Analyzing Local Party Manifestos in Multi-Level Democracies

Full text

ANALYZING LOCAL PARTY MANIFESTOS IN
MULTI-LEVEL DEMOCRACIES?
martin gross†& michael jankowski¶
August 9,2017
abstract
Extracting party positions from election manifestos is a well-established procedure in
comparative political science research. The potential of analyzing local party man-
ifestos, however, still needs to be investigated. In this article, we show that local
manifestos are a fruitful source for research on party competition, but that these doc-
uments also impose certain methodological challenges. We discuss these challenges
and demonstrate by analyzing more than 400 local party manifestos in Germany that
these documents can be used to estimate party positions. We find, however, sub-
stantive differences in the results produced by different methods. While Wordscores
extracts meaningful ideological positions from local manifestos, Wordfish highlights
that manifestos also differ with regard to the extent national policies are addressed at
the local level.
?Authors are listed alphabetically. Both authors contributed equally to all work.
†Mannheim Centre for European Social Research (MZES), University of Mannheim,
martin.gross@mzes.uni-mannheim.de.
¶University of Oldenburg, michael.jankowski@uol.de.
1

introduction 2
1 introduction
Party manifestos constitute a fruitful resource for political science research as they
comprise the most important policy positions of party supporters and party elites
on a broad range of issues. Consequently, it is only natural that scholars in vari-
ous sub-fields of political science frequently rely upon information extracted from
party manifestos. Most commonly, the information contained in manifestos is used
to position parties along political dimensions such as the left-right-scale. Approaches
for extracting party positions from manifestos range from hand-coding (as, e.g., in
the CMP/MARPOR project; Budge et al.,2001;Klingemann et al.,2006;Merz, Regel
and Lewandowski,2016) over semi-supervised content analysis (such as ‘Wordscores’;
Laver, Benoit and Garry 2003;Lowe 2008) to fully unsupervised methods (most promi-
nently: ‘Wordfish’; Slapin and Proksch 2008).1
As interesting as national party manifestos are, their characteristics can also be un-
satisfactory for the purpose of estimating party positions. Harmel et al. (2016,1), for
example, argue that manifestos are “written for two audiences: one internal and the
other external”. As a consequence, some scholars have argued that party manifestos
often mask intra-party heterogeneity (e.g., Jankowski, Schneider and Tepe,2016,1).
Moreover, Bräuninger and Giger (2016) demonstrate that parties have incentives to
remain ambiguous in their manifestos in order to attract different voters (see also Lo,
Proksch and Slapin,2014). Existing research has responded to this problem of explor-
ing the degree of intra-party heterogeneity by, for instance, estimating the position of
regional parties. Bräuninger and Debus (2012), for example, rely on party manifestos
from the German Länder to estimate the position of parties at the regional level. Giger,
Müller and Debus (2011) use a similar approach for analyzing party positions in Swiss
1See Grimmer and Stewart (2013), Laver (2014)orWilkerson and Casas (2017), among others, for a
discussion and comparison of these techniques.

introduction 3
cantons. This research shows that parties are indeed heterogeneous and that positions
vary systematically with certain socio-economic characteristics (see also Müller 2009,
2013 for the German case). Moreover, Clark and Bennie (2016) highlight how parties
strategically adjust their manifestos at the sub-national level in multi-level systems.
Yet, despite this research, there is only a very small number of studies which have
tried to use local party manifestos as a resource for political science research (Aga-
soster,2001;Debus and Gross,2016). Furthermore, this lack of information about the
positions of local parties is problematic if we want to analyze local policy outcomes or
coalition formation at the local level. Simply imputing the policy positions of parties
from the national level might not be sufficient and is potentially misleading. There-
fore, the purpose of this paper is to close this gap by presenting several ways how to
estimate policy positions of local parties using information in their manifestos.
Analyzing local party manifestos constitutes a worthwhile research project from a
theoretical, methodological, and empirical perspective. First, compared to national
and regional politics, the dynamics of local party competition are strongly understud-
ied. By analyzing local party manifestos, we can directly test if and how theoretical
approaches developed and tested in the context of national party competition can be
transferred to the local level. In particular, we know very little about the question of
whether policy dimensions identified at the national level are also relevant at the local
level. In fact, some debates about the polarization of local politics directly address the
question of how relevant ideological differences are at the local level (see, e.g., Kjær
and Elklit,2010;Loughlin, Hendriks and Lidström,2011). By extracting the underly-
ing dimensions structuring local party competition in Germany, using more than 400
local party manifestos in 74 German cities with at least 100,000 inhabitants, this article
directly contributes to this debate.

introduction 4
Second and closely related to the previous point, it is an important methodological
question if and how existing approaches for extracting positions from party mani-
festos can be applied to the local level. We discuss several challenges which might
hamper the application of commonly used scaling methods—such as Wordscores or
Wordfish—to local party manifestos. We argue that local party manifestos constitute
a least-likely case for the application of scaling methods, compared to national and
regional party manifestos. Nonetheless, we demonstrate that these methods can be
successfully used to identify meaningful political dimensions if one takes the special
characteristics of local party manifestos and of the scaling methods into consideration.
Finally, we also provide a first empirical analysis on how position estimates of local
parties can be explained by socio-economic and political variables, thus contributing
to the growing literature on the explanation of sub-national parties’ position-taking
(see, e.g., Giger, Müller and Debus,2011;Müller,2009,2013).
By applying semi- and unsupervised scaling methods, we demonstrate that the esti-
mated policy positions depend heavily on the chosen methods. Local party positions
on a general left-right dimension estimated by Wordscores are equivalent to the order
of parties from left to right on the national level. Applying Wordfish, we do not detect
parties’ left-right positions. Yet, this is not a drawback of Wordfish, because we are
able to identify an important additional dimension of local politics which differenti-
ates parties with regard to their emphasis on local and national policies. This finding
highlights that local party competition is substantially different from party compe-
tition on the regional and national level. It also emphasizes that both methods are
no substitutes for each other, but produce strongly different results. We additionally
show that particularly smaller parties—that also successfully compete on the regional
and national level—often exploit local elections to address policy issues which fall
into the realm of national or regional politics. Therefore, this paper does not only

assessing local party positions: the story so far 5
provide new insights into the dynamics of local party competition, but it also shows
new opportunities for future research.
The article proceeds as follows. The next section briefly describes our motivation to
analyze local party manifestos by reflecting on existing research on the assessment of
local party positions. The third section presents different approaches for estimating
local party positions with a specific focus on text scaling methods. Additionally, we
highlight three challenges scholars are facing when analyzing local party manifestos.
In the fourth section, we first present the data we use. Secondly, we present the
estimations of local party positions using two different text scaling methods and we
discuss potential limitations of these approaches. Section five is devoted to a first
explanation of local party positions using political as well as socio-economic variables.
The final section concludes.
2 assessing local party positions: the story so
far
Positions of local parties predominantly play a role in explaining coalition formations
and policy outcomes at the local level. Several studies in a variety of European coun-
tries have been conducted to investigate which factors increase the likelihood of poten-
tial coalitions to be formed following an election (see, e.g., Bäck,2003;Denters,1985;
Laver, Rallings and Thrasher,1998;Debus and Gross,2016;Serritzlew, Skjæveland and
Blom-Hansen,2008;Serritzlew, Blom-Hansen and Skjæveland,2010;Skjæveland, Ser-
ritzlew and Blom-Hansen,2007;Steunenberg,1992). Since policy-seeking approaches
are a prominent strand of the literature on government formations in multi-level sys-

assessing local party positions: the story so far 6
tems (see, e.g., Mueller,2009), estimating local parties’ policy positions is crucial for
the understanding of political processes at the local level.
Scholars used a variety of methods and data sources to place local political parties
on the general left-right dimension, which is widely considered to also be decisive
for local party competition. For example, Laver, Rallings and Thrasher (1998) use
expert survey data on national parties’ left-right positions in Great Britain as proxy
for local parties’ left-right orientations. Denters (1985) and Steunenberg (1992) do the
same for Dutch parties, whereas Goerres and Tepe (2013) use the order of German
parties on the left-right dimension on the national level as a proxy for the ideological
preferences of local parties’ in German cities. Yet, research on sub-national party
positions and intra-party heterogeneity in multi-level systems tells us that parties on
the sub-national level can take on different policy stances than their national parties
(see, e.g., Bernauer and Bräuninger,2009;Giger, Müller and Debus,2011;Müller,2009,
2013). Therefore, using national party positions as proxies for political actors’ policy
orientations on the local level might be misleading.
Debus, Knill and Tosun (2013) use regional parties’ positions as proxies for lo-
cal parties’ positions in municipalities and counties in the German state of Baden-
Württemberg. Yet, this approach is not without problems. Particularly in Baden-
Württemberg, independent local lists are electorally strong and win a large number
of seats in local parliaments. Since these lists are not competing in regional or national
elections, there is no way to assess their policy positions by solely relying on regional
or national data. Ideally, a reliable measure of local party positions should be derived
directly from information on the local level.
As Skjæveland, Serritzlew and Blom-Hansen (2007) note, “the obvious solution to
the problem of identifying local policy positions” is to conduct a survey asking lo-
cal political actors about their policy orientations on several policy dimensions. Bäck

