Measuring Partisanship as a Social Identity
in Multi-Party Systems
Alexa Bankert, Stony Brook University
Leonie Huddy, Stony Brook University
Martin Rosema, University of Twente
An earlier version of this paper was presented at the annual meeting of the Midwest Political Science
Association, 16-19 April, 2015 Chicago, USA. We wish to thank the directors of the LISS, Swedish
Citizen Panel, and the BES for the opportunity to include items in their ongoing surveys. We owe a
considerable debt to Stanley Feldman for his insightful comments on this research.
Abstract
Partisanship remains a powerful influence on democratic political behavior (Green et al 2002; Brader
and Tucker 2009; Brader, Tucker and Duell 2013; Dalton and Weldon 2007). But there remains a
lively debate on its nature and origins: Is it largely instrumental in nature and shaped by party
performance and issues stances? Or is it basically a long-standing expressive identity held in place by
biased reasoning and strong emotion (Fernandez-Vazquez 2014; Green et al 2002; Huddy et al
2015)? The expressive approach is grounded in social identity theory and has generated a multi-item
scale of partisan identity. In this study, we assess the virtues of this scale in a European context
examining its measurement properties and predictive validity in national surveys conducted in the
Netherlands, Sweden, and the U.K. Using a latent variable model, we show that an eight-item
partisan identity scale provides greater information about partisan intensity than a standard singleitem
and has the same measurement properties across the three countries. In addition, the identity
scale better predicts in-party voting and political participation than the single partisan strength item.
It also accounts for far more variance in political outcomes than an ideological intensity measure
(based on both left-right self-placement and agreement with the party on key issues), providing
initial support for an expressive approach to partisanship in several European democracies. We
conclude with a discussion of a short four-item version of the partisan identity scale that could be
readily adopted in diverse national election surveys.
1
Partisanship remains a powerful influence on political behavior within developed and
developing democracies (Green et al 2002; Brader and Tucker 2009; Brader, Tucker and Duell 2013;
Dalton and Weldon 2007). In the U.S., partisanship has increased in strength in recent years and
continues to wield impressive influence on a range of political behavior including vote choice, voter
turnout, and electoral campaign activity (Huddy et al 2015; Nicholson 2012). In other developed
democracies, levels of partisanship have declined (Dalton 2000) but partisanship remains a potent
political force nonetheless (Bartle and Bellucci 2009; Holmberg 2007).
Instrumental and Expressive Partisanship
If the influence of partisanship on political behavior is not in dispute, there remains a lively
debate on its nature and origins. From a normative democratic perspective, the extent to which
partisanship reflects issue preferences, a reasoned and informed understanding of the parties’
positions, and is responsive to ongoing events and political leadership remain central concerns. We
refer to partisanship grounded in this type of responsive and informed deliberation as instrumental.
As a test of instrumental partisanship, researchers have examined its origins in long standing socioeconomic
cleavages and contemporary forces such as leader evaluations and issue proximity (Dalton
and Weldon 2007; Garzia 2013). Garzia (2013) reports, for example, a decline in recent decades in
the link between social cleavages and partisanship in the U.K., Germany, Italy and the Netherlands.
He also reports a corresponding increase in the link between partisanship and leader evaluations. In
another strand of work, researchers have examined the degree to which partisans are aware of
ideological shifts in a party’s platform. Adams et al (2011) find no shift in public awareness when a
party’s platform changes whereas Fernandez-Vazquez (2014) reports change but finds that it falls
well short of the full magnitude of a party’s ideological shift. Based on this accumulated research, it
appears that in various democratic nations partisanship is at least modestly responsive to
2
contemporary forces such as changing leaders and shifting party positions, lending support to an
instrumental model.
In the U.S, another approach to partisanship has gained credence. We refer to this as an
expressive identity model in which partisanship is viewed as a social identity that motivates defense of
the party in order to maintain its positive standing. Expressive partisanship generates political
activity in support of the party, rests on biased political reasoning, and leads to vilification of outparties.
Most importantly these activities and mental processes reinforce partisanship resulting in a
relatively stable political identity. Such psychological processes serve to defend a partisan identity,
making it less likely a partisan will modify their partisanship in response to a party’s shift on issues or
poor performance in office especially when such attacks emanate from partisan rivals (Greene 2002).
A partisan identity is likely to strengthen over time as a young voter consistently supports one party
over others in successive elections (Dalton and Walden 2007). The expressive approach to
partisanship is grounded in social identity theory (Green et al 2002; Huddy et al 2015).
The expressive partisanship model has generated a new multi-item measure of partisan
identity in the United States (Huddy et al 2015). From an identity standpoint, measuring gradations
in social identity strength is crucial to identifying the individuals who are most likely to engage in
expressive partisan activities such as taking action in defense of the party (Huddy 2001; 2013). As a
consequence, psychologists typically measure social identities with multiple items to create a finegrained
scale of identity strength. Several identity scales have been developed to assess partisan
identity in the U.S. Steven Greene (2002, 2004) developed a ten-item scale of partisan social identity,
based on Mael and Tetrick’s (1992) Identification with a Psychological Group Scale, which had good
measurement properties and proved to be a better predictor than the standard partisanship measure
of a range of political variables including political involvement. Huddy and colleagues (2015)
developed a four-item scale with items that tap identity importance and sense of party belonging (see
3
also Huddy and Khatib 2007). This scale is also a better predictor than the standard partisanship
measure of campaign activity and emotional reactions to partisan threat and reassurance. In the
current study, we developed the Identification with a Psychological Group (IDPG) scale created by
Mael and Tetrick (1992), and adapted by Green for partisan identity, for inclusion in national surveys
in the U.K., Sweden, and the Netherlands.
Multi-item partisan identity scales have proven to be quite effective in predicting political
outcomes in the U.S., suggesting that the traditional partisanship measure underestimates the
partisan nature of political activity. But there are reasons to question whether or not such a partisan
identity scale is needed and will be equally successful in predicting political outcomes in Europe.
Partisan identity may be more complex in Europe than in the U.S. because of the existence of multiparty
systems that could dampen loyalty to any one party (Niemi et al. 1991; Weisberg 1980).
Moreover, frequent coalitional governments aligned along ideological lines may also blur loyalty to a
single party (Hagevi 2015; Meffert et al. 2012; Gonzalez et al. 2008). To determine the value of a
partisan identity scale in Europe, we examine several key properties of the scale.
First, the measurement properties of the new multi-item scale will determine its utility
beyond the U.S. It should differentiate equally well among low, middling, and high levels of identity.
It is important for the scale to differentiate among partisans across the full range of partisan
intensity to ensure that it works in European nations that differ in levels of partisan intensity. It is
especially critical to capture good variation at low levels of partisanship in light of evidence that
partisanship has declined in Europe in recent decades (Hagevi 2015; Dalton 2000). It is also
important to determine whether the partisan identity scale has the same meaning and properties
across countries even when such countries exhibit different levels of partisanship. Ideally, the scale
can provide a standardized measure of partisan intensity to remedy past inconsistencies in the
measurement of partisanship in the European context (Johnston 2006).
4
Second, a multi-item partisan identity scale is an unneeded luxury if it has no greater
predictive validity than the existing single-item. Partisan intensity is often measured more finely in
European countries than in the U.S. because the typical partisanship question contains a slightly
greater number of points than the 2 (i.e., strong, not so strong) employed in the U.S. For this reason,
the single partisan strength item employed in European countries may detect considerable variance
in political outcomes, making the new multi-item partisan identity scale unnecessary. It is therefore,
important to demonstrate that the partisan identity scale is a better predictor than the extant singleitem
of key political variables such as in-party voting and working on the party’s behalf.
