot reported at The generic congressional ballot is not without issue as a measure of vote choice, since its emphasis on party labels erodes the significance of incumbent candidates in local districts, in which responses tend to adhere to partisan evaluations (e.g. McGhee and Baldassare 2004). 14

15 However we know that the generic ballot becomes an increasingly strong predictor of election outcomes as the election day nears (Erikson and Wlezien 2002; n.d.) and are therefore confident that the question taps national support for each party at a given point in time, while it may not fully capture dynamics in the local districts samples. We examined the robustness of our findings using two additional covariates; public policy preferences and congressional approval. Our measure of the public s preference for more liberal or more conservative government is Stimson s measure of public policy mood (Stimson 1991; Stimson et al. 1995). Our congressional approval series is from Durr et al. (1997), extended by Ramirez (2009) and updated through 2008 with data from Gallup, ABC/WaPo, CBS/NYT and NBC/WSJ polls. In total, the series is composed of 997 survey items. Results The application of the dyad ratios algorithm suggests that there is a strong underlying latent dimension in evaluations of the issue competence of parties in the U.S. For each item, the algorithm calculates how much variance is shared with the latent dimension and the strength of its loading onto this dimension. This loading is interpretable as a factor score. The latent dimension of valence explains 66.8% of all variance in the observed public ratings of issue competence of the Democratic Party and 50.84% for the Republicans. 4 These figures are substantively high. Stimson s (1991) analysis of left-right mood in policy preference sin the US found a first dimension accounting for some 27% of the variance, and 14% of variance for the second. A comparison of the variance explained across the item selection in the measure (with valence issues, with position issues and without economic issues) suggests that the index is substantively distinct if composed of valence or position issues, and uninfluenced by the removal of economic issues, shown in Table 1. 15

16 - Table 1 about here - If all issue ratings were functions of partisan bias, there would be almost identical correspondence in the variance explained. Using the valence index, therefore, the dyad ratios algorithm provides a value of valence for each party in each time point. The values are presented in Figure 1 using annual data between 1944 and Figure 1 about here - The measure reflects the relative strength/cohesiveness of the parties and their reputations for governing. For most of the time period between 1945 and 1980, the Democrats held a relative advantage over the Republicans in their competence rating, suffering occasional dips such as around 1953, 1967 through 1970 (coinciding with escalation of the Vietnam War), 1979 through 1981 (towards the end of the Carter presidency and the troubled 1970s, as Reagan and the Republicans started to gain ascendance), with ratings of the Republicans occasionally peaking for short periods (during the mid- to late 1950s under President Eisenhower, in the late 1960s under the pre-watergate President Nixon). Then, from the 1980s onwards, the index exhibits the gradual strengthening of Republican valence. The Democrats gain the ascendancy in competence ratings again around 1996 after the Republican Party had taken control of the House in the wave election of 1994, and rise sharply under the presidency of George W. Bush, before falling precipitously back after taking the presidency in the aftermath of the economic bailout. In the recent time points, then, the series depict the decline of Republican competence under George W. Bush, and latter a sharp decline in Democratic competence under the presidency of Barack Obama. Further to this, valence can be compared alongside the generic congressional ballot, using quarterly data for the period between 1980 and 2009, presented in Figure 2. The series exhibit strong correspondence over time for both the Republican and 16

17 Democrat series, indicating that issue valence may have the potential to offer an important explanatory variable in accounting for vote choice. - Figure 2 about here - Using this estimation of valence, we can explore whether common variation extends across the spectrum of policy domains which would be indicative of transfer or generalised competence heuristics. Here issue categories are aggregated thematically, allowing for sufficient sample sizes, and are estimated as sub-moods using the dyad ratios algorithm. Table 2 presents the correlations of the available issue sub-moods with valence for each party. - Table 2 about here - The results show that in all cases, issues from across each policy domain (e.g. foreign affairs, economy) correlate significantly with valence for both political parties. The associations are strongly significant for the Democrats, and significant but less strongly so, on average, for the Republicans, which can account for the slightly lower proportion of variance explained by valence for the Republican Party. These findings provide support for our theoretical expectation that variation in competence ratings is common across issues, indicative of a process of common transfer, or heuristics. Fractional Integration It is of substantive importance to establish whether shocks persist in competence evaluations. If shocks to the reputation of parties due to good or bad performance last indefinitely, public opinion on valence is an integrated process, and if shocks dissipate quickly, it is stationary. This determines the degree to which parties are forgiven for past mistakes and mismanagement. This is also important in methodological terms. The idea of fractional integration (Granger, 1980) is increasingly relevant to the study of political behaviour (e.g. Box-Steffensmeier and Smith 1996; 1998; Byers et al. 17