assessing local party positions: the story so far 7
(2003,2008) did such a survey among local councillors in 49 municipalities located in
one Swedish region. Skjæveland, Serritzlew and Blom-Hansen (2007) measured local
policy positions by conducting an expert survey of local councillors in all Danish mu-
nicipalities. Egner, Sweeting and Klok (2013) provide the largest cross-sectional data
set on local actors’ policy positions based on a survey among a sample of councillors
in municipalities with more than 10,000 inhabitants in 16 countries. Yet, this data is
only available for a specific point in time and empirical analyses are thus restricted to
answering cross-sectional research questions.
In order to obtain data on local actors’ policy positions over time, two promising
proposals have been made in recent studies: (i) using electoral results of local elections,
and (ii) using local party manifestos. Nyhuis (2017) shows the possibility to infer
local actors’ policy preferences from election results in elections with multiple votes.
Assuming that voters spread their votes across party lists that are ideologically similar,
these patterns can be exploited to infer local actors’ policy positions. Although only
presenting results from one election in one German municipality, Nyhuis’ approach
can be applied over time and for different municipalities in different countries, as
long as voters have the possibility to spread their votes across multiple party lists.
However, since the method depends on a specific electoral system it is only applicable
for a very small number of cases which limits its applicability. Debus and Gross (2016)
transfer an already existing approach of estimating policy positions on the national
and regional level to the local level by inferring local actors’ policy preferences based
on information in local parties’ manifestos. They apply the ‘Wordscores’ technique to
estimate local parties’ policy positions in 29 large cities in the German state of North
Rhine-Westphalia between 1999 and 2009 on a general left-right, a societal, and an
economic dimension, demonstrating that the policy positions of party families and of
party branches of the same party competing in different cities differ. This paper shares

methodological approaches and challenges 8
their assumption that local party manifestos are a valuable source for the estimation
of local party positions (see below).
3 methodological approaches and challenges
Having discussed how existing research has addressed the challenges of estimating lo-
cal party positions, we now turn to the methodological approaches which might seem
suitable for analyzing local party positions in general. For this purpose, the following
subsection briefly discusses differences between expert-, hand-, and (semi-)automatic
computer-coding approaches, and we specifically discuss the potential challenges im-
posed by local party manifestos for hand- and computer-coding approaches. Then, we
elaborate in more detail why local party manifestos pose a challenge for text scaling
methods and we present several solutions to these challenges.
3.1 Expert-, Hand-, and Computer-Coding
The importance of locating parties in a low-dimensional policy space is widely ac-
knowledged in political science research (Laver,2014). Estimating these positions is,
however, challenging as positions cannot be directly observed. While many different
methodological approaches have been developed to estimate party positions, three
of them stand out: (1) expert surveys (see, e.g., Benoit and Laver,2006;Bakker et al.,
2015), (2) hand-coding of party manifestos (Budge et al.,2001;Klingemann et al.,2006),
and (3) (semi-)automatic computer-coding (Laver, Benoit and Garry,2003;Slapin and

methodological approaches and challenges 9
Proksch,2008).2While all of these approaches have certain advantages and short-
comings (see Laver,2014), we argue in this section that automatic content analysis of
local party manifestos is a promising approach for estimating policy positions of local
political actors.
expert surveys At the national level, expert surveys are a valuable source for es-
timating party positions. Experts possess detailed knowledge about parties and their
positions on various issues. Moreover, expert surveys are probably less affected by the
strategic behavior of parties. For example, extremist parties might use a less-radical
wording in their manifestos to appeal to a wider audience. Compared to manifesto
based approaches, experts can integrate various information into their judgment, thus
counterbalancing parties’ strategic behavior in drafting manifestos. As sophisticated
as these expert surveys are, there are several weaknesses that have to be addressed.
First, conducting expert surveys is expensive. To obtain a representative sample of
local actors’ policy positions most of the time requires the financial resources of a
cross-national research group (Egner, Sweeting and Klok,2013). Secondly, and re-
lated to the first point, expert surveys on the local level only provide data of local
actors’ policy positions at a specific point in time. Unlike on the national and Euro-
pean level, there do not exist time-series cross-national data sets on local actors’ policy
orientations, thus rendering it impossible to analyze trends over time. Thirdly, even
for experts it might be challenging to provide reliable information on policy positions
of small local parties that might compete for the first time for seats in local councils,
not to mention the myriad of independent local lists (Reiser and Holtmann,2008).
Fourthly, existing research ex ante assumes that only a general left-right dimension is
decisive for local party competition. Hence, all local expert surveys solely ask council-
2Moreover, surveys of voters and politicians are also used to extract party positions (for an overview,
see Mair,2001).

methodological approaches and challenges 10
lors to position themselves on a general left-right dimension. Yet, this is a very strong
assumption and it remains an open empirical question if actually only one policy
dimension is decisive for local party competition.
hand-coding Hand-coding of party manifestos, most prominently conducted by
the CMP / MARPOR project (Budge et al.,2001;Klingemann et al.,2006), is also a
well-established approach for estimating policy positions. Just like automatic content
analysis, it is based on the analysis of party manifestos. The heart of the MARPOR
project constitutes a coding scheme of 56 categories. ‘Quasi sentences’ in party mani-
festos are assigned to these categories by coders. These categories are then used to es-
timate party positions on political dimensions. As with expert surveys, this approach
has a high reliability as manifestos can be coded quite precisely.3Yet, hand-coding
seems to be unfeasible for analyzing local party manifestos for at least two reasons.
First, hand-coding is very expensive and time-consuming. While these costs are well
justified for generating positions of parties at the national level, time and money are
probably too scarce to apply hand-coding techniques to the analysis of local party
positions. Additionally, the costs are going to be a lot higher because the number of
party manifestos drastically increases at the local level since there are more relevant
political actors and hundreds of municipalities that might have to be coded (even
though local party manifestos are usually a lot shorter compared to national party
manifestos). A second drawback of hand-coding is that the well-defined categories
from the national level cannot be directly transferred to the context of local politics,
meaning that a completely new coding scheme would have to be developed for the
3The reliability is high with regard to the aggregated measures produced by this coding scheme (Laver,
2014,217), but the CMP approach has also been subject to criticism. For example, Däubler et al. (2012)
demonstrate that the focus on ‘quasi-sentences’ causes intercoder unreliability (see also Mikhaylov,
Laver and Benoit 2012). Flentje, König and Marbach (2017) discuss the problem of implausible ‘leapfrog’
shifts in party positions derived from the CMP project.

methodological approaches and challenges 11
analysis of local politics (Agasoster,2001). Again, this makes the application of hand-
coding for local party manifestos a challenging task. To be clear, we are not saying
that local manifestos cannot (or should not) be hand-coded. Rather, we doubt that
the high effort necessary for doing so is rarely feasible.4This holds even more true,
since (semi-)automatic coding procedures, as we will discuss in the following, might
constitute a fruitful and less costly alternative.
text scaling Computer-assisted methods for analyzing texts have become in-
creasingly popular in the last fifteen years (for an overview, see, e.g., Grimmer and
Stewart,2013). These approaches are promising as they can provide estimates which
are comparable to the methods described above, but require much less resources. This
is also the reason why we think that these approaches are the most feasible for analyz-
ing local party manifestos. Both for expert surveys and for hand-coding approaches,
we have argued that extracting policy positions for the wide variety of municipalities
and parties is rather unrealistic given the high amount of resources which have to
be spent. In contrast, with (semi-)automatic text scaling approaches, only the party
manifestos have to be collected and transferred to a computer-readable format.5The
drawback of these approaches is, however, that they do not guarantee to provide valid
estimates. Among others, Grimmer and Stewart (2013) as well as Proksch and Slapin
(2009) describe a wide variety of potential pitfalls when it comes to the application
of automatic content analysis. Therefore, estimating positions from local party man-
ifestos using solely statistical approaches which rely on word frequencies, requires
4Recently, Benoit et al. (2016) demonstrate that using crowd-sourced text analysis might circumvent
many of these problems. In particular, crowd-sourcing is quick, reproducible and it reduces costs.
Applying crowd-sourced text analysis to local party manifestos is, however, beyond the scope of this
article.
5We note here that the collection of party manifestos at the local level is more challenging than at the
national or regional level since local parties and independent local lists sometimes are less profession-
alized than their regional or national counterparts, thus not always providing scholars with a party
manifesto.