Third, it is also important to demonstrate that the partisan identity scale has theoretical
validity. To that end, it is important to assess whether it better predicts political outcomes than an
indicator of instrumental partisanship such as issue-based ideological indicators. This brings us back
to the debate over whether partisanship is largely instrumental or expressive in nature. The multiitem
partisan identity items are crafted to reflect identity and in that sense explicitly measure identity.
But instrumental factors such as the advancement or protection of ideological values or specific
policy positions may also influence the identity scale. It is thus important to differentiate between
instrumental and expressive partisanship.
Partisanship in Europe
There is an obvious need to standardize the measurement of partisanship in Europe. The
Comparative Study of Electoral Systems (CSES) asks respondents’ in numerous countries to indicate
whether they are close to a party, or closer to one than another, using a standard set of questions
(Dalton and Weldon 2007). But this approach has not been widely adopted in single-country
election surveys in Europe where the wording, as well as the specific combination of items used to
gauge partisan attachments, varies across countries. The creation of a standardized scale is obviously
complicated by language differences which make it difficult to use precisely the same wording to
5
measure partisanship and draw valid conclusions about differences in the absolute value of
partisanship in different countries (Campbell and Valen 1961). A multi-item scale may help to evade
some of these problems by relying on a number of different items to measure the same concept.
But the development of a multi-item scale confronts several measurement hurdles that need
to be crossed before a partisan identity scale can be widely adopted across different European
polities. The first hurdle is to create a scale that identifies partisan strength equally well at low and
high levels of partisan identification. This is especially important in studying the effects of
partisanship across the full range of political behavior. Strong partisanship is a likely precursor of
effortful political behaviors such as working for a candidate. In contrast, an accurate assessment of
low levels of partisanship may be needed to predict less effortful political activity such as voting for
the party of advocating that others do so. Second, the scale should have the same measurement
properties in different countries in order to identify equally well individuals with high and low levels
of partisanship in multi-nation datasets, regardless of nationality.
To gauge the properties of a partisan identity scale, we rely on latent trait models to assess
the ability of items to differentiate equally across all levels of partisan strength and to do so in the
same way across countries. We thus regard partisanship as a latent construct that cannot be directly
measured (for similar approaches, see Green, Palmquist, and Schickler 2002; Green and Palmquist
1990).
Latent Trait Models as a Scale Assessment Tool
Item Response Theory
We first draw on Item Response Theory (IRT) to analyze the measurement properties of a
partisan identity scale. In essence, the IRT model determines how well each item in a scale measures
an underlying latent construct. This is broken down into several steps. In the first stage, the IRT
model estimates an item response function for each of an item’s response options. The item
6
response function can be expressed as the probability of agreeing with a response category on a given
item and is based on both an individual’s score on the latent trait and attributes of the items used to
assess the latent trait (see Alen and Yen 1979; de Ayala 2009).
In the case of ordinal variables, a response function is calculated for each option using the
Graded Response Model (Samejima 1969). More formally, the item response function can be
expressed as:
𝐼 𝑥 𝑗
(𝜃) = {−
∂2 lnpk
∂θ2
} 𝑝 𝑘 ,
where θ denotes the latent trait, x refers to the ordinal manifest variable with j possible response
categories and Pk refers to the probability of an individual picking a particular response category. A
response function is estimated for each response option and can be plotted, graphing the underlying
latent trait on the x axis and the probability of picking that option for a specific item on the y axis.
An item’s response functions are typically plotted on the same graph to indicate how well each
response option measures the underlying latent trait. Such depictions help to identify items that are
good at detecting high but not low, low but not high, and middling levels of the trait. In a next step,
item response functions are summated to obtain the information function for each item in the scale.
The item information function (IIF) represents the information provided by a specific item x across
the range of the latent trait.
Graphing the information function is particularly useful as a way to visually examine each
item’s ability to cover the latent trait’s range while discriminating among its various levels. In general,
a scale item with narrowly peaked response functions at locations across the latent trait’s range is
highly discriminating and provides considerable information about the trait overall (de Ayala
2009:230). In essence, a peak in an item’s information function indicates that a specific level of the
underlying trait is captured with precision. Thus, polytomous items tend to provide more
information about the latent trait than dichotomous items.
7
Finally, the individual item information functions can be further summed to generate an
information function for the entire scale, indicating how well it discriminates among values of the
latent trait:
𝐼(𝜃) = ∑ 𝐼𝑗(𝜃)𝐿
𝑗=1 .
Invariance Analysis
To test the scale’s invariance we employ a series of confirmatory factor analyses (CFA) to
determine if the underlying latent trait is measured in the same way across countries. Invariance is
defined here as "whether or not, under different conditions of observing and studying phenomena,
measurement operations yield measures of the same attribute" (Horn and McArdle 1992:117).
Measurement invariance is established by comparing the fit of a series of hierarchical CFA models
with increasingly stringent equality constraints (Cheung and Rensvold 1999; Vandenberg and Lance
2000).
The most basic type of invariance, configural invariance, is established if the same items load on
the same factor across countries. Put differently, configural variance ensures that items measure the
same construct in all countries. Once configural invariance is established, the next most stringent
form of invariance is metric invariance which assumes that all items load on the same factor (as in
configural) but that their loadings are constrained to be equal in each country. Metric invariance is
established if the metric invariance model is no worse a fit to the data than the baseline model that
establishes configural invariance (Hirschfeld and Brachel 2014; Asparouhov and Muthén 2014).
Evidence for metric invariance indicates “…that the people in different nations understand the
items in the same way” (Davidov 2009: 69). While metric invariance allows researchers to compare
analytic models across countries, it does not guarantee that observed differences in the mean levels
of the scale reflect differences in the mean levels of the underlying latent trait. Thus, the final step in
a hierarchical invariance analysis involves testing for scalar invariance. In this model, factor loadings
8
and the intercepts are constrained to be equal. If the fit of the scalar variance model is not
significantly worse than the fit of the metric invariance model, there is evidence that the scale’s
scores can be compared across countries.
Hypotheses
We first examine the measurement properties of the partisan identity scale in the
Netherlands, Sweden and the U.K. and expect each scale item to provide more complete
information about partisan strength than the traditional single party identification item. In terms of
the IRT analysis, this means each item’s information function will be more peaked and contain
greater information than the standard single measure of partisan strength. In the three European
multi-party systems under study, we expect both lower and higher levels of partisan intensity to
remain less well detected when measured with the traditional item
Second, we examine the scale’s configural, metric, and scalar invariance in the three
countries. We expect the partisan identity scale to exhibit all three types of invariance, which means
that the fit of the metric invariance model will be no worse than the fit of the configural model, and
that the fit of the scalar model will be no worse than that of the metric model.
Third, we test the partisan identity scale’s predictive validity. We expect partisan identity to
more powerfully predict in-party voting and political engagement than the traditional party
identification item. We also expect the partisan identity scale to better predict political behavior than
a multi-item indicator of ideological intensity in support of an expressive model of partisanship. If
successful, this last test also helps to rule out the possibility that the identity scale has greater
predictive validity than the traditional single-item because of its better measurement properties
9
Methodology
Sample
Netherlands. We report on data from the Dutch Parliamentary 2012 Election Studies conducted with
data from the Longitudinal Internet Studies for the Social Sciences (LISS) panel. The LISS contains
4,691 households, entailing 8,000 individuals, drawn as a true probability sample of households in
the national population register maintained by Statistics Netherlands. Non-computer households are
provided with a computer and internet connection and the panel members complete monthly online
surveys and receive payment for each completed questionnaire. We examine data from 4,691
respondents interviewed in August 2012 which included the partisan identity items, after the national
election in September 2012, and again as part of a values and politics module in December 2012.1
Sweden. Swedish data were drawn from the Swedish Citizen Panel, a largely opt-in online panel run
by the Laboratory of Opinion Research (LORE) at the University of Gothenburg.2
We utilize data
from Panel 8 which was conducted between the 14th of November 2013 and 18th of December
2013, and add-on Panel 8-2 conducted between December 10, 2013 and January 7, 2014. 16,130
panelists were invited to take the Panel 8 survey and 9,279 completed it for a completion rate of
64%. The add-on Panel 8-2 was sent out to 2,000 panelists, of which 1,496 answered the survey. The
partisan identity module had a split sample and was run in two steps: 2,000 respondents in Citizen
Panel 8 and 2,000 in Citizen Panel 8-2. Of these 4,000 respondents, 2,464 completed the battery of
items.3
1 Roughly 5,000 individuals responded to each wave (5,195 for the pre, 5,225 for post, and 5,732 for the values
module). The 9/12 and 12/12 waves contained additional items needed for this analysis.