18 1997; 2000; Lebo et al. 2000; Clarke and Lebo 2003), prompting some debate (e.g. Wlezien 2000; Pickup 2010). Fractional integration suggests that time series can be long-memoried : where some effects of an initial innovation persist for an extended duration of time, while others are subject to much faster rates of decay. As was noted earlier, the aggregation of heterogenous processes at the individual level (where public opinion for some individuals is highly autoregressive, or persistent) can give rise to fractional integration at the aggregate level (Granger and Joyeaux 1980; Box- Steffensmeier and Smith 1996). The degree of fractional integration of each series can be estimated with the Robinson (1995) multivariate semiparametric method, which calculates the value d for the mood index of valence. The idea of fractional integration (Granger 1980) relaxes the assumption that the order of integration, d, is an integer, where the value d is a real number between 0 and 1. Table 3 presents these estimates, and the estimates for the covariates used in subsequent analysis in the paper. - Table 3 about here - The values reported in Table 3 suggest that our measure of valence is a fractionally integrated process for both parties, with the d being equal to 0.89 for the Democrats and 0.76 for the Republicans, with the t-statistic indicating this value is significantly different from zero. This indicates that shocks to valence are long-memoried, meaning that party reputations are slow to recover from shocks, and that this is more the case for valence than for similar disturbances to macro-partisanship or public policy mood, especially for the Democratic Party. The estimated order of fractional integration of valence for each party is also more persistent than for the version of the index that consists of position issues, which more closely mirrors the value of d for macropartisanship. These findings are quite intuitive: party identification and liberalconservative attitudes (via public policy mood, and position) are expected to revert 18

19 quickly to their long-run equilibrium, whereas valence is expected to be rather more persistent, carrying forward the value of shocks about competence into the future. Note that generic congressional ballot exhibits far less persistence in its dynamics as the effects of shocks to vote choice dissipate much more quickly than for macropartisanship, or valence. The estimation of the order of fractional integration, d, points to the need for fractional differencing to transform or pre-whiten the series prior to exploring their dynamic interrelationships. This modelling decision enables us to avoid threats to inference from spurious regressions with near- and fractionally integrated data (see De Boef and Granato 1999; Lebo et al. 2000), as well as avoiding over-differencing of series, which can lead to introduction of a moving average parameter. Valence and Macro-partisanship We next explore the degree to which it is plausible that valence may be serving as a proxy for partisanship, and the degree to which common variation in competence evaluations may be attributable to partisan bias. To explore these possibilities, we use tests for Granger-causation on the fractionally differenced series. The estimation of Granger-causation between time series (Granger 1969) tests whether past values of a variable x improve prediction of another variable y relative to prediction of y from past values of itself alone (see Freeman 1983 for a review of applications in political science). The Granger approach is useful for determining whether past values of issue competence improve prediction of current values of macro-partisanship, controlling for past values of macro-partisanship, and whether prior past of macro-partisanship improve prediction of current values of Valence, controlling for prior values of Valence. This relates to the temporal ordering of variables, even if these move together over time, to determine whether issue competence tends to lead party 19

20 identification or whether party identification tends to lead issue competence. As such, the tests for Granger causation enable us to demonstrate weak exogeneity if one variable contemporaneously occurs before the other. To test for Granger causation, a vector autoregression framework is used, with the number of lags selected according to the Akaike information criterion (AIC). The results for the null hypotheses, that Valence does not lead macro-partisanship, and macro-partisanship does not lead Valence, are presented in Table 4. - Table 4 about here The tests for Granger causation between Valence and macro-partisanship are encouraging. Valence leads macro-partisanship in the model for each party, rather than vice versa, suggesting that in the temporal ordering of variables, Valence improves prediction of macro-partisanship, but macro-partisanship does not improve prediction of Valence. This does not suggest that Valence is not influenced by macropartisanship (we expect there to be a theoretical and empirical relationship) but that the stronger temporal relationship is from Valence to macro-partisanship, controlling for prior values of the dependent variable, in the aggregate data that is available to us. Furthermore, using tests for Granger causation, we can scrutinise our selection of issues. We removed position issues from the measure of valence because we anticipated greater partisan bias, or endogeneity, for issues on which the parties and their voters are strongly aligned. We would expect therefore macro-partisanship to lead position issues but to have less statistical prediction upon valence issues. Support for this expectation is provided in Table 5, which reports the tests for Granger causation between macro-partisanship and the index comprised of position issues. - Table 5 about here - 20