methodological approaches and challenges 12
special attention to validating and interpreting the estimated positions. In the next
section, we describe several challenges scholars have to be aware of in order to apply
computer-based approaches to local party manifestos. While some of these problems
are well-known problems of computerized text analysis, we argue that these problems
can be particularly severe for the analysis of local party manifestos.
3.2 Local Party Manifestos: A Least-Likely Case for Text Scaling?
In this subsection, we describe three challenges which are specifically related to the
analysis of local party manifestos and which might hamper the application of com-
monly used text scaling methods, such as Wordscores or Wordfish and its variants, to
local party manifestos.
3.2.1 Comparability
One of the most important assumptions in text scaling is the assumption of a com-
parable data generation process of the analyzed text. When text documents are not
comparable, estimated differences in the positions might simply reflect different word-
ings based on the type of document instead of different party positions. Bluntly put,
press releases and transcripts of legislative speeches are in many ways different to a
party manifesto and these differences are reflected in the wording of these texts.
In contrast to the national level, local party manifestos are more likely to be incom-
parable. While all documents analyzed in this article are the official manifestos of a
party at the local level, there is a wide variation in the ‘style’ of the manifestos. Some
of the manifestos closely resemble the structure and style of national or regional party
manifestos, while others resemble advertising flyers and only display the most impor-
tant issues in the form of bullet points. The large heterogeneity in how the party

methodological approaches and challenges 13
manifestos are written is also reflected in the different manifesto length, ranging from
less than 100 to more than 10,000 words (see Figure 1).
Figure 1: Boxplot showing the Length of Local Party Manifestos for Different Parties
Note: Y-axis is on log10-scale. Parties are sorted according to the average manifesto length.
The data set comprises all publicly available party manifestos from 74 large cities in Germany.
Only manifestos from the most recent elections are included. The number of observations is
N=540.
Closely related to this issue is the fact that not all manifestos follow the same struc-
ture. This is an important point, as many analyses of party manifestos rely on scaling
only specific sections of the party manifesto to estimate parties’ positions on certain
issues (e.g. Slapin and Proksch,2008,712-714). In local party manifestos, however,
not every party has a section about the economy, while such a section is an essential
part of every manifesto at the national and regional level. Therefore, sub-analyses of
party manifestos that only focus on specific sections will be much more challenging
and time-consuming at the local level.

methodological approaches and challenges 14
3.2.2 Level of Professionalism
Party manifestos at the national or regional level are usually the product of a long-
lasting intra-party negotiation process, written by political experts and experienced
politicians. Furthermore, spin doctors and political consultants strengthen the strate-
gic wording used in manifestos to appeal to a larger or specifically targeted audience
(Däubler,2012). In a nutshell, national and regional party manifestos usually reflect
high levels of professionalism. Again, such assumptions about party manifestos are
unlikely to hold at the local level. Most importantly, parties usually have fewer re-
sources at the local level (Egner, Sweeting and Klok,2013). Especially smaller parties
often consist of only a handful of (active) party members without much financial re-
sources. In contrast, national parties competing at the local level have much more
resources and often more active party members which allows them to write more pro-
fessional manifestos. We expect these differences to have an impact on the estimated
party positions, as more professional party manifestos are likely to address more,
but also different topics. Moreover, the exact wording used in the manifestos also
depends on the professionalism of the party manifesto. This might produce results
which highlight differences between smaller and larger parties instead of ideological
differences between parties.
3.2.3 Agenda Effects and Clustering
Another aspect which might affect the results of automatic content analyses are ‘agenda
effects’. In time-series analyses of party manifestos, agenda effects refer to the prob-
lem that certain topics are more prevalent in a certain election than in another. If that
is the case, automatic text scaling approaches might not measure a change in the posi-
tion of a party, but rather a difference in the issues discussed in the party manifestos
between Electiontand Electiont−1(Proksch and Slapin,2009,332).

methodological approaches and challenges 15
Even if scholars do not analyze time-series data, these effects might also occur in
the case of analyzing local party manifestos. Comparing the positions of parties in
multiple cities, we are dealing with clustered data. The problem is probably even
more severe than with time-series data. Almost by definition, local party manifestos
are likely to contain lots of words which only occur in a very specific local context.
Just think of the name of a street, the name of urban districts, the city name itself,
and so on. Moreover, party manifestos from a city close to the North Sea will contain
lots of words related to this geographic characteristic. Cities in which wine-growing
constitutes a large part of the local economy will discuss words related to this issue
more frequently. Several more examples can be found. The problem of these ‘local
words’ is that we might end up scaling differences between cities instead of ideological
party positions. Put differently, parties from the same city might appear more similar
in their positions due to a common ‘local’ vocabulary used in their party manifestos.
3.3 Addressing the Challenges
The above described challenges suggest that local party manifestos constitute a least-
likely case for the successful application of text scaling approaches compared to re-
gional or national party manifestos. Some of these problems, however, can be directly
addressed by researchers and, thus, at least reduce the potential bias. Here we discuss
how we approach these issues in our empirical analysis.
Agenda effects can be mitigated by requiring a word to occur in a certain number of
manifestos or by manually removing ‘problematic’ words from party manifestos. We
use both approaches in our analysis as we manually remove all city names from the
party manifestos and by requiring that a word has to occur in at least ten manifestos
in order to be included in the analysis. While this seems like a very strict exclusion

methodological approaches and challenges 16
standard, it should be noted that we analyze almost 500 manifestos, and excluding
words which appear in less than ten party manifestos means that a word has to appear
only in approximately two per cent of all manifestos. Since we have no case where
more than nine parties have run for election, this procedure is likely to remove words
which are only relevant for certain cities. Varying the word threshold from seven
to 100 – thus following the suggestion of Proksch and Slapin (2009) to only include
words that are mentioned in at least 20 per cent of all documents – does not change the
robustness of the estimations (see Appendix). Hence, we proceed by only excluding
words that are mentioned in less than ten party manifestos.
Addressing the comparability and professionalism of local manifestos is more prob-
lematic. In order to guarantee a minimum of comparability between the documents,
we exclude all party manifestos with less than 1,000 words (see, e.g., Klemmensen,
Hobolt and Hansen,2007;Klingelhöfer and Müller,2015). Further, as described in
more detail below, we analyze local party competition only in cities with more than
100,000 inhabitants where we can safely assume a higher level of professionalism of
local politics than in smaller municipalities.
3.4 Methods Used in the Analysis
The previous section discussed certain challenges for analyzing local party manifestos
using text scaling approaches. We also discussed potential ways to address these
challenges. In this section, we now briefly discuss which text scaling methods we use
in the empirical analysis.

methodological approaches and challenges 17
3.4.1 Wordscores
The first approach which we use is the Wordscores technique (Laver, Benoit and Garry,
2003;Lowe,2008). In contrast to fully unsupervised scaling methods, Wordscores is
an a priori approach as it allows the researcher to determine the dimension on which
political actors are scaled. The basic idea of Wordscores is to compare the frequency
distribution of words of documents whose programmatic position is known (‘refer-
ence texts’) with the frequency distribution of texts whose programmatic position is
unknown (‘virgin texts’). ‘Known’ programmatic positions are usually derived by us-
ing expert judgments on parties’ positions on specific policy dimensions (so-called
‘reference scores’). The application of Wordscores assumes that political actors do not
use words randomly in their election manifestos, but rather that they send “ideolog-
ical signals” (Pappi and Shikano,2004) to their voters by using some words more
frequently than others. Wordscores estimates the policy-area-specific position of a
document by comparing the relative word frequency of ‘reference texts’ to the word
distribution of ‘virgin texts’. Every word a party uses in a text is treated as a position
of the party on an a priori specified scale and the average position of all words used
in a text indicates the position of the party.
Given the fact that word frequencies in ‘reference texts’ and ‘virgin texts’ are used
for estimating party positions, only words which appear in these documents can be
used for scaling. Put differently, words which only appear in local party manifestos—
which are our ‘virgin texts’—and not in the ‘reference texts’ do not contribute to the
position of a party. Therefore, party positions estimated by Wordscores are probably
less affected by words which appear only in a very specific local context since these
words are less likely to be included in the ‘reference texts’. On the other hand, this
might also hold true for words which are important in the local context, but not
discussed at other levels of the political system, such as ‘playground’. This short

