The Ties That Bind: Text Similarities and Conditional Diffusion among Parties

Abstract

Comparative analyses of party policy diffusion are only just emerging. To better understand the conditions under which diffusion occurs, this article argues that three heuristics – availability, representativeness and anchoring – shape parties' efforts to gather information (from elsewhere), leading to differing diffusion effects. The study operationalizes the outcome as textual similarity of party manifestos in nineteen Western democracies from 1960 to 2016, applying a text-as-data approach and machine translation. Analyzing dyads, it assesses how commonalities and sender/receiver attributes impact diffusion. It finds that there is little room for cross-border diffusion as successful parties stick to their old program. Beyond the still-prevailing domestic context, ‘learning from cultural reference groups’ in a region is most important. In addition, diffusion appears within EP factions and transnational party organizations independently of the success/loss of the sender. The analysis thus sheds light on (un-)favorable conditions for party policy diffusion and paves the way for future studies applying machine translation and quantitative text analyses.

Full text

ARTICLE
The Ties That Bind: Text Similarities and Conditional
Diffusion among Parties
Nils Düpont1* and Martin Rachuj2
1
Collaborative Research Center 1342, Global Dynamics of Social Policy, University of Bremen, Germany and
2
Department of
Political Science & Communication Studies, University of Greifswald, Germany
*Corresponding author. E-mail: duepont@uni-bremen.de
(Received 20 February 2020; revised 9 July 2020; accepted 9 September 2020)
Abstract
Comparative analyses of party policy diffusion are only just emerging. To better understand the conditions
under which diffusion occurs, this article argues that three heuristics –availability, representativeness and
anchoring –shape parties’efforts to gather information (from elsewhere), leading to differing diffusion
effects. The study operationalizes the outcome as textual similarity of party manifestos in nineteen
Western democracies from 1960 to 2016, applying a text-as-data approach and machine translation.
Analyzing dyads, it assesses how commonalities and sender/receiver attributes impact diffusion. It
finds that there is little room for cross-border diffusion as successful parties stick to their old program.
Beyond the still-prevailing domestic context, ‘learning from cultural reference groups’in a region is
most important. In addition, diffusion appears within EP factions and transnational party organizations
independently of the success/loss of the sender. The analysis thus sheds light on (un-)favorable conditions
for party policy diffusion and paves the way for future studies applying machine translation and quanti-
tative text analyses.
Keywords: diffusion; party policy; text similarity; text-as-data approach; machine translation
Approaching Party Policy Diffusion
In the 1990s the ‘Third Way’swept through Europe, fundamentally transforming Social
Democratic parties. The emergence of green issues in the 1970s and 1980s had a profound impact
on party manifestos in the Western world. Exchange among parties is quite an old phenomenon;
they are more than simply functional responses. European parties in particular have a long his-
tory of cross-country interactions, starting with the First International in the 1860s. Closely
aligned with the development of the European Union (EU), the professionalization of trans-
national party cooperation led to institutionalized platforms for the cross-border exchange of
ideas, political guidelines and strategies. In addition to such ‘landmark events’, there are also
smaller-scale anecdotes about party policy diffusion. For instance, the Norwegian Kristelig
Folkeparti was a pioneer for Christian Democratic parties in Scandinavia, and the Finnish
Christian League’s first electoral program turned out to be ‘virtually a verbatim translation of
that of the Norwegian Christians’(Arter 1980, 146).
Comparative analyses of party policy diffusion, however, gained momentum rather late: emer-
ging from studies about ‘what moves parties?’(cf. Adams 2012), Böhmelt et al. found that parties
respond to the left–right positions of (larger) governing parties in foreign countries (Böhmelt
et al. 2016; Böhmelt et al. 2017). Still, they concluded that future studies should ‘identify
© The Author(s) 2021. Published by Cambridge University Press. This is an Open Access article, distributed under the terms of the Creative
Commons Attribution-NonCommercial-ShareAlike licence (http://creativecommons.org/licenses/by-nc-sa/4.0/), which permits non-commer-
cial re-use, distribution, and reproduction in any medium, provided the same Creative Commons licence is included and the original work is
properly cited. The written permission of Cambridge University Press must be obtained for commercial re-use.
British Journal of Political Science (2021), page 1 of 18
doi:10.1017/S0007123420000617

conditions under which party-policy diffusion effects are stronger or weaker’(Böhmelt et al. 2016,
407; emphasis in original).
In this vein, we argue that parties adapt ideas, rhetoric or style from other party manifestos,
resulting in increased text similarity. How similar they are, however, depends on which instances
are considered, which is in turn shaped by three heuristics: availability, representativeness and
anchoring. Applying heuristics means effort reduction, examining fewer cues and ‘integrating
less information’(Shah and Oppenheimer 2008, 209). Commonalities –the ties that bind –affect
availability, while vote gains or losses, a sender and receiver attribute, mirror representativeness
and anchoring. Differing diffusion effects occur if parties apply these heuristics when filtering
relevant instances.
Building on a basic definition of diffusion, we shed light on how commonalities and heuristics
are interlinked. Methodologically, we opt for a dyadic analysis, assessing the impact of linkages
and sender/receiver attributes on text similarity as the outcome of diffusion. Moving beyond
left–right positions, we choose a text-as-data approach and apply machine translation to estimate
the textual similarity of multilingual party manifestos in nineteen Western democracies from
1960 to 2016 based on the Manifesto Corpus (Krause et al. 2018). Testing different ties, our
results uncover (un-)favorable conditions for party policy diffusion. We find there is less room
for ideas from abroad as parties stick to their old program –especially if they gained votes.
Unexpectedly, cross-border diffusion is also less relevant when parties lost, despite being in
need (‘anchoring’). Beyond the still prevailing domestic impact, text similarity is highest
among parties from the same ‘family of nations’(Castles 1998). What Simmons and Elkins
(2004, 175) labeled ‘learning from cultural reference groups’mirrors a level of ‘effort reduction’
that particularly eases diffusion. Diffusion from and among government parties is partly driven
by the representativeness heuristic –that is, text similarity is higher if the sender gained votes.
Both are of minor importance though. Instead, diffusion takes place within factions in the
European Parliament (EP) and in transnational party organizations; but ‘success’does not matter
for diffusion among rather ideologically like-minded parties.
While our focus here is on manifestos as a whole, the ‘surface’of texts, we also show the poten-
tial of quantitative text analysis and machine translation paving the way for prospective analyses
of the content that diffuses. Finally, our results suggest a stronger focus on regional party cooper-
ation for unfolding the pathways of diffusion among parties.
Diffusion Among Parties –A Dyadic Approach
Rogers’(2003, 5) famous definition of diffusion is our cornerstone for approaching diffusion
among parties: ‘the process in which an innovation is communicated through certain channels
over time among the members of a social system’. Adapting the definition to political parties
and party competition has several implications. It puts the spotlight on parties being both a
sender and receiver at times, with an exchange taking place among parties that are somehow
‘connected’, that is, within dyads sharing a tie.
1
Inherently, the exchange of ideas entails a process
of convergence (Rogers 2003,5–6), and we cannot observe diffusion per se but only its outcome
(here, the pairwise similarity of manifestos). The outcome, in turn, is affected by commonalities;
and certain characteristics of the sender and receiver may ease or hinder the spread of ideas
(Rogers 2003,15–16).
Besides, people and organizations often rely on heuristics when processing information and
making decisions (Gigerenzer and Gaissmaier 2011, 451). Applying heuristics means examining
fewer cues and alternatives, applying less effort to retrieve and weigh information, and ‘integrat-
ing less information’(Shah and Oppenheimer 2008, 209). Combining both perspectives, on an
1
A different approach to capturing diffusion relies on estimating spatial lags; see Neumayer and Plümper (2016), however,
for the meaning and difficulties associated with specifying the weighing matrix W.
2 Nils Düpont and Martin Rachuj