21 The results reveal Granger causation from macro-partisanship to the position issues index for Democrats, as well as Granger causation from positional issues to macropartisanship for Republican partisanship. One further means of assessing the effects of partisanship on valence is within a multivariate framework. This also allows us to test theoretical expectations about valence, namely that valence should be explained by presidential performance, especially for the party of the incumbent, by economic evaluations, by macropartisanship and also by salient political and/or economic events. We then, finally, estimate the effects of valence on generic ballot controlling for the determinants of valence, to establish whether valence simply mediates existing performance indicators or provides additional explanatory power. A performance model of valence To explore the dynamic interrelationships between variables, we use a multivariate ARFIMA framework (e.g. Box-Steffensmier and Smith 1998; Box-Steffensmier et al. 2009), where the model of valence for each party takes the form: d VALENCE t = α 0 + β 1 d MACRO-PARTISANSHIP t + β 2 d PERSONAL ECONOMIC EXPECTATIONS t + β 3 d PERSONAL ECONOMIC EXPECTATIONS t *PRESIDENT t + β 4 d PRESIDENTIAL APPROVAL t + β 5 d PRESIDENTIAL APPROVAL t *PRESIDENT t + β 6 PRESIDENT t + φ 7 EVENTS t + μ t This is where valence is modelled via a performance model comprising party identification, economic evaluations, presidential approval and events, and where we expect presidential approval to operate when the party has a president in office. The model also includes an intercept (α 0 ) and the disturbance (μ t ), where μ t = ρμ t + θε t-1 + ε t in which the noise structure is modelled as a function of an autoregressive term (ρ) and the moving average (θ) of the error term (~ i.i.d. N(0, ζ 2 )). 5 21

22 The data are transformed ( pre-whitened ) by fractional differencing each series to make it stationary using the d s estimated with Robinson s (1995) semiparametric method reported in Table 3. We then explored whether an ARFIMA(p, d, q) framework was an appropriate specification to control for the presence of autocorrelation or moving averages in the data, testing for up to three autoregressive and three moving average parameters. We find that an ARFIMA(2, d, 0) specification best accounts for valence for the Democratic Party and that a (0, d, 0) model for the Republican Party. The results for the ARFIMA (p, d, q) model of valence are reported in Table 6. - Table 6 about here - The results indicate that the effect of macro-partisanship (the percent Democratic of party identifiers) on valence is significant in the expected direction for each of the parties. These effects are not insubstantial in size: a one point increase in partisan affiliation is associated with a 0.7 point increase in valence for the Republican Party and a 0.4 point increase for the Democratic Party. These findings suggest that the link between valence and partisanship is of a greater order for the Republicans. In terms of the link between presidential approval and party valence, through which presidential rates might be transferred onto evaluations of party performance, there is a negative and significant effect for the Democratic Party. This is perhaps not surprising given that for twenty years of the twenty-nine period there was a Republican President, leading positive approvals of Republican presidents to erode competence evaluations of the Democratic Party. This is confirmed upon inspection of the interaction of presidential approval with party control of the presidency for the Democrats, with the effect becoming positive and significant. The results do not 22

23 indicate, however, a link between presidential approval for Republican valence, whether in general or for incumbents in particular. Turning to the effect of economic evaluations, the model for the Republican Party suggests that the effect of egocentric economic evaluations is positive and significant, but also that the effect of economic evaluations on valence is reduced when the party holds the presidency, or greater when the party is out of power, with the effect of the interaction being negative and significant. The interaction of economic evaluations with party control is similarly negative and significant for the Democratic Party. This is an interesting finding, and suggests that whereas valence is moved by presidential approval for the Democrats when a party has a president in office, voters turn to other performance indicators, in this case economic evaluations, to inform their evaluations of valence in opposition. Economic evaluations affect mood (Durr 1993), macro-partisanship (Mackuen et al. 1989; Erikson et al. 2002), presidential approval (Mackuen et al. 1992), congressional approval (Durr etal. 1997), and trust in government (Chanley et al. 2000), and so this mechanism may reflect the transmission of a range of influences towards the challenger party. We find that there is evidence that the effects of major political events can have significant and lasting consequences for valence (tested in the model using a step intervention). Specifically, the model indicates that the federal government shutdown of 1996 had a positive and near significant effect on Democratic valence, while the bailout passed in 2008 was linked to a negative and significant effect on party valence. For the Republicans, the Persian Gulf War of 1991 was associated with a drop in party valence, while the effect of the events of September 11 and the start of combat operations in Afghanistan were associated with a positive and near-significant effect for Republican valence. Thus, there is evidence that events can have direct and 23