methodological approaches and challenges 18
discussion points to the most important step of an analysis using Wordscores: the
selection of ‘reference texts’ (Laver, Benoit and Garry,2003). These documents should
be of the same character as the ‘virgin texts’ regarding the structure and syntax of the
text, as well as the terminology used by political actors. Additionally, to increase the
validity and reliability of Wordscores estimates, ‘reference texts’ should cover the time
period and the entire range of the ideological dimension under study, and the number
of scored words and unique words in the ‘virgin texts’ should be high (Bräuninger,
Debus and Müller 2013,19;Klemmensen, Hobolt and Hansen 2007;Hjorth et al. 2015;
Laver, Benoit and Garry 2003,313-314). All these factors point to the direction that
Wordscores works best when both ‘reference texts’ and ‘virgin texts’ are relatively
long (Klemmensen, Hobolt and Hansen,2007). Manifestos, however, can be short—
and this is one peculiarity of party manifestos on the local level (see Figure 1)—and
there is no way scholars interested in local party positions and eager to use automated
text analysis techniques can circumvent this fact. Put differently, “if the texts in which
we are genuinely interested are short, then they are short and we just have to make
the best of the situation in which we find ourselves” (Laver, Benoit and Garry,2003,
315).
To apply Wordscores selecting the ‘right’ ‘reference texts’ and ‘reference scores’ is
a crucial decision. ‘Reference texts’ should provide a sufficient vocabulary which can
be used to estimate the position of virgin texts. Therefore, we opt for using national
party manifestos from the German general election in 2013 as ‘reference texts’. This
decision has several advantages. First, national party manifestos are long and thus
provide a comprehensive vocabulary. Second, parties are most likely to use their
national manifestos to send ideological signals to the electorate. Therefore, the words
in these manifestos provide a meaningful foundation for positioning parties on several
policy dimensions. Third, relying on expert surveys we have a detailed understanding

methodological approaches and challenges 19
of party positions at the national level and we can use these expert judgments as
meaningful ‘reference scores’. Finally, using manifestos from the national level allows
us to have common reference texts for all parties from different regions and cities. We
are thus able to position all manifestos in the same ideological space which makes the
position estimates comparable between cities.
We use national party manifestos of AfD, CDU, FDP, Green Party, NPD, Pirate
Party, SPD, and Left Party for the 2013 German national elections as ‘reference texts’
to comprise the ideological range of party positions as much as possible. 6The
national elections in 2013 are the closest ones to the most recent local elections in the
cities under study (2013-2016). As ‘reference scores’, we assign expert judgments on
parties’ general left-right, economic, and societal positions to the ‘reference texts’ by
using the Chapel Hill Expert Survey (CHES) 2014 (Bakker et al.,2015). 7
Admittedly, these advantages of using the national party manifestos as ‘reference
texts’ also have some minor drawbacks. Most importantly, one could argue that using
regional party manifestos as ‘reference texts’ might be more suitable as these party
manifestos are more likely to include words which are only relevant in a very specific
regional context. Put differently, when using national party manifestos we are likely
to lose many words which are only addressed at the local or regional level. However,
we consider this loss of information as rather unproblematic since such regional or
local words are probably less relevant for positioning parties on an ideological scale.
Furthermore, when using regional party manifestos we would also lose many of the
previously mentioned advantages of using national party manifestos. We would no
longer be able to position all parties in the same ideological space and we do not
6This data is provided by the Political Documents Archive (www.polidoc.net; see Benoit, Bräuninger
and Debus 2009).
7CDU and CSU are treated separately in the CHES survey; yet, both parties ran with a joint manifesto
for the general elections in 2013. We therefore use the averaged scores of experts’ evaluations of their
positions.

methodological approaches and challenges 20
possess precise information about the positions of regional parties. The only com-
prehensive data set on regional party positions in Germany comes from Bräuninger
and Debus (2012) and these positions are already estimated using Wordscores. Lastly,
some parties relevant for local party competition do not write manifestos for all re-
gional elections which reduces the ideological heterogeneity of the vocabulary in some
regions and makes party positions less comparable.
3.4.2 Wordfish
In addition to the semi-supervised approach of Wordscores, we use an unsuper-
vised scaling technique. Unsupervised techniques are usually described as a poste-
riori (Benoit and Laver,2006), since we do not have prior knowledge about which
dimensions will be observed in the document. Therefore, unsupervised methods
put more emphasis on an in-depth interpretation of the estimated dimensions. Since
the estimated dimensions are not predetermined by the researcher, it is also possible
that unsupervised methods discover dimensions in documents which are only of little
interest for political scientists or which do not reflect ideological differences between par-
ties. As Grimmer and Stewart (2013,293) put it, unsupervised methods “seize upon
the primary variation in language across actors”, but this variation is not necessarily
ideological. As a consequence, it is an open question whether meaningful dimensions
can be extracted from local party manifestos using unsupervised methods. In the fol-
lowing, we apply the Wordfish technique, which is the most prominent unsupervised
text scaling method in political science research (Slapin and Proksch,2008).

scaling local party manifestos 21
4 scaling local party manifestos
We begin with the presentation of the data we are going to use in the following
analyses. Then, we present the estimation results of local parties’ policy positions,
first estimated with Wordscores, then using Wordfish.8Finally, we discuss potential
limitations of the two applied techniques for the scaling of local party manifestos.
4.1 Data
The empirical analysis focuses on the estimation of local party positions in 74 German
cities with more than 100,000 inhabitants. We focus on large cities in Germany for sev-
eral reasons. First, recent research indicates that local party competition in Germany is
highly politicized in specific issue areas. For example, Debus, Knill and Tosun (2013)
show that registration fees for same-sex unions in cities located in the German state
of Baden-Württemberg are primarily determined by parties’ preferences in district or
municipal councils. Parties with socially progressive positions adopt lower levels of
extra fees for same-sex unions than parties with more conservative positions. Yet, the
authors are not directly measuring local parties’ policy preferences on societal issues,
they rather use the vote share of socially progressive parties in district and municipal
councils as a proxy for the ideological positions of local parties.9Secondly, whereas
local politics in small municipalities is mainly shaped by independent local lists, party
competition in Germany is more and more politicized as the number of inhabitants
is increasing, i.e. parties that run for regional and national elections also compete in
8All analyses have been conducted using the quanteda-package for R, version 0.9.9-25 (Benoit et al.,
2017).
9Debus, Knill and Tosun (2013) define socially progressive parties based on their positions extracted
from regional party manifestos.

scaling local party manifestos 22
larger cities with independent local lists (Brearey,1989;Fried,1976). Thirdly, particu-
larly local councils in large cities are equivalents to national and regional parliaments
where party competition is structured along the lines of majority and minority fac-
tions (see, e.g., Egner,2015;Gunlicks,1986). Fourthly, for a first application of text
scaling techniques to local party manifestos, we strive for the inclusion of as many
party manifestos as possible, and due to the more professionalized local politics en-
vironment in larger cities, data collection of manifestos has been easier than in small
municipalities since party manifestos have been more frequently available. We ac-
knowledge, however, that our results are thus probably not representative for smaller
municipalities and we intend to address this issue in future research.
We collected all publicly available party manifestos from all parties running for local
councils, including also small parties and independent local lists. For many cases,
however, only the most recent party manifestos were available. Therefore, while our
complete corpus consists of more than 1,000 manifestos, we only include manifestos
from the most recent local elections in the respective cities (2013-2016). This reduces
the number of documents to 540 (see Figure 1). Moreover, to guarantee a certain
amount of comparability in the length of the manifestos and to increase the reliability
of the estimated policy positions, we only include manifestos with more than 1,000
words (see, e.g., Klemmensen, Hobolt and Hansen,2007;Klingelhöfer and Müller,
2015). This leaves us with a corpus of 478 local party manifestos.
We applied common pre-processing steps to the manifestos (for a summary on these
steps see, e.g., Lucas et al.,2015). This included stemming, converting all text to lower-
case, removing of punctuation and numbers, as well as deleting all words that occur
in less than ten documents. Furthermore, and additionally to the use of the German
dictionary in the quanteda-package as stopwords, we treated all city names, party
labels and specific words only common to specific regions in Germany as stopwords.

scaling local party manifestos 23
To explain and cross-validate the extracted policy positions of local political actors,
we additionally collected socio-economic and political indicators of cities and parties
under study.
4.2 Wordscores
The results of the Wordscores estimations on each of the three CHES dimensions are
displayed in Figure 2. The pattern observed in Figure 2suggests that Wordscores can
be applied to local party manifestos for estimating party positions at the local level.
For example, the rank order of parties from left to right bears high resemblance to the
rank order of these parties at the national level with the Left Party being identified
as the most leftist party followed by the Greens, the Social Democrats (SPD), the
Christian Democrats (CDU/CSU), the Liberal Party (FDP), and finally the right-wing
populist party AfD.10 Also the more detailed results for the economic and societal
dimension show a high degree of face validity. The right-wing populist AfD is on par
with CDU/CSU and FDP on the economic dimension, which is plausible since the
parties’ focus is on societal issues such as migration and identity politics (Jankowski,
Schneider and Tepe,2016). In contrast, the Left Party which is heavily concerned
with the issue of redistribution is consistently estimated to be the leftists party at
the local level on the economic dimension. The face validity of the results is further
corroborated by party positions on the cultural/societal dimension, which uses CHES’
GAL-TAN index as ‘reference scores’. Here, the Greens are most often estimated to
be the most leftist party. The AfD is the most right-wing party on this dimension by a
large margin (Arzheimer,2015). Again, these results are in line with our knowledge
about party competition at upper levels of Germany’s multi-level system.
10 This corroborates the findings of Debus and Gross (2016) for the left-right positions of local parties in
North Rhine-Westphalia.