abstract level one may hypothesize that the lower the effort expended to recognize and process
information from other parties (elsewhere), the more likely a diffusion of ideas, rhetoric or
style becomes.
More specifically, we argue that the ties that bind and vote gains/losses (that is, sender/receiver
attributes) resemble the availability, representativeness and anchoring heuristic that shape diffu-
sion. All three, though to different degrees, help reduce efforts by affecting the visibility or imme-
diateness of relevant instances, the ease of adaptation or the wealth of information that needs to
be processed. ‘Filtering’which instances are taken into account (and which are ignored), they lead
to differing diffusion effects.
Regarding commonalities, we will consider the role of several ties. In line with what Simmons
and Elkins (2004,175)labeled‘learning through communication’, we look at governing parties
abroad, EP factions and transnational party organizations. Echoing their ‘learning from cultural ref-
erence groups’, we consider Castles’(1998)‘family of nations’. Finally, arguing for a diffusional per-
spective even on domestic party competition, we take competitors and a party’s past into account.
A Diffusional Perspective on Party Competition
Previous research on party competition and party policy change has almost exclusively focused
on the domestic context. While there is abundant evidence and spatial theories on (rational)
party behavior, it has seldom been viewed from a diffusional perspective. This is despite ‘diffu-
sional’evidence: parties respond to competitors’moves at the last election (Adams and
Somer-Topcu 2009; Williams 2015); they adapt ideas to fight ‘newcomers’(Meguid 2008); like
a‘predator’, they exploit ‘successful choices made by other agents’(Laver and Sergenti 2012,
134); and they ‘adjust their issue attention in response to (lagged) changes in other parties’
issue attention’(Green-Pedersen and Mortensen 2015, 748). Admittedly, the rationale for observ-
ing and responding to other parties is different for the domestic context: parties compete with
each other for votes, and they do not compete with other parties elsewhere. Party policy makers
may therefore have a different motivation to adopt ideas, rhetoric, or style from one sender or the
other. Yet each of the above-mentioned findings resembles Roger’s fundamental definition of dif-
fusion processes, backing the idea that linkages and sender/receiver attributes matter even in the
domestic context.
But as Kayser (2007) has convincingly argued, party competition is no longer purely domestic.
Parties respond to external impacts like economic globalization (for example, Adams, Haupt and
Stoll 2009; Ezrow and Hellwig 2014; Haupt 2010), and a ‘Europeanization’of party politics is tak-
ing place (Nanou and Dorussen 2013; Somer-Topcu and Zar 2014). Consequently, Böhmelt et al.
find indications of ‘party policy diffusion’, that is, parties respond to the left–right positions of
(larger) governing parties in foreign countries (Böhmelt et al. 2016; Böhmelt et al. 2017). This,
again, resembles the basic definition: diffusion is more likely if a (larger) governing party (that
is, a sender attribute) is ‘connected’(that is, shares a tie) with other parties.
To uncover (un-)favorable conditions for party policy diffusion, we take a dyadic approach
proposed for analyzing the diffusion of public policies (Gilardi and Füglister 2008; Volden
2006). It is advantageous because each entity ‘is, in turn, allowed to be the potential “receiver”
and “sender”of a policy, and independent variables can measure the characteristics of both
[…] as well as their relationships’(Gilardi and Füglister 2008, 415). Arguing that party compe-
tition can fruitfully be seen from a diffusional perspective brings us back to our cornerstone and
the question of how heuristics are interlinked with commonalities and sender/receiver attributes.
A Dyadic Approach to Diffusion among Parties
In line with Budge (1994, 452), we assume that parties act in an uncertain environment. Like any
decision maker, parties can gather information from the past and elsewhere to look for
British Journal of Political Science 3

inspiration and ideas, ‘draw their lessons’(Rose 1991) and thus reduce uncertainty about the
‘best’strategy or policy offer. If parties were rational learners, they would collect information
about all alternatives, and have full analytical capabilities and resources. These conditions are
rarely met though, and people and organizations have been found to often rely on heuristics –
a strategy ‘that ignores part of the information’and reduces effort (Gigerenzer and Gaissmaier
2011, 454). In an uncertain environment with limited resources, applying heuristics can be an
efficient solution though, leading to ‘satisficing decisions’(Simon 1993, 397–98). Typically,
three heuristics matter for (policy) learning (Meseguer 2006, 41; Simmons and Elkins 2004,
175; Tversky and Kahneman 1973; Tversky and Kahneman 1974): availability, representativeness
and anchoring. Each is discussed in more detail below.
(1) The availability heuristic: Commonalities resemble the availability or perceptibility of
information, that is, ‘the ease with which relevant instances come to mind’(Tversky
and Kahneman 1973, 207). Parties applying this heuristic are ‘not guided by what they
are able to compute, but by what they happen to see at a given moment’(Kahneman
2003, 1469). They would thus favour immediate, ‘visible’cues even if limited (Shah
and Oppenheimer 2008, 214). If domestic competitors come to mind more easily than
ideas from governing parties elsewhere, the former should have a stronger impact on
the outcome. Likewise, if information from members of the same EP faction is easier
to gather and adapt than information from any other party with which the focal party
has nothing in common, diffusion should be greater among the former. But even if infor-
mation easily comes to mind and the number of instances is narrowed down through
commonalities, certain attributes may further condition diffusion.
(2) Applying the representativeness heuristic means overemphasizing information from ‘suc-
cessful’parties –an attribute of the sender. The logic is that ‘observers expect the statistics
of a sample to closely resemble (or “represent”) the corresponding population parameters,
even when the sample is small’(Kahneman and Frederick 2002, 49). A victory or loss is an
observable signal of a relative (dis-)advantage of a manifesto. Despite not being represen-
tative of all alternatives, parties considering only such instances further reduce their efforts
(Shah and Oppenheimer 2008, 215), making diffusion more likely.
(3) Finally, applying the anchoring heuristic means parties adjust their evaluation of alterna-
tives in light of a salient and accessible value –an anchor (Shah and Oppenheimer 2008,
211). The most salient trigger for parties is surely electoral defeat (Mair 1983, 408;
Panebianco 1988, 243), which is a receiver attribute. Simply ‘recycling’the same manifesto
with which a party lost seems a rather bad idea. Conversely, there is less need to change a
successful manifesto, so –consequently –there is less room for diffusion. Applying this
heuristic further reduces a party’s effort by eliminating alternatives (Shah and
Oppenheimer 2008, 214).
In short, heuristics lead to ‘heterogeneity in exposure’and ‘heterogeneity in responsiveness’
(Neumayer and Plümper 2012). As a result, differing party policy diffusion effects occur when
heuristics guide parties in their search for information and their efforts to reduce uncertainty.
As commonalities, vote gains of the sender and losses of the receiver make diffusion more likely,
we expect text similarity to be higher if parties share a tie while being conditional on the perform-
ance of the sending or receiving party.
Differing Characteristics of Commonalities, Differing Diffusion Effects
Applying heuristics means reduced effort, examining fewer cues and taking less information into
account. While a dyadic setup allows us to explicitly test the impact of commonalities, party pol-
icy diffusion research is still in its infancy regarding linkages. We therefore consider six ties,
4 Nils Düpont and Martin Rachuj

