We May Disagree, but We All Love the BJP: Populists’ Networks and Targeting Opportunities on Twitter

Abstract With the rise of social media, politicians face new challenges and opportunities in presenting their messages online to large and diverse constituencies. Populists especially have capitalized on social media to achieve wide electoral success. This paper explores populists’ opportunities to target their constituencies online using the case of India’s Bharatiya Janata Party (BJP) on Twitter. First, the structure of the BJP’s network on Twitter is analyzed through social network analysis. Second, a content analysis is used to examine the messages of important accounts in different networks clusters. The findings suggest the ruling party in India has the opportunity to employ microtargeting to siloed clusters within its network of supporters. This practice could damage democratic discourse by reducing the possibility of debate in the public square.


Introduction
As the world's largest political party, India's Bharatiya Janata Party (BJP) must maintain the support of a vast, heterogenous, and fractured electorate.The BJP is not alone in facing this challenge: politicians seeking to build broad coalitions of supporters often need to partition their messages to appeal to diverse constituencies with conflicting interests. 1But social media has brought both challenges and opportunities for messaging that reaches multiple audiences.On the one hand, varied self-presentation is more difficult because of the structure of social media, which collapses multiple contexts and coalesces disparate audiences with distinct interests into one space. 2On the other hand, social media tools allow candidates to narrowly target voters based on data they've collected without messages being transmitted to other members of the electorate. 3hile the impact of social media on targeted messaging is an open question for all political parties, populists are uniquely affected because of their reliance on social media.Social media has become the primary medium of communication for populist parties. 4Populists have shunned traditional media, seeking unmediated and direct access to their constituents as part of their self-perception as advocates of "the people." 5 As a result, Schaub and Morisi 6 find internet use, among other factors, causes higher support for populist parties.Therefore, it is no surprise that the rise of social media has been accompanied by a populist wave. 7he surge of populism, however, is degrading democratic discourse. 8The recent rise of right-wing populism in particular has sparked increased nationalism and xenophobia, encouraged Manichean worldviews, and blurred lines between truth and fiction. 9To understand the nature of this threat and how to best address it, we must first examine how populists can target their constituencies online.How does the structure of populists' networks of supporters inform their targeting strategies?On the one hand, populists may receive support from an unsegmented mass of constituents online united by vague antagonisms who care little about specific issues and nuanced discourse.On the other hand, populists may target siloed communities, which could create an electorate of segregated groups, narrowly focused and only caring about issues tailor-made to them.In other words, do populist movements broadly target one coalesced online constituency through vague messaging or do these movements appeal to specific, individualized grievances of multiple insulated communities?Both dynamics coarsen democratic culture, but the dangers they pose and the solutions they require are distinct.
This study will examine populists' opportunities to target online by analyzing the composition of the social networks of party members and their supporters as well as the messages populists transmit to their constituencies.Specifically, I will examine the networks of BJP officials and allies connected to the account of Prime Minister Narendra Modi using a social network analysis.I argue that the presence of distinct constituencies in the BJP's network provides the party the opportunity to use microtargeting to target siloed communities within its network of supporters.Such a desegregated electorate could have dangerous implications for democratic discourse. 10his form of targeting allows citizens to support a party based on certain issues while ignoring or remaining ignorant on other stances they might oppose, reducing the possibility of debate in the public square.

Populists on social media
Populists often seek direct and unmediated access to their constituencies. 11Social media inherently promotes this kind of disintermediation. 12Bimber 13 predicted that political communication would transform because of unmediated interactions between citizens and political institutions online, allowing constituents to be influenced by not only traditional media, but information spread on the internet.Today, both established political groups and emerging movements appeal to and mobilize constituencies through their messaging on social media, bypassing news voices. 14urthermore, many populist movements have shunned traditional media as part of a constructed "elite."They allege that these outlets conspire with politicians and corporations and convey false messages that help preserve the corrupt establishment. 15Social media allows populists to maintain this criticism while still communicating with their constituents, circumventing gatekeepers who might seek to censor their messaging. 16rnst et al. 17 find parties prefer to use populist content on social media than on talk shows.Traditional media can also be misleading and targeted toward particular constituencies, but the presence of third-party mediators on these platforms has led populists to prefer alternatives.Beyond a desire to avoid mediation, social media is also far less expensive than traditional media and Twitter in particular is seen as an efficient and cheap platform for campaigning. 18Furthermore, Bartlett 19 contributes that "the short acerbic nature of populist messages works well in this medium" (94).
While social media does provide an unmediated connection to constituents, its horizontal, nonhierarchical nature causes context collapse, which is the flattening of multiple audiences into one. 20In fact, Meyrowitz 21 first formulated the concept of context collapse to refer to the lack of contextual distinctions on electronic media.He used the case of American politician Stokely Carmichael who employed different styles of communication to address white and black audiences but flopped on television because he was unable to choose between the two styles.Online, context collapse is even more difficult to overcome: the variable self-presentation, or "code switching," necessary to maintain diverse constituencies is prevented by the requirement to present a singular and verifiable identity. 22Furthermore, the accessibility and horizontality of social media platforms limits a political group's ability to fully control how its messages are used. 23hen populists use social media, they must weigh the benefits and risks of these platforms.Access to their constituents online allows for their message to be received without mediation, which might suggest populists seek to excite their base through more antagonistic and inflammatory rhetoric. 24Simultaneously, context collapse implies that all internet users have access to all content.Context collapse might prevent populists, or any other politician, from sharing messages that might alienate certain supporters. 25The concept of online echo chambers seems to challenge the importance of context collapse: if online users intentionally segregate themselves along partisan lines, then access may not matter if constituents are only amenable to information with which they agree.However, Barber a 26 finds past studies "have overestimated the degree of ideological segregation in social-media usage".Furthermore, Dubois and Blank 27 find that only a small segment of their study-those who lacked political interest and a diversity of social media platforms-were in echo chambers.Therefore, context collapse remains a serious obstacle to politicians who seek to target audiences through partitioned messaging.This research will explore opportunities available to populists to navigate context collapse.