scaling local party manifestos 24
Finally and also in line with our expectations, the heterogeneity of the categories
‘local list’ (comprising independent local lists that only run for local elections) and
‘small party’ (comprising very small parties which are usually unsuccessful at the
regional or national level) is reflected by their large variation of positions on all three
dimensions. Independent local lists are more frequently estimated to take a position
close to CDU/CSU and FDP – and thus appear to be rather conservative—which is in
line with previous research (Reiser and Holtmann,2008).

scaling local party manifestos 25
Figure 2: Boxplots of Wordscores Results
Note: Boxplot of local party positions estimated using Wordscores with national party mani-
festos as ‘reference texts’ and CHES party positions as ‘reference scores’. Upper panel shows
results for the general left-right dimension. Middle panel shows economic positions of local
parties. Lower panel uses the GAL-TAN index as ‘reference scores’, which measures positions
on cultural and societal issues. Boxplots are ordered based on the position of the median party
manifesto.

scaling local party manifestos 26
4.3 Wordfish
The following Figure 3shows the results from the Wordfish analysis. The left panel of
Figure 3displays the results aggregated for each party using boxplots and the right
panel shows the position of each manifesto for all parties including 95% confidence
intervals (horizontal lines). The results are apparently very different compared to the
Wordscores analysis. Most importantly, the Left Party and the right-wing AfD are
positioned in close proximity to each other on the right end of the dimension. The
manifestos of FDP, Greens, SPD, and CDU/CSU, in contrast, form a large block on
the left side of the dimension. Within this block, SPD and CDU/CSU are closer to the
left of the dimension than the Green Party and the FDP. Obviously, this dimension
does not reflect the general left-right pattern. Yet, the positioning of parties in certain
blocks suggests that positions follow a certain pattern and are not randomly created.
Figure 3: Results of the Wordfish Analysis
Note: Left panel displays the distribution of party manifestos on the Wordfish dimension using
boxplots. Right panel shows positions of each party manifesto in the analysis on the Wordfish
dimension. Horizontal lines are 95% confidence intervals. Boxplots are ordered based on the
position of the median party manifesto.

scaling local party manifestos 27
Using Wordfish, the substantive meaning of the dimension can be interpreted not
only by investigating the position of parties, but by estimating the contribution of
each word to the positioning of the manifestos on the estimated dimension (β). In
Figure 4these word contributions are displayed against the word fixed effects (ψ) and
the 20 words with the highest and lowest β-values are highlighted. These are words
which can be informative for interpreting the substantive meaning of the dimension.
Based on these words, we interpret the dimension estimated by Wordfish as a ‘local
vs. national policies’-dimension. In particular the AfD and the Left Party use words
related to national policies and politics, whereas other parties much more frequently
use words specifically related to local policies and politics. This focus on national
issues is accompanied by a rather critical and sometimes populist tone. For example,
some of the words with the highest contribution are words such as ‘fascism’ (‘Faschis-
mus’), ‘weapons’ (‘Waffen’), ‘foreign policies’ (‘Außenpolitik’). On the other end of
the dimension, we find several words which have clear connections to local politics,
including words such as ‘parking lot situation’ (‘Parkplatzsituation’) and city tourism
(‘Städtetourismus’).
This dimension highlights that in contrast to the national level the primary vari-
ation in word usage in local manifestos does not seem to be ideological. Wordfish
cannot extract a meaningful left-right dimension since the “ideological dominance as-
sumption” (Grimmer and Stewart,2013,291) of unsupervised methods does not hold.
Instead, party competition seems to be structured by the question of how much par-
ties address issues which are substantially relevant at the local level. Small and rather
populist parties seem to exploit local elections in order to address national or even Eu-
ropean topics. A strategy which might be in line with the ‘second-orderness’ of less
important elections in which at least some voters cast a vote based on the political
situation at the national level.

determinants of local party positions 28
Figure 4: Word Weights and Word Fixed Effects
5 determinants of local party positions
In the following, we present a first explanation of the estimated policy positions of
local parties by running several regression models including party dummies, socio-
economic and political variables, as well as several control variables. In a first step, we
offer an analysis of the positions on the local vs. national policies dimension extracted
by Wordfish.Then we provide a comparative analysis for the left-right dimension ex-

determinants of local party positions 29
tracted by Wordscores. In other words, the first analysis offers new insights about
the question of which factors influence the ‘localness’ of party manifestos, while the
second analysis follows existing approaches to validate the positions of parties on the
left-right scale (see, e.g., Giger, Müller and Debus,2011;Müller,2009,2013).
5.1 Explaining the Local vs. National Conflict
We use the estimated positions of CDU/CSU as reference category since the results
in Figure 3showed that CDU/CSU are on the ‘local’ end of the local vs. national
policies dimension. We expect all other parties to position themselves more on the
‘national’ end of the dimension compared to CDU/CSU. As socio-economic variables,
and in order to control for recent events like the Euro crisis and the refugee crisis
that might also play a role in local parties’ election campaigns since municipalities
and cities also have to deal with the consequences of both crises, we include the
share of foreigners, the share of unemployed people and the logarithmic number of
inhabitants. For the share of foreigners and unemployed people, we expect a positive
correlation with local parties’ focus on national politics because policies associated
with high unemployment rates or the financial help for the integration of foreigners
are often located at the national and not at the local level, and parties might see the
national level as being responsible for the numbers of foreigners and unemployed
people.
The political context local parties compete in is measured by using a dummy vari-
able indicating the party affiliation of the mayor 11, a dummy variable indicating if
the mayor is from the same party, and the logarithmic number of seats won by parties
11 The dummy variable distinguishes between mayors from small parties (FDP and Greens) and mayors
with no party affiliation (independent mayors), with mayors from CDU/CSU and SPD as reference
category. Note that neither the Left Party nor the AfD holds a mayoral position in the cases under
study.

determinants of local party positions 30
in the previous election. We use the number of seats a party obtained in the previous
legislative period in the city council as an indicator for the experience in local politics.
The assumption is that with more active politicians from the local level parties have
more knowledge about local problems. Put differently, with more local politicians we
expect that parties are becoming more like a local party instead of being just the local
organization of a national party. We also expect, however, that this effect should be
weaker the more local politicians a party has, meaning that we expect that it matters
more whether a party has two instead of only one local politician instead of 15 or
14. Therefore, we use the logarithmic number of seats in the previous city council as
indicator.12 Finally, we control for year effects and the logarithmic number of words
in the manifestos. We expect local parties to talk more about national politics in their
local manifestos in the year of the national elections in 2013, and we assume that the
longer manifestos are, the higher is the likelihood that local parties also address is-
sues that are not only purely related to local politics. Additionally, even though we
are only focusing on cities with at least 100,000 inhabitants, we control for the loga-
rithmic number of inhabitants since city population ranges from a little bit more than
100,000 to over a million inhabitants.
The data is structured in a hierarchical way because parties are clustered in cities.
We account for this clustering by estimating several linear mixed-effect models with
cities at the upper level and parties at the lower level. We use cluster-corrected stan-
dard errors to capture potential city-specific peculiarities. 13 Overall, the empirical
analysis of local party positions comprises 478 local election manifestos in 74 German
cities with at least 100,000 inhabitants.
12 As a robustness check, and in order to not lose too many cases due to parties which did not hold
any seats in the previous council, we additionally calculated the logarithmic number of seats won by
parties in the previous election, imputing a zero for all parties which did not have a seat in the local
council in the previous election. The substantial findings remain the same.
13 Using restricted maximum likelihood instead of clustered standard errors does not change the results.