mostly drawing on previous research: (1) a party’s past, (2) domestic competitors, (3) governing
parties abroad, (4) EP factions, (5) transnational party organizations and (6) other parties of the
‘same family of nations’. Each commonality affects the availability or ‘ease of coming to mind’,
the amount of information that needs to be processed and the ease of adaptation in terms of com-
parability. Each tie features characteristics that –more or less –reduce a party’s effort when con-
sidering ‘only’those instances rather than information from all (unrelated) parties.
(1) To start with, the easiest and cheapest approach is for a party to simply ‘recycle’its last
program rather than engage in a purposive search for information. It is a prime example
of the availability heuristic, most likely being the first instance that comes to mind. In this
way, parties also ensure their coherence and secure ‘their’issue ownership (Budge 2015).
Little can be learned from a single instance, though; especially not if the party lost.
Mirroring the anchoring heuristic, a party’s willingness to change indeed seems higher
after losses (Schumacher et al. 2015; Somer-Topcu 2009). If parties were more open to
diffusion in these circumstances, text similarity should be higher for any other tie.
(2) Broadening the view, the next instances that come to mind are domestic competitors.
There is ample evidence that parties observe and respond to each other. The domestic
context is identical, and parties ‘know’each other, which makes it quite easy to adapt
ideas, rhetoric or style. Yet some effort is required as the number of instances increases.
And adaptation is recommended for reasons of credibility and avoiding the image of a
‘poor copycat’(cf. Budge 1994; Laver and Sergenti 2012).
It is easy to see an ordering of ‘recycling’and domestic competitors, with the former presumably
having the strongest impact on text similarity (unless a party completely rewrites its entire mani-
festo). Turning to cross-border ties, this is less predictable. Undoubtedly, the efforts increase once
broadening the view beyond borders: the number of instances explodes when considering all
other parties elsewhere and the compatibility of contexts declines. Language borders arise, and
other electoral systems with different imperatives (Cox 1990), diverging party systems (Lipset
and Rokkan 1967) and different institutions (Castles 1998) need to be considered when thinking
about adapting ideas, rhetoric or style. For this reason, any commonality among parties helps
reduce the efforts by filtering and ‘ignoring’parties with which the focal party has nothing in
common, and instead learning and adapting from those with which they share a tie. Inspired
by previous research, we consider four additional, cross-border ties.
(3) Böhmelt et al. (2016) found that parties respond to the left–right positions of governing
parties elsewhere, arguing that parties focus on incumbency as a signal of ‘success’.
Resembling the representativeness heuristic, in light of our dyadic approach this would
mean diffusion from any governing party to everyone else. Yet an alternative reading
would be diffusion among governing parties. The idea is that ‘[ f ]requent intergovernmen-
tal meetings at multiple official levels can transmit information to policy makers about
“what works”in other settings’(Simmons and Elkins 2004, 175), particularly in the
European context with regular and formalized meetings. In this view, ‘being in govern-
ment’creates a tie and ‘success’becomes a sender attribute. While it certainly reduces
efforts, such meetings bring together ideologically diverse parties from heterogeneous
contexts, and communication is likely about anything but party policies.
(4) Factions in the EP depict another link (Senninger, Bischof and Ezrow 2020). Exposed to
members of like-minded parties from abroad, EP factions –already starting in the 1950s
in the Common Assembly –represent an institutionalized platform for the exchange of
information. Being a long-term member of an EP faction increases the likelihood of dif-
fusion as regularity and stability in communication reduces uncertainty and deepens the
British Journal of Political Science 5

trustworthiness of the sender. Although member parties still have heterogeneous back-
grounds, adaptation becomes easier through at least partial ideological congruence.
(5) The development of European parties closely resembles European integration (Ladrech
2006; Mittag 2006) but is not limited to it. Alongside integration, transnational party orga-
nizations (TPOs) have been founded. Despite some overlap with EP factions, these are not
restricted to EU members (for example, the US Democrats and Conservatives became
members of the European Democrat Union in 1982). Furthermore, parties may join dif-
ferent TPOs. Three Christian Democratic organizations, for example, existed side by side
throughout the 1970s and 1980s; some parties (like the German CDU/CSU) were mem-
bers of all three of them. Several TPOs exist, often along vaguely defined lines of party
families, linking parties from around the globe. These federations provide institutionalized
channels for the diffusion of ideas, platforms and strategies as well. While EP factions may
be founded for reasons other than ‘like-mindedness’(for example, access to resources),
TPOs operate at a more ideational level, which simplifies adaptation.
(6) Previous cross-border ties reflect ‘learning through communication’(Simmons and Elkins
2004, 175), while public policy studies often found a geographic pattern (Meseguer 2006,
41) as proximity and similarity in contexts are favoured in policy learning (Rose 1991,
13–15). The reason is that ‘[l]earning takes place at least partially through analogy, and
lessons are viewed as more relevant the extent to which a foreign case is viewed as
analogous’(Simmons and Elkins 2004, 175). Geographic proximity has been used as
an indicator as often as it has been criticized as an oversimplification (Maggetti and
Gilardi 2016). We therefore return to Castles’(1998) notion of ‘families of nations’to cap-
ture ‘learning from cultural reference groups’(Simmons and Elkins 2004, 175). Parties in
the same region presumably come to mind first after leaving the domestic context.
Recognizing and carrying over insights from regional parties is easier precisely because
of their countries’proximity, common history, linguistic affinity and similarities of the
political system. This commonality may therefore play a particular role in party policy dif-
fusion providing the right amount of effort reduction vs. wealth of information to still
arrive at ‘satisficing decisions’.
In summary, we started with a famous definition of diffusion, arguing that even domestic party
competition can be seen from a diffusional point of view. This perspective sheds light on the role
of heuristics and commonalities. Applying heuristics helps parties reduce efforts when gathering
and processing information from other parties (elsewhere). As they shape which instances are
taken into account, and which parties are ‘ignored’, the availability heuristic in particular is inter-
linked with the ties that bind. Resembling the representativeness and anchoring heuristic, vote
gains or losses of the sender and receiver may further condition diffusion.
As party policy diffusion research is still in its infancy regarding linkages, a dyadic approach is
well suited to test the impact of commonalities and sender/receiver attributes on the outcome of
diffusion, the pairwise text similarity. We identified two domestic and four cross-border ties that
will be put to the test. Each commonality features certain characteristics that –more or less –
reduce a party’s effort. Thus heuristics lead to heterogeneity in exposure and responsiveness
and, consequently, to differing diffusion effects. While it is partly indeterminate which ties are
more important, we can nevertheless summarize the following expectations:
(1) Availability: In principle, we expect a higher level of text similarity for domestic ties than
cross-border ones; still, we expect some ordering of the latter depending on the (dis-)simi-
larity of contexts and the ‘amount’of effort reduction it reflects.
(2) Representativeness: More is adapted, and text similarity is higher, if the sender gained
votes.
6 Nils Düpont and Martin Rachuj