2 The sample of respondents in this wave is mixed with 70 % representing opt-in panelists and 30 % selected via
probability-based population methods. The opt-in panel was recruited through internet advertising on the websites of
newspapers, Twitter, Facebook and blogs.
3 The sample of Panel 8 is a mixed, but mainly self-recruited, sample: 70 % is opt-in, while 30 % is probability based
recruitment from population samples. The second sample was drawn using a probability-based recruitment method only.
10
United Kingdom. Data for the U.K were taken from the 2015 British Election Study (BES), an online
panel study conducted by YouGov.4
We draw on data from pre-election wave 3 of the BES,
conducted between September 19, 2014 and October 17, 2014 and pre-election wave 4, conducted
in March 2015. In total, 27,839 respondents participated in wave 3 and 6,141 were randomly
assigned to a module that included the partisan identity items. In wave 4, 16,629 respondents
participated and 3,500 of them completed the partisan identity module. The more numerous data
from wave 3 are used to examine the measurement properties of the partisan identity scale in the
U.K. and data from wave 4, which includes key political dependent variables, are used to assess its
predictive validity.
Measures
Partisanship Strength: The partisanship question was asked differently in each of the three countries,
underscoring the lack of uniformity in its assessment. In the Netherlands, respondents in the preelection
survey were first asked if they supported a party. If yes, they were asked how strongly they
supported it (very strongly, strongly, not so strongly). If they did not support a party, they were
asked if they were attracted to one and if yes how strongly (very strongly, fairly strongly, not so
strongly). In Sweden, respondents were asked if they felt close to a particular political party. If they
named a party, they were then asked if they felt very close, rather close, or not very close. U.K.
respondents were asked “Generally speaking, do you think of yourself as Labour, Conservative,
Liberal Democrat or what?” If no party was provided, respondents were asked “Do you generally
think of yourself as a little closer to one of the parties than the others? If yes, which party?”
Respondents who listed a party in response to either question were then asked “Would you call
4 Panel members are recruited from various different sources such as advertising on a wide range of website. For each
recruited panel member, YouGov records a large number of socio-demographic information in order to ensure a
nationally representative adult sample in terms of age, gender, and social class. Everyone taking part in a YouGov survey
receives a modest cash incentive for their participation.
11
yourself very strong, fairly strong, or not very strong [partisan]?” The partisan strength item thus
had at least three categories in each country.
Comparable and large numbers of respondents indicated some level of partisanship. In the
Netherlands, 90% of respondents indicated a preference for a party (supporters, attracted, had voted
for a party in the last election), in Sweden, 91% indicated that they were close to a party, and 86% of
those in the U.K. indicated a party preference.5
It is difficult to compare partisan strength with this
measure, however, because of the differing way in which it was asked and constructed in each
nation. In the U.K., partisans and those close to a party were asked partisan strength in a single
question. In contrast, the Dutch were asked a series of differing questions to get at partisan strength.
We made the traditional partisanship measure comparable across countries through the creation of a
three-item measure in Sweden and the U.K., ranging from very close/very strong to not very
close/not very strong. As seen in Table 1, these numbers are comparable indicating that a majority
or near-majority placed themselves in the middle of the strength scale. In the Netherlands, we
created a three-level partisanship measure (supporter, not a supporter but attracted to a party,
neither supporter nor attracted but had voted for the party in the last election). When constructed in
this way, the number of supporters in the Netherlands is comparable to those who said they were
very close or strong partisans in Sweden and the U.K., but overall the responses suggest lower levels
of affiliation in the Netherlands. From this three-level measure, we created a 0-1 measure of partisan
strength with 1 representing the strongest partisan. 6
5 These numbers are based on the entire sample in each country, not just those who received the partisan identity
module.
6 A five-item strength measure (very strong supporter, strong supporter, not strong supporter, adherent, voted for a
party) provides more information than the three-item strength measure on in-party voting in the Netherlands, and its
coefficient is comparable in size to that of partisan identity. But it does not account as well as partisan identity for
political participation.
12
Table 1: Partisan Strength By Country (Traditional Partisanship Measure)
Netherlands
(Supporter)
Sweden
(Close)
United
Kingdom
(Strong)
Supporter/very close/very strong 24 17 20
Attracted/rather close/fairly strong 47 63 48
Voted for Party/not very close/not
very strong
29 20 32
N 4,680 2,405 5,882
Note: Entries are the percentage of respondents who provided a party preference in each country. Numbers are
calculated for those asked the partisan identity questions. Numbers for the U.K. are based on wave 3. See text for details.
Partisan Identity Items
The partisan identity items were asked of respondents who had indicated a party preference
in response to the standard partisanship question. This resulted in 4,691 respondents who
completed the partisan identity battery in the Netherlands, and 2,464 in Sweden.7
In the BES, a
randomly selected 25% of those with a party preference were asked and 5,954 completed the
partisan identity battery in wave 3 and 24% of respondents with a party preference were randomly
assigned to the partisan identity module resulting in an effective sample of 3,500 respondents8
The partisan identity index is composed of eight items such as “I have a lot in common with other
supporters of this party” and “When I speak about this party, I usually say ‘we’ instead of ‘they.’”
Unfortunately, the response options differ between the UK (“agree-disagree”) and the Netherlands
and Sweden (frequency response format). Table 2 provides wording and responses to all 8 partisan
identity questions in each of the three countries (for wording in Dutch and Swedish see Table A1,
7 Respondents who did not indicate any party preferences were not asked the partisan identity questions. In Sweden this
included those who said they did not feel close to any party, did not want to, or did not answer the question resulting in
the omission of 254 respondents. In the Netherlands, 499 respondents were dropped because they had never voted or
left the question blank. In wave 3 of the BES, roughly 13% (3616) of respondents in the full sample did not provide a
party preference. Similarly, 13% of respondents (2234) in wave 4 of the BES did not indicate a party preference.
8
UK respondents were also asked the identity battery in wave 1 but there are fewer cases allotted to the module and we
thus rely on data from wave 3 for the IRT analysis. However, since only wave 4 contains a battery of questions regarding
political engagement, we utilize data from wave 4 for the regression analyses.
13
Online Appendix). There is considerable variance across countries in response to the partisan
identity questions (although this is also coterminous with differing response formats). Partisan
strength is highest in the U.K., followed by Sweden, and then the Netherlands. For example, when
asked if they say ‘we’ rather than ‘they’ when talking about their party, only 25% of those in the BES
strongly disagree whereas 79% of the Dutch and 65% of Swedes say they never feel this way.
When asked if they feel connected with someone who supports their party, 57% of wave 3 BES
respondents agree; 27% of Swedes and 16% of the Dutch say they feel this way always or often. We
created an additive partisan identity scale from these eight items. The scale ranges from 0 to 1 with 0
representing no party identity and 1 representing the strongest identity. All three data sets include
the complete eight item partisan index.