Microtargeting versus context collusion
Political actors, including populists, can best take advantage of social media as a mouthpiece while accounting for context collapse through various messaging strategies.Two messaging strategies politicians use online are microtargeting and context collusion.
Microtargeting is the means by which politicians learn more about voters' characteristics and ideologies and narrowly target these constituents online without messages being transmitted to others. 28Therefore, microtargeting can help political actors maintain various self-presentations to different actors by tailoring messages that only specific constituencies receive. 29Borgesius et al. 30 note, for example, that voters often have specific interests in certain policy areas, while other aspects of a politician's platform may alienate them.Political parties can microtarget voters with messages that related to their specific policy interests, while excluding other information. 31This strategy reflects a continuation of a widespread and longstanding practice of political segmentation, wherein political parties in multi-party systems target constituencies with distinct messages to build diverse coalitions. 32However, more traditional political marketing strategies can be undermined by the context collapse that occurs on social media. 33esearchers have primarily studied how microtargeting has been enabled by modern information technology, which permits political actors to collect immense individual data on voters, use sophisticated data mining tools, and implement this information strategically in campaigns. 34Halpern 35 argues Trump won the 2016 election through his use of the Battleground Optimizer Path to Victory analytical model.Trump's campaign and the Conservative Party during Brexit both used psychometric studies, which can infer sensitive personal information through analyzing "likes" on Facebook. 36This specified targeting can expose voters to manipulation.According to Gorton, 37 microtargeting can turn "citizens into objects of manipulation and undermines the public sphere by thwarting public deliberation, aggravating political polarization, and facilitating the spread of misinformation".In fact, targeted information can still be impactful even if it is false, which is the case for much populist messaging. 38While microtargeting using sophisticated software by political parties has been primarily identified in the US and Europe, the logic of intentionally targeting multiple audiences separately is applicable across online contexts. 39lternatively, political actors may seek to overcome context collapse not by microtargeting specific constituencies but by coalescing multiple groups into one.Context collusion, where a political actor intentionally brings together diverse contexts, can also help to overcome the consequences of context collapse.Politicians can convey messages that focus on the greatest common factor-attributes individuals from different contexts share-to emphasize their similarities and understate their differences. 40On social media, politicians and other popular users often emphasize messaging that is vague and relatable to a vast majority of their followers while users with few followers represent more specific interests. 41This inclusive content is termed by Engesser et al. 42 as "personal action frames."This content is "not so much based on established social groups, memberships, and substantive ideologies but rather on flexible political identifications." 43Populists, who emphasize a vague conflict between the "people" and the "elite" may be well-suited to capitalize on context collusion by coalescing diverse groups of people around anti-elite messaging.In contrast, Gonawela et al. 44 sees populist messaging-particularly that of right-wing populists-as often less collaborative, finding that "political actors may benefit from in-group coalescence around antagonistic messaging" which can serve "as a call to arms for online collaboration for those ideologically aligned".Either way, these approaches to overcoming context collapse require mobilizing a dense, connected base of supporters rather than seeking to maintain a fragmented one with multiple distinct audiences.Whether populists on social media target a broad, unsegmented base through vague messaging or multiple insulated constituencies by appealing to their specific, individualized grievances is the main concern of the paper.