(3) Anchoring: There is less space for (cross-border) diffusion if a party previously won and
more if they lost.
In this way, our analysis contributes to better understanding the ‘rationale’and conditions of
party policy diffusion effects. Comparing party manifestos across time and space is not an
easy task though, as we have to overcome language barriers. For this reason, we now present
our methodological approach to assessing whether diffusion accounts for the (dis-)similarity of
election manifestos.
Measuring Text Similarity, Ties and Attributes
Previous research on party policy diffusion has provided valuable insights using left–right posi-
tions (for example, Adams and Somer-Topcu 2009; Böhmelt et al. 2016; Williams 2015), more
specifically the well-known RILE (Budge and Klingemann 2001). In order to move forward,
we present a text-as-data approach. This method is superior to left–right scores in several
respects. As Benoit, Laver and Mikhaylov (2009, 497–99) note, the Manifesto group’s coding
scheme per se and any derived index are but one of many possible realizations coded from
the actual text. Sticking to the ‘text generated by the authors’helps us to shorten the pathway
of inference by circumventing all issues related to the measurement instrument, the coding
and the scaling (Benoit, Laver and Mikhaylov 2009, 498). For example, parties may be mislead-
ingly similar on a left–right axis but differ in their framing, as became apparent in the debate on
welfare chauvinism (Schumacher and van Kersbergen 2016). To illustrate our point, consider
Australia’s National Party 2010, which may have been influenced by New Zealand’s
Progressive Party and its 2008 manifesto (≈regional diffusion). Both score very similarly on
the RILE (−0.47 and 0); when looking at the distribution across the categories that make up
the RILE, both address very different topics, however. Unsurprisingly, we find a very low text
similarity of 0.19 on a scale from 0 to 1; the mean within the English-speaking family of nations
is 0.43 (SD = 0.15), with an overall mean of 0.27 (SD = 0.13).
2
Above all, since left–right scores are
based on pre-selected categories, they miss the diffusion of new issues and further aspects of man-
ifestos such as references to the past and future, to track records or policy pledges (Dolezal et al.
2018; Jahn 2014).
For this reason, we use the Manifesto Corpus (Krause et al. 2018) and look at party programs
in nineteen Western countries.
3
Still, we must overcome language barriers before we can compare
text similarity. Machine translation has matured as a feasible option for comparative analyses but
has been seldom applied.
4
As a pilot, our analysis shows the potential of text-as-data approaches
combined with machine translation to gain insights into party policy diffusion, and paves the way
for prospective analyses of the content that diffuses.
The Dependent Variable: Estimating Pairwise Text Similarities
We define our dependent variable as the text similarity of both party manifestos within a dyad.
Every dyad resembles the most recent information available to the focal party since its last elec-
tion, making the dyad directed. We thus assume, for example, that Swedish parties competing in
the 2010 election would only look at instances that became available during their inter-election
2
There is a negative but weak correlation of text similarity and the absolute distance of RILE positions (Pearson’sr=
−0.16). Including it in our regressions as a sensitivity check does not alter the results, which supports our notion that
text-as-data approaches are better suited to analyzing diffusion among parties (cf. Appendix).
3
Namely: Australia, Canada, Ireland, New Zealand, United Kingdom, United States, Denmark, Finland, Norway, Sweden,
Austria, Belgium, France, Germany, Italy, Netherlands, Switzerland, Portugal and Spain. Due to data availability, we had to
omit Iceland, Luxembourg and Greece.
4
But see Proksch et al. (2019), who use machine-translated dictionaries to uncover sentiment in legislative speeches.
British Journal of Political Science 7

period since 2006 and with which they share a tie. Representing a ‘bag-of-words approach’
(Grimmer and Stewart 2013), we choose cosine similarity as our measure.
5
Derived as the cosine
of the angle of two vectors, its metric is on a familiar scale from 0 to 1; it takes the overlap and
frequency of words into account, and it is independent from document length (Bär, Zesch and
Gurevych 2015, 16; Huang 2008, 51). To estimate text similarities, we use machine translation
to create one document-feature matrix (DFM) in English that contains all party manifestos.
We then apply common procedures to construct a DFM (Reber 2018): we first split n-grams
into unigrams that may be present from translating compound words because ‘n-grams do little
to enhance performance’(Grimmer and Stewart 2013, 272). Secondly, we remove English ‘stop-
words’, as they –by definition –do not contain topical content (de Vries, Schoonvelde and
Schumacher 2018, 421). Thirdly, we trim the DFM, removing terms that occur in less than 1
and more than 99 per cent of the documents (Grimmer and Stewart 2013, 273). Finally, we nor-
malize the DFM, turning term occurrences into relative frequencies (Welbers, van Atteveldt and
Benoit 2017, 253–54), and opt for cosine similarity as our dependent variable.
Constructing such a DFM in English is not trivial, given the many languages in which party
manifestos are written. Therefore, we now briefly present our translation approach that leads to
this DFM. Afterwards we operationalize our independent variables.
Excursus: A Cost-Effective Approach for Overcoming Language Barriers
Cross-lingual analyses still face the obstacle that one would need a large amount of resources to
translate all documents into one common language for analysis. Professional human translation is
seldom feasible, given the amount of text and assets it would require. Machine translation is thus
an alternative. It is cheaper (although not free) and faster, and recent advances make it a viable
option. It has also been noted that working with translated DFMs is sufficient for analytical pur-
poses in most cases (de Vries, Schoonvelde and Schumacher 2018; Lucas et al. 2015; Reber 2018).
Reducing the amount of text being translated down to ≈5–10 per cent, this becomes all the more
interesting for small-scale projects and pilot studies. As text similarity measures rest on a DFM,
one needs to tokenize all manifestos and have the features translated only once. As for any auto-
mated text method, however, Grimmer and Stewart (2013, 271) remind us that validation is key.
We summarize our translation approach here; the Appendix contains a comprehensive
discussion.
6
Opting for an intermediary solution in terms of resources, we take two ‘paths’. First, we create
cost-effective ‘feature-translated DFMs’in which the features are translated to English only once.
We then compare them to DFMs based on a random sample (≈20 per cent) of full-text transla-
tions for each language. This method allows us to contrast the feature-translated DFMs with the
full-text ones to ensure the former still capture the essence of the full texts. The feature-translated
DFMs can then be combined into a large DFM in English to perform analyses. We assess their
equivalence by:
•looking at cosine and Jaccard similarity,
7
which ranges from 0.65 for Finnish to 0.80 for
German and Portuguese with an overall mean of 0.74; the Jaccard coefficient ranges from
0.37 for Finnish to 0.57 for Catalan with an overall mean of 0.50;
5
In the Appendix, we discuss Jaccard similarity as an alternative measure. Re-running our regressions, the results reveal the
same ordering of ties, which supports our substantive conclusions.
6
The translation was done with Google Translate in January 2019. Note that we are not validating machine translation per
se. It has been shown that full-text machine translation has caught up to human-translated texts (de Vries, Schoonvelde and
Schumacher 2018; Le and Schuster 2016; Lotz and van Rensburg 2014; Lucas et al. 2015). If human translations were the ‘gold
standard’, we instead compare the ‘silver standard’of full-text machine translation to the ‘bronze standard’of DFM transla-
tions. To our knowledge, only Reber (2018) provides first evidence of the suitability of this approach.
7
Jaccard similarity solely quantifies the union of two sets and therefore gives lower scores (cf. Appendix for a discussion).
8 Nils Düpont and Martin Rachuj

•looking at the vocabulary, which shows a slightly reduced number of types in the feature-
translated DFM, ranging from −10.1 per cent for Catalan to only −2.4 per cent for
Galician with an overall mean of −4.8 per cent;
•inspecting whether the pattern of pairwise similarities that we find across all sampled full-
text translated documents can be detected by the feature-translated DFM as well. Here,
Pearson’s correlations range from 0.77 for Norwegian to 0.99 for Portuguese, with an overall
mean of 0.95.
Our assessment shows that feature-translated DFMs are sufficiently equivalent, show an ample
overlap in the vocabulary, and that no language systematically deviates due to the translation.
Most importantly, these DFMs can detect the same patterns of pairwise similarities. This
makes us confident that DFM translation is indeed a viable, cost-effective option, especially for
pilot studies and small-scale projects, and that the combined DFM allows us to compare party
manifestos across space and time.
Independent Variables: Capturing Ties and Attributes
There has been a lively debate in public policy studies about capturing diffusion processes at the
indicator level (cf. Maggetti and Gilardi 2016; Neumayer and Plümper 2016). A dyadic approach
allows us to explicitly test the impact of commonalities on the outcome of diffusion. We argue
that six linkages and two attributes resemble the heuristics that shape diffusion, and lead to het-
erogeneity in exposure and responsiveness –and, consequently, to differing diffusion effects.
In Table 1 we summarize our operationalizations of the independent variables, denoting party
ias the receiver and party jas the sender. Each tie represents a previously discussed commonality
in the focal dyad. Vote gains or losses of the sender and receiver capture the representativeness
and anchoring heuristic, signaling the success or disadvantage of a manifesto.
8
Methodology
To account for the complex structure of the data, multilevel modeling is an appropriate approach
(Rabe-Hesketh and Skrondal 2012). Gilardi and Füglister (2008, 425) suggest including three ran-
dom intercepts in dyadic settings, one for each state plus time. Yet parties are nested in countries
and elections. Other analyses indicate that unobserved election-specific factors are more import-
ant than the country or party level, though (Adams and Somer-Topcu 2009, 836; Lacewell 2017,
451; Meyer 2013). We therefore include two non-hierarchical random intercepts to control for
peculiarities of the ‘sender election’and ‘receiver election’.
9
We further include decade fixed
effects to deal with temporal trends and shocks (Plümper and Neumayer 2010). The oil crisis
in the 1970s, the economic crisis in the early 1990s or the fall of the Iron Curtain may explain
text similarity due to ‘independent problem solving’(Holzinger and Knill 2005, 786) as parties
responded to these new circumstances in a similar manner. For our analysis, we focus on 162
mainstream and niche parties, analyzing 105,575 directed party dyads in nineteen Western coun-
tries from 1960 to 2016.
10
A descriptive account of the (in-)dependent variable(s) can be found in
the Appendix.
8
In the Appendix we report additional models using weighted vote gains/losses to account for the critique that a certain
gain/loss may ‘mean’different things for small and large parties. The interaction effects are slightly less pronounced, but still
support our conclusions.
9
Models with random intercepts for parties hardly differ, which confirms our conclusions (cf. Appendix). Furthermore,
initial tests with more complex nesting structures often failed to converge and the variance component was close to zero,
indicating that grouping at these levels was of no use (Hox and Wijngaards-de Meij 2015, 135).
10
Due to gaps in the corpus we ‘interpolated’a few missing documents by replacing them with their previous program,
assuming the old text was still valid.
British Journal of Political Science 9

