Abstract
In this research, we explored how foundational issues can become subtext in political discourse by studying how statehood was debated in social media prior to two plebiscites in Puerto Rico. Historically in Puerto Rico, local parties are divided on this issue instead of along the more common conservative versus liberal division that is found in most parts of the United States. We collected the corpus of Twitter communication by members of Puerto Rico’s Legislature for the term of 2017-2021. Using latent topic modeling techniques, we classified the political discussion along party lines. Surprisingly, statehood was not a major topic in our model when using the full corpus of data. However, when we filtered the data to include only those tweets discussing statehood and sub-sampled them by major party, the sub-topics within statehood communication became clear and coherent, as was the partisan divide. Ultimately, while there is a clear division about statehood between the parties, the issue has become so intrinsic to the political sphere in Puerto Rico that it no longer commanded significant attention in the political discourse during this period.
Introduction
A significant topic in recent U.S. political discourse is the possibility of Puerto Rico becoming a state. Much of the national news coverage of the statehood debate focuses on how a Democratic takeover of Congress could make this possible. If Puerto Rico becomes the 51st state, it will have two senators and around five representatives in the House. With the assumption that most of the seats would be filled by Democrats, the national media coverage often focuses on a possible long-term power shift in favor of the Democratic Party. Largely absent from the mainland media discourse, though, is whether Puerto Ricans actually support statehood. Even further removed from mainland discourse is the fact that the primary dividing line between the parties in Puerto Rico is whether they support statehood versus maintaining the status quo rather than the conservative/liberal spectrum.
The issue of statehood, often referred to as the status debate because it includes more than simply statehood (status quo and independence), has caused sharp divisions in the population. This has never been more evident than in Puerto Rico’s 2020 election. There was a plebiscite on the ballot aimed at gauging citizen support for statehood. The plebiscite results were close to a 50/50 split. However, that is only a portion of the narrative. In this research, we consider the importance of status and review the political discourse concerning statehood on Twitter and consider how significant this issue is to political leaders. We use latent topic modeling to characterize legislators’ communication and examine the nature and distribution of the independence movement among political leaders. This allows us to step beyond the assumption that the political conversation between voters and politicians is constrained by the statehood debate.
Latent variable modeling methods, such as latent topic models have become increasingly popular in social media political communication research. These methods are similar to factor analysis or latent class analysis in that this general family of procedures models a larger set of data to identify a smaller set of concepts. In this case, it is a corpus of observed words being modeled as a function of a much smaller number of latent topics. Support for topic modeling comes from the idea that using the actual correlation between words to identify topics is a better measure than relying on traditional dictionary based methods of content analysis that are fraught with error because they rely on pre-determined lists that may mischaracterize the communication, or completely miss parts of it.
Our approach here is straightforward. We build on the most common validation check which Ying et al. (2021) called “bespoke validation”. Such validation is more or less a thorough check on face validity by checking predictive, convergent, and hypothesis validity. Ying et al. (2021) call for a more systematic approach and offer such. We think their approach is a good one. To execute our test, we scraped the universe of tweets from every member of the Puerto Rican legislature with an active account across the entirety of their current terms beginning in January 2017, until the day before the 2020 status plebiscite. Surprisingly, the model did not indicate that status was a primary topic. That said, when we filter the data to include only those tweets with statehood words and divide the data by major party - the Popular Democratic Party (PPD) and the New Progressive Party (PNP) - the subtopics within status discussion become clear and coherent. The PNP’s support for statehood and the PPD’s resistance is evident; and the models identify a set of meaningful, more nuanced discussions about statehood.
While the parties and historic political system may have been built around the status issue, our research suggests that the actual focus of the domestic politics have moved beyond that question. While status is there, the discussions are more limited and less focused. It is now an assumed position of the parties that has become more of a subtext to the political discourse than a driving issue for the parties. Statehood is baked into the partisan division. In many ways, the ideological division is one that is wrapped up in identity - a colonial and anti-colonial identity. This is the subtext, and then other issues rise to the top under that assumed context. Conventional wisdom suggests that political elites often use the status debate as a distraction tactic from other more immediate issues, but our results indicate to the contrary. The status debate in Puerto Rico is not overtly raised as a smokescreen, but rather is more firmly embedded in the partisan division - like a valence issue that does not need require constant in-depth discussion.