The Partisan Identity Scale
Scale Measurement Properties (IRT) Since each item in the partisan identity scale contains four response
categories (i.e. rarely/never, sometimes, often and always), we apply a Non- Rasch Model for
ordered polytomous data, namely the Graded Response Model (Samejima 1974)9
. Based on the
estimated item response function for each item’s response categories, we generated an information
function for each item and then for the scale as a whole within each country. Figure 1 contains the
graphical representation of each item information function by country. Graphs were created using
the ltm package (Rizopoulos 2006) in R (see Figure A1 in the Online Appendix for item response
functions).
Dichotomous items tend to generate unimodal item information functions in IRT but
polytomous items, as in this scale, tend to have multiple peaks, as seen in Figure 1. The multiple
peaks occur because adjacent response functions for a specific item are combined to form an overall
9 The Graded Response Model is an extension of the two-parameter logistic (2PL) model for graded response data in the
sense that it allows the discrimination parameter α to vary across items.
14
item information curve. The peaks in Figure 1 represent the location on the underlying partisan
identity trait at which an item provides the greatest amount of information. For example, an
information function that peaks in the middle of the latent trait suggests an item with considerable
ability to distinguish middling from higher and lower levels of identity.
15
Table 2: Partisan Identity Items by Country
Netherlands Sweden United Kingdom (Wave 3)
Always Often
Some-
times
Never Always Often
Some-
times
Never
Strongly
agree
Agree
Dis-
agree
Strongly
disagree
When I speak about this party, I
usually say “we” instead of
“they”.
2 5 14 79 5 10 20 65 7 18 50 25
I am interested in what other
people think about this party.
3 22 51 24 7 30 45 19 12 61 21 6
When people criticize this party,
it feels like a personal insult. *
1 3 23 74 1 6 33 59 5 22 50 23
I have a lot in common with
other supporters of this party.
3 26 46 25 2 33 50 14 12 65 18 5
If this party does badly in
opinion polls, my day is ruined.
1 5 29 65 1 3 20 77 2 12 55 31
When I meet someone who
supports this party, I feel
connected with this person. *
2 14 44 40 3 24 52 21 6 51 32 11
When I speak about this party, I
refer to them as “my party”. *
1 6 17 76 3 7 17 73 5 15 54 26
When people praise this party, it
makes me feel good. *
5 24 44 28 6 21 42 30 6 47 34 13
N 4,691 2,464 5,954
x̄ (St. Err) 1.66 (0.5) 1.79 (0.5) 2.30 (0.57)
Alpha 0.86 0.83 0.88
Traditional Partisan Strength
x̄ (St. Err) 1.94 (0.72) 1.97 (0.61) 1.89 (0.71)
Note: Entries are percentages. * Items included in the short four-item partisan identity scale.
16
Whereas the amount of information provided by an item represents its ability to differentiate
between partisans of different strength, the distance between each peak suggests how much of the
range of partisan identity is covered by an item. If the distance between two adjacent peaks is large,
the item provides less information about levels of partisan identity in between peaks. When the
distance is smaller and the peaks are located more closely together on the latent trait continuum, the
loss of information becomes less severe. Thus, ideal items should cover a wide range of partisan
identity strength and be able to discriminate effectively among different levels. To compare the
information provided by each item in the partisan identity scale with the traditional three-point
strength measure, we added the item information function for the latter to each figure. Figure 1,
Panels A, B, and C depict the item information function for all eight scale items and the traditional
strength measure in the Netherlands, Sweden, and U.K. respectively.
Our first hypothesis concerns the ability of the items in the partisan identity scale to provide
more information than the standard single measure about partisan strength across its full range.
Figure 1 demonstrates that the eight items supplement each other to cover a broad range of the
underlying partisan identity trait in each country. The individual items vary in amount of information
as well as in their ability to capture high or low levels of partisan identity (equivalent to an item’s
level of difficulty in IRT parlance).
17
Figure 1: Partisan Identity Item Information Functions
A. Netherlands
B. Sweden
18
C. United Kingdom
In all three countries, two items provide especially good coverage of lower levels of partisan
identity: “When I meet someone who supports this party, I feel connected”, indicated in purple, and
“When people praise this party it makes me feel good” indicated in black. A third item, “I have a lot
in common with other supporters of this party,” indicated in dark blue, also provides reasonable
information at lower levels of partisan identity. In contrast, the item “When I speak about them, I
refer to them as my party” provides good coverage of higher levels of partisan identity in all three
countries.
Combining these items into a scale helps to compensate for weaknesses in any one item. For
example, there is a large gap in the U.K. (Panel C, Figure 1) between 0 and 1 on the partisan
identification continuum for the item “When I meet someone who supports this party, I feel
connected.” This gap is covered, however, by the item “When I speak about this party, I refer to
them as ‘my party’”. The remaining four items in the scale vary in the amount of information they
provide about partisan identity. The item “I am interested in what other people think about this
party” (red line) is by far the weakest, providing little information and failing to discriminate among
19
those at low, middling or high levels of partisan identity. Thus, the combination of several items
provides the scale with the ability to cover a broad range of party identification levels while
effectively differentiating among them.
Interestingly, the information function for the weakest of the 8 items (“I am interested…”)
looks very similar to the information function for the traditional partisan strength measure, depicted
as a black broken line in Figure 1. This demonstrates that the simple three-point measure provides
little information and poorly discriminates across the range of the latent partisan identity trait, in
support of our first hypothesis. In the Netherlands (Panel A), the traditional strength measure is
characterized by a wide bell curve ranging from -2 to 2 on the latent trait continuum. This shape
suggests that the measure captures party identification strength at both the low and high end
respectively but that it does not provide a great deal of information relative to the scale items.
Partisan strength performs a little better in Sweden (Panel B) but provides similarly modest
information in the U.K. – as shown in Panel C. The single strength items performs best in Sweden
where it provides additional information on both sides of 0, suggesting that it does discriminate
between weak and strong partisans to a greater degree than in the Netherlands or the U.K. Overall,
the partisan identity scale measures partisan identity well across its range, a distinction that is
captured far more poorly by the traditional single-item of partisan strength.
Cross-National Partisan Identity Scale Invariance
We next consider the partisan identity scale’s cross-national properties in a test of our
second hypothesis, beginning with the scale’s configural invariance, the weakest level of invariance.
We then move to assess more stringent invariance tests (Davidov 2009). We expected the partisan
identity scale to exhibit all three levels of invariance We conducted a multi-group confirmatory
20
factor analysis (CFA) using the lavaan package in R.10
Lavaan provides several fit measures such as
the Chi-square statistic, the Tucker- Lewis Index (TLI), the Comparative Fit Index (CFI), as well as
gamma-hat and the Root Mean Square Error of Approximation (RMSEA) which we present for
each invariance model in Table 3.
Table 3: Scale Invariance Fit Measures
Model
Chi-
Square
Degrees
of
Freedom
TLI CFI RMSEA
Adj.
Gamma
Hat
Configural Invariance 386.119 48 0.997 0.998 0.045 1.000
Metric Invariance 542.325 62 0.996 0.997 0.048 1.000
Scalar Invariance 3566.214 92 0.983 0.981 0.105 1.000
The fit indices for the configural invariance model presented in the first row of Table 3
indicate a good model fit. The TLI, CFI are above the cutoff value of 0.95 (Hu and Bentler 1999)
and the RMSEA value is below 0.05, a threshold commonly to determine model fit (Kenny et al.
2014; Browne and Cudeck 1993). The Chi-square statistic is relatively high given the degrees of
freedom but in contrast to other fit indices the test is more sensitive to overall sample size,
differences in sample sizes between groups, non-normality and model complexity (e.g. Hu and
Bentler 1999, 1998; Bentler and Bonnet 1980). Hence, we rely on the alternative fit indices to
evaluate the model performance. From this vantage point, we cannot reject the configural model
and conclude that the partisan identity scale measures the same construct in the Swedish, Dutch and
British sample with all items loading on one factor in each country.