In sum, we argue that a text-as-data approach can overcome the shortcomings of left–right
scores, and use pairwise text similarities as the dependent variable when analyzing diffusion
among parties. Furthermore, we explicitly capture ties that bind and sender/receiver attributes.
We are now able to assess whether diffusion processes become manifest in (dis-)similarities of
election programs.
The Ties that Bind and Conditional Diffusion
We argue that three heuristics shape parties’efforts when gathering information (from else-
where), leading to heterogeneity in exposure and responsiveness and, consequently, to differing
diffusion effects. Rather than looking at all information, commonalities and sender/receiver attri-
butes affect which instances are taken into account. Looking at the text where the diffusion of
ideas, rhetoric or style materializes, we test whether (and in which way) ties and sender/receiver
attributes explain the outcome of diffusion. Especially for cross-border ties we thus shed light on
(un-)favorable conditions for party policy diffusion. Table 2 reports the results for seven regres-
sion models. Given abundant evidence that domestic party competition still accounts for the
lion’s share of the explanation –but also considering our notion to view it from a diffusional per-
spective –the ‘purely’domestic Model 1 serves as the baseline to which each cross-border tie is
then added.
Availability
The most important finding for the domestic context is the effect of Recycling on pairwise text
similarities; it has the largest effect size in all models. Parties do not reinvent the wheel every
time they draft a new manifesto. They instead drop parts of the text to make space for new
ideas while keeping much of the old text. This ensures coherence and ‘issue ownership’in the
long run (Budge 2015). Changing too much would also distract voters, giving them a hazy signal
about what the party actually stands for. However, this means there is restricted room for new
ideas adapted from elsewhere.
Text similarity is higher if both parties are competitors, which corroborates earlier
findings that parties observe and respond to each other (Adams and Somer-Topcu 2009;
Table 1. Operationalization of commonalities and attributes
Tie/Attribute Operationalization Source(s)
Recycling Dummy: 1 if iis linked to itself at the past election (that is,
i
t=
i
t−1
)
Jahn et al. (2018)
Competitors Dummy: 1 if jwas a domestic competitor at the past
election
Jahn et al. (2018)
From governments Dummy: 1 if jis in government during the overlapping
inter-election period but not i
Jahn et al. (2018)
Among governments Dummy: 1 if iand jare both government parties during the
overlapping inter-election period
Jahn et al. (2018)
EP factions Dummy: 1 if iand jare members of the same EP faction
during the overlapping inter-election period
Update and own extension of
Warntjen, Hix and Crombez
(2008)
Transnational party
organizations
Dummy: 1 if iand jare members of the same TPO during
the overlapping inter-election period
Update and own extension of Mittag
(2006)
Family of nations Dummy: 1 if iand jare in the same ‘family of nations’Castles (1998)
Vote gain/loss
j
j’s past or previous vote gain/loss (in percentage points) Jahn et al. (2018)
Vote gain/loss
i
i’s past or previous vote gain/loss (in percentage points) Jahn et al. (2018)
Note: for cross-border diffusion, the past vote gain/loss of the sender is relevant (that is, Δvote share
it =
vote share
it
–vote share
it−1
). For the
domestic context, the previous gain/loss applies (that is, Δvote share
it−1=
vote share
it−1
–vote share
it−2
).
10 Nils Düpont and Martin Rachuj

Table 2. The impact of commonalities on text similarity
1234567
‘Recycling’0.406*** 0.406*** 0.404*** 0.400*** 0.399*** 0.359*** 0.351***
(0.003) (0.003) (0.003) (0.003) (0.003) (0.002) (0.002)
Competitors 0.228*** 0.228*** 0.227*** 0.228*** 0.228*** 0.180*** 0.181***
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
From governments −0.001* 0.005***
(0.001) (0.001)
Among governments 0.016*** 0.016***
(0.001) (0.001)
EP factions 0.017*** 0.004**
(0.001) (0.001)
Transnational party organizations 0.019***
(0.001)
0.013***
(0.001)
Family of nations 0.064*** 0.063***
(0.001) (0.001)
Intercept 0.270*** 0.271*** 0.267*** 0.269*** 0.269*** 0.256*** 0.250***
(0.008) (0.008) (0.008) (0.008) (0.008) (0.008) (0.008)
Decade FEs Yes Yes Yes Yes Yes Yes Yes
Random parts
Var: elecid.i (Intercept) 0.002 0.002 0.002 0.002 0.002 0.002 0.002
Var: elecid.j (Intercept) 0.003 0.003 0.003 0.003 0.003 0.003 0.003
Var: Residual 0.007 0.007 0.007 0.007 0.007 0.007 0.006
Num. groups elecid.i 290 290 290 290 290 290 290
Num. groups elecid.j 290 290 290 290 290 290 290
AIC −218,994 −218,983 −219,498 −219,158 −219,337 −229,179 −229,920
BIC −218,889 −218,868 −219,383 −219,044 −219,222 −229,064 −229,767
LL 109,508 109,503 109,761 109,591 109,680 114,601 114,976
Obs. 105,575 105,575 105,575 105,575 105,575 105,575 105,575
Note: multilevel models with non-hierarchical random intercepts for elections. Decade fixed effects included but not shown. *** p < 0.001, ** p < 0.01, * p < 0.05.
British Journal of Political Science 11