24 immediate effects on party valence. However, key to understanding the nature of valence is whether valence is simply the product of events and other performance indicators, as a mediating variable in the vote choice, or whether it adds explanation to the electoral preferences of voters. This is the subject of the final analysis. A valence model of generic ballot Following from the specification of the performance model of valence, our model of vote choice for each party takes the form: d VOTE t = α 0 + β 1 d MACRO-PARTISANSHIP t + β 2 d PERSONAL ECONOMIC EXPECTATIONS t + β 3 d PERSONAL ECONOMIC EXPECTATIONS t *PRESIDENT t + β 4 d PRESIDENTIAL APPROVAL t + β 5 d PRESIDENTIAL APPROVAL t *PRESIDENT t + β 6 d VALENCE t + β 7 d VALENCE t *PRESIDENT t + β 8 PRESIDENT t + φ 9 EVENTS + ε t The model includes an intercept (α 0 ) and stochastic error term (ε t ), (~N(0, ζ 2 ). Once again, the data are pre-whitened by fractional differencing by d and we explored whether an ARIMA(p, d, q) framework was an appropriate specification, testing for up to three autoregressive and three moving average parameters. Since we did not find evidence of presence of these parameters, we conduct ordinary least squares regression of the transformed series. The results for the OLS regression model of vote are reported in Table 7. In each party model (A and B) we report the performance baseline model from Table 6 in Model 1, and then introduce valence in Model 2 and interact this with party control of the presidency in Model 3, before introducing a selection of event dummies (Model 4). - Table 7 about here - The results indicate first of all that the effect of macro-partisanship (the percent Democratic of party identifiers) is significant in the expected direction in both of the base models, being negative and significant for the Republican Party and positive and 24

25 significant for the Democratic Party. When valence is introduced to the model, however, the effect of macro-partisanship ceases to be significant at the 95 confidence level for the Republicans. Interestingly, the effect of partisan identification on vote remains significant in each of the Democratic models. This is consistent with our findings in the model of valence in Table 6, where the effect of partisanship on valence was stronger for the Republican Party, suggesting that some effects of partisan identification are subsumed in the measure of valence for the Republicans. However, if valence is acting as a mediating variable for partisanship in the Republican model, it is exerting an additive effect in the Democrat model, whereby both valence and macro-partisanship are significant in models 2, 3 and 4. The increase in model fit between models 1 and 2 (from 0.29 to 0.31 in the Republican model, and from 0.39 to 0.41 in the Democrat model) suggests that valence is providing additional explanatory power, and not simply functioning as a substitute for macropartisanship. The effect of valence on vote is positive and significant for both the Republicans and Democrats. Furthermore, the size of the coefficient is substantial, with a one point increase in valence leading to a 0.5 percent increase for the Republicans and a 0.3 per cent increase for the Democrats in their two-party share of the vote in our final model (Model 4). When the interaction of valence with party control of the presidency is introduced, however, the effect of the interaction is found to be negative and significant for the Republican Party in our final model (*and approaching significance in Model 3), indicating that the effect of party valence on party vote is lower when the party holds the presidency. This is reminiscent of the earlier finding that the effects of economic evaluations were greater for the challenger party where presidential approval is stronger for the incumbent. Here, in the 25

26 Republican vote model, it again appears that voters turn to valence judgements in their performance assessment on the vote for the party in opposition where they can no longer turn to presidential approval to translate a performance assessment on their vote choice. The absence of effects for the Democratic Party might be attributed to the relatively shorter duration of Democratic presidential control, and The pattern for the effect of economic evaluations in the base model is similar to that for valence, with the effect being positive and significant for both the Republican and Democratic vote share with the interaction being negative and significant, indicating that the effect of economic evaluations on party vote is lower when the party holds the presidency. Once valence and its interaction with presidential control are introduced into the model (Model 3), however, the interaction of economic evaluations with presidential control ceases to have a negative and significant effect on vote. With the exception of the effect of the 1990 federal government shutdown on Democratic vote, events are not significant in our model of vote, suggesting their effects are transmitted through other measures, such as valence. In both cases, the introduction of event variables increases the coefficient for valence, supporting the interpretation that events have persistent effects on voters evaluations of parties on valence, and in turn this has a significant effect at the ballot box. We checked our results for robustness, introducing variables (congressional approval and public policy mood) and by alternating personal economic expectations with national economic expectations and consumer sentiment. In all cases the substantive conclusions remained identical, and the statistical power of valence on generic ballot was greater. Therefore the results presented in Table 7 represent the most conservative estimate of valence across potential model specification. 26