Measuring Party Brand Personality in Two- and Multiparty Systems.
Master Thesis
Research Master of Communication Science Graduate School of Communication
University of Amsterdam
Luca Sebastian Meister Student ID: 10393315 University of Amsterdam Supervisor: dr. Yphtach Lelkes Date of Submission: 11 July 2014
Acknowledgements
In the process of this thesis, I have received lots of support and
encouragement for my work in many different ways. In the following, I would like to thank some persons that were particularly supportive.
My Supervisor, dr. Yphtach Lelkes,
For always being critical with my work and thereby helping me to make most out of it.
My friend and fellow student, Niklas Johannes,
For his valuable methodological advice and his bad jokes.
My parents, Petra and Jörg-Uwe Meister
For their unconditional support throughout my academic career.
But, most of all
My grandfather, Dietrich Heinrich Max Meister,
Abstract
Over the last two decades, political communication research has increasingly devoted attention to the phenomenon of political party branding. Smith (2009) was one of the first to conceptualise and test a construct he called the party brand personality (PBP) scale in order to make party branding measurable. However, his concept exhibits limitations as it solely deals with PBP in the context of the two-party system of the UK. However, multi-party systems are largely different, leading to the question in what respects party brand personality differs in multiparty systems compared to two-party systems. Emanating from the scale developed by Smith, two additional dimensions have been added in order to account for differences in party competition among various political systems. To test these newly introduced dimensions, a comparative study in two fundamentally different political systems was employed, comparing the US to Germany. Using Confirmatory Factor Analysis (CFA) the scale has been tested and respecified for both countries. Results do not only show that party systems matter for the conceptualisation of PBP. In
particular, one of the additional dimensions, Cooperativeness, helps to explain the overarching construct and shows that PBP should account for the different political market in which parties compete. As an additional finding, party identification proved to affect the way people evaluated political parties. Summing up, this paper discusses the results’ implications for future research on political party brands.
Keywords: political branding, political parties, structural equation modelling, confirmatory factor
Introduction
Speaking of political actors as brands has become prevalent over the past decades. Principles from branding are increasingly applied to politics and some researchers even discuss political branding as a new quality of political marketing (Adolphsen, 2010; Scammell, 2007). In general, political branding is considered a response to changing electoral markets (Reeves & de Chernatony, 2006). These are largely down to the societal changes in Western countries over the post-war period and led to voter dealignment and the decline in party mass membership (Dalton & Wattenberg, 2002). As a response, parties have permanently sought for new ways of reaching their audiences using advice of outside experts from the fields of marketing and branding (Farrell & Webb, 2002).
A brand is best described as a psychological concept, individually constructed in people’s minds when thinking about an organisation and/or its product (Keller, 1993). A
political brand as defined by Smith and French is “an associative network of interconnected
political information and attitudes held in memory and accessible when stimulated from the memory of the voter” (Smith & French, 2009, p. 212).
Although the normative implications of political branding for democracy have been widely discussed1, only few researchers have attempted to measure the construct to make it more tangible. Drawing back on Aaker’s (1997) seminal paper measuring the brand personality of consumer brands, Gareth Smith (2009) adapted the original scale to political parties.2 Nevertheless, Smith’s party brand personality (PBP) scale bears some limitations. Most importantly, it fails to address how different party systems influence PBP. As for now, the construct has only been conceptualised for and tested in the context of a two party system. !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
1 The impact of political branding on democracy is best captured by Scammell (1999), Lilleker (2005) as well as Marsh and Fawcett (2011).
2 Aaker deliberately used the term “personality” in her research, even though the term is mainly connoted with persons and not abstract and intangible constructs such as organisations and/or corporations. The idea behind this was that respondents apparently found it easier to describe and characterise brands when imagining them as persons. This principle has been picked up by Smith who asked respondents to think of political parties as persons.
This is why the main research question of this paper is: In what respects does party
brand personality differ in multiparty systems compared to two-party systems? To address
this question, this work intends to make a more institutional approach at measuring PBP in two- and multiparty systems in a comparative way. According to Sartori (1976), party systems are defined by how parties compete and interact. This not only translates into different roles parties assume within those systems but also into different political cultures (e.g. Lijphart, 1999). The latter is crucial for people’s perception and evaluation of political parties and thus also affects the construct of PBP. In short: Different party systems might affect the way people think about a party’s brand personality. Consequentially, this paper seeks to find out whether PBP has to be measured differently in two- and multiparty systems, considering the differences regarding party competition between the systems.
Just as brand personality is a very important proxy for purchase decisions of consumer goods, PBP in a similar vein is a proxy for people’s voting decisions. According to Smith, it is one of the key determinants for voting choice (Smith, 2009). Research on PBP thus inevitably means delving into the most crucial questions of political science and political marketing: Why do people vote for certain parties? And what can parties do about their image – defined by implicit and explicit ways of communicating – to be more appealing to voters? Since PBP is about parties’ image as a latent form of communication, it is inevitably a matter for political communication science. These latent forms are for example informed by how party representatives communicate with their target audiences and their competitors. As Smith (2009) argues, PBP is not only determined by what parties communicate but also by how they do it. This is why PBP has the potential to make the often-neglected emotional and intangible aspects of vote decisions tangible.
The paper’s theoretical section consists of two parts. The first deals with the relation of politics and branding, with particular respect to Smith’s (2009) seminal paper. The second picks up one of its main limitations and focuses on the implications of different party systems
for measuring party brand personality (PBP). The main contribution of this paper is to make an attempt at adding aspects from political science theory to and existing construct from the field of political branding and measure them accordingly.
Branding and politics
This section first discusses to what extent political parties are brands and second, introduces the construct of political brand personality, which will be subject to this study.
Brands and branding
A brand’s defining criterion is to “act as a shortcut to consumer choice, enabling differentiation between broadly similar products” (Scammell, 2007, p. 177). It is so to say “’the psychological representation’ of a product or organi[s]ation” (ibid.).
Branding, then, is the process of creating these psychological representations and shortcuts. It is a sub-discipline of marketing. Compared to other marketing concepts, branding particularly stresses the symbolic value and the soft aspects of a product (Marsh & Fawcett, 2011; Scammell, 2007). It “is a process of creating identity for a product; in other words, creating consumer equity and thus contributing to the greater uptake of a product in the marketplace” (Basu & Wang, 2009, p. 78). Thus, it goes beyond what previous forms of marketing aimed at: the mere fabrication of a certain image. Branding is about creating an emotional added-value that signifies more than the actual value of the product or organisation in order to make it more appealing to and recognisable for its (potential) customers (Marsh & Fawcett, 2011). Consequentially, a “brand therefore does not constitute what a product is but what the customers perceive it to be” (Adolphsen, 2010, p. 31).