Green-Pedersen and Mortensen 2015; Laver and Sergenti 2012; Meguid 2008; Williams 2015). It
also shows that party manifestos as a text genre are more cohesive within countries.
11
Two hypothetical examples may ease the interpretation: (1) a party adds previously unused
terms or (2) it aligns its emphasis of a common word. First, imagine a draft and a document
to learn from, each consisting of fifty unique terms mentioned once (that is, no overlap) and
the frequency remains constant: under ceteris paribus conditions, the effect of Recycling is equiva-
lent to adding ≈20 to 25 so far unused words from its old program or ≈10 to 15 terms from
competitors. For the second case, imagine two documents with ten unique terms fixed, each
of which is mentioned five times. In addition, both share one common word, mentioned
twenty-five times in the template but initially only once in the draft. The effect of Recycling is
equivalent to increasing the mentions to ≈7 to 8 times, or ≈4 times to align with a competitor.
12
In reality, parties adjust both the vocabulary and emphasis, which makes interpreting cosine simi-
larity quite abstract.
Turning to cross-border ties, each commonality has a positive effect on the pairwise similarity
of manifestos. Two findings are notable. First, their effect is smaller, indicating that there is little
impact of new ideas from abroad. Secondly, ties differ, and an ordering appears whereby Family
of Nations has the largest effect. Using Simmons and Elkins’(2004, 175) nomenclature, ‘learning
from cultural reference groups’seems more important than ‘learning through communication’.
To start with, we find that diffusion among governments is more relevant than diffusion from
government parties. While ‘intergovernmental meetings […] transmit information to policy
makers about “what works”in other settings’(Simmons and Elkins 2004, 175), their impact is
very small. Likewise, government parties per se signal ‘what works’. Still, both types of diffusion
have a rather subtle and long-term impact on party programs.
EP factions and transnational party federations bring together like-minded parties. Joining an
EP faction is rational for reasons other than the exchange of ideas though, for example access to
resources and seats in committees. The development of the European’s People Party and their
Christian Democratic member parties’struggle over whether (and how) to include conservative
parties (Jansen and van Hecke 2011, chap. 3) exemplifies that EP factions increase exchange, but
not necessarily (ideological) compatibility. EP factions matter, providing instances that more eas-
ily come to mind (Senninger, Bischof and Ezrow 2020), but they still have a limited effect on the
diffusion of ideas, rhetoric or style.
13
While transnational party organizations operate at a more
ideational level than EP factions, they serve the same purpose. Depending on the assumed nesting
structure (cf. Appendix) the effects of EP Factions and Transnational Party Organizations are
often on par. Returning to the hypothetical scenario, their effects are equivalent to adding ≈1
to 2 adapted terms. What looks negligible at first sight amounts to a couple of paragraphs if
for every fiftieth word, two previously unused terms are added, given that the ‘average’manifesto
in the corpus consists of 2,830 unique terms.
Supporting our expectation that ‘learning from cultural reference groups’plays a crucial role in
party policy diffusion, Family of Nations has the highest impact of all cross-border ties. The
regional context provides more instances to learn from than the domestic one but is still access-
ible, providing ‘analogous experience’due to cultural similarity. The effect of Family of Nations is
equivalent to adding ≈3 to 4 unused terms or increasing the emphasis to 3 in our hypothetical
examples. This is six times less than Recycling but three times more than any other cross-border
tie. When considering other parties abroad, such instances seem to particularly represent the
11
We further interpret this as criterion validity of our dependent variable: documents of the same original language are still
more similar when analyzing their translated, English version. Meanwhile, in multilingual countries like Switzerland or
Canada, text similarity is higher among domestic competitors. Thus there is no bias towards higher similarity simply due
to language.
12
See the Appendix for a third scenario of a party replacing terms.
13
Yet the outcome –increasing convergence of platforms –confirms Nanou and Dorussen’s(2013) notion that European
integration constrains the ‘menu’on offer in the long run.
12 Nils Düpont and Martin Rachuj

right mix of effort reduction, wealth of information and level of compatibility. In other words, it
is less effort for the Swedish Social Democrats to gain insights and adapt ideas from the
Norwegian Conservatives than the Spanish Socialist Workers’Party despite sitting together in
the same EP group.
Representativeness and Anchoring
So far, we have only looked at how commonalities and the availability heuristic shape diffusion.
Turning to sender/receiver attributes, we introduce an interaction term. This allows us to assess
whether diffusion is conditional on the success of the sender signaling an advantage of its ideas
(representativeness), the ‘desperation’of a party and its willingness to change (anchoring), or
whether instances ‘just’need to be available.
We predict values of text similarity under ceteris paribus conditions, holding all covariates at
zero except one linkage over levels of Vote Gains/Losses (Figure 1). As a baseline, assume both
parties share one commonality. The degree of adaptation then depends on the success of the
sender or ‘desperation’of the receiver.
For Recycling we find a consistent pattern in line with previous research: the more a party
gained, the more it keeps of its past program (Schumacher et al. 2015; Somer-Topcu 2009).
Conversely, text similarity is lower if it (seriously) lost. Interestingly, this is the only conditional
effect we find for the ‘anchoring heuristic’(lower graphs in Figure 1). We expected that parties
would look for inspiration elsewhere when they are in need. However, relevant instances are only
considered to the extent they are available, but not assessed in terms of (dis-)advantage.
Severe losses often challenge the balance of power between factions, leading to internal struggle
(Harmel et al. 1995; Harmel and Tan 2003). Such circumstances force parties to engage in ‘soul
searching’, which leaves them less open to cross-border diffusion.
The picture looks different for ‘representativeness’(upper graphs in Figure 1). As expected,
text similarity is higher if the sender gained votes, and lower if it lost. This is especially true
for competitors and slightly less so for diffusion from government parties. The former mirrors
Laver and Sergenti’s‘predator rule’that ‘exploiting successful choices made by other agents’is
an efficient strategy in an uncertain environment (Laver and Sergenti 2012, 134). The latter
lends support to Böhmelt et al.’s (2017) finding that a governing party signaling ‘success’
makes diffusion more likely. In terms of our hypothetical example, for both the difference
between a sender that gained vs. lost 5 percentage points is roughly equivalent to adapting ≈1
to 2 words more or less for every fiftieth type.
As EP factions and transnational party organizations connect quite ideologically like-minded
parties, the almost ‘non-conditionality’on the sender attribute indicates that ‘success’or ‘disad-
vantage’are irrelevant categories for diffusion among sister parties. Surprisingly, for diffusion
among government parties, and even more for diffusion in the regional context, we find the
reverse effect. Here, text similarity is higher if the sender lost and lower if it gained. When
considering weighted vote gains/losses, emphasizing highly visible instances of loss or success
(cf. Appendix), the same overall pattern emerges, though the latter two effects vanish. For
now, we can only speculate about the ‘rationale’for aligning one’s own text with an unsuccessful
one. Beyond availability, ‘success’may simply be misinterpreted when considering ‘analogous
experience’from the region. A second reading comes to mind though: Böhmelt et al. (2016) esti-
mated spatial lags using weighing matrices, which implies measuring the weighted sum of stimuli
from abroad. Analyzing dyads focuses on commonalities in a one-on-one setting instead.
Combining both results may point to ‘herd behavior’(Levi-Faur 2002): a sufficient number of
parties opting for the same idea is required before other parties respond to it. The resulting
S-shaped curve of the number of adopters is a typical pattern for the diffusion of ideas
(Rogers 2003). It may well be the case that ‘herd behavior’takes place in the regional context
–hence the largest cross-border effect for availability –even if it turned out to be less
British Journal of Political Science 13