determinants of local party positions 31
The results of three estimated models in which the position of a party manifesto
on the Wordfish dimension–which measures the relative emphasis parties put on na-
tional politics in their manifestos—denotes the dependent variable are displayed in
the left panel of Table 1. Model 1only accounts for differences between parties and
additionally includes three control variables. This model confirms the results from
the visual analysis of the estimated positions: Compared to CDU/CSU, the relative
emphasis put on national policies and politics is much higher for the Left Party and
the right-wing populist AfD. The same holds true for the liberal FDP (in two out of
three models) and the ecologist Green party, but to a much smaller extent. Even the
difference between the SPD and CDU/CSU is statistically significant, but the effect
size is rather negligible in this case. This pattern holds for all estimated models. Our
finding from the visual inspection regarding small parties is also confirmed. On av-
erage, these parties position themselves more on the ‘national’-side of the dimension.
Yet, the considerably larger standard errors should be acknowledged which reflect
the heterogeneity of parties within this group. Finally, independent local lists are also
more ‘national’ than CDU/CSU manifestos. This might be the case because indepen-
dent local lists not holding the mayoral office are more likely to focus on ‘national’
rather than local issues, compared to campaigns when the mayor has no party affili-
ation (see below). Moreover, local lists might be founded due to dissatisfaction with
the policies of national parties. Therefore, these local lists also have incentives to ad-
dress at least some national issues in their manifestos in order to send a signal to the
electorate how they differ from established parties. It should be noted, however, that
the effect size for local lists is substantially smaller compared to the ‘small parties’
category.
In Models 2and 3, we additionally include several socio-economic and political
variables varying at the city level. The share of foreigners yields positive and statis-

determinants of local party positions 32
tically significant effects in both model specifications, while the unemployment rate
is only significant in the third model. In both cases, the effect is in the expected di-
rection: The higher the number of unemployed people and the higher the number of
foreigners living in a city, respectively, the more are local parties’ manifestos charac-
terized by national rather than local policy issues. Interestingly, we find that in cities
with an independent mayor the relative emphasis of local politics is higher in parties’
election manifestos. Parties know at the time of drafting their manifestos for local
council elections (a) if they have to compete against an independent local list already
holding the mayor’s office, or (b) if an independent candidate has realistic chances
to gain the mayor’s office when council and mayoral elections are held concurrently.
Assuming that independent candidates and independent local lists primarily focus on
local policies in their election campaigns if one of the two scenarios (a) or (b) is true
(see, e.g., Reiser and Holtmann,2008), other parties should have strong incentives to
also focus more on local topics. Likewise model 3also suggests that when the mayor
is from the same party the manifesto will be more local, but the effect is only signifi-
cant at p < .1and the effect size is rather small. Finally, a party’s experience in local
politics matters for the explanation of their party positions on the local vs. national
policies dimension. The effect of the variable is negative and statistically significant.
The effect size is also quite large, given the fact that the variable has a range between
0and 3.66. The more seats a party obtained in the previous local council elections,
the more the manifesto is characterized by local policy issues.
Regarding the effect of the three control variables we see that the number of inhab-
itants does not have a statistically significant effect. The election year dummies do
not display robust results across all models. Yet, for all years the relative emphasis
on national policies is lower compared to the year 2013, thus lending support to our
assumption that local parties focus more on non-local issues in a national election

determinants of local party positions 33
year. Therefore, national election campaigns probably have an impact on how local
parties draft their manifestos. This assumption is further supported by the fact that
parties are most visible located to the ‘local’ side of the dimension in 2015, which is
the year most distant from the national elections in 2013 and 2017. Furthermore, the
results indicate that the longer local party manifestos are, the more do parties also
focus on national policies. Lastly, the share of cross-city variance on the total variance
amounts to 17 and 21 per cent, respectively.

determinants of local party positions 34
Table 1: Determinants of Local Party Manifesto Positions on Wordfish and Wordscores Di-
mension
Wordfish Wordscores
Local vs. National Policies General Left-Right
(1) (2) (3) (1) (2) (3)
Parties
AfD 1.901∗∗∗ 1.895∗∗∗ 2.137∗∗∗ 2.159∗∗∗
(0.103) (0.104) (0.340) (0.342)
FDP 0.553∗∗∗ 0.553∗∗∗ 0.0838 0.134 0.140 0.564†
(0.059) (0.059) (0.082) (0.188) (0.189) (0.316)
Greens 0.754∗∗∗ 0.756∗∗∗ 0.460∗∗∗ -1.652∗∗∗ -1.660∗∗∗ -1.470∗∗∗
(0.058) (0.059) (0.074) (0.217) (0.216) (0.277)
Left Party 2.524∗∗∗ 2.531∗∗∗ 1.978∗∗∗ -1.956∗∗∗ -1.959∗∗∗ -1.509∗∗∗
(0.098) (0.096) (0.115) (0.221) (0.220) (0.378)
Local Lists 0.934∗∗∗ 0.935∗∗∗ 0.268∗-0.0771 -0.0343 0.496
(0.115) (0.117) (0.128) (0.369) (0.371) (0.624)
SPD 0.155∗∗ 0.157∗∗ 0.211∗∗ -0.824∗∗∗ -0.826∗∗∗ -0.928∗∗
(0.057) (0.057) (0.068) (0.228) (0.227) (0.287)
Small Parties 1.700∗∗∗ 1.703∗∗∗ 1.034∗∗∗ -1.330∗∗∗ -1.319∗∗∗ -0.320
(0.076) (0.076) (0.153) (0.246) (0.245) (0.476)
Socio-Economic Variables
% Foreigners 2.012∗∗∗ 1.507∗∗ -0.502 -0.744
(0.590) (0.552) (1.087) (1.117)
% Unemployed 1.800 3.138∗4.978†5.180∗
(1.573) (1.481) (2.646) (2.450)
Political Variables
Mayor = Greens / FDP -0.0883 -0.234
(0.076) (0.147)
Mayor = Independent -0.344∗-0.131
(0.152) (0.253)
Mayor is from the same party -0.141†0.177
(0.079) (0.282)
Seats in councilt−1(log) -0.328∗∗∗ 0.273
(0.050) (0.175)
Control Variables
Year = 2014 -0.157 -0.261†-0.0946 -0.536 -0.425 0.487
(0.165) (0.139) (0.075) (0.761) (0.767) (0.388)
Year = 2015 0.000938 -0.307†-0.379∗∗ 0.195 -0.0380 0.724†
(0.161) (0.158) (0.116) (0.736) (0.758) (0.419)
Year = 2016 -0.106 -0.244†-0.179∗-0.414 -0.283 0.481
(0.158) (0.141) (0.079) (0.732) (0.743) (0.390)
Population (log) 0.0517 0.00675 0.0384 0.239†0.204†0.180
(0.060) (0.067) (0.064) (0.129) (0.122) (0.123)
Manifesto length (log) 0.146∗∗∗ 0.143∗∗∗ 0.199∗∗∗ -0.369∗-0.358∗-0.310∗
(0.039) (0.039) (0.040) (0.143) (0.142) (0.137)
Constant -0.740∗∗∗ -0.741∗∗∗ -0.846∗∗∗ 0.370∗∗∗ 0.369∗∗∗ 0.252∗∗∗
(0.036) (0.037) (0.048) (0.065) (0.065) (0.069)
Observations 478 478 397 478 478 397
Intraclass Correlation Coefficient .206 .170 .206 .015 .010 .000
Notes: Table displays results from three linear mixed-effects models for each of the two scaling methods
(Wordfish and Wordscores). Cluster-robust standard errors at the level of each city are shown in
parentheses. Reference category for parties is CDU/CSU. Reference category for party of the mayor is
CDU/CSU/SPD. Reference category for year effects is 2013.†p<0.10,∗p<0.05,∗∗ p<0.01,∗∗∗
p<0.001.
Our assumptions about the effects of previous experience in local politics also im-
ply that we can expect an interaction effect between parties and the number of seats
held in the previous city council. We have argued that the number of local politicians
should matter most when the number of local politicians is rather small. But some

determinants of local party positions 35
parties, most importantly SPD and CDU/CSU, most of the time hold a large number
of seats in the city council. Therefore, the effect should only be visible for smaller par-
ties. Bluntly put, we expect that parties are becoming more local when they possess
more local politicians. To test for this assumption, we estimated Model 3again, this
time including an interaction effect between parties and the logged number of seats
in the city council. The effects are displayed in Figure 5. We can clearly see how the
number of seats in the city council affects parties differently. In particular the Left
Party and the group of local parties show strong interaction effects, whereas other
parties show only weaker or no effects. Note that the x- and y-axis are different for
each plot. The x-axis only includes the observed range of seats for a party and the
y-axis is adjusted for each party to account for the different overall position on the
dimension. Hence, it is important to note that even after accounting for the number
of local politicians the Left Party remains the party with the lowest amount of relative
emphasis on local policy issues.14
14 The predicted position of the left party with more than 15 previous local council members is 0.5and
thus higher than the predicted value for all other parties with no previous local council members,
except for the heterogeneous group of ‘small parties’.