Political parties as brands
Political parties are brands because they feature two main aspects of how Adolphsen (2010) defines a brand. I refer to them as emotional narratives and trust-building.
Particularly in Western Europe, parties developed along social cleavages and established certain ideologies accordingly (Inglehart, 1977). These ideologies – just like
brands – act as emotional narratives, social unifiers and also help to differentiate parties from their opponents (Adolphsen, 2010). As Smith and French (2009, p. 212) argue, parties developed “core brand values” that voters relate to them. For example, Social Democratic parties might be more strongly associated with social justice, whilst Liberal parties’ core brand value is individualism and freedom (ibid.). Despite programmatic convergence among parties in many Western countries, ideologies remain a highly important brand facet (Katz & Mair, 1995; Mair, Müller, & Plasser, 2004). Another branding strategy parties employ is linking the party to people’s lifestyles (Adolphsen, 2010; Smith & French, 2009). The idea behind it: When buying a specific product, some people do this to express a certain identity (Smith & French, 2009). Consumerism becomes an act of self-expression. Applied to politics, parties’ agendas also become emotional-laden. This means that people who consider environmentalism to be relevant to their identity might be inclined to vote for an eco-friendly party. However, some researchers reject this notion and argue that politics is not comparable to commercial markets (e.g. Henneberg, 2006; O'Shaughnessy & Henneberg, 2007).
As for trust-building, a party’s most essential brand values are important. Just as with commercial brands, voters’ relationship with political parties is defined by commitment and trust (Marsh & Fawcett, 2011). If parties either fail to live up to the promises once made or do not communicate as unitary actors (e.g. by sending inconsistent messages), voters might withdraw their support. In contrast, if they deliver over a longer period of time, voters might more easily “forgive” and keep supporting the party. The more exciting and unique a political brand is, the more easily it is forgiven (Smith & French, 2009).
Summing up, “[p]olitical parties are the ultimate brands” (Burkitt, 2002, p. 176 cited in Scammell 2007). This is because they serve as helpful heuristics in a complicated market and provide “cohesion, recognition and predictability” (Smith & French, 2009, p. 212). Thereby, they make information seeking and processing easier for voters, which makes brands a valuable aid when people make voting decisions (Popkin, Gorman, Phillips, & Smith, 1976;
Schneider, 2004). Learning about how people perceive of political parties as brands thus also helps to learn more about how they form voting decisions (Cwalina & Falkowski, 2008; Scammell, 1999). This is why this paper takes a consumer-oriented perspective to political branding. According to Smith and French (2009), this approach assumes that voters behave like consumers. They acquire and process information, create opinions and make decisions about voting for parties just as they do with consumer brands.
Measuring political branding: party brand personality
Despite all theoretical elaborations on the relation between branding and politics, only few studies attempted to measure the concept. Schneider (2004) remains one of the few so far, having measured brand perceptions of political parties and politicians in Germany. Despite also drawing back on the work of Aaker (1997), he did not amend the original scale to political parties. Except for the work of Schneider, research on parties as brands is scarce for multiparty systems.
Tested in the (then) two-party system of the UK3 and based on literature from branding and psychology, Smith (2009) developed a construct called Party Brand Personality (PBP). Personality, in this respect, is an observer’s individual sense making “to describe the ‘inner’ characteristics of another person” (Allen & Olson, 1995, p. 392). Brand personality, according to Aaker (1997, p. 347) then is a “set of human characteristics associated with a brand.” Consequentially, PBP denotes a selection of human characteristics that people associate with certain political parties. People’s main motivation to individually construct PBP lies with the fact that many voters consider acquiring political information as costly (Downs, 1957). Learning about parties (and their perceived personalities) reduces the risk of making wrong (voting) decisions and facilitates future (voting) decisions (Smith, 2009). !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
3 By the time Smith conducted his research on PBP in the UK, the country featured an effective two-party system with a single-party government (Mellows-Facer, 2006). However, after the 2010 general elections, the United Kingdom saw the formation of a coalition government between two parties (Conservatives and Liberal Democrats) for the first time since the end of WWII. This fact is highly relevant for this studies’ case selection and will be dealt with below.
According to Smith (2009), four antecedents determine PBP. These are events, political actors, advertising and brand users/endorsers.4 The impact of these four factors on the construct is moderated by party identification. “The more partisan a voter, the more positive their perception of the personality of the party should be”, he argues (Smith, 2009, p. 218). Here it may be noted that PBP is not the only factor influencing people’s voting intention. As Newman (Newman, 1999) notes, however, it is just one of five main predictors for voting behaviour.
The shortcoming of PBP as now conceptualised
Despite its unarguable achievement for making political branding more tangible, the current conceptualisation of PBP exhibits one major limitation, which this paper seeks to pick up.
As Smith mentions himself, “differing political cultures may well produce differing personality dimensions than those of the United Kingdom” (Smith, 2009, p. 225). Most democratic Continental European countries feature multiparty governments and thus also have different political cultures compared to the UK with its tradition of an effective two-party system (e.g. Mair, 2002). Following the argumentation in the introduction, this alone would be reason enough to test the scale in a different context. But even more important is the fact that research in the field is mostly about applying theories and concepts from branding to politics without accounting for the peculiarities of the political market compared to consumer markets. Although there are ample arguments for why political parties are brands, this does not mean that political and commercial markets are comparable in all respects. Even more, political systems (and thus markets) also highly differ. This is why this paper seeks to add an institutional dimension to the existing PBP construct, integrating both aspects from branding and politics in the scale and acknowledging for differences between political markets.
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
4 Due to parsimony, these four antecedents will not be explained in detail. For more information see Smith (2009).
Party brand personality and party systems
As touched upon in the previous section, Smith’s (2009) approach to conceptualising brand personality fails to account for the differences among party systems. The second theoretical section of this paper seeks to point out why different electoral and party systems matter for the conceptualisation of party brand personality.
The main argument of this section is that party competition in multiparty systems is much more driven by cooperation, whilst two-party systems feature parties with a more competitive nature. This is why I introduce and test Competitiveness and Cooperativeness as additional dimensions to the construct developed by Smith (2009) in the final subsection of this part. First, however, it is important to discuss the differences between party systems and how these relate to different ways of party competition and political cultures.
Electoral and party systems
Electoral systems shape party systems and how parties compete within these (Riker, 1982). In its essence, Duverger’s law posits that plurality (also called majority) voting systems lead to two-party systems and proportional representation systems to multiparty systems (Duverger, 1963).56 As a result, two-party systems mostly breed single-party governments whereas proportional representation systems generally lead to coalition/multiparty governments (e.g. Sartori, 1976).