Figure 1. Conditional effect of sender and receiver attributes on text similarity
Note: predictions with 90 per cent confidence intervals, adjusting for all other covariates and assuming RE = 0. The bottom graphs show the kernel density of observed data for Vote Gains/Losses.
14 Nils Düpont and Martin Rachuj

advantageous. This underlines that it is foremost the regional context where diffusion takes place,
but future research should explore the underlying mechanisms.
In summary, our analysis shows that party policy diffusion is stronger if certain conditions are
met (and less relevant otherwise). The domestic context prevails, but even here diffusion takes
place with parties adapting ideas, rhetoric or style from competitors –especially when the latter
gained votes. Parties change less of their manifesto if they were successful on their own (anchor-
ing), which leaves less room for new ideas from abroad. Despite being in need, and against expec-
tations, cross-border diffusion plays less of a role if a party (seriously) lost and instead may
engage in internal struggles.
Our results complement Böhmelt et al.’s analyses regarding (un-)favorable conditions for
party policy diffusion. Once looking abroad, the number of instances explodes that could provide
information about the ‘best’strategy or policy offer. Given limited resources, people and organi-
zations tend to apply heuristics to ‘filter’them, leading to differing diffusion effects. More specif-
ically, we find diffusion among and from government parties to be less relevant. Instead, sitting
together in an EP faction or joining a transnational party organization helps narrow down the
instances. Because aligning with sister parties is unconditional on the sender attribute, diffusion
among rather ideologically like-minded parties is not subject to assessing the (dis-)advantage of
what diffuses. Still, ‘learning from cultural reference groups’is most important. Considering
‘analogous experience’from the region provides a particular mix of effort reduction vis-à-vis
wealth of information and compatibility while still arriving at ‘satisficing decisions’–at times
even ignoring disadvantages.
Conclusion
We contribute to the emerging research on party policy diffusion by analyzing the outcome of
diffusion as the pairwise similarity of party manifestos in nineteen Western democracies from
1960 to 2016. To this end, we overcame language barriers by using machine translation. Our
study shows the potential of text-as-data approaches paving the way for refined analyses of the
content of diffusion. As a first step, we took a bird’s eye perspective looking at entire documents
and cost-effective ‘feature-translated’DFMs. For this reason, we are likely underestimating the
role of party policy diffusion. Having established that machine translation is a suitable tool, a
logical next step is zooming in on a reduced number of manifestos –focusing on one region,
for example –to conduct full-text translations. This would allow for detecting and tracking spe-
cific terms and concepts and more advanced text analyses exploiting word embeddings and styl-
istic features.
Beyond our methodological approach, we started from a basic definition of diffusion, arguing
that even domestic party competition can be seen from a diffusional perspective. This sheds light
on the role of commonalities and sender/receiver attributes and how they link to heuristics.
Heuristics are an efficient strategy in an uncertain environment with limited resources that reduce
efforts when gathering and processing information (from elsewhere). Commonalities among par-
ties help narrow down the instances taken into account, in turn leading to heterogeneity in expos-
ure and responsiveness. The results are stronger and weaker diffusion effects.
Testing six linkages featuring different degrees of availability and effort reduction, the domestic
context still prevails for explaining the outcome of diffusion. Beyond borders, we find that diffu-
sion takes place mostly in the regional context. Parties adapt ideas, rhetoric or style from parties
within the same family of nations –at times even ignoring any disadvantages of doing so.
Membership in the same EP faction or joining a transnational party organization also increases
the likelihood of diffusion. However, categories like ‘success’or ‘disadvantage’do not matter for
diffusion among rather ideologically like-minded parties.
The ‘verbatim translation’of the Finnish Christian League’s program exemplifies that it is
regional diffusion that matters most. Commonalities of countries within a family of nations
British Journal of Political Science 15

perfectly fit the bias and heuristics of policy makers to arrive at ‘satisficing decisions’. Our results
call for a closer inspection of the mechanisms, though, as it seems that we witness ‘herd behavior’
rather than ‘rational learning’. To conclude, heuristics help to understand (un-)favorable condi-
tions for party policy diffusion; it is the ties that bind and sender/receiver attributes that shape
diffusion to different degrees.
Data availability statement. The data, replication instructions, codebook and additional files (cf. Düpont and Rachuj
2020) can be found in Harvard Dataverse at: https://doi.org/10.7910/DVN/ZJGHKK.
Supplementary material. Online appendices are available at https://doi.org/10.1017/S0007123420000617.
Acknowledgements. We would like to thank Stefan Müller, Roman Senninger and discussants from the 2019 Manifesto
User Conference (Berlin), the 2019 ‘Jahrestagung des Arbeitskreises für Handlungs- und Entscheidungstheorie’(Bremen),
the 2019 QTA-DUB Workshop (Dublin) and the 2019 EPSA Conference (Belfast) for their comments on earlier versions
of this article. We also thank the editors and anonymous reviewers of the British Journal of Political Science for their useful
feedback throughout the review process.
Financial Support. Nils Düpont appreciates funding from the Deutsche Forschungsgemeinschaft (DFG, German Research
Foundation) via Collaborative Research Center (SFB) 1342 Global Dynamics of Social Policy (project A01) at the University
of Bremen. Martin Rachuj appreciates funding from the University of Greifswald as a Bogislaw Fellow.
Conflicts of interest. We declare no potential conflicts of interest with respect to the research, authorship, and/or publica-
tion of this article.
References
Adams J (2012) Causes and electoral consequences of party policy shifts in multiparty elections: theoretical results and
empirical evidence. Annual Review of Political Science 15(1), 401–419.
Adams J, Haupt AB and Stoll H (2009) What moves parties? The role of public opinion and global economic conditions in
Western Europe. Comparative Political Studies 42(5), 611–639.
Adams J and Somer-Topcu Z (2009) Policy adjustment by parties in response to rival parties’policy shifts: spatial theory and
the dynamics of party competition in twenty-five post-war democracies. British Journal of Political Science 39(4), 825–846.
Arter D (1980) The Finnish Christian League: party or ‘anti-party’?Scandinavian Political Studies 3(2), 143–162.
Bär D, Zesch T and Gurevych I (2015) Composing Measures for Computing Text Similarity. Technical Report
TUD-CS-2015-0017.
Benoit K, Laver M and Mikhaylov S (2009) Treating words as data with error: uncertainty in text statements of policy posi-
tions. American Journal of Political Science 53(2), 495–513.
Böhmelt T et al. (2016) Party policy diffusion. American Political Science Review 110(2), 397–410.
Böhmelt T et al. (2017) Why dominant governing parties are cross-nationally influential. International Studies Quarterly 61
(4), 749–759.
Budge I (1994) A new spatial theory of party competition: uncertainty, ideology and policy equilibria viewed comparatively
and temporally. British Journal of Political Science 24(4), 443–467.
Budge I (2015) Issue emphases, saliency theory and issue ownership: a historical and conceptual analysis. West European
Politics 38(4), 761–777.
Budge I and Klingemann H-D (2001) Finally! Comparative over-time mapping of party policy movement. In Budge I,
Klingemann H-D, Volkens A, Bara JL and Tanenbaum E (eds), Mapping Policy Preferences: Estimates for Parties,
Electors, and Governments, 1945–1998. Oxford: Oxford University Press, pp. 19–50.
Castles FG (1998) Comparative Public Policy: Patterns of Post-war Transformation. Cheltenham: Elgar.
Cox GW (1990) Centripetal and centrifugal incentives in electoral systems. American Journal of Political Science 34 (4), 903–935.
de Vries E, Schoonvelde M and Schumacher G (2018) No longer lost in translation: evidence that Google Translate works
for comparative bag-of-words text applications. Political Analysis 26(4), 417–430.
Dolezal M et al. (2018) Beyond salience and position taking: how political parties communicate through their manifestos.
Party Politics 24 (3), 240–252.
Düpont N and Rachuj M (2020) Replication data for: The ties that bind: text similarities and the conditional diffusion of
party policies. Harvard Dataverse. https://doi.org/10.7910/DVN/ZJGHKK.
Ezrow L and Hellwig T (2014) Responding to voters or responding to markets? Political parties and public opinion in an era
of globalization. International Studies Quarterly 58(4), 816–827.
Gigerenzer G and Gaissmaier W (2011) Heuristic decision making. Annual Review of Psychology 62, 451–482.
16 Nils Düpont and Martin Rachuj