Social network and content analyses
I use two methods to determine the targeting strategies available to populists on social media.First, I determine whether distinct neighborhoods exist within the social network and what accounts comprise those clusters through a social network analysis.Then, using a content analysis, I seek to unravel what distinguishes these neighborhoods and whether they are subject to distinct messages.I selected the BJP as my case study because Modi and others in the party are not only populists but rely heavily on social media for their popularity and success. 45Furthermore, India is a large and heterogeneous democracy, which means, to achieve broad electoral success, politicians must address context collapse for their message to appeal to a sufficient diversity of constituencies.Lastly, the structural conditions of India are conducive for targeting on social media to have a substantive effect on constituencies because most citizens have internet access, and the country is among the largest and fastest-growing digital markets.
First, I conducted a social network analysis of friend connections on Twitter, originating from India's populist Prime Minister Narendra Modi's account to investigate whether the BJP is faced with a siloed or coalesced constituency on social media.On Twitter, friends are defined as the accounts a user is following.To observe this social network, I collected two degrees of Modi's Twitter network: his friends and his friends' friends.I accessed this data through the Twitter application programming interface (API), which provides researchers and app developers access to a database containing most past activity on Twitter.Using "Tweepy," a python library with commands for accessing the Twitter API, I collected the screen names (or handles) of the 2,351 accounts Modi followed.However, the Twitter API imposes a rate limit which only allows a small amount of data collection every 15 minutes.Therefore, collecting the friends of Modi's friends, which amount to over 2 million accounts, was impossible through Tweepy.I worked with a group of coders who built a Twitter scraping tool using python libraries Selenium and Beautiful Soup to collect the second degree of Modi's Twitter network.This tool was able to avoid Tweepy's rate limits and is capable of scraping Twitter friends unlike popular tools such as Twint.
If the study aimed to exclusively examine Modi's targeting strategies, collecting data on his Twitter followers would have been more instructive than on his friends.However, this study does not narrowly focus on Modi, but instead seeks to capture the network of BJP officials and allies on the platform.While Modi is used as the locus of the analysis as the leader of the BJP, he only represents one node within the network.Because of their large followings, the top accounts represented in this analysis are not primary recipients of messages but rather secondary disseminators of messages to the wider constituency of BJP supporters on Twitter.Furthermore, Modi does not have control over the accounts that follow him, which represent a wide range of supporters and opponents.In contrast, Sinha 46 finds that Modi strategically follows accounts whose messages he hopes to amplify without attaching himself directly to their rhetoric.
Having collected the data, I visualized the network using Gephi, an open-source network analysis and visualization software.The data was organized into nodes and edges.In this network, nodes-or main objects in a relationship-represent each individual account.The edges are connections or ties between various nodes, which, in this case, are connections between Modi, each of his friends, and their friends.Using both nodes and edges, Gephi presents the full network as a visual map, which is called a sociogram.However, visualizing the entire network of 2 million nodes would have both been impossible for Gephi to process as well as visually incomprehensible.Therefore, the coders calculated the 5,000 users followed by the most other accounts within the network as a whole, which I then input into Gephi.While these accounts are just a sample, they are the most influential nodes and, therefore, disseminate the most influential messaging.
To best visualize this data, I used the layout Yifan Hu, which is specifically designed to better distinguish clusters and can handle large networks.I used Gephi's tools to examine different properties of the network, such as the level of interconnectedness between accounts, the relative importance of different accounts, and the presence or absence of "neighborhoods," which are clusters of nodes with deep ties to each other.
Having determined the most important nodes and neighborhoods through social network analysis, I conducted a content analysis of tweets to determine the primary messages of the most important users in each cluster.From each of the four main neighborhoods present in the network as revealed by the sociogram, I selected the six accounts with the highest eigenvector centrality for analysis.As data for content analysis, I collected ten tweets from each of these accounts, ultimately amounting to 240 total tweets.
I created a standardized procedure to collect tweets.Using Twitter's advanced search feature, I recorded one tweet from each of the 24 important nodes on the 1st and 15th of each month, spanning from October 2020 until February 2021.If there was no relevant tweet on these days, I expanded my search by one day in both directions until one appeared.I deemed tweets irrelevant if they referred to family, sports, or other topics unrelated to politics.
Then I hand-coded the tweets to determine what defines and, potentially, differentiates neighborhoods present in the sociogram.The codebook I applied to each tweet had four primary sections.The first asked for general information about the tweets, such as its publication date and the number of likes and retweets.The second section related to general topics present in the tweets, specifically Anti-Muslim content, negative reference to the Indian National Congress (INC), pro-Hindutva content, pro-BJP content, or reference to the farmers' protests.The third section regarded the presence, tone, and concreteness of a policy discussion.The final section determined the presence or absence of populist features in the tweets, especially relating to positive or negative language and us-versus-them rhetoric.The codebook was pre-determined based on themes common in other content analyses of Modi and other populists on Twitter. 47Furthermore, tweets could fall in more than one category.For example, if a tweet specifically praised the BJP while also disparaging the INC, it would be coded as including both pro-BJP and anti-INC content.
I conducted an inter-rater reliability test to ensure responses were consistent regardless of coder bias.I recruited nine people without significant topical knowledge to apply the codebook to ten randomly selected tweets from my sample.To measure agreement among the ten raters, including myself, I used Krippendorff's Alpha.I selected this reliability coefficient over other options because it functions for all measurement levels (that is, ordinal, interval, nominal) and is not limited by sample size (that is, the number of questions in the codebook) or the number of raters.Krippendorff's Alpha is measured on a scale of 0-1, where 1 indicates perfect agreement and 0 indicates no agreement.In this case, the analysis found an overall Krippendorff's Alpha of 0.72. 48Section one of the codebook was excluded from the analysis as well as other metrics in the second, third, and fourth sections that did not have alphas over 0.667.
Having collected data using the content analysis form, I conducted statistical hypothesis tests in R. To determine whether the differences in frequencies of each variable from the content analysis form was significant depending on its neighborhood, I conducted a series of chi-squared tests.I also created tables of proportions (1) to analyze groups generally categorized as "us" and "them" by different modules and (2) to examine policies most frequently mentioned in each module.