Plurality and proportional representation systems have been designed to fulfil different purposes, as described by Lijphart (1999). According to him, the former’s main goal is to create stable government and ensure the rule of majority. This is why plurality systems (mostly) breed so-called majoritarian democracies. Proportional representation systems, in !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
5 In the following section I will use plurality/two-party systems and proportional representation/multiparty systems as synonyms. I am well aware of the fact that despite Duverger’s law there are few exceptions from the rule such as Canada and India, who feature multiple parties in parliament despite having plurality systems (Riker, 1982). However, Duverger does not claim causality between electoral systems and types of party systems. Rather, he acknowledges a strong correlation. Taking this into account, I will nevertheless treat these pairs to be synonymous.
6 For a critical discussion of the theses of Maurice Duverger, see amongst others the works of Riker (1982) and Schlesinger and Schlesinger (2006).
contrast, highlight rights of minorities. Moreover, parliament is intended to depict the
different societal groups in relation to their size. Since bargaining processes in these systems are signified by compromise, Lijphart calls it consensus democracy (ibid.).
Different political cultures are both cause and effect of the various founding principles of majoritarian and proportional representation systems. As a result, people have different conceptions of the role of parties in a given system and how they should act. This is why I argue that these translate into different ways of evaluating political parties. As individuals’ evaluation of parties is crucial to measuring party brand personality, I hypothesise the construct has to be measured differently in different party systems. Drawing on Sartori’s definition of a party system as a “system of interactions resulting from inter-party
competition” (1976, p. 44), the differences in party competition among the two types of party systems will be described.
Party competition in two- vs. multiparty systems
The main difference between majoritarian and proportional representation systems lies with how parties interact and compete. Two-party systems are mainly defined by competition between parties but multiparty systems are characterised by both competition and co-operation (Dahl, 1966). This particularly holds for two arenas: the electoral and
parliamentary. Since these two arenas are the ones most visible to citizens (e.g., compared to the factional arena), they are also crucial for shaping parties’ images and, consequentially, how people think of these parties as brands.
As for the electoral arena, elections in two-party systems are zero-sum: The win of Party A inevitably implies the loss of Party B (Downs, 1957). In multiparty systems, the composition of the to-be government is not entirely clear on Election Day. In fact, it is not elections but inter-party negotiations after elections that determine future government
coalitions (Müller, 2005). To illustrate this: Party A (receiving 40% of the seat share), which has lost elections in absolute terms to Party B (45%) can still be considered a winner in the
end. In case A manages to form a coalition with Party C (15%) outnumbering B, the latter then ends up in opposition despite having won the biggest share of seats in Parliament.
This is also why both parliamentary systems heavily differ regarding the competition
prior to elections. As governments in majoritarian systems are either replaced entirely or
remain in office, campaigns are often defined by a rather confrontational style (Mair, 1996). The same goes for the policies the two main parties advocate prior to elections. In
majoritarian democracies, the two main parties present policy programmes between which voters then chose in general elections (Downs, 1957). Simply put: As there is no post-election bargaining, voters know what policies they get when going to the voting booth.7
Campaigning in proportional-representation systems is less black and white. This is mainly down to the fact that all office-seeking parties want to have as many coalition options at hands as possible. Especially in times of voter-dealignment, decreasing party membership and increasing number of floating voters8, these parties are forced being able to forge new allies at any time (Katz & Mair, 1995). As a result, negative campaigning is less prevalent in multiparty systems and campaigns in general feature slightly milder tonality (Hansen & Pedersen, 2008).9 The coalitional uncertainties prior to elections also imply that voters do not know exactly what policies they will get in the end. In multiparty systems, they can express a preference by voting for a given party. However, the actual government programme is shaped in inter-party negotiations of the coalition parties after elections (Müller, 2005).
Leaving the electoral arena, parties enter the parliamentary arena. Again, two-party systems invoke a very straightforward logic with two opposing camps. Their interaction is
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
7 This is a simplified view. Even in a prototypical two-party system such as the US, single-party governments cannot necessarily put their policy agenda into practice after assuming office. This, however, is down to the bicameral institutional layout of most Western parliamentary systems, which allows for legislative gridlock under certain circumstances (e.g. Heller, 2007).
8 For more information about these socio-political transitions, please see the work of Inglehart (1977).
9 For more information on negative campaigning in multiparty systems as opposed to two-party systems, see for example Walter et al. (2014)
highly competitive and confrontational. Proposals of Party A are very likely to be rebutted in their entirety by Party B and vice versa.10
Even though multiparty systems are also characterised by the competition between government and opposition, they also require some degree of cooperation. Since coalition governments are less stable than those in two-party systems (Mair, 2002), office-seeking parties increasingly overcome ideological hindrances to become more versatile (Mair et al., 2004). Today’s fierce opponent might be a valuable power option in the future. By
maintaining ties with as many parties as possible, the bargaining position of parties improves and they become a pivotal player in the party system (Müller, 2005).
The dimensions of party brand personality in two- and multiparty systems
Based on the line of reasoning about the two types of party systems, this paper’s main hypotheses pick up the differences between two- and multiparty systems and thus states:
H1: Party brand personality (PBP) has to be measured differently in two- than in multiparty systems.
Competitiveness11 particularly refers to how parties compete in two-party systems as described above. To test the explanatory power of this dimension, I added personality traits such as ambitious, competitive as used by Parke et al. (2004) as well as risk-taking and others to the original set of items (see Figure 1). Due to the rather confrontational character of party competition in majoritarian democracies, I assume that this dimension particularly helps explain the overarching construct in two-party systems. Hypotheses 2 thus states:
H2: In a two-party system, Competitiveness is more helpful explaining the variance in party brand personality than in a multiparty system.
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
10 Here it may be acknowledged that faction discipline is less prevalent in two- than in multiparty systems (e.g. Cox & McCubbins, 2005). As a consequence, Republicans might well vote for a Democrat proposal and vice versa. Nevertheless, this does not undermine the very character of two strictly opposing camps in two-party systems in general.
11 It is important to note that dimension Competitiveness and Cooperativeness are not semantic differentials. This is also underlined by the two dimensions’ indicators, which are certainly not semantic differentials. All indicators have been measured independently, using a five-point Likert scale (G. Smith, 2009).
Cooperativeness, in contrast, is expected to have a higher explanatory power for multiparty systems. Because of their consensus-driven political culture, citizens in these countries are expected to value compromise and inter-party cooperation more than their counterparts in majoritarian democracies. This is why hypothesis 2 reads as follows:
H3: In a multiparty system, Cooperativeness is more helpful explaining the variance in party brand personality than in a two-party system.
Items for cooperativeness include altruistic (Capara, Barbaranelli, & Zimbardo, 1999) as well as mediating, accessible and willing to compromise. Figure 1 displays the conceptualised model, which will be tested in the following section.