determinants of local party positions 36
Figure 5: Interaction between Parties and Number of Seats in the Previous City Council
Notes: Y-Axis displays predicted values on the Wordfish dimension based on a linear mixed-effects
model. Figure is based on Model 3from Table 1with an additional interaction between parties and
parties’ number of seats in the previous city council. 95% confidence intervals are shown as shaded
area around the predictions.
5.2 Explaining Left-Right Positions
While the previous analysis focused on a completely new dimension of local party
competition which has not been identified by existing research, we now turn to the
more established analysis of explaining general left-right positions of parties using
socio-economic and political variables as predictors. In the following, we estimate
identical models as for the Wordfish regression analysis, but this time we use the
Wordscores positions on the general left-right dimension as dependent variable. The
results are displayed in the right panel of Table 1.
The results demonstrate that only very few variables have a significant effect on
the placement of parties on the left-right dimension. Party dummies still capture the
main differences between the parties as could already been seen from the descriptive

conclusion 37
analysis. Almost all of the remaining variables are insignificant and only display
very weak effects. The unemployment rate is the only variable with a strong positive
effect, meaning that parties tend to be more right-wing in cities with higher levels
of unemployment. These results regarding the effect of unemployment rate on party
positions are in line with previous analyses of this type which also find that higher
unemployment rates are positively correlated with more right-wing positions (Giger,
Müller and Debus,2011;Müller,2013).
The absence of other effects should not necessarily be interpreted as a weakness
of the models. There are few reasons to assume why, for example, the party of the
mayor should have an effect on the left-right placement of parties. The same holds
true for the year dummies as we would not expect parties to be systematically more
left- or right-wing in certain years.
6 conclusion
Like their national and regional counterparts, political actors at the local level can
have different views on specific policy issues and they distinguish themselves based
on their programmatic orientations by joining different parties (or independent local
lists). Local politics gets more and more politicized, particularly in larger municipal-
ities (see, e.g., Egner,2015;Kjær and Elklit,2010), and local governments have con-
siderable powers in shaping the political agenda in municipalities and cities (see, e.g.,
Bäck,2005;Heinelt and Hlepas,2006). As recent research shows, the party affiliation
of local politicians does make a difference for policy preferences and policy outcomes
(see, e.g., Serritzlew,2003;Gerber and Hopkins,2011;Debus, Knill and Tosun,2013;

conclusion 38
Goerres and Tepe,2013). These results, however, heavily rely on the possibility to
accurately assess local actors’ policy positions.
In this paper, we presented two ways of supervised (Wordscores) and unsupervised
(Wordfish) approaches to estimate local party positions, solely using information from
local party manifestos. The approaches demonstrate that they can be used to extract
meaningful estimates of local party positions. Both methods have identified substan-
tially meaningful but different dimensions. Our results suggest that Wordscores can
be used to place parties on ideologically meaningful left-right dimensions (general,
economic, and cultural left-right). This is even more important as Wordfish does not
seem to be able to fulfill this task. Instead, Wordfish identifies a dimension which cap-
tures the amount of ‘national policies’ addressed in local party manifestos. Of course,
such a dimension is non-ideological in the sense that these differences do not directly
inform us about the ideological position of a party. However, as noted by Grimmer
and Stewart (2013,293), identifying such a non-ideological dimension is not a short-
coming of unsupervised scaling methods. Instead, it is a meaningful dimension with
a substantive, though non-ideological, meaning. Moreover, this dimension might pro-
vide more detailed insights into the dynamics of local party competition. A simple
interpretation of this dimension could be, for example, that it measures the “second-
orderness” of a local party manifesto. Recent debates about second-order elections
have discussed the question of to what extent sub-national elections can be classi-
fied as being “second-order” (Marien, Dassonneville and Hooghe,2015). Our results
can speak to this debate from a different perspective. Existing research on second-
order elections focuses on the motives of voters (Reif and Schmitt,1980), while our
results provide interesting insights from the perspective of the political supply-side
by showing that local parties not only address local policy issues but also topics that
are primarily related to regional, national and European politics. Just like some voters

References 39
only care about national politics, it seems like that some parties at the local level are
no different. Analyzing this aspect in more detail can provide new insights into the
question of why certain elections are perceived as more second-order than others.
As with any other scaling method, our results are at least partly subject to one’s
individual interpretation. This holds especially true for the case of local party mani-
festos as we have only very little chance to cross-validate our position estimates. It is
important, however, to highlight that this lack of cross-validation measures is one of
the fundamental motivations for this research project. Our article does not present an-
other method for extracting positions for a case where several other well-established
measures of party positions already exist. Instead, we applied methods for extracting
positions to a case where we currently lack systematic information on party positions.
Our regression analyses of these estimates demonstrate that – for the case of Word-
fish – almost all effects support our interpretation of the underlying dimension and –
for the case of Wordscores – that the effects resemble patterns observed in previous
studies focusing on different levels of the German political system.
More generally, we hope that our work inspires future research to explore other data
sources for estimating party positions in local politics. Our article only provides a first
step into this direction. It still remains an open question how much local positions
matter for policy outcomes. Our estimates allow for a more detailed analysis for
explaining local policy outcomes compared to other studies which have used party
positions from the regional or national level as proxies.
references
Agasoster, Bodil. 2001. A framework for analysing local party policy emphases in Scot-
land. In Estimating the Policy Positions of Political Actors, ed. Michael Laver. London
and New York: Routledge pp. 76–89.

References 40
Arzheimer, Kai. 2015. “The AfD: Finally a Successful Right-Wing Populist Eurosceptic
Party for Germany?” West European Politics 38(3):535–556.
Bakker, Ryan, Catherine de Vries, Erika Edwards, Liesbet Hooghe, Seth Jolly, Gary
Marks, Jonathan Polk, Jan Rovny, Marco Steenbergen and Milada Anna Vachudova.
2015. “Measuring party positions in Europe: The Chapel Hill expert survey trend
file, 1999–2010.” Party Politics 21(1):143–152.
Benoit, Kenneth, Drew Conway, Benjamin E. Lauderdale, Michael Laver and Slava
Mikhaylov. 2016. “Crowd-sourced Text Analysis: Reproducible and Agile Produc-
tion of Political Data.” American Political Science Review 110(2):278–295.
Benoit, Kenneth, Kohei Watanabe, Paul Nulty, Adam Obeng, Haiyan Wang, Benjamin
Lauderdale and Will Lowe. 2017.quanteda: Quantitative Analysis of Textual Data. R
package version 0.9.9-25.
URL: http://quanteda.io
Benoit, Kenneth and Michael Laver. 2006.Party Policy in Modern Democracies. London:
Routledge.
Benoit, Kenneth, Thomas Bräuninger and Marc Debus. 2009. “Challenges for Esti-
mating Policy Preferences: Announcing an Open Access Archive of Political Docu-
ments.” German Politics 18(3):441–454.
Bernauer, Julian and Thomas Bräuninger. 2009. “Intra-Party Preference Heterogeneity
and Faction Membership in the 15th German Bundestag: A Computational Text
Analysis of Parliamentary Speeches.” German Politics 18(3):385–402.
Brearey, Patricia. 1989. City Coalitions in West Germany: A Case-Study of Nordrhein-
Westfalen. In Political parties and coalitions in European local government, ed. Colin
Mellors and Bert Pijnenburg. London [u.a.]: Routledge pp. 275–300.
Bräuninger, Thomas and Marc Debus. 2012.Parteienwettbewerb in den deutschen Bun-
desländern. Wiesbaden: VS Verlag für Sozialwissenschaften.
Bräuninger, Thomas, Marc Debus and Jochen Müller. 2013. “Estimating Policy Posi-
tions of Political Actors Across Countries and Time.” MZES Arbeitspapiere - Working
Papers (153).
Bräuninger, Thomas and Nathalie Giger. 2016. “Strategic Ambiguity of Party Positions
in Multi-Party Competition.” Political Science Research and Methods (online first).
Budge, Ian, Hans-Dieter Klingemann, Andrea Volkens, Judith Bara and Eric Tanen-
baum, eds. 2001.Mapping Policy Preferences: Estimates for Parties, Electors, and Gov-
ernments 1945-1998. Oxford: Oxford University Press.
Bäck, Hanna. 2003. “Explaining and predicting coalition outcomes: Conclusions from
studying data on local coalitions.” European Journal of Political Research 42(4):441–472.

References 41
Bäck, Hanna. 2008. “Intra-Party Politics and Coalition Formation: Evidence from
Swedish Local Government.” Party Politics 14(1):71–89.
Bäck, Henry. 2005. The institutional setting of local political leadership and commu-
nity involvement. In Urban Governance and Democracy, ed. Michael Haus, Hubert
Heinelt and Murray Stewart. London, New York: Routledge pp. 65–101.
Clark, Alistair and Lynn Bennie. 2016. “Parties, mandates and multilevel politics:
Subnational variation in British general election manifestos.” Party Politics (online
first).
Debus, Marc, Christoph Knill and Jale Tosun. 2013. “Registration Fees for Same-
Sex Unions, Local Party Politics and Societal Demand.” Local Government Studies
39(6):756–776.
Debus, Marc and Martin Gross. 2016. “Coalition formation at the local level: Institu-
tional constraints, party policy conflict, and office-seeking political parties.” Party
Politics 22(6):835–846.
Denters, Bas. 1985. “Towards a Conditional Model of Coalition Behaviour.” European
Journal of Political Research 13(3):295–309.
Däubler, Thomas. 2012. “The Preparation and Use of Election Manifestos: Learning
from the Irish Case.” Irish Political Studies 27(1):51–70.
Däubler, Thomas, Kenneth Benoit, Slava Mikhaylov and Michael Laver. 2012. “Natural
Sentences as Valid Units for Coded Political Texts.” British Journal of Political Science
42(4):937–951.
Egner, Björn. 2015. “Parliaments in Disguise? How German Councillors perceive
Local Councils.” Local Government Studies 41(2):183–201.
Egner, Björn, David Sweeting and Pieter-Jan Klok, eds. 2013.Local Councillors in Europe.
Wiesbaden: Springer VS.
Flentje, Jan-Erik, Thomas König and Moritz Marbach. 2017. “Assessing the Validity
of the Manifesto Common Space Scores.” Electoral Studies 47:25–35.
Fried, Robert C. 1976. “Party and Policy in West German Cities.” American Political
Science Review 70(1):11–24.
Gerber, Elisabeth R and Daniel J Hopkins. 2011. “When Mayors Matter: Estimating the
Impact of Mayoral Partisanship on City Policy.” American Journal of Political Science
55(2):326–339.
Giger, Nathalie, Jochen Müller and Marc Debus. 2011. “Die Bedeutung des regionalen
Kontexts für die programmatische Positionierung von Schweizer Kantonalparteien.”
Swiss Political Science Review 17(3):259–285.