Modi and the BJP
India's democracy is complex and has a diverse and fractured electorate.However, the Lok Sabha-the lower house of India's bicameral parliament-is not as fractured as one would expect.The BJP holds 56% of seats, the INC 10%, and all other parties each hold below 5%. 49As the party that led India's movement for independence from Britain, the INC commanded India's national politics for decades.However, the BJP won resoundingly in the 2014 general elections which saw Modi sworn in as Prime Minister, replacing Manmohan Singh of the INC.While the BJP previously led a coalition government from 1998 to 2004, Modi's dominant victory marked a decisive shift for India's democracy.The BJP is now not only India's largest party, but the largest political party globally, with over 110 million members.
The BJP maintains consistent symbolic and grassroots support from the Rashtriya Swayamsevak Sangh (RSS), a Hindu nationalist, far-right volunteer organization.Since its founding in 1925, the RSS has propagated the ideology of Hindutva ("Hindu-ness"): all Indians are ultimately Hindus because their holy land is in the region.Sikhs, Buddhists, and Jains, whose religions originated south of the Indus River, are thus all considered Hindus.Muslims and Christians are excluded from this identity, seen as "foreigners" and "invaders" brought to the subcontinent by the Mughal and British empires respectively, which both subjugated Hindus.Championing Hindutva, the RSS has sought to undermine secular government in favor of an ethnicized Hindu nation-state. 50Its ideology has inspired a large body of political, cultural, and militant organizations, which the RSS leads under the umbrella of the Sangh Parivar ("Sangh Family").
In the past, the BJP's close association with the RSS limited its prospects for broad popularity.However, the party's strategic use of social media and increasing embrace of populism have coincided with its meteoric rise to national political dominance.Modi's own transformation reflects the evolution of the BJP and Hindu nationalism.Modi first gained national prominence as a Hindu nationalist ideologue.He had supported the destruction of the Babri Mosque in Ayodhya as a member of the Ram Janmabhoomi movement in 1992 and was complicit in the slaughter of 3,000 Muslims in the 2002 communal riots in Gujarat, where he was governor at the time. 51While this image had already won him the steadfast support of firebrand Hindu nationalists, in 2014, Modi broadened his political appeal through skillful use of social media.Employing a populist style, Modi became highly interactive with his followers, shunned conventional media, and vowed to rid India of "elite" corruption. 52owever, while there is broad consensus regarding the importance of social media to Modi's success and the prevalence of populism in his rhetoric, there is some discord as to the targeting strategies he uses to build his base online.Rao 53 describes Modi's base as segmented, incorporating distinct audiences with different policy preferences, echoing the microtargeting strategies discussed earlier.By analyzing hashtags used by his supporters, Rao finds that, while some Twitter users follow Modi for his emphasis on development and technocratic government, others support him because of his overt identification with Hindu culture.Pal 54 presents a somewhat distinct picture of Modi's messaging strategy online, finding that most of his tweets are "banal" and positive messages that allow Modi to coalesce multiple audiences into one context.
While the subsequent analysis sheds light on Modi's own rhetoric, it expands on past literature by emphasizing the broader messaging and network surrounding his Twitter presence.Although Modi himself uses more banal rhetoric, many accounts he follows post "anti-Muslim, anti-opposition tweets, morphed photos and videos, jokes and cartoons," that present "half-truths and untruths, fake news, rumors, and slander against Modi's opponents." 55Ultimately, Modi is only one actor in the much broader BJP and Hindu nationalist movement.Pal and Rao examined Modi's own targeting strategies, but this analysis seeks to determine the messaging and targeting strategies available to the BJP's network more broadly.