Research design and method Selection of cases
Since this study aims at a comparison of PBP in the context of two- and multiparty systems, it is about choosing countries, which come closest to either archetype.12 The selection of cases was thus guided by three criteria: First of all, the number of parties in parliament and government, to ensure differences regarding party system and political culture. Second, countries’ voting systems had to be in line with Duverger’s law: majority voting for the two-party system and proportional-representation for the multitwo-party system country. And, third, as for multiparty systems, a rather low number of parties were preferred since this required respondents to evaluate fewer parties.
This is why I selected the USA and Germany as countries to study. The United States feature a two-party system with a single-party government. In combination with their majoritarian voting system, the US come close to the archetype of a majoritarian democracy as discussed above (e.g. Lijphart, 1994). The Federal Republic of Germany, in contrast, always had multiple parties in parliament since 1949 and always featured a coalition government except for the time between 1957 to 1961. As a result of the 2013 general elections, only four factions13 are represented in the German parliament (Egeler, 2013), which makes Germany suitable for this study, as it requires respondents to evaluate fewer parties (compared to many other European multiparty systems). Even though Germany features a mixed voting system with aspects of both majoritarian and proportional representation systems, the allocation of seats in parliament is mainly determined by the second vote, called !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
12 For his study, Smith (2009) assessed PBP for British parties. However, since the 2010 general elections, the UK features a two-and-a-half party system with a coalition government between Liberal Democrats and Conservatives (G. Smith & Spotswood, 2013). As pointed out above, more than two parties in parliament and coalition governments are not a feature of the archetype of a majoritarian democracy.
13 Even though the German Bundestag features four factions, these comprise five parties in total. Here it may be noted that the CDU’s sister party, the Christian-Democratic Union (CSU), with which the CDU forms a faction in parliament, has not been considered. This is due to the principle of parsimony and the fact that the CDU is running for office in 15 out of the 16 German federal states, whereas the CSU is only competing in Bavaria. Since people’s perceptions of CDU and CSU might diverge due to slightly different political agendas and the greater relevance of the CDU, I chose for only querying respondents about the CDU.
Zweitstimme (Nohlen, 2004). This second vote follows the principle of proportional
representation, including a five per cent threshold (ibid.). A proportional representation-guided voting system in combination with a multiple but limited number of parties in parliament render Germany the ideal country to study the perceptions of PBP in multiparty systems.
Questionnaire design
As for the core section on PBP, all character traits featured in Smith’s (2009) final scale were included with the additional eleven items for cooperativeness and competitiveness as depicted above. In total, respondents were asked to assess 45 personality trait items. Compared to Smith (2009), the question wording has been slightly amended. People were explicitly asked to think of the parties as persons and indicate how much they (dis)agreed the respective personality traits were descriptive of the parties. This has been done to both increase the results validity and to make sure people were evaluating the party as a whole instead of evaluating individual political actors.14 Also, in contrast to Smith’s 5-point scales, 7-point scales were used to obtain more accurate data.15
After constructing the first draft of the questionnaire, seven cognitive interviews were conducted. Following the guidelines of Caspar et al. (1999), respondents were first asked to take the self-administered survey (for both English and German) in the presence of the interviewer and comment on improvements. Afterwards they were purposively asked about potential issues regarding the questionnaire and/or survey, respectively. This not only proved helpful for matters of question wording and item presentation. Since all seven interviewees
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
14 Unfortunately, the original question wording used by Smith could not be obtained. The complete questionnaires for both the US and Germany can be found in the appendix.
15 The original 5-point Likert scales were amended to 7-point scales. As Colman et al. (1997), the latter might provide more accurate data since people avoid choosing extreme values in general. Offering two more scale points, prevents them from giving neutral answers and accounts for more nuanced evaluations. Despite these differences, Dawes (2008) argues that 5- and 7-point scales are still comparable, which is important for this study.
were native German-speakers who were fluent in English, the interviews also helped to rule out misunderstandings and translation issues from English to German.
As for the US, respondents were asked to rate the Democratic and Republican party. German participants had to assess the four parties represented in the German Bundestag, which are Christlich Demokratische Union Deutschlands (CDU), Sozialdemokratische Partei
Deutschlands (SPD), Die Linke (the Left) and Bündnis 90/Die Grünen (the Greens). In order
to keep the survey rather short and to avoid dropouts, German respondents were asked to evaluate two randomly selected parties out of the four to make the German survey more comparable to the US survey. To prevent order effects, trait batteries were randomised just as the order of items within the batteries.
The sample
U.S. data was collected via crowdsourcing platform Amazon Mechanical Turk. Between 23 to 24 May 2014, a total of 182 competed the questionnaire, of which 169 have been included in the analyses.16 45.6 per cent of those 169 were female and 54.4 per cent were male. Participants’ age ranged from 20 to 69 years (M = 35.0; SD = 11.73).17 Regarding their educational background, amongst others 5.9 per cent finished high school, 15.4 per cent have been to college but left without a degree. 42.6 per cent of US respondents hold a Bachelor’s degree with another 15.4 per cent holding a Master’s. Just as the mean age slightly deviated from the mean age reported in the 2010 US Census (37.4), male respondents are overrepresented in the sample compared to the 49.2 per cent reported in the Census (U.S. Census Bureau, 2011). Most notably, however, is the overrepresentation of highly educated respondents as the US Census reports merely 18.2 per cent of the population holding a bachelor’s and another 10.9 per cent holding a postgraduate degree.
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
16 The excluded cases are outliers that had to be deleted to meet the requirements for the analysis as outlined by Kline (2010). This will be addressed in the following section. Since for both countries less than ten per cent of the cases were deleted, this seems feasible.
17 The survey required respondents to be eligible to vote, meaning 18 years of age or older for both the US and Germany.
Regarding Germany, data was collected via snowball sampling between 15 to 31 May 2014. In total, 191 respondents completed the survey, 174 were considered in the analyses; 47.1 per cent of which were female and 52.9 per cent were male. They were between 18 and 76 years of age (M = 31.6; SD = 12.62). Regarding education, overall, German respondents were educated above average with only 7.5 per cent of the sample not having obtained the general qualification for higher education, called Abitur. 15.5 per cent of the German respondents to have terminated education with Abitur. The largest groups either hold a Bachelor’s (36.8%) or Master’s degree (37.4%). Just as for the US, men are also slightly overrepresented in the German sample compared to 49.1 per cent in the population (Willand, 2013). The mean age was considerably lower compared to the population (M = 43.7). Regarding education, the German sample also features an overrepresentation of highly educated people. On aggregate, almost three in four respondents hold a degree that it comparable to the Bachelor’s degree in the US system, whilst this holds for only 19.3 per cent of the population (ibid.). Summing up, both samples are not representative. Nevertheless, since the same groups are overrepresented in both samples, I argue that comparability is given.