References 42
Goerres, Achim and Markus Tepe. 2013. “Für die Kleinen ist uns nichts zu teuer?
Kindergartengebühren und ihre Determinanten in Deutschlands 95 bevölkerungsre-
ichsten Städten zwischen 2007 und 2010.” dms - der moderne Staat 6(1):169–190.
Grimmer, Justin and Brandon M. Stewart. 2013. “Text as Data: The Promise and
Pitfalls of Automatic Content Analysis Methods for Political Texts.” Political Analysis
21(3):267–297.
Gunlicks, Arthur B. 1986.Local Government in the German Federal System. Durham:
Duke Univ. Press.
Harmel, Robert, Alexander C Tan, Kenneth Janda and Jason Matthew Smith. 2016.
“Manifestos and the two faces of parties.” Party Politics (Online First).
Heinelt, Hubert and Nikolaos-K Hlepas. 2006. Typologies of Local Government Sys-
tems. In The European Mayor., ed. Henry Bäck, Hubert Heinelt and Annick Magnier.
Wiesbaden: VS Verlag für Sozialwissenschaften pp. 21–42.
Hjorth, Frederik, Robert Klemmensen, Sara Hobolt, Martin Ejnar Hansen and Peter
Kurrild-Klitgaard. 2015. “Computers, coders, and voters: Comparing automated
methods for estimating party positions.” Research & Politics April-June.
Jankowski, Michael, Sebastian Schneider and Markus Tepe. 2016. “Ideological alter-
native? Analyzing Alternative für Deutschland candidates’ ideal points via black
box scaling.” Party Politics (online first).
Kjær, Ulrik and Jorgen Elklit. 2010. “Party politicisation of local councils: Cultural
or institutional explanations for trends in Denmark, 1966-2005.” European Journal of
Political Research 49(3):337–358.
Klemmensen, Robert, Sara Binzer Hobolt and Martin Ejnar Hansen. 2007. “Estimating
policy positions using political texts: An evaluation of the Wordscores approach.”
Electoral Studies 26(4):746–755.
Klingelhöfer, Tristan and Jochen Müller. 2015. “Consociational rational coalitions:
Norm-based government formation and the case of the Dutch provinces.” Acta Po-
litica 50(1):101–124.
Klingemann, Hans-Dieter, Andrea Volkens, Judith Bara, Ian Budge and Michael D.
McDonald, eds. 2006.Mapping Policy Preferences II: Estimates for Parties, Electors,
and Governments in Eastern Europe, European Union and OECD, 1990-2003. Oxford:
Oxford University Press.
Laver, Michael. 2014. “Measuring Policy Positions in Political Space.” Annual Review
of Political Science 17:207–223.
Laver, Michael, Colin Rallings and Michael Thrasher. 1998. “Policy Payoffs in Local
Government.” British Journal of Political Science 28(2):333–353.

References 43
Laver, Michael, Kenneth Benoit and John Garry. 2003. “Extracting Policy Posi-
tions from Political Texts Using Words as Data.” American Political Science Review
97(2):311–331.
Lo, James, Sven-Oliver Proksch and Jonathan B. Slapin. 2014. “Ideological Clarity
in Multiparty Competition: A New Measure and Test Using Election Manifestos.”
British Journal of Political Science .
Loughlin, John, Frank Hendriks and Anders Lidström, eds. 2011.The Oxford Handbook
of Local and Regional Democracy in Europe. Oxford: Oxford University Press.
Lowe, Will. 2008. “Understanding Wordscores.” Political Analysis 16(4):356–371.
Lucas, Christopher, Richard A. Nielsen, Margaret E. Roberts, Brandon M. Stewart,
Alex Storer and Dustin Tingley. 2015. “Computer-Assisted Text Analysis for Com-
parative Politics.” Political Analysis 23(2):254.
Mair, Peter. 2001. Searching for the positions of political actors. A review of ap-
proaches and a critical evaluation of expert surveys. In Estimating the Policy Positions
of Political Actors, ed. Michael Laver. London and New York: Routledge pp. 10–30.
Marien, Sofie, Ruth Dassonneville and Marc Hooghe. 2015. “How Second Order
Are Local Elections? Voting Motives and Party Preferences in Belgian Municipal
Elections.” Local Government Studies 41(6):898–916.
Merz, Nicolas, Sven Regel and Jirka Lewandowski. 2016. “The Manifesto Corpus: A
new resource for research on political parties and quantitative text analysis.” Re-
search & Politics 3(2).
Mikhaylov, Slava, Michael Laver and Kenneth Benoit. 2012. “Coder Reliability and
Misclassification in the Human Coding of Party Manifestos.” Political Analysis
20(1):78–91.
Mueller, Wolfgang C. 2009. Government Formation. In The SAGE Handbook of Com-
parative Politics, ed. Todd Landmann and Neil Robinson. Los Angeles, CA: SAGE
pp. 227–245.
Müller, Jochen. 2009. “The Impact of the Socio-Economic Context on the Länder
Parties’ Policy Positions.” German Politics 18(3):365–384.
Müller, Jochen. 2013. “On a Short Leash? Sub-National Party Positions between
Regional Context and National Party Unity.” Journal of Elections, Public Opinion and
Parties 23(2):177–199.
Nyhuis, Dominic. 2017. “Estimating policy positions of local parties in elections with
multi-vote ballots.” Local Government Studies 43(3):475–498.

References 44
Pappi, Franz Urban and Susumu Shikano. 2004. “Ideologische Signale in den Wahlpro-
grammen der deutschen Bundestagsparteien 1980 bis 2002.” MZES Working Papers,
Mannheimer Zentrum für Europäische Sozialforschung (76).
Proksch, Sven-Oliver and Jonathan B. Slapin. 2009. “How to Avoid Pitfalls in Statistical
Analysis of Political Texts: The Case of Germany.” German Politics 18(3):323–344.
Reif, Karlheinz and Hermann Schmitt. 1980. “Nine Second-Order Elections: A Con-
ceptual Framework for the Analysis of European Election Results.” European Journal
of Political Research 8(1):3–44.
Reiser, Marion and Everhard Holtmann, eds. 2008.Farewell to the Party Model? Inde-
pendent Local Lists in East and West European Countries. Wiesbaden: VS Verlag für
Sozialwissenschaften.
Serritzlew, Soren. 2003. “Shaping Local Councillor Preferences: Party Politics, Com-
mittee Structure and Social Background1.” Scandinavian Political Studies 26(4):327–
348.
Serritzlew, Søren, Asbjorn Skjæveland and Jens Blom-Hansen. 2008. “Explaining Over-
sized Coalitions: Empirical Evidence from Local Governments.” The Journal of Leg-
islative Studies 14(4):421–450.
Serritzlew, Søren, Jens Blom-Hansen and Asbjorn Skjæveland. 2010. “Portfolio Allo-
cation or Policy Horizons? Determinants of Coalition Formation in Danish Local
Government.” Local Government Studies 36(6):843–866.
Skjæveland, Asbjorn, Søren Serritzlew and Jens Blom-Hansen. 2007. “Theories of
coalition formation: An empirical test using data from Denmark.” European Journal
of Political Research 46(5):721–745.
Slapin, Jonathan B. and Sven-Oliver Proksch. 2008. “A Scaling Model for Estimat-
ing Time-Series Party Positions from Texts.” American Journal of Political Science
52(3):705–722.
Steunenberg, Bernard. 1992. “Coalition theories: Empirical evidence for Dutch munic-
ipalities.” European Journal of Political Research 22(3):245–278.
Wilkerson, John and Andreu Casas. 2017. “Large-Scale Computerized Text Analysis in
Political Science: Opportunities and Challenges.” Annual Review of Political Science
20:230–251.