Social network analysis
Analyzing the structure of the BJP's network on Twitter can reveal how information travels across the network and whether it is composed of siloed neighborhoods or is densely connected.To conduct this analysis, I focused on three key measures: how interconnected different actors in the network are (that is, density), whether there is structural evidence of distinct neighborhoods (that is, modularity), and if so, who the most influential actors in each neighborhood are who could influence messaging (that is, eigenvector centrality).
In social network analysis, network density is the ratio of actual connections to all possible connections among actors.The values of this variable range from 0 to 1, a value of 1 meaning all dyads-or pairs of nodes-in the network are connected and value of 0 meaning not a single connection is present.In general, we might expect networks that are less dense to signify siloed communities because a completely dense network suggests every node is connected to every other node and, therefore, distinct neighborhoods (that is, siloed clusters within the network) would not be possible.In the case of this network, the analysis shows density to be .016,which means 1.6% of possible connections are present.
There is no set rule as to whether a network density is low or high, as the interpretation is dependent on the type of network and its size.While a density of .016suggests very few nodes are densely connected with most other nodes, the larger the network the more difficult it becomes for every node to be connected.As a result of this limitation, network density alone does not prove that the BJP faces heterogeneous yet distinct neighborhoods within their online network that can be microtargeted.Still, low density does produce a more difficult setting for context collusion, where multiple distinct communities in an interconnected, dense network are targeted with vague rhetoric.
Modularity measures the presence and internal coherence of distinct modules in a network.High modularity suggests dense connections between nodes within neighborhoods but sparse connections between nodes in other neighborhoods.This measure takes values from range À1 to 1, where a value of 1 indicates no connections exist between nodes in different modules.This means information cannot pass directly between these disconnected neighborhoods, suggesting siloing is present within the network.A value of À1, in contrast, indicates nodes' connections (or edges) have the same likelihood to be connected with any particular node in the network.Such a value means all information flows to the entire network easily and immediately, indicating no siloing is present within the network.
In Modi's network, the analysis shows modularity to be 0.271.This modularity score suggests the network has a moderately strong community structure: there are significant connections between neighborhoods, but the network is more siloed than united.This finding provides evidence for the presence of distinct, siloed neighborhoods within the network.
Having determined the network's modularity, I analyzed the characteristics and structure of the modules themselves.Modules (that is, neighborhoods) are clusters of nodes densely connected to one another and less connected to other regions of the network.Using this method, Gephi found four primary modules, labeled 0, 1, 2, and 3. Figure 1 is a sociogram of the network color-coded to represent the four modules.As demonstrated by Table 1, nodes were relatively evenly proportioned between the four modules.Had modules been less evenly proportioned such that one or two modules comprised most nodes, the analysis would likely not indicate siloing within the network because the smaller clusters may only have been insignificant exceptions to an otherwise interconnected network. 56Overall, the presence of four modules within the network provides some support for the notion of segmented network but does not imply that these neighborhoods are differentiated by their political messaging.Without a content analysis of the messages of accounts in each module, segmentation could have been caused by factors irrelevant to this research, such as geography.
I selected specific nodes to sample their tweets for content analysis because machine coding would not capture sufficient nuances of meaning.To decide the appropriate nodes to focus on, I determined their relative influence by calculating their eigenvector centrality.Eigenvector centrality takes into account not only a nodes' volume of connections but also its nearness to other well-connected nodes.Table 2 presents the six accounts with the highest eigenvector centrality in each neighborhood. 57These nodes, compared to other nodes in their neighborhood, have the most connections to other nodes and have the highest proximity to other well-connected nodes in the network.I sampled tweets from these accounts for the proceeding content analysis.
Module 3 clearly contains the nodes with the highest eigenvector centrality, which suggests these nodes are most densely connected to accounts in their neighborhood and may, relative to other accounts, have more connections to other neighborhoods.This fact is also reflected in Figure 2, a sociogram of the network filtered by eigenvector centrality, in which far more nodes from module 3 are represented.The six nodes most connected in module 3 are the accounts of the BJP's most prominent political figures, including Modi and Minister of Home Affairs Amit Shah.These nodes' high centrality and identity suggests module 3 may contain most establishment BJP politicians.The most important nodes in module 0, conversely, generally have the lowest eigenvector centrality compared to other modules.Based on the accounts represented, important nodes in this neighborhood are seemingly less-partisan figures, including India's Minister of External Affairs, who is focused on foreign policy, and several "non-partisan" news networks and reporters.
Separately, the most important nodes in modules 1 and 2 generally have intermediate levels of eigenvector centrality compared to other modules.Important accounts in these  neighborhoods are partisan figures but do not have the same notoriety as establishment politicians in module 3.However, differences between identities of important nodes in module 1 versus module 2 are not as immediately discernible.Both modules primarily consist of younger, upstart BJP politicians, many of whom are explicitly associated with the RSS.However, module 2 also includes Kangana Ranaut (a famous actress who gained popularity among far-right Indians for her polarizing tweets) and @ShriRamTeerth, the informational account for the construction of the Ram Temple in Ayodhya.Although these accounts are not those of BJP politicians, they espouse a similar Hindu supremacist ideology and, therefore, can still be seen as part of a coherent neighborhood.Module 1 also includes several right-wing media pundits who are sympathetic to the RSS.While the important nodes in modules 0, 1, and 2 have lower eigenvector centrality than those in module 3, each contains a greater portion of the overall accounts in the network than module 3, suggesting these neighborhoods may have a less centralized, more horizontal structure.While this discussion provides a rough characterization of the most important nodes, it does not speak to the identities of the majority of accounts in each neighborhood.Most of the accounts included in the network are not politicians or media figures, but laypeople who have small to moderate followings among BJP allies.The preceding section describes the social network analysis that revealed several characteristics of the BJP's network.First, the network's low density suggests few nodes are connected to one another and, therefore, the network is more organized into isolated silos than deeply interconnected.Second, an analysis of modularity found the network's structure was moderately siloed and contained four distinct neighborhoods.Having determined the most important nodes in each network by calculating their eigenvector centrality, the next section describes the content analysis of these accounts' tweets to determine whether political preferences contributed to this segmentation.