Ironically, there can be a low emphasis on important issues when those topics are so fixed and foundational, that they are no longer effective for voter appeal and not worth expended resources (Page 1976). According to Robert Anderson (Anderson 1988) Puerto Rico’s party system is not defined in terms of struggle among social classes or ideological controversies, but on a particular type of colonial dependency. Anderson argues that the status problem makes Puerto Rico politically interesting, because it defines parties individual electoral behaviour (p. 3). Since the origins of the modern political parties, these have been divided between “autonomism” and “assimilationism”.
Therefore, in Puerto Rico, where the status issue is baked into the partisan division, it is understandable how the issue would be a subtext rather than focus. Indeed, the rise in focus of domestic issues appears to have made some of the alternative parties more appealing while forcing the issues related to status into a less prominent position (Caban 2021). The partisan division itself speaks to the status issue, and where the party brand is still strong, the partisan leanings are more resilient as long as the parties do not contradict their established position (Lupu 2016). We consider this context and our measures below.
The Puerto Rican context - yes or no to statehood?
Puerto Rico is an ideal case to explore topic modeling in politicians communication, since political parties are the product of a divide on a single issue, the status of the relationship between the Island and United States (Morera Colón et al. 1993) Puerto Rico has been an unincorporated territory of U.S. since the end of the Spanish-American War. Until 1948, American presidents appointed the governor, when Congress amended the 1917 Jones Act to allow islander elect their own governor. However, international and local pressures to decolonize pushed Congress to create in 1952 the Commonwealth of Puerto Rico. This political arrangement provided greater local autonomy over the administration of the Island’s internal affairs and the creation of an elective local government. However, the Commonwealth did not alter the fundamental colonial relationship between the United States and Puerto Rico. Puerto Rico remains an unincorporated U.S. territory over which the U.S. Congress has plenary power (Fuentes-Rohwer 2008). Thus, Puerto Rico’s unique political configuration provides residents with limited democratic power over the decisions that affect their day-to-day lives. For example, Congress controls policy over the issues of immigration, foreign investment, infrastructure, maritime and air transportation, social well-fare, security, and many others.
The two main political parties which dominate the local election cycles are the New Progressive Party (PNP Spanish acronym) and the Popular Democratic Party (PPD Spanish acronym) (Cámara-Fuertes 2004). The PNP favors changing the current political status in favor of becoming the 51st state of the United States, while the PPD favors maintaining the current political status. A third minority party is the Puerto Rican Independence Party (PIP Spanish acronym). However, it has failed to win a majority in the legislature or the governorship. * The PNP and PPD have alternated power for the past 50 years. Given the limited areas of policy that the local government has power over, the local political parties have focused on electoral competition without providing clear ideological cues to larger aggregate political interests (Meléndez Vélez 1998). Thus, parties in Puerto Rico are what Kirchheimer (Kirchheimer 1966) called, “catch all” parties. These parties are not constrained ideologically, but instead offer a confusing mix of ideas(Meléndez Vélez 1998). For example, within the PNP party, its leadership is divided by conservatives such as Jennifer Gonzalez, the 20th Resident Commissioner of Puerto Rico, † and progressives governors such as Ricardo Rosselló and Pedro Pierluisi.
The Commonwealth was thought to be a transitional status towards either full integration, independence or greater autonomy (Morera Colón et al. 1993). Since the creation of the Commonwealth, Puerto Rico has held six status plebiscites. ‡ These have been in 2020, 2017, 2012, 1998, 1993, and 1967. None of the plebiscites were supported by a Congress law that made the results self-executing. § Furthermore, only the 1967 and 2020 plebiscites had clear results. In 1967, the majority (60.4%) of Puerto Ricans voted in favor of continuing with current status. In 2020, a smaller majority (52%) voted in favor of statehood. While the plebiscites of 2012 and 2017 also saw statehood as the wining option, but these elections were plagued by controversy regarding the interpretation of voters intentions. ¶ Nevertheless, the status issue became a fundamental political talking point during the legislative term of 2017-2021.
During the term of 2017-2021, the pro-statehood PNP party controlled the majority of the legislature seats and the executive. On February 2017 the Legislature and Governor approved a law to conduct a plebiscite on June 2017. At the time, the intention was to include only decolonizing options, namely statehood, or alternatively, independent status. However, the U.S. Department of Justice did not approve of the language and this became an opportunity for the opposition party, the PPD, to call for a boycott. The boycott resulted in a turnout of barely 23% of the eligible voters, which became an excuse for Congress to remain inactive on the issue. The Resident Commissioner introduced the H.R.6246 (115th): Puerto Rico Admission Act of 2018 which died shortly after being introduced. This prompted the local legislature to pass another law in May 2020, to conduct a second plebiscite during the general elections on November 2020. This plebiscite simply asked voters if they favored statehood yes or no. In this plebiscite, the turnout was 55.02% of the registered voters, and 52.52% of the voters favored statehood.