The fit indices for metric invariance are presented in the second row of Table 3. In this
model, factor loadings for all items are constrained to be equal across countries. Once again, we
10 Due to similar wording of some items, we allowed the error terms of certain pairs of items to covary. These pairs are:
"When I speak about this party, I usually say ‘we’ instead of ‘they’” and "When I speak about this party, I refer to them
as "my party."; "When people criticize this party, it feels like a personal insult" and "When people praise this party, it
makes me feel good” as well as "I have a lot in common with other supporters of this party" and "When I meet
someone who supports this party, I feel connected with this person."
21
cannot reject the metric model because there are minimal changes in the fit indices between it and
the configural model (e.g. ΔCFI = 0.001; ΔTLI= -0.001). In fact, the RMSEA increases by as little as
0.003 and the Gamma-Hat remains constant. The Chi-square difference again indicates a significant
increase in this model (p < .05) but, as noted, large sample sizes can generate large Chi-square
differences. Thus, we do not apply the Chi-square difference test as a measure of fit (Cheung and
Rensvold 2002; Dovidov 2009). Table 4 summarizes the invariant factor loadings for the Dutch,
Swedish and British samples. All factor loadings are substantial and significant. These results suggest
that the partisan identity scale has the same metric across countries. In other words, a unit increase
in the partisan identity scale means the same thing in the Netherlands, Sweden and the U.K.
Table 4: Metric Invariance in the Partisan Identity Scale
Scale Item
Partisan Identity
Factor Loading
When I speak about this party, I usually say “we” instead of
“they”.
1.000
I am interested in what other people think about this party. 0.967
When people criticize this party, it feels like a personal insult. 1.117
I have a lot in common with other supporters of this party. 1.102
If this party does badly in opinion polls, my day is ruined. 1.056
When I meet someone who supports this party, I feel
connected.
1.309
When I speak about this party, I refer to them as “my party”. 1.231
When people praise this party, it makes me feel good. 1.368
Note: Entries are unstandardized factor loadings in the metric invariance model in which loadings are constrained to be
equivalent across countries. All coefficients are significantly different from 0 (p<.01).
Finally, we test the most stringent model: scalar invariance. In this CFA model, the
intercepts of the eight scale items are constrained to be equal across countries (in addition to prior
constraints placed on the factor loadings). This model tests whether a given observed value of the
partisan identity scale indicates the same level of the latent partisan identity trait in each country. In
other words, if the scale is arrayed from 0 to 1, zero would mean the complete absence of identity
and 1 would indicate maximum identity strength across countries. Fit indices presented in Table 3
22
provide mixed evidence on this point. The CFI, TFI, and the Gamma-Hat values indicate a good
fit. The RMSEA value, however, is above the cutoff point of 0.05.
It is difficult to advocate cross-national use of the partisan identity scale in the absence of
metric invariance and we turn to an alternative test. Following Oberski (2014), we examine
invariance sensitivity, “…the likely impact of measurement differences on substantive comparisons
of interest” (Oberski 2014:3). Sensitivity analyses are used to supplement the results of traditional
invariance tests which often rely on arbitrary cutoff lines (for a thorough critique of cutoff values for
fit measures, see Barrett 2007). We thus compute the EPC-interest which is a measure of the
expected change in the parameter of interest, partisan identity in this case, when freeing a particular
equality constraint. With the EPC-interest we can evaluate whether it is feasible to compare partisan
identity means across countries by estimating the change in partisan identity if certain invariance
restrictions (such as equivalence constraints on a scale item’s intercept) are removed. Put differently,
the EPC-interest evaluates directly whether a violation of measurement invariance also leads to
biased estimates of partisan identity in different countries (Oberski 2014).
In the previous invariance analyses, RMSEA was above the cutoff point indicating that
the scalar invariance model was a poor fit to the data. Table 5 shows the EPC-interest values that
exceeded .01 when the scalar invariance restrictions of equal factor loadings and intercepts are
relaxed. As can be seen in Table 5, very few items shift in terms of the mean value of the latent
partisan identity trait. Overall, dropping the equivalence restrictions on item 1 (“When I talk about
this party, I say ‘we’ instead of ‘they’.”) in the U.K. and the Sweden increases very slightly the mean
value of latent partisan identity in the Netherlands and Sweden. Dropping the intercept equivalence
restrictions for item 8 (“When people praise this party, it makes me feel good.”) in the U.K.
decreases the latent partisan identity score slightly in Sweden and the Netherlands, whereas dropping
the same restriction in Sweden leads to a slight decrease in latent partisan identity in Sweden and a
23
slight increase in the Netherlands. Finally dropping the intercept equivalence restriction in the
Netherlands for item 6 (“When I meet someone who supports this party, I feel connected.”) slightly
decreased the mean value of partisan identity in Netherlands. Overall, these changes are minor in
magnitude with all EPC-interest values at or below 0.028 in absolute value. This number is
significantly smaller than the latent mean differences of partisan identity across countries (shown in
Figure A2 in the Online Appendix) indicating that substantive conclusions regarding the comparison
of partisan identity across countries are not changed by potential model misspecifications. Even
when the requirements of the metric invariance model are relaxed the magnitude of partisan identity
remains relatively constant. Overall, these results provide evidence for our claim that the partisan
identity scale exhibits features of scalar invariance, in addition to configural and metric invariance.
This means that the partisan identity scale works similarly in all three countries.
Table 5: Changes in Mean Level of Latent Partisan Identity
Equivalence Assumption Dropped For:
Item 1 in
UK sample
Item 1 in
Sweden
sample
Item 8 in
UK sample
Item 8 in
Sweden
sample
Item 6 in
Netherlands
sample
Sweden 0.025 0.028 -0.018 -0.015 --
Netherlands 0.019 0.017 -0.020 0.003 -0.013
Note: Entries are EPC-interest values, or expected change in the mean value of partisan identity in each country, as the
equivalence constraint on the listed item’s intercept and factor loading is removed. The partisan identity scale ranges
from 0 to 1.
Partisan Identity and Political Behavior
Our third hypothesis is that partisan identity better predicts political behavior than the
traditional single-item measure of partisan strength and a multi-item scale of ideological strength in
the three countries under study. To determine the identity scale’s predictive validity, we examine inparty
voting, and political participation. In-party voting is an especially tough dependent variable on
which to determine the influence of partisan identity because it is a binary choice that may be well
explained by simple party preference.
24
Measures
In-Party Voting. Surveys were conducted at different stages of the electoral cycle in each country. In
the Netherlands, respondents were asked if they had “voted in the most recent parliamentary
elections held on September 20, 2012.” 85% of Dutch respondents said they had. They were then
asked which party they had voted for and were scored 1 if they voted for their party. In Sweden and
the U.K., respondents were asked prospective questions about their likely vote. In Sweden,
respondents were asked “How likely are you to vote in next year’s Swedish election?” and regardless
of whether they intended to vote or not were asked “Which party do you intend to vote for in next
year’s Swedish national election?” 91% said they were very likely to vote. Those intending to vote
for their party were coded 1 on the in-party vote variable. In the U.K., respondents were asked “And
if there were a U.K. General Election tomorrow, which party would you vote for?” In-party vote
was coded 1 for those who indicated that they would vote for their party. In-party voting was
highest in Sweden (88%), intermediate in the UK (76%)11
, and lowest in the Netherlands (61%).