Content analysis
How do we know from the content of tweets, whether the neighborhoods identified in the social network analysis contain distinct political messages?Content analysis allows us to distinguish these neighborhoods based on three dimensions.First, I look at the prevalence of topics and themes which I expected to see in Hindu nationalist/BJP rhetoric, including current events and potentially divisive identity-based discussions.Second, I analyze which neighborhoods most frequently discuss policy and which policy issues are most prevalent in each.Finally, I examine the populist features of the tweets, including language tone as well as the prevalence and variety of us-versus-them rhetoric.
Overall, this content analysis indicates there is a significant difference between modules across all three of these dimensions.The overwhelming evidence of differences between modules provides strongly suggests that the BJP's network is comprised of siloed communities with distinct political preferences, creating an opportunity for microtargeting.
For clarity, I have assigned labels to the four modules that will be used in lieu of their number for the remainder of the paper, except in figures and tables.These labels were determined based on the distinguishing characteristics of important nodes in each module revealed in the content analysis, though some accounts may not fully fit the classifications.These characteristics and labels are presented in Table 3.
I selected several topics and themes, which I expected to see in Hindu nationalist/BJP rhetoric.Discussion that is anti-INC, Anti-Muslim, Pro-Hindu, Pro-BJP, or regards the farmers' protests is relevant for analyzing the presence siloing because neighborhoods that invoke such topics are willing to antagonize certain constituencies that prefer inclusive rhetoric.Figure 3 displays each topics' frequency of occurrence by neighborhood.In the following graphs, count refers to the number of tweets including that specific content and "modularity class" is equivalent to "neighborhood." The Mainstream Media and Partisan Establishment Networks rarely use anti-INC, Anti-Muslim, or Pro-Hindu rhetoric.Based on the generally nonpartisan nature of accounts in the Mainstream Media Network, the low rate of divisive themes and topics is unsurprising.However, at first glance, the similar absence of such rhetoric in the Partisan Establishment Network may be unexpected, considering the accounts' overtly partisan nature.One explanation relates to strategic factors: Gonawela et al. 58 argues the RSS's extensive grassroots networks on the ground and online have already secured Modi's right-wing base.Therefore, social media for Modi-and, by extension, other establishment BJP politicians-is a "place for reserved populism," showing restraint in deploying divisive rhetoric. 59In contrast to these neighborhoods, both the Far-right Partisan network and the Hindu supremacist network have much higher rates of anti-Muslim, pro-Hindu, and anti-INC discussion.Therefore, these neighborhoods, which include young BJP politicians associated with the RSS and other far-right figures, may represent the RSS-associated networks that practice a more "unreserved populism," seeking to captivate the party's right-wing base.Furthermore, the Hindu Supremacist Network, uses significantly more pro-Hindu rhetoric than even the Far-Right Partisan network, suggesting subtle but distinct messaging differences in between these modules as well.
For each of these topics, I conducted a separate chi-square test to determine whether there is evidence of a relationship between modules and the prevalence of these topics.Based on these tests, anti-INC, Anti-Muslim, and Pro-Hindu discussion all have p-values less than 0.0001.Therefore, there is less than a 0.01% chance that the number of tweets containing these topics in each neighborhood differs so greatly from the expected counts if there was no relationship at all between modules and the prevalence of these topics.Therefore, there is significant evidence the prevalence of these topics is, in fact, dependent on the neighborhood.
The other two topics I coded, however, do not differ significantly across neighborhoods.Mention of the farmers' protest and pro-BJP discussion both had p-values greater than 0.05, which means there is a reasonably high possibility the number of tweets containing these topics in each neighborhood differed from the expected counts by random chance alone.Therefore, there is no evidence to suggest that, in these cases, the frequency of discussion depends entirely on the module and instead suggests the network's segmentation is not limitless.In the case of pro-BJP discussion, while modules may differ on certain topics, they are all ultimately part of the BJP's network and, though their political preferences differ, favor the party equally.Furthermore, the farmers' protest was an important issue across the political spectrum.The lack of significance here may not reflect a lack of difference between neighborhoods, but rather a shortcoming of the coding category, which did not capture the tone or way the protests were discussed (Figure 4).I also sought to determine whether neighborhoods differed in the rate of policy discussion and which policies are most prevalent.Unlike previous measures, the Far-right Partisan Network and the Partisan Establishment Network are most similar in their frequency of policy discussion.While these networks have different policy emphases, they contain more politicians than the other neighborhoods and are, therefore, more likely to foreground policy.Accordingly, the low frequency of policy discussion in the Hindu Supremacist Network reflects the lack of politicians within the neighborhood.Instead, celebrity and organizational accounts focus on Hindu empowerment and antagonism of out-groups without mentioning specific policy proposals.Based on a chi-square test, frequency of policy discussion is significantly different between neighborhoods (p ¼ 0.0002).This result provides further evidence for the presence of siloed neighborhoods with distinct preferences (Figure 5).