It’s worth highlighting that local opinions for or against statehood become insignificant if Congress does not show willingness to take action. Meléndez (1991) makes it evident that the dynamics surrounding Puerto Rico’s status referendums diverge between the US and Puerto Rico. Since the 1960s, Congress hasn’t granted approval for any of Puerto Rico’s status referendums. Within Congress, there’s limited interest in discussing the possibility of independence, and there’s minimal disagreement concerning the present Commonwealth status. However, this doesn’t hold true for statehood, which is a contentious topic causing strong disputes. While apprehensions about the partisan consequences of statehood are clear, deeper reservations regarding statehood persist among Congressional Republicans and the broader American conservative spectrum, particularly the more biased factions (Santiago et al. 2023; Whitaker and Giersch 2021). In recent years, discussions in Congress regarding the status question have followed a predictable pattern. Initial discussions praise Puerto Ricans for their loyalty to the nation and their service in the US military. Subsequently, debates intensify, sometimes revealing biased perspectives motivated by racism and xenophobia within a substantial portion of the American right (Barreto 2016). This deeply casts doubt on the effectiveness of pushes for statehood. As long as Congressional Republicans perceive Puerto Rico statehood to favor Democrats in Congress – adding two Democratic Senators and more House members, statehood is unlikely prospect.
Not only is the status matter met with indifference upon its arrival in Congress, but the party framework within Puerto Rico is also undergoing significant changes, a factor certain to exert its influence on the status discourse within the territory. The literature addressing Puerto Rico’s contemporary political structure highlights the stability of its party system (Rivera et al. 1991). However, the combination of the island’s dire fiscal crisis, the aftermath of Hurricane Maria, and the considerable emigration from the island has dismantled this stability, rendering it a thing of the past. The legitimacy of the prior political establishment is now at risk (Melendez and Venator-Santiago 2018; Vargas-Ramos 2018). The repercussions are evident in the outcomes of the two most recent elections: in 2016, support for the PPD and PNP plummeted to 81%, and in 2020, it dwindled to a mere 65% of the popular vote.
In sum, the political status of Puerto Rico has been one of the most salient issues within the Island’s politics. It has shaped partisan politics. Here, we seek to examine how it shapes elite partisan discourse. Using topic-modeling, we explore the nature of the political debate and classify the words used by political leaders during this time-period into the dominant topics of the political communication. We begin with a discussion of latent topic modeling and then measure the elite political discourse in Puerto Rico through Twitter.
Topic modeling in political science
Political scientists have long attempted to create models for understanding political discourse and conflict. Identifying to which topics that people or politicians pay attention, or even the nature of the mix of those topics in the discourse, can be fundamental in understanding the political sphere (Riker 1996; Mayhew 1974). Social science is often faced with determining what events or ideas are the subject of political communications by inferring from the relationships and structures in the data (Ying et al. 2021). Isolating and organizing the latent variables in the data to create a cognitive structure or structures is called topic modeling (Vayansky and Kumar 2020). The earliest iterations of this method involved counting words and finding sets of words in and across documents that that help identify and explain the content of a document so that the topic model would then come to represent the document (Blei et al. 2003).
Finding latent variables in data can be explanatory for political phenomena ranging from election results (Lowi 1964), to policy change (Kingdon 1995), to issue evolution (Carmines and Stimson 1989). Historically, identifying patterns in written data has been labor-intensive, as it required human-coders trained in highly specific systematic methodology. While human-coders can be more nuanced, they may lack consistency without frequent checks. This required significant investment in measures of intercoder reliability. The end result is that modeling subject or topic patterns in political science was, at least initially, used for sizable projects involving large datasets (e.g., Adler and Wilkerson 2013). While costly, the approach has had a significant impact on research and theory (Quinn et al. 2010).
The increasing adoption of computers and algorithms to the process has opened a wider range of possible communication to measure and with more sophisticated approaches that used probabilistic modeling to determine the topics. It has also decreased costs while increasing reliability (King and Lowe 2003). Today, topic modeling is largely seen as an automated process and in social science the topics are used as measures (Ying et al. 2021). Topic modeling is particularly useful measure and describe content moving quickly though the increasingly dynamic digital networks of communication in the political sphere. Social media conversations and topics shift quickly and irregularly. It is unsurprising that unsupervised probabilistic topic models have emerged as a particularly popular strategy for analysis since their introduction to political science (Quinn et al. 2010). Scholars have developed online dictionaries and digital classification systems which can and do create efficiencies and consistencies in measurement, though they may lack the subtlety of human-coders (Vayansky and Kumar 2020). However, the tradeoff seems beneficial.