Political Participation. Political participation was gauged differently in each country. In the
Netherlands, respondents were asked if they had raised a political issue or influenced politicians or
government over the past five years in one of four ways: (1) by making use of a political party or
organization, (2) through participation in a government-organized public hearing, discussion or
citizen participation meeting, (3) contacting a politician or civil servant, or (4) participating in a
political discussion or campaign by Internet, e-mail or SMS. These questions were included in the
post-election values module administered in late 2012 and early 2013. 12% had performed at least
one activity. A scale was created by adding all four items and rescaling form 0 to 1.
In Sweden, only a subset of respondents was randomly assigned to receive the political
participation questions. In total, 915 Swedish respondents answered four questions concerning
11
These percentages remain stable across wave 3 and wave 4.
25
whether they had ever undertaken certain political actions: (1) contacted a politician, (2) given/raised
money to/for a political organization, (3) contacted a civil servant, or (4) attended a political rally.
All four questions were asked in each of the three survey waves. All 12 items (4 items in 3 waves)
were additively combined and rescaled on a 0 to 1 scale.
In the U.K., all respondents in wave 4 were asked whether in the last four weeks they visited
the website of a candidate or party and whether they signed up or officially registered online to help
a party or candidate in their campaign and if so, which party. Both items were coded 1 if the
respondent (1) visited their in-party’s website and (2) signed up online to help their in-party. Wave 4
also included questions on whether the respondent had read or found information in the last four
weeks tweeted by (1) political parties or candidates, (2) a personal acquaintance, or (3) others, such
as a commentator, journalist, or activist. The same three questions were repeated for election
information obtained on Facebook. Five additional questions asked whether the respondent has
shared political information through (1) Facebook, (2) Twitter, (3) email, (4) instant messaging or (5)
another website or online platform. All 13 questions were combined additively and rescaled on a 0
to 1 scale. These questions clearly reflect a heterogeneous set of activities and time frames across
countries. Data from all three countries is combined in the following analysis and dummy variables
for country are included as a control for such differences in the nature of political participation. We
also analyze each country’s data separately (shown in the Online Appendix) and report any observed
differences in the text.
Ideological Intensity. We constructed an ideological intensity scale that reflects the respondent’s
ideological strength and alignment with the in-party so that higher values reflect ideological intensity
and consistency with the in-party’s ideological stance. In the U.K., this scale consists of 5 left-right
values such as “Government should redistribute incomes” and “Management will always try to get
the better of employees” (all items are listed in Table A2 in the Online Appendix). All five items
26
were combined on a -.5 to +.5 scale, with -.5 representing the most right-leaning position and .5
representing the most left-leaning position, and then folded around 0. Respondents whose overall
score was at odds with their party preference were given a score of 0. Thus, a Labour supporter who
strongly agreed with all five left-leaning items received an ideological intensity score of 1. A Labour
supporter who disagreed on average with the left-leaning items received an ideological intensity
score of 0.
In Sweden, the ideological intensity scale was created from five ideologically tinged issues,
including the reduction of the public sector, lowering taxes, and increasing unemployment benefits.
The same steps were followed as in the U.K. The items were combined to reflect a left-right
dimension and then folded around 0 to reflect intensity. Once again, those whose average score
conflicted with their party’s left-right stance received a score of 0. In the Netherlands, we did not
have a multi-item ideological scale and used the respondent’s self-placement on a left-right
dimension to assess ideological intensity. Those whose left-right placement conflicted with that of
their party received a score of 0.
Determinants of In-Party Voting
The third hypothesis predicts stronger effects of partisan identity than the traditional
measure or ideological intensity on political behaviour. We begin by analysing the determinants of
in-party voting in all three countries combined, consistent with the partisan identity scale’s
demonstrated measurement invariance, using a logistic regression model. The results are shown in
Table 6. Similar results are obtained when the data are analysed within each country (Table A5,
Online Appendix).
The first column in Table 6 estimates an equation in which in-party voting is regressed on
partisan identity; the second column contains the same analysis replacing partisan identity with the
single-item measure of partisan strength. Partisan identity has a larger effect than partisan strength
27
on in-party voting although both coefficients are significant and both forms of partisanship
dramatically increase the likelihood of voting for one’s party. As we had noted at the outset, the
binary nature of vote choice makes it a difficult test of the greater political effects of partisan
identity. .Nonetheless, when the predicted values of in-party voting are plotted across the range of
the partisan identity scale for each country, with other values in the equation in Table 6 set at their
mean, it is apparent that in-party voting better predicts voting for the party than partisan strength.
Table 6: Determinants of In-Party Voting and Political Participation
In-Party Vote Political Participation
1 2 3 4
Party Identity 2.91 (0.22) --- 0.28 (0.01) --
Partisan strength --- 2.04 (0.12) --- 0.12 (0.00)
Ideological intensity 0.21 (0.08) 0.08 (0.08) 0.06 (0.00) 0.07 (0.00)
Country
Netherlands 0.00 (0.11) -0.52 (0.08) 0.07 (0.00) 0.01 (0.00)
Sweden 1.87 (0.21) 1.50 (0.18) 0.38 (0.01) 0.31 (0.01)
Country X Party Identity
Netherlands X Party Identity 0.26 (0.32) --- -0.11 (0.02) --
Sweden X Party Identity 0.36 (0.76) --- 0.11 (0.04) --
Country X Partisan Strength
Netherlands X Partisan Str. --- -0.14 (0.16) --- -0.05 (0.01)
Sweden X Partisan Strength --- -0.17 (0.37) --- 0.09 (0.02)
Education 0.11 (0.09) 0.03 (0.09) 0.09 (0.00) 0.09 (0.00)
Gender (Female) 0.00 (0.05) 0.04 (0.05) -0.02 (0.00) -0.02 (0.00)
Age 0.03 (0.01) 0.02 (0.01) -0.00 (0.00) -0.00 (0.00)
Employed 0.15 (0.05) 0.13 (0.06) -0.00 (0.00) -0.00 (0.00)
Constant -0.58 (0.17) -0.11 (0.16) -0.04 (0.01) 0.01 (0.01)
Pseudo R- squared/ R- squared 0.10 0.12 0.33 0.32
N 8,459 8,459 7,962 7,962
Note: Logistic regression was used to analyze in-party vote; political participation is analyzed with OLS regression. The
U.K. is the omitted country in columns 1 and 2. Bolded coefficients are significant at p < 0.05 in a two- tailed test. All
variables are scaled between 0 and 1 for ease of interpretation, except for age which is measured in decades.
As seen in Figure 4, the probability of voting for one’s party ranges from a low of
approximately .45 in the Netherlands and .5 in the U.K. at the lowest levels of partisan identity to a
high of .9 for those at the highest levels. This means that someone with the highest level of partisan
identity is almost certain to vote for their party. In contrast, the probability of in-party voting
changes somewhat less dramatically across the range of partisan strength, ranging from a low of .4 to
28
a high of .75 across the range of strength in the Netherlands, and just under .6 to above.8 in the
U.K. The effect of both partisan strength and identity on in-party voting is reduced in Sweden
where voter partisan loyalty was high across the board, as seen in the top panel of Figure 4. There is
one other noteworthy aspect of the trends depicted in Figure 4. Voting for one’s party increases
rapidly at lower levels of partisan identity and tends to decelerates at higher levels. This trend is less
apparent for partisan strength, suggesting that the partisan identity scale’s ability to better detect
partisanship across its range improves its predictive validity.