Along with analyzing the frequency of policy discussions, I examined the policy issues themselves.Table 4 describes how, in their tweets that contain policy, each neighborhood's discussion is distributed across a range of policy issues.There are significant disparities in the proportions between neighborhoods, which provides strong evidence of distinct policy preferences.The Far-right Partisan Network and the Hindu Supremacist Network continue to mirror one another in their more divisive tendencies, both consistently discussing corruption and religious issues, which are generally anti-Muslim or pro-Hindu.Likewise, the Mainstream Media and Partisan Establishment Networks both frequently mention economic development, an issue more palatable to centrists than Islamophobia.Despite these similarities, each neighborhood has distinctive attributes, suggesting every neighborhood contains different messages.For instance, only the Mainstream Media Network consistently discusses foreign policy and the Hindu Supremacist Network discusses security issues far more than the Far-right Partisan Network.The prevalence of security policy discussion in Hindu Supremacist Network may be explained by important accounts' desire to project a need for increased security to protect the Hindu state against not only Muslims specifically, but also against Pakistan, China, and other adversaries with conflicting ideologies.Ultimately, neighborhoods share support for the BJP while containing distinct policy preferences.
Apart from the topics and policies discussed by each neighborhood, I also sought to determine whether some neighborhoods were deploying populism as a discursive style.In their discourse, populists construct an "us-versus-them" divide between "the people" and "the elite" or between specific in-groups and out-groups. 60Furthermore, negative language is key to many populist discourses as populists often use antagonistic rhetoric to foster a staunch, unwavering base of support for a particular politician. 61Therefore, analyzing us-versus-them rhetoric and language tone provides insight into whether different neighborhoods are relying on populism as a discursive style more than others.
Based on the analysis, the Far-right Partisan Network and the Hindu Supremacist Network use more negative language.This finding further indicates these two modules, who exhibit an "unreserved populism," target their constituencies using antagonistic rhetoric, emphasizing divisive issues that elicit emotional responses.In contrast, tweets from the Partisan Establishment Network are overwhelmingly positive in their tone.One explanation is that, as establishment politicians in India's ruling party, these accounts may seek to present an image of successful governance by focusing more on the country's positive features than on emphasizing any faults.The Mainstream Media Network uses similarly little negative language but is far more neutral overall.This tonal neutrality reflects the projected political "neutrality" of many important accounts in the neighborhood, which includes mainstream media sources, reporters, and foreign policy officials.A chi-square test of language tone found the difference between the frequency of negative language used by accounts in the modules was statistically significant (p < 2.2e-16) (Figure 6).
The analysis of us-versus-them rhetoric reflects similar trends as language tone.The Far-right Partisan Network and the Hindu Supremacist Network use far more us-versus-them rhetoric than the Mainstream Media and Partisan Establishment networks, though the Far-right Partisan Network is unique in that nearly all tweets contain such rhetoric.In fact, there is a significant difference between the frequency of use of us-versus-them rhetoric in the former two modules (p ¼ 0.009) which suggests these oftensimilar modules are distinguishable in their discursive style.Overall, a chi-square test of us-versus-them rhetoric found a significant difference between the frequency of use of us-versus-them rhetoric depending on the module (p < 2.2e-16).
The groups defined as part of the "us," presented in Table 5, also provide evidence of siloed neighborhoods containing distinct messages.For instance, while both the Farright Partisan Network and the Hindu Supremacist Network most often define the "us" as BJP supporters and appeal to Hindus frequently, the latter is far more likely to also invoke nationalists explicitly as part of the in-group.In the little us-versus-them rhetoric present in the Mainstream Media and Partisan Establishment Networks, there is not a single allusion to Hindus as defining the "us" and essentially none to nationalists.These definitions of the "us" may be based on situational factors, namely the identities of accounts in each module.In other words, these definitions may reflect the social groups that make up each module-accounts in the Far-right Partisan Network and the Hindu Supremacist Network are far more open about their Hindu nationalist identities whereas those in the Mainstream Media and Partisan Establishment networks are more comfortably described as BJP supporters or Indians.The groups defined as part of the "them," presented in Table 6, are more varied than the "us," but also suggest different messages are present in each module.While the Farright Partisan Network and the Hindu Supremacist Network and the Partisan Establishment Network all refer most to political elites and opponents as the "them," each module is distinctive in which other outgroups it emphasizes.For instance, though both the Far-right Partisan Network and the Hindu Supremacist Network shun Muslims, accounts in the Hindu Supremacist Network also consistently castigate secularists and leftists.This finding reflects the neighborhood's primary motivation of championing Hinduism, to which both Islam and secularism are threats.In contrast, the Mainstream Media and Partisan Establishment Networks rarely mention Muslims and never secularists, but instead emphasize Pakistan and-in the case of the Partisan Establishment Network-the West far more often.Defining foreign entities as part of the "them" can be perceived as far less partisan and divisive than pointing out specific groups in the domestic population as out-groups.Still, the Partisan Establishment   Network is clearly more partisan than the Mainstream Media Network because of its opposition to political opponents.Nevertheless, these differences all support the notion of siloed neighborhoods receiving distinct messages.Overall, this content analysis provides strong evidence that different messages and political preferences are present in each neighborhood.In their themes, tone, and use of populist rhetoric, there are significant differences between the messages of important accounts in the four neighborhoods.Consequently, the BJP's network appears poised for microtargeting where messages are narrowly transmitted to certain voters without being received by others.