The method
Whilst Smith’s design was exploratory, using Principal Components Analysis (PCA) in order to define the factorial structure for PBP, this paper emanates from a more theoretical approach. Consequently, it is about testing (and ideally confirming) the eight hypothesised dimensions for PBP. This is why a Structural Equation Model (SEM) in form of a Confirmatory Factor Analysis (CFA) appears to be suitable for this work. First of all, to test H1, a multi-group comparison CFA was conducted. In case respecification should not lead to one model sufficiently fitting both groups, the CFA models have been respecified for both countries independently. Then, a second-order factor was introduced to the models. This was done to measure the degree to which the dimensions hypothesised explain PBP and to test H2
and H3. Finally, multi-group analyses were conducted to assess the impact of party identification on PBP. All analyses have been carried out using IBM AMOS 20 software.
Data handling
In order to simplify the following analysis, variables for how the total of six parties scored on the 45 character traits were grouped for both the US and Germany. This form of parcelling was necessary to allow for a CFA comparing the two groups.18
Before running the analysis, I checked for variance, skewness, kurtosis and multivariate critical ratio. According to Kline (2010) this is necessary to meet the requirements for a CFA. As for variance, the highest value should not exceed the lowest one by a factor ≥ 10. Bounds for skewness were ±1, for kurtosis ±3 and for multivariate kurtosis 6 (Kline, 2010).
First, data was checked for on aggregate and then for the separate groups. On aggregate, no problems occurred. For both the US and Germany most variables have been within bounds, with values for skewness between ± 1 and kurtosis between ±3. Consequentially, these variables are normally distributed and therefore fulfil the requirements as noted by Kline (2010). However, for the variables Ambitious, Hardworking and Confident, skewness was slightly out of bounds for Germany. Considering the rather high multivariate critical ratio for Germany (12.696), square-root transformations of the three variables were carried out in order to correct for skewness and kurtosis. Since this did not help improving the fit indices, bootstrapping was used to obtain reliable standard errors.
As for the US, none of the variables exceeded the boundaries for skewness, variance and kurtosis. Nevertheless, multivariate critical ratio was too high (24.217). Since transformations of single variables were not recommendable, bootstrapping was used just as !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
18 Even though data has been parcelled, querying respondents about specific parties beforehand was necessary. First, people find it easier to assess tangible than intangible constructs, e.g. concrete parties rather than parties on aggregate. Second, when evaluating political parties on an aggregate level, other confounding factors might have come into play, such as political cynicism and political efficacy. To avoid those concepts, I argue that asking people about particular parties is more appropriate. Going in line with the arguments of Little et al. (2002), parcelling was necessary.
in the case of Germany. To meet the prerequisite of normally distributed data, a total of 30 outliers were deleted in both data sets, which accounts for less than ten per cent of each data set and is thus still acceptable, according to Kline (2010). Subsequently, checks for multicollinearity and linearity were conducted and yielded sufficient results for both countries. As a rule of thumb (Kline, 2010), none of the variables should be correlated with another by more than .85, which was not the case.
Procedures and results
To make the process of respecification more comprehensible, section 5.1 briefly outlines its underlying principles.
Motivating the respecification procedures. As mentioned above, the theory-driven
approach of this paper recommends performing a Confirmatory Factor Analysis (CFA) rather than an Exploratory Factor Analysis (EFA). The former technique requires hypotheses to be formulated first and then to be tested – ideally to be confirmed – by the data (e.g. Kline, 2010). This translates into two fundamental guiding principles when respecifying the models in the following. First, all applied changes have to be theoretically justifiable. Even though Modification Indices (M.I.) in AMOS were consulted for tentative adaptions, they were only put into practice in case they were theoretically sound. Second, even if changes recommended by M.I. seemed theoretically feasible, they were only employed when found to significantly improve model fit. This was both checked for via p values of the standardised estimates in AMOS and χ2-difference test, which both had to be significant to indicate substantial improvements regarding model fit. 19
The possibilities for respecifying CFA models include five main steps as described by Kline (2010). They follow two principles: First, try to uphold unidimensionality20 and second, !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
19 In both cases, p ≤ .05 was the threshold for significance.
20 According to Kline, the principle of unidimensionality is violated when “any indicator loads on ≥ 2 factors or if its error term is assumed to covary with that of another factor” (Kline, 2010, p. 115). As he also argues, in multidimensional models testing for discriminant and convergent validity might be less precise (ibid.).
try to avoid (premature) deletion of indicators. Before respecifying the model, the first step involves checking for convergent and discriminant validity. According to Kline (2010), discriminant validity requires factor covariances to be ≤ .85. If this is not the case, merging factors shall be considered. Convergent validity requires indicators to feature standardised factor loadings of at least γ* = .7 (also meaning R2 ≤ .5).. If this is not the case, badly fitting items might be removed later on (step five).
The second step involves adding error correlations within factors. Correlations among error terms are based on the assumption that two indicators share something that is not accounted for in the model, i.e. that is not measured in the model (Kline, 2010). As long as these are added only within factors, this does not violate the principle of unidimensionality, meaning that indicators (and their error terms) load on one factor only.
Thirdly, cross-factor loadings may be added. The underlying assumption is that an indicator might well inform more than one factor within the model, as it is crucial for measuring more than one of the latent constructs within the model. Consequentially, as a fourth step, error correlations across factors might be added. Here it may be noted that both the second and third step of respecification violate the assumption of unidimensionality. Nevertheless, as Kline (2010) argues, this is still better than removing indicators, which inevitably denotes losing data. Thus he considers the deletion of indicators – the fifth and final step – ultima ratio when respecifying CFA models. It also has to be kept in mind that factors have to have at least two indicators (Kenny, 1979).
Motivating the interpretation of fit indices. When interpreting model fit, “there is
no statistical ‘gold standard’ in SEM” (Kline, 2010, p. 190). In the course of the following analyses, I will resort to three indicators, which are widely used to interpret Structural Equation Models such as this CFA. These are: the Comparative Fit Index (CFI), the Root !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
anyway, allowing for cross-factor loadings might be a way to find a factorial structure that fits the data best. Of course under the condition that all changes introduced are theoretically justifiable and significantly help to improve model fit.
Mean Square Error of Approximation (RMSEA) and χ2. All three are used to assess model fit in the following. However, I will briefly discuss pros and cons of each measure and motivate why I resort to χ2 as ultimate indicator of model fit.
CFI and RMSEA are both approximate fit indices (Kline, 2010). The advantage of CFI, compared to other fit indices, is that it is not influenced by sample size, which might be useful with a comparably small sample size as in this study (e.g. Bentler, 1989; Cheung & Rensvold, 2002). RMSEA, in contrast, is largely unaffected by degrees of freedom (df), which makes it a good complementary index to CFI.