The ability to discern topic patterns from the data itself, along with the explosion in online records of a wide range of political communication, has greatly expanded the range of study. These studies can give a window into the nature and penetration of the political and policy agendas (Greene and Cross 2017). This is particularly useful when the policies or positions are latent or not otherwise clear from a more superficial view of the communication. Topic modeling can bring out the latent ideological conflicts that could otherwise be missed (Slapin and Proksch 2010), since algorithms can detect the latent semantic structure (Blei and Lafferty 2006; Blei et al. 2003).
Despite their increasing sophistication, topic models still require fine tuning and testing (Agrawal et al. 2018). Neither humans or algorithms can solve the more global structuring question. Classification is difficult without some idea of the nature and substance of the topics before beginning the process. However, this deductive approach may introduce bias. Some scholars have tried a more inductive approach to address that issue using observed data to help create the topic structures so there is no need for a preexisting taxonomy (Quinn et al. 2010). Either approach needs to be validated (Grimmer and Stewart 2013). Here, we consider the most common validation check of topic model results, “bespoke validation” which examines the face validity by checking predictive, convergent, and hypothesis validity. We also whether the researcher benefits from using deductive approaches built on prior knowledge or benefits from relying solely on the inductive approach where the topic structures are constructed only from the observations themselves.
In this research, we explore the specifications for a topic model in the Puerto Rican context. Ultimately, latent topic modeling allows the data to speak by defining its own organizational structure. Instead of having the researchers presume the categories, the data from taken from the Puerto Rican political communication will define the structure and determine the classification of the topics. We presumed that much of the election discourse was centered on statehood, and since the historic positions of the parties on statehood are well known, we believed we had a good idea of what the model should find. Our model should classify the discussion of these topics on party lines. However, the data drawn from Twitter, did not reflect those assumptions. Statehood was not a topic in our model of the full corpus of data. We discuss our model and results below.
Data, descriptives, and analytical strategy
Before collecting data we decided on parameters of the legislator communications we intended to capture and explore. As stated above, there is more than a decade of research on how individual members of the U.S. Congress use of platforms such as Facebook and Twitter (for early research, see Gainous and Wagner 2014; Peterson 2012; Straus et al. 2013) but to the best of our knowledge there are no large scale data driven studies, or for that matter small n qualitative analyses, examining how members of Puerto Rico’s legislative assembly use Twitter to publicly communicate. ∥ Our data here fill that gap. We proceed assuming that research suggesting that elites are predisposed to communicate via soundbites in traditional media (Bennett 2016; Graber 1976) is the case in the Puerto Rican context as well. This makes social media platforms such as Twitter a particularly appealing mode of communication for them. To capture the broadest snapshot of communications, using software that interacts with Twitter’s Application Programming Interface (API), we scraped the universe of tweets from every member with an active account across the entirety of their current terms beginning in January 2017, until the day before the 2020 status plebiscite.
The legislature consists of an upper house, the Senate composed of 30 senators.∗∗ During the analyzed term, the partisan makeup was as follows: New Progressive Party (PNP) = 21 members, Popular Democratic Party (PPD) = 7 members, Puerto Rican Independence (PIP) = 1 member, and Independent = 1 member. The lower house, the House of Representatives, consisting of 51 representatives. The partisan makeup here was: Popular Democratic Party (PPD) = 16 members, New Progressive Party (PNP) = 34 members, and Puerto Rican Independence Party (PIP) = 1 member. Our data come from the 52 legislators who had active Twitter accounts (26 candidates from the House and 26 from the Senate). Senators were more likely to have Twitter accounts. The partisan distribution was as follows: Popular Democratic Party (PPD) = 13 members, New Progressive Party (PNP) = 35 members, Puerto Rican Independence (PIP) = 2 members, and Independent = 2 members. We focus our analysis here on tweets from the two major parties, the PPD and the PNP.
These candidates tweeted 168,636 times across the 3 year period our data cover. The PPD accounts for 47,558 of those tweets, while the PNP accounts for nearly twice as many with 93,298 tweets. The most common words used by the PPD and PNP are presented in Figure 1.†† Given that the PNP tweets roughly twice as much as the PPD, the figures includes words appearing at least 1000 times for the former and 500 times for the latter. Interestingly, words indicative of statehood discussion do not appear here at all (i.e. estadidad and estado). Instead