Conclusion
With social media, politicians have been faced with new challenges and opportunities in partitioning and presenting their message online.Populists, especially, have capitalized on the reach and absence of traditional gatekeepers on social media to achieve wide electoral success.However, whether they choose to address diverse interests among voters by targeting siloed constituencies with tailored messages or coalescing distinct constituencies through vague rhetoric that can appeal to the greatest common factor can have profound implications for their democracies.The former reduces the space for debate in the public square and breeds an environment of misconceptions and pedaled falsehoods.The latter, however, also limits the opportunity for specific policy discussion, as politicians rely on vague rhetorical and emotional appeals.While both outcomes are damaging for democracy, taking action to protect democracy requires understanding the nature of the affliction.This study examined a populist party in India, the BJP, as a case study, mapping the BJP's network to determine whether it represented a siloed or coalesced constituency, each providing different opportunities for targeting.
The study's results contributed to the study of context collapse-the flattening of multiple audiences into one context that occurs online-by presenting a potential context for the use of microtargeting.The findings suggest the BJP is well-positioned to employ microtargeting as its means of addressing context collapse.While its approach may not rely on modern information technology, it appears different influential actors in the BJP are positioned to target distinct audiences with different policy issues, topics, and rhetorical styles.Four neighborhoods are present within the network, and each appears to contain distinct political and rhetorical preferences, ranging from positive, inclusive, and policy-oriented to negative, exclusive, and vague.In contrast, had the BJP's network been dense and unsegmented, this structure would create the opportunity for context collusion.In that case, the BJP could practice context collusion by deliberately bringing together constituencies whose intensity of belief varies on issues of religion, corruption, and development by focusing instead on vague appeals to "the people."While this approach also seems efficacious, the results suggest the suitability of microtargeting for the BJP to build a broad base of support in India.
While this study demonstrates the opportunity for microtargeting in the BJP's network, future research can examine whether this network structure was the specific intent of BJP actors.The network is partitioned into distinct neighborhoods, but this siloing is not necessarily the product of BJP strategies.Alternatively, influential accounts in the network may espouse distinct political preferences without a coordinated top-down party strategy.This explanation would suggest the presence of ideological diversity among influential BJP members and that the distinct neighborhoods of supporters may have developed around their differing political preferences more organically from the bottom up.Connecting the segmentation of the network to a concerted BJP targeting strategy would require further research.Interviews and digital ethnography of top BJP officials and the other important actors highlighted in this study could reveal causal relationships between targeting strategies and online social network structure.
Still, the study's findings have significant implications not only for Indian democracy but all countries with populist parties.In India, the BJP has become the world's largest party and dominates parliament.To avoid the consequences of context collapse with such a large constituency, microtargeting siloed constituencies on social media can be an effective strategy.While this approach may appear more opportune for the BJP, the opportunities for context collusion and microtargeting may depend on several factors not captured in this study, which is focused exclusively on one case.Structural factors (like the availability of internet), situational factors (like specificities of a country's politics or society), and strategic factors (deliberate preferences) may all lead to different choices in how populists target their constituencies online.Future studies, therefore, may examine the populists' networks on social media with a comparative perspective.How do structural, situational, and strategic factors create the context for distinct targeting choices?While these questions remain, this study has presented a new approach to studying populists' networks on social media and their targeting opportunities.Analyzing language and rhetoric is a central element of understanding messaging strategies, but it alone cannot reveal the context and opportunities for targeting.Examining the composition of populists' networks can provide new ways of observing populism's coarsening of democratic discourse and assist in efforts to strategically encourage exposure and engagement with a diversity of ideas in online spaces.

Figure 3 .
Figure 3. Frequency of occurrence of anti-Congress, anti-Muslim, and pro-Hindu discussion by neighborhood.

Figure 4 .
Figure 4. Frequency of occurrence of Pro-BJP and farmers' protest discussion by neighborhood.

Figure 5 .
Figure 5. Frequency of policy discussion by neighborhood.

Figure 6 .
Figure 6.Distribution of language tone and frequency of us-versus-them rhetoric in each module.

Table 1 .
Percentage of the network's total nodes in each module.

Table 2 .
Nodes with the highest eigenvector centrality in each module.

Table 3 .
Characteristics of the important nodes in each module and module labels.

Table 4 .
Percentage of policy-containing tweets mentioning each policy issue by module.

Table 5 .
Percentage of tweets that imply an "us" who is constructed as part of an us-versus-them dichotomy by module.

Table 6 .
Percentage of tweets that imply a "them" who is constructed as part of an us-versus-them dichotomy by module.