The main drawback of approximate fit indices compared to χ2 is that their thresholds applied for hypothesis-testing are only rules of thumb (Iacobucci, 2010; Kline, 2010). Some researchers even argue that these thresholds are completely arbitrary and not generalizable, which would make them inappropriate for determining model fit (Barrett, 2007; H. W. Marsh, Wen, & Hau, 2004). Instead, only χ2, df and p values should be reported (Hayduk, Cummings, & Boadu, 2007). χ2, in contrast, is the only inferential fit index in SEM and the only index for which clear prerequisites are defined (Iacobucci, 2010). Admittedly, χ2 is very sensitive to sample size as it then is more likely to be significant, which indicates poor fit and generally more easily leads to the conclusion that the model does not fit the data (ibid.). With data sets of rather small N ≤ 400 (as it is here the case), this shall not be a problem as Kline (2010) states.21 The main advantage of χ2 compared to other indices, however, is that it provides a tool for improving model fit: χ2-difference test. In contrast to approximate fit indices, χ2 -difference test enables us to make sense of whether respecifications significantly improve model fit. Due to the aforementioned drawbacks of approximate fit indices, I resort to χ2 as !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
21 As some researchers argue, despite significant χ2, some models might be still considered to fit the data. More precisely, if χ2/df ≤ 3 (see Iacobucci, 2010; Weber & Federico, 2007). Kline (2010, p. 204) refers to this as
“normed chi-square” and rejects this notion due to the lack of well-defined guidelines for the maximum value of
χ2/df. Even more, he criticises the use of df as a denominator in order to reduce sample size sensitivity of χ2. According to him, however, df is not related to sample size at all. This undermines the statistical soundness of this measure, which he urges not to apply (2010).
main indicator of model fit. However, as Kline (2010) argues, CFI and RMSEA may well provide valuable additional information for assessing model fit.
Much research employing SEM recommends cutoff values for CFI to be ≥ .90 and RMSEA should be around .05, with the lower boundary of the confidence interval being around .05 and the higher one not exceeding .1 (e.g. Hu & Bentler, 1995; Weber & Federico, 2007). However, in their paper on how to best avoid Type I and II errors, Hu and Bentler (1999) argue for stricter rules. The threshold for CFI should be ≥ .95 and RMSEA close to .06 with a confidence interval between .05 and .1. This is why the latter cutoff values are used in this study.
Results
CFA with multi-group comparison. First, the hypothesised model has to be
estimated for both countries simultaneously as a CFA model with multi-group comparisons. Only in case this model does not yield sufficient fit, there is reason to assume that the initial eight-factor structure does not fit both countries equally well. Subsequently, the models have to be estimated and respecified separately. In the course of finding a model that fits both the samples for the US and Germany, it is important to note that respecifications have only been undertaken, when suggested for both countries. Therefore, Modification Indices (M.I.) in AMOS were consulted. For a detailed look at all respecified CFA Models, please see the Appendix.
Table 1
Model Fit Indices of the Multi-Group Respecification Process.
χ2 df p CFI RMSEA 90%CI
Original Model 4119.75 1833 < .001 .86 .060 .059, .063 Model with Error Correlations 4073.60 1828 < .001 .86 .060 .058, .062 Model with Cross-Factor Loadings 4002.14 1826 < .001 .86 .059 .057, .062 Model with Cross-Factor Error Correlations 3856.87 1814 < .001 .87 .057 .055, .060 Model without insufficiently fitting items - - - -
When estimating the original model for both countries, maximum-likelihood estimation converged to an acceptable solution (no Heywood cases), but with bad model fit: χ2 (1833) = 4119.75, p. < .001; CFI = .86, RMSEA = .060, 90%CI [.058, .063]. Table 1 features the overall results of the respecification process.
As a first measure, error correlations between indicators within factors have been added. χ2-difference test reported significantly improved model fit (χ2-diff (5) = 46.155, p. < .001). The other fit indices exhibit only slight changes with CFI (= .86) and RMSEA (= .060, 90%CI [.058, .062]). In order to further increase model fit and to avoid premature deletion of indicators for aforementioned reasons, one cross-factor loading and several cross-factor error correlations have been added in two separate steps. Adding the cross-factor loading significantly improved model fit (χ2-diff (2) = 71.456, p. < .001) and other fit indices such as CFI (= .86) and RMSEA (= .059, 90%CI [.057, .062]). The same can be said about the six error correlations across factors. Again, χ2-difference test yielded significant enhancement of model fit with χ2-diff (12) = 145.273, p. < .001. Also, values for CFI and RMSEA slightly changed (CFI = .87, RMSEA = .057, 90%CI [.055, .060]. Here it may be noted that even though M.I. suggested only very few improvements for both countries simultaneously, they did indeed feature a wide array of respecifications for either one of the two groups. All in all, M.I. do already point at the fact that it might be best to fit the models separately.
Despite the improvements, model fit is still insufficient since χ2 (1814) = 3856.869, p. < .001. As a final step, dropping indicators was considered. However, checking the two models, this does not seem an option. As for Germany, more than a third of the indicators do not fulfil the requirements mentioned above, which indicates problems with convergent validity. 22 For the US, in contrast, it is just two indicators (honest: γ* = .33 and reliable: γ* = .31)23 and these are particularly well informing the model for Germany (honest: γ* = .84 and !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
22 The severe differences between countries regarding convergent – and also discriminant – validity will be addressed more detailed below.
reliable: γ* = .73). Model fit for the final structure measuring PBP for the both countries
concurrently is χ2 (1814) = 3856.869, p. < .001; CFI = .87, RMSEA = .057, 90%CI [.055, .060]. Due to its general insufficient fit, this is why H1 can be retained: Party Brand Personality has to be measured differently for the US and Germany.
Measuring party brand personality in Germany. To answer H2 and H3, both models have be tested and respecified separately. This and the next section discuss the respecification process and justify the changes made compared to the original model. Here it is important to note that even in case of insufficient fit, indicators of the dimensions Competitiveness and Cooperativeness have not been deleted in the process of respecification. This has been done to ensure comparability across the two country-specific models in order to assess the hypotheses.
Before testing the model for Germany, reliability checks have been conducted using SPSS. Results for six of the eight dimensions were good (Competitiveness, Cronbach’s α = .78) to very good (Honesty, α = .88; Spirited, α = .82; Image, α = .88; Leadership, α = .85; Cooperativeness, α = .86). Results for Toughness (α = .53) and Uniqueness (α = .62) were substantially lower, which might be down to their comparatively low number of indicators. Nevertheless, reliability checks already pointed to issues with those two dimensions, which also have been found in the course of respecification as mentioned below.24
Table 2
Model Fit Indices of the Respecification Process for Germany.
χ2 df p CFI RMSEA 90%CI
Original Model 1990.58 917 < .001 .77 .082 .077, .087 Model with Error Correlations 1757.31 898 < .001 .82 .074 .069, .080 Model with Cross-Factor Loadings 1600.80 891 < .001 .85 .068 .062, .073 Model with Cross-Factor Error Correlations 1562.20 887 < .001 .86 .066 .061, .072 Model without insufficiently fitting items 918.24 485 < .001 .89 .072 .065, .079
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
24 Due to this paper’s theory-driven approach, no items were dropped before conducting the CFA as recommended by Kline (2010).
When estimating the model, maximum likelihood estimation converged to an acceptable solution (no Heywood cases) but with poor model fit: χ2 (917) = 1990.58, p. < .001; CFI = .77, RMSEA = .082, 90%CI [.077, .087]. Table 2 features the results of the respecification process for Germany.
Checking for convergent validity, a total of 19 indicators did not sufficiently inform the factors with standardised factor loadings below .7. As for discriminant validity, covariances between Leadership, Toughness and Competitiveness were among the highest.25 Since merging these to one factor also made sense theoretically as respondents might have well perceived indicators to measure a similar construct, they have been merged. χ2-difference test, however, proved the six-factor structure to fit the data significantly worse than the original eight-factor structure with χ2-diff (3) = 32.54, p. < .001. This is why the initial model has been retained. To improve model fit, modification indices recommended adding error correlations between indicators within factors. The respecified model fit significantly better (χ2-diff (19) = 233.27, p. < .001) and also other parameters improved as CFI = .82; RMSEA = .074, 90%CI [.069, .080].
Subsequently, cross-factor loadings have been added. Especially since reliability of the dimensions toughness and uniqueness were considerably low (see above), it might well be that those dimensions are informed by indicators of the other six factors. Or – the other way round – that these indicators load on some of the other six dimensions. For example,
outdoorsy and masculine both were linked to Image, as they theoretically fit together with
most other indicators of that dimension and were descriptive for Image in general. Also, the covariance between the respective dimensions Image and Toughness was considerably high as well. In contrast, even though recommended by M.I., cross-factor loadings from sincere on Competitiveness have not been added. First, this is because the dimensions Honesty and Competitiveness have been hypothesized to measure rather different personality traits. !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
25 Covariances between were: Leadership and Toughness (.89), Leadership and Competitiveness (.89), Toughness and Competitiveness (.65).
Second, the trait sincere for example did not match the indicators of Competitiveness theoretically. Again, model fit significantly improved with χ2-diff (7) = 156.51, p. < .001 and values for CFI and RMSEA also improved (CFI = .85; RMSEA = .068, 90%CI [.062, .073]). Apparently, several indicators inform more than one dimension/factor in the model measuring party brand personality.
Since the respecified model is not unidimensional anymore, error correlations between error terms across factors have been included as well. This is to account for indicators’ shared variance, which has not been measured in the current model. This significantly enhanced model fit (χ2-diff (4) = 38.60, p. < .001). Also, other measures for model fit improved such as CFI = .86; RMSEA = .066, 90%CI [.061, .072]. Even though the respecification process so far increased model fit significantly, several items remain to have insufficient explanatory power for the general construct. This is why in a final step, badly fitting indicators were removed.
As this paper’s goal is first to develop a scale measuring PBP in the context of a multiparty system and then compare it to the scale of a two-party system, a more parsimonious scale is to be preferred. This is why in total 11 indicators have been deleted, as they did not sufficiently load on one of the eight factors.26 After respecification, the final CFA model comprises an eight-factor structure with 34 indicators. (see Figure 2 for a simplified presentation of the final CFA model).27 Even though CFI improved, approximation indices were still not sufficient (CFI = .89; RMSEA = .072, 90%CI [.065, .079]). Most importantly, !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
26 Badly fitting items of Cooperativeness and Competitiveness are: competitive (.35), risk-taking (.30), altruistic (.33), willing to compromise (.41) and compliant (.30). As mentioned above, they were not deleted to allow for comparison regarding HII and HIII. In case indicators would have been deleted, the final model would have comprised an eight-factor structure with 29 indicators. Model fit therefore was (χ2 (342) = 603.27, p < .001); (CFI = .92,); RMSEA = .066, 90%CI [.058, .074]. Even though this model would fit significantly better, it still would not fit the data.
27 Even though not meeting the requirements for sufficient fit, the indicator unique has not been deleted. This is because factors require at least two indicators (Kline, 2010) and after deleting original, unique and independent were the last remaining two indicators for the dimension Uniqueness. Since none of the two remaining indicators loaded on one of the other seven factors in the model, I decided to maintain the factorial structure with
however, χ2 was still significant (χ2 (485) = 918.238, p < .001), which is not acceptable. This is why it has to be concluded that the respecified model still does not fit the data.
Figure 2: Simplified Model of Party Brand Personality as Respecified for Germany.
Note: Cross-Factor Loadings of Indicators are added in bold to the Dimensions they also load on.
Measuring party brand personality in the US. In the case of the US, reliability
checks reported excellent values for all eight subscales with Honesty, α = .92; Spirited, α = .95; Image, α = .96; Leadership, α = .97; Toughness, α = .92; Uniqueness, α = .92; and Competitiveness, α = .92 as well as Cooperativeness, α = .92. Nevertheless, these high values might already point at the issues with discriminant validity to be mentioned below.
Estimating the hypothesised model, maximum likelihood estimation converged to an acceptable solution (no Heywood cases) but with insufficient model fit: χ2 (917) = 2153.909,
p. < .001; CFI = .89, RMSEA = .090, 90%CI [.085, .095] (see also Table 3).
Except for honest (γ*= .36) and reliable (γ* = .33), all indicators converged on the factors they were supposed to. In contrast, discriminant validity proved to be an issue since
virtually all factors featured covariances of above .85. This is why as a first step of respecification, merging factors has been considered. In total, four combinations of factors were merged. Drawing on χ2-difference test, the models with merged factors were reported them to fit the data significantly worse, this is why the original model has been retained. In contrast, adding within-factor error correlations, significantly improved model fit (χ2-diff (15) = 257.216, p. < .001). Also, other measures for model fit slightly improved such as CFI = .91; RMSEA = .081, 90%CI [.076, .086].
Table 3
Model Fit Indices of the Respecification Process for the US.
χ2 df p CFI RMSEA 90%CI
Original Model 2153.91 917 < .001 .89 .090 .085, .095 Model with Error Correlations 1896.69 902 < .001 .91 .081 .076, .086 Model with Cross-Factor Loadings 1800.67 899 < .001 .92 .077 .072, .082 Model with Cross-Factor Error Correlations 1701.42 891 < .001 .93 .074 .068, .079 Model without insufficiently fitting items 1516.69 808 < .001 .93 .072 .067, .078
Since some items seemed to load on several factors, cross-factor loadings were included where theoretically reasonable. This holds for risk-taking loading on both the dimensions Image and Cooperativeness, as this indicator can be thought of as informing a certain image as well as someone’s way of cooperating. In contrast, risk-taking could not be thought of as informing Uniqueness. Again, χ2-difference test yielded significantly improved model fit (χ2-diff (3) = 96.022, p. < .001) with CFI = .92; RMSEA = .077, 90%CI [.072, .082]. Since principle of unidimensionality has been violated anyway, error correlations across factors were included, which also significantly improved model fit (χ2-diff (8) = 99.254, p. < .001). Other fit indices just marginally improved, such as CFI = .93; RMSEA = .074, 90%CI [.068, .079]. In the final step of respecification, two insufficiently fitting indicators were dropped, Honest and Reliable. Even though the resulting model fits better, overall, it still does not help to sufficiently inform the factorial structure for the US sample, which is underlined by rather weak values for RMSEA – despite almost sufficient CFI
figures. Model fit of the resulting factorial structure is χ2 (808) = 1516.686, p. < .001; CFI = .93, RMSEA = .072, 90%CI [.067, .078].
Having respecified the CFA models for both countries separately and looking at the fundamentally different structures for both models, this justifies the measures taken in hindsight. Apparently, the differences between the US and Germany regarding PBP are too strong to find a model that explains the concept for both countries. This is why the two processes of respecification lead me to strengthen my conclusion that H1 can be retained. Figure 3 features a simplified scheme of the model as respecified for the US.
Figure 3: Simplified Model of Party Brand Personality as Respecified for the US.
Note: Cross-Factor Loadings of Indicators are added in bold to the Dimensions they also load on.
Comparing PBP as measured for the US and Germany
To test H2 and H3, a second-order factor representing the overarching construct of PBP was introduced to the final respecified model of each country. By checking the loadings of the eight first order factors on the second-order factor, it could be quantified how well each of
them helps explaining the variance in the overarching construct. Both models with second-order factor fit worse than the respecified CFA models as values for CFI and RMSEA indicate. For Germany model fit is χ2 (505) = 1157.984, p. < .001; CFI = .83, RMSEA = .086, 90%CI [.080, .093] and for the US χ2 (828) = 1723.456, p. < .001; CFI = .92, RMSEA = .080, 90%CI [.075, .086].
As for Competitiveness, the factor loading of the given dimension in the US model (γ* = .83) clearly exceeds the factor loading of the factor Competitiveness in the German model (γ* = .62) (see Table 4).28 Consequentially, Competitiveness has higher explanatory power in the US model than in the model for Germany. However, since discriminant validity is not given for the US model due to high covariances among factors, the comparably high explanatory power might be blamed on this. Also, the factor loading for this dimension is by far the smallest out of the eight dimensions for the US. Nevertheless, results suggest that the dimension Competitiveness does inform the PBP better for the two-party system of the US than for the multiparty system of Germany. This is why H2 can be retained.
The results for Cooperativeness suggest rather contrary conclusions. Other than expected, cooperativeness did not have higher explanatory power in the context of the multiparty system (γ* = .84) compared to the two-party system (γ* = .97). This is why HIII has to be rejected. Still, it may be noted that absence of discriminant validity for the US model might be one of the reasons for the high factor score weight.
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
28 Here it might be mentioned that the factor loading of γ* = .62 from Competitiveness onto PBP is partly down to the fact that the indicator risk-taking also informs two other dimensions, which are Image (γ* = .33) and Cooperativeness (γ* = .07). In a respecified model for Germany, in which also indicators for Competitiveness and Cooperativeness had been dropped in order to find the best fitting model, Competitiveness loaded substantially lower on PBP (γ* = .40) and two of its items were dropped due to bad fit (competitive and
Table 4:
A Comparison of Factor Scores on the Second-Order Factor Party Brand Personality (PBP).
PBP dimensions USA Germany
Honesty .97 .89 Spirited .99 .95 Image .99 .84 Leadership .94 .75 Toughness .91 .77 Uniqueness .96 .79 Competitiveness .83 .62 Cooperativeness .97 .84
Note: Reported figures are standardised estimates (γ*).
Nevertheless, the results presented in Table 4 suggest two other interesting findings. First, the explanatory power of Cooperativeness in the case of Germany clearly exceeds the one of Competitiveness. Even though factor loadings are not higher than in the US as initially hypothesised, results clearly demonstrate that German respondents find Cooperativeness much more useful to describe a party’s (brand) personality than Competitiveness. The differences for the US model are certainly less striking.
Second, for both countries the Cooperativeness dimension was among the top four factors in terms of explanatory power regarding the overarching construct of PBP. This indicates that the ways party compete and interact might well help to describe their personality and subsequently also help to explain how they are perceived as brands.
The impact of party identification on PBP
Having tested the three hypotheses, it is now to briefly address the role of a party identification (PID). As Smith (2009) already stated, people do evaluate parties’ brand personality differently based on their own PID. For example, people identifying with the Republican Party might evaluate it more positive than the ones who think of themselves as Democrats and vice versa.
To address this, based on the respecified, country-specific models29, a multi-group comparison CFA was conducted to find out whether PID matters. Since respondents for both countries evaluated two parties each, two groups were composed for this analysis: Respondents who evaluated two parties not in line with their PID and respondents who evaluated “their” party together with one they did not identify with. In case the respecification process should not lead to a model satisfactorily fitting both groups, it can be deducted that PID does matter for how people evaluated other parties’ PBP. Just as above, respecifications were only conducted when recommended by M.I. for both groups.
As for Germany, Table 5 sums up the results of the respecification process. Modification Indices did neither suggest error correlations within or across factors, nor cross-factor loadings. However, five indicators did not sufficiently fit in both models and were deleted accordingly.30 Looking at χ2 and the approximate fit indices, this considerably enhanced model fit. Nevertheless, all indices indicated insufficient model fit. Apparently, both groups feature a different factorial structure, which leads to the conclusion that PID does affect how German respondents evaluated parties’ brand personality.
Table 5:
Model Fit Indices of the Multi-Group Comparison regarding PID (Germany).
χ2 df p CFI RMSEA 90%CI
Model as previously respecified 1587.35 971 < .001 .83 .061 .055, .066 Model without insufficiently fitting items 1050.11 684 < .001 .88 .056 .049, .062
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
29 Since it has already been shown that PBP has to be measured differently for the US and Germany and the fact that this is a form of a post-hoc test, I draw back on the respecified models and not the ones initially
hypothesised.
30 First mentioning values for the “evaluated own and one other party” model and then for the “evaluated two other parties” model, these indicators were friendly (γ* = .34, γ* = .37), successful (γ* = .22, γ* = .42),
compliant (γ* = .11, γ* = .42), risk-taking (γ* = .31, γ* = .33), competitive (γ* = .28, γ* = .39). Values for the
model in which respondents evaluated two parties other than their PID were mentioned second in brackets. Also, unique and masculine did not sufficiently fit but were not deleted. Since these items loaded on factors with just two indicators, they have not been deleted according to the rules defined by Kline (see above).