Gilardi F and Füglister K (2008) Empirical modeling of policy diffusion in federal states: the dyadic approach. Swiss Political
Science Review 14(3), 413–450.
Green-Pedersen C and Mortensen PB (2015) Avoidance and engagement: issue competition in multiparty systems. Political
Studies 63(4), 747–764.
Grimmer J and Stewart BM (2013) Text as data: the promise and pitfalls of automatic content analysis methods for political
texts. Political Analysis 21(3), 267–297.
Harmel R and Tan AC (2003) Party actors and party change: does factional dominance matter? European Journal of Political
Research 42(3), 409–424.
Harmel R et al. (1995) Performance, leadership, factions and party change: an empirical analysis. West European Politics
18(1), 1–33.
Haupt AB (2010) Parties’responses to economic globalization. What is left for left and right for right? Party Politics 16(1), 5–27.
Holzinger K and Knill C (2005) Causes and conditions of cross-national policy convergence. Journal of European Public
Policy 12(5), 775–796.
Hox J and Wijngaards-de Meij L (2015) The multilevel regression model. In Best H and Wolf C (eds), The SAGE Handbook
of Regression Analysis and Causal Inference. Los Angeles, CA: Sage, pp. 133–151.
Huang A (2008) Similarity Measures for Text Document Clustering. Proceedings of the Sixth New Zealand Computer Science
Research Student Conference (NZCSRSC2008). Christchurch, New Zealand.
Jahn D (2014) What is left and right in comparative politics? A response to Simon Franzmann. Party Politics 20(2), 297–301.
Jahn D et al. (2018) PIP –Parties, Institutions & Preferences: PIP Collection [Version 2018-02].
Jansen T and van Hecke S (2011). At Europe’s Service: The Origins and Evolution of the European People’sParty. Berlin: Springer.
Kahneman D (2003). Maps of bounded rationality: psychology for behavioral economics. American Economic Review 93(5),
1449–1475.
Kahneman D and Frederick S (2002) Representativeness revisited: attribute substitution in intuitive judgment. In Gilovich T,
Griffin DW and Kahneman D (eds), Heuristics and Biases: The Psychology of Intuitive Judgment. Cambridge: Cambridge
University Press, pp. 49–81.
Kayser MA (2007) How domestic is domestic politics? Globalization and elections. Annual Review of Political Science 10,
341–362.
Krause W et al. (2018) Manifesto Corpus: Version: 2018-2.
Lacewell OP (2017) Beyond policy positions: how party type conditions programmatic responses to globalization pressures.
Party Politics 23(4), 448–460.
Ladrech R (2006) The European Union and political parties. In Katz RS and Crotty W (eds), Handbook of Party Politics.
London: Sage, pp. 492–498.
Laver M and Sergenti E (2012) Party Competition: An Agent-Based Model. Princeton, NJ: Princeton University Press.
Le QV and Schuster M (2016) A Neural Network for Machine Translation, at Production Scale. Available from https://ai.
googleblog.com/2016/09/a-neural-network-for-machine.html (accessed 13 February 2019).
Levi-Faur D (2002) Herding towards a New Convention: On Herds, Shepherds, and Lost Sheep in the Liberalization of the
Telecommunications and Electricity Industries. Paper presented at the workshop “Theories of Regulation”, Nuffield
College, Oxford.
Lipset SM and Rokkan S eds (1967) Party Systems and Voter Alignments: Cross-National Perspectives. New York: Free Press.
Lotz S and van Rensburg A (2014) Translation technology explored: has a three-year maturation period done Google
Translate any good? Stellenbosch Papers in Linguistics Plus 43, 235–259.
Lucas C et al. (2015) Computer-assisted text analysis for comparative politics. Political Analysis 23(2), 254–277.
Maggetti M and Gilardi F (2016) Problems (and solutions) in the measurement of policy diffusion mechanisms. Journal of
Public Policy 36(1), 87–107.
Mair P (1983) Adaptation and control: towards an understanding of party and party system change. In Daalder H and Mair P
(eds), Western European Party Systems: Continuity & Change. London: Sage, pp. 405–429.
Meguid BM (2008) Party Competition Between Unequals: Strategies and Electoral Fortunes in Western Europe. Cambridge:
Cambridge University Press.
Meseguer C (2006) Rational learning and bounded learning in the diffusion of policy innovations. Rationality and Society
18(1), 35–66.
Meyer TM (2013). Constraints on Party Policy Change. Colchester: ECPR Press.
Mittag J ed. (2006). Politische Parteien und Europäische Integration: Entwicklung und Perspektiven Transnationaler
Parteienkooperation in Europa [Political Parties and European Integration: Development and Perspectives of
Transnational Party Cooperation in Europe]. Essen: Klartext.
Nanou K and Dorussen H (2013) European integration and electoral democracy: how the European Union constrains party
competition in the member states. European Journal of Political Research 52(1), 71–93.
Neumayer E and Plümper T (2012) Conditional spatial policy dependence: theory and model specification. Comparative
Political Studies 45(7), 819–849.
Neumayer E and Plümper T (2016) W. Political Science Research and Methods 4(1), 175–193.
British Journal of Political Science 17

Panebianco A (1988) Political Parties: Organization and Power. Cambridge: Cambridge University Press.
Plümper T and Neumayer E (2010) Model specification in the analysis of spatial dependence. European Journal of Political
Research 49(3), 418–442.
Proksch S-O et al. (2019) Multilingual sentiment analysis: a new approach to measuring conflict in legislative speeches.
Legislative Studies Quarterly 44(1), 97–131.
Rabe-Hesketh S and Skrondal A (2012) Multilevel and Longitudinal Modeling Using Stata: Volume I: Continuous Responses.
College Station, TX: Stata Press.
Reber U (2018) Overcoming language barriers: assessing the potential of machine translation and topic modeling for the
comparative analysis of multilingual text corpora. Communication Methods and Measures 3(4), 1–24.
Rogers EM (2003) Diffusion of Innovations. New York: Free Press.
Rose R (1991). What is lesson-drawing? Journal of Public Policy 11(1), 3–30.
Schumacher G and van Kersbergen K (2016) Do mainstream parties adapt to the welfare chauvinism of populist parties?
Party Politics 22(3), 300–312.
Schumacher G et al. (2015) How aspiration to office conditions the impact of government participation on party platform
change. American Journal of Political Science 59(4), 1040–1054.
Senninger R, Bischof D and Ezrow L (2020) How transnational party alliances influence national parties’policies. Political
Science Research and Methods. doi: 10.31235/osf.io/rkgmc.
Shah AK and Oppenheimer DM (2008) Heuristics made easy: an effort-reduction framework. Psychological Bulletin 134(2),
207–222.
Simmons BA and Elkins Z (2004) The globalization of liberalization: policy diffusion in the international political economy.
American Political Science Review 98(1), 171–189.
Simon HA (1993) Decision making: rational, nonrational, and irrational. Educational Administration Quarterly 29(3), 392–411.
Somer-Topcu Z (2009) Timely decisions: the effects of past national elections on party policy change. Journal of Politics 71
(1), 238–248.
Somer-Topcu Z and Zar ME (2014) European parliamentary elections and national party policy change. Comparative
Political Studies 47(6), 878–902.
Tversky A and Kahneman D (1973) Availability: a heuristic for judging frequency and probability. Cognitive Psychology 5(2),
207–232.
Tversky A and Kahneman D (1974) Judgment under uncertainty: heuristics and biases. Science 185(4157), 1124–1131.
Volden C (2006) States as policy laboratories: emulating success in the children’s health insurance program. American
Journal of Political Science 50(2), 294–312.
Warntjen A, Hix S and Crombez C (2008) The party political make-up of EU legislative bodies. Journal of European Public
Policy 15(8), 1243–1253.
Welbers K, van Atteveldt W and Benoit K (2017) Text analysis in R. Communication Methods and Measures 11(4), 245–265.
Williams LK (2015) It’s all relative: spatial positioning of parties and ideological shifts. European Journal of Political Research
54(1), 141–159.
Cite this article: Düpont N, Rachuj M (2021). The Ties That Bind: Text Similarities and Conditional Diffusion among
Parties. British Journal of Political Science 1–18. https://doi.org/10.1017/S0007123420000617
18 Nils Düpont and Martin Rachuj