Figure 4: Predicted Probability of In-Party Voting
Figure 5: Predicted Level of Political Participation
0
.2.4.6.8
Predictions
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1
Party Identity
UK Netherlands
Sweden
by Party Identity
Predicted Levels of Political Participation
0
.2.4.6.8
Predictions
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1
Partisan Strength
UK Netherlands
Sweden
by Partisan Strength
Predicted Levels of Political Participation
29
Holding a strong ideological position that is consistent with one’s party also significantly
increases in-party voting, as seen in Column 1 of Table 6, although this effect evaporates in column
2 when paired in the equation with partisan strength. This suggests that the political effects of
partisan strength but not partisan identity overlap with that of ideological intensity. In that sense,
partisan identity may provide a conceptually cleaner, less instrumental measure of partisanship than
the traditional partisan strength item. The effects of ideological intensity are also far weaker than
those of partisan identity: The probability of in-party voting changes only from .75 (0.01) to .78
(0.01) across the range of ideological intensity. In contrast, the probability of voting for one’s party
ranges from a low of .54 (0.01) at the lowest levels of partisan identity to a high of .96 (0.00) for
those at the highest levels (predicted probabilities of in-party vote by country are shown in Table A7
in the Online Appendix). Overall, these results provide evidence for an expressive approach to
partisanship in several European contexts.
Determinants of Political Participation
The analysis of political participation unfolds in parallel to that of in-party voting, although
analyses are estimated using ordinary least squares regression, making it possible to directly compare
the size of parameter coefficients. As seen in columns 3 and 4 of Table 6, both partisan identity and
strength are significant predictors of political engagement, although the effects of partisan identity
are roughly twice as large as those of strength. Predicted levels of participation across the range of
party identity and partisan strength (based on equations in columns 3 and 4) are shown in Figure 5.
This figure makes clear the greater power of partisan identity than partisan strength to drive political
participation. The effects of party identity and strength also differ by country. A negative interaction
between the Netherlands and party identity and a positive interaction between Sweden and party
identity, as seen in column 3 of Table 6, indicates the greater power of partisan identity in Sweden
30
and reduced effects in the Netherlands12
. Similar interactions exist between the two countries and
partisan strength. These trends are depicted in Figure 5, showing the dramatic effect of partisan
identity on participation in Sweden where participation goes from .37 at the lowest level of partisan
identity to .75 at the highest. In contrast, the effects of partisan strength in Sweden are more muted
with participation ranging from .36 to .57 across the range of partisan strength. Figure 5 also makes
clear that the differing time frames within which political activity was assessed (“ever” in Sweden
compared to shorter time frames in the U.K. and Netherlands) affected mean levels of reported
activity, not surprisingly.
Ideological intensity has a small significant effect on participation but its coefficient (.06) is
far smaller than the coefficients for the two partisan variables (.28 for identity and .12 for strength).
As noted, ideological intensity is better measured by multi-item issue/value scales in the U.K. and
Sweden than in the Netherlands where it was based on a single self-placement left-right intensity
measure. This may have led to an underestimation of the effects of ideological intensity in the
combined analyses shown in Table 6 because the estimated effects of ideological intensity are greater
in Sweden and the U.K. than in the Netherlands (Table A5, Online Appendix). Nonetheless, the
coefficient for partisan identity is more than twice the size of that for ideological intensity in the
Netherlands and almost 3.5 times as large in Sweden. These findings provides ample support for
hypothesis 3 and the expressive partisan model.
Finally, to underscore the additional predictive power of the partisan identity measure, we
regressed in-party voting and participation on partisan identity, country, the interaction between
country and partisan identity and basic demographics at each level of partisan strength (i.e., “not so
strong”, “fairly strong” and “very strong”). Partisan identity predicted in-party vote and participation
12 We re-ran the analysis of political participation for the Netherlands, using an ordered probit model since the political
participation variable has only 5 points in the Dutch data set. Regardless of the model specification, however, the results
remain robust.
31
at each level of partisan strength providing additional information about the dependent variable
(analyses shown in Table A6 in the Online Appendix). Among weak, moderate, and strong partisans,
partisan identity has a significant positive effect on in-party voting and political participation.
Conclusion and Discussion
In this paper, we revisit the measurement and conception of partisanship, examining the
properties of a multi-item scale of partisan identity in three European nations. While Greene (1999,
2002, 2004) and others (Green et al 2002; Huddy et al 2015) have argued for a multi-item approach
to the measurement of partisanship, most national election studies continue to rely on a single
measure of partisan intensity. The greater predictive validity of the identity scale, especially in
accounting for political participation but also in-party voting, provides a powerful rationale for the
broad adoption of a multi-item scale of partisan identity. The ability of the partisan identity scale to
better detect than the traditional item weak, middling, and strong levels of partisanship also provides
a strong reason for its inclusion in political behavior research. The scale may prove to be especially
useful in experimental studies in which many experimental treatments depend on the strength of
partisan identity.
Many surveys have space limitations and may lack sufficient space for an eight-item scale.
We examined a shorter four-item scale that performed especially well in differentiating across the
range of partisan identity. The four items were chosen based on the information they conveyed
about partisan identity in the IRT analyses. We re-ran all analyses presented in this manuscript with
the shorter scale comprised of these items (marked with an asterisk in Table 1). These results are
shown in Tables A7-A9 (Online Appendix) and confirm that the predictive power of the partisan
identity scale does not depend solely on its finely differentiated measurement. The short scale attains
results very similar to those shown in Table 6.
32
In sum, our results underscore the value of standardizing the measurement of partisanship in
diverse polities. This standardization already occurs within the CSES in which respondents are asked
the following question regardless of country: “Do you usually think of yourself as close to any
particular party?” If the answer is yes, respondents are asked “Which party do you feel closest to?,”
and then “Do you feel very close to this party, somewhat close, or not very close?” If the answer to
the closeness question is no, respondents are asked “Do you feel yourself a little closer to one of the
political parties than the others?” In most countries in the CSES, this series of questions garners a
majority of respondents who feel at least closer to one party than another, although there are a few
countries in which this does not occur. To implement the partisan identity scale, we recommend
employing the CSES questions to identify those with a party affiliation and then follow up with the
eight (or four) partisan identity questions. In the current research, party affiliation was especially low
in the Netherlands where respondents were asked about being a party supporter, adherent, or having
voted for a party in order to obtain a respondent’s party affiliation. But even in the Netherlands, the
close and closer questions in the 2010 CSES identified 70% of the electorate as having a party
affiliation. The social identity approach to partisanship holds considerable potential for the study of
comparative politics more broadly. In the U.S., a similar partisan identity scale predicts political
behavior in a political system that is dominated by two parties and characterized by strong levels of
partisanship (Huddy et al 2015). In the three European countries studied in this research, the
partisan identity scale also predicted political behavior despite varied levels of national partisan
strength and number of political parties. We have focused to a considerable degree in this research
on the promising measurement properties of the partisan identity scale. But the study also hints at
its conceptual heft. Partisan identity was a far better predictor of political behavior than a measure of
ideological intensity, at odds with an instrumental model of partisanship. Partisan identity was also
an especially good predictor of political activity in support of the expressive approach.
33
In conclusion, we urge the adoption of a multi-item partisan identity scale in political
behavior research to better assess the degree to which citizens are attached to their political party.
This is especially important at a time of declining partisanship in Europe. We have demonstrated
that strong identifiers remain loyal to their party and continue to support it electorally. At the highest
levels of partisan identity, this electoral support is almost universal regardless of country. But it is
also the case, that in-party voting was much reduced at lower levels of identity. It is thus important
to track partisan identity over time with the best measures available in order to understand why and
how it changes in response to political events. The accurate measurement of partisanship is also
important from a normative democratic perspective. In the current research, strong partisan
identities increased political activity in all three countries. If partisanship is part of the glue that ties
citizens to their electoral system, the decline of partisanship is cause for some considerable concern
and a worthy topic of further investigation. In the end, a robust measure of partisan identity that is
identical across countries may prove to be an important theoretical and empirical addition to
research on the study of partisanship and political engagement.
References
Adams, James, Lawrence Ezrow, and Zeynep Somer‐Topcu. (2011). “Is Anybody Listening?
Evidence That Voters Do Not Respond to European Parties’ Policy Statements During
Elections.” American Journal of Political Science 55 (2): 370–382. doi:10.1111/j.1540‐
5907.2010.00489.x.
Allen, Mary J., and Wendy M. Yen. (2001). Introduction to measurement theory. Waveland Press.
Asparouhov, T., & Muthén, B. (2014). Multiple-group factor analysis alignment. Structural Equation
Modeling: A Multidisciplinary Journal, 21(4), 495-508.
Bartle, J., & Bellucci, P. (Eds.). (2014). Political parties and partisanship: Social identity and
individual attitudes. Routledge.
34
Bentler, P. M., & Bonnet, D. G. (1980). Significance tests and goodness-of-fit in the analysis of
covariance structures. Psychological Bulletin, 88, 588–606.
Brader, Ted A., and Joshua A. Tucker. (2009). "What's left behind when the party's over: survey
experiments on the effects of partisan cues in Putin's Russia." Politics & Policy 37.4: 843-
868.
Brader, Ted, Joshua A. Tucker, and Dominik Duell. (2012). “Which Parties Can Lead Opinion?
Experimental Evidence on Partisan Cue Taking in Multiparty Democracies.” Comparative
Political Studies (October 14).
Browne, M. W., & Cudeck, R. (1993). Alternative ways of assessing model fit. Sage Focus Editions,
154, 136-136.
Campbell, Angus, and Henry Valen. (1966). "Party Identification in Norway and the United States."
In Elections and the Political Order, ed. Angus Campbell, Philip Converse, Warren Miller,
and Donald Stokes. New York: Wiley.
Cheung, G. W., & Rensvold, R. B. (2002). Evaluating goodness-of-fit indexes for testing
measurement invariance. Structural equation modeling, 9(2), 233-255.
Cheung, G. W., & Rensvold, R. B. (1999). Testing factorial invariance across groups: A
reconceptualization and proposed new method. Journal of management, 25(1), 1-27.
Dalton, R. J., & Wattenberg, M. P. (Eds.). (2000). Parties Without Partisans: Political Change in
Advanced Industrial Democracies. Comparative Politics. Oxford Scholarship Online.
Dalton, Russell J., and Steven Weldon. (2007)."Partisanship and party system institutionalization."
Party Politics 13.2: 179-196.
Davidov, E. (2009). Measurement equivalence of nationalism and constructive patriotism in the
ISSP: 34 countries in a comparative perspective. Political Analysis, 17(1).
De Ayala, Rafael Jaime. (2013).Theory and practice of item response theory. Guilford Publications.
35
Fernandez-Vazquez, Pablo. (2014). "And Yet It Moves: The Effect of Election Platforms on Party
Policy Images." Comparative Political Studies 47.14: 1919-1944.
Garzia, D. (2013). Changing parties, changing partisans: the personalization of partisan attachments
in Western Europe. Political Psychology, 34(1), 67.
González, R., Manzi, J., Saiz, J. L., Brewer, M., Tezanos‐Pinto, D., Torres, D., & Aldunate, N.
(2008). Interparty attitudes in Chile: Coalitions as superordinate social identities. Political
Psychology, 29(1), 93-118.
Green, D. P., & Palmquist, B. (1990). Of artifacts and partisan instability. American Journal of
Political Science, 872-902.
Green, Donald, Bradley Palmquist, and Eric Schickler. (2002). Partisan Hearts and Minds: Political
Parties and the Social Identity of Voters. Yale University Press.
Greene, Steven. (1999). “Understanding Party Identification: A Social Identity Approach”, Political
Psychology 20: 393‐403.
Greene, Steven. (2002). “The Social‐Psychological Measurement of Partisanship”, Political Behavior
24(3):171‐197.
Greene, Steven. (2004). “Social Identity Theory and Political Identification”, Social Science
Quarterly 85 (1): 138‐153.
Hagevi, Magnus. (2015). "Bloc Identification in Multi-Party Systems: The Case of the Swedish TwoBloc
System." West European Politics 38.1: 73-92.
Hirschfeld, Gerrit, and Ruth von Brachel. (2014) "Multiple-Group confirmatory factor analysis in
R–A tutorial in measurement invariance with continuous and ordinal indicators." Practical
Assessment, Research & Evaluation 19.7: 2.
Holmberg, S. (1994). Party identification compared across the Atlantic. Elections at home and
abroad: Essays in honor of Warren E. Miller, 93-121.
36
Horn, J. L., & McArdle, J. J. (1992). A practical and theoretical guide to measurement invariance in
aging research. Experimental Aging Research, 18, 117–144
Hu, L. T., & Bentler, P. M. (1999). Cutoff criteria for fit indexes in covariance structure analysis:
Conventional criteria versus new alternatives. Structural equation modeling: a
multidisciplinary journal, 6(1), 1-55.
Hu, L. T., & Bentler, P. M. (1998). Fit indices in covariance structure modeling: Sensitivity to
underparameterized model misspecification. Psychological methods, 3(4), 424.
Huddy, Leonie. (2001). “From Social to Political Identity: A Critical Examination of Social Identity
Theory.” Political Psychology. 22: 127‐156.
Huddy, Leonie. (2013). “From Group Identity to Political Commitment and Cohesion.” In Oxford
Handbook of Political Psychology, eds. Leonie Huddy, David O. Sears, and Robert Jervis.
New York: Oxford University Press, 737–73.
Huddy, L., & Khatib, N. (2007). American patriotism, national identity, and political involvement.
American Journal of Political Science, 51(1), 63-77.
Huddy, L., Mason, L., & Aarøe, L. (2015). Expressive partisanship: Campaign involvement, political
emotion, and partisan identity. American Political Science Review, 109(01), 1-17.
Johnston, Richard. (2006). Party Identification: Unmoved Mover or Sum of Preferences?” Annual
Review of Political Science 9 (1): 329–351. doi:10.1146/annurev.polisci.9.062404.170523.
Kenny, D. A., Kaniskan, B., & McCoach, D. B. (2014). The performance of RMSEA in models with
small degrees of freedom. Sociological Methods & Research, 0049124114543236.
Mael, F. A., & Tetrick, L. E. (1992). Identifying organizational identification. Educational and
psychological measurement, 52(4), 813-824.
37
Meffert, M. F., Gschwend, T., & Schütze, N. (2009). Coalition preferences in multiparty systems.
Presented at Annual Conference of the International Society of Political Psychology, Dublin,
Ireland.
Nicholson, Stephen P. (2012). “Polarizing Cues”, American Journal of Political Science 56(1): 52‐
66.
Niemi, R. G., & Jennings, M. K. (1991). Issues and inheritance in the formation of party
identification. American Journal of Political Science, 970-988.
Netemeyer, Richard G., William O. Bearden, and Subhash Sharma (2003), Scaling Procedures: Issues
and Applications, Newbury Park, CA: Sage Publications, Inc.
Oberski, D. L. (2014). Evaluating sensitivity of parameters of interest to measurement invariance in
latent variable models. Political Analysis, 22(1), 45-60.
Rizopoulos, Dimitris. (2006). "ltm: An R package for latent variable modeling and item response
theory analyses." Journal of Statistical Software 17.5: 1-25.
Samejima, F. (1969). Estimation of latent ability using a response pattern of graded scores.
Psychometrika monograph supplement.
Samejima, F. (1974). Normal ogive model on the continuous response level in the multidimensional
latent space. Psychometrika, 39(1), 111-121.
Vandenberg, R. J., & Lance, C. E. (2000). A review and synthesis of the measurement invariance
literature: Suggestions, practices, and recommendations for organizational research.
Organizational research methods, 3(1), 4-70.
Weisberg, H. F. (1980). A multidimensional conceptualization of party identification. Political
Behavior, 2(1), 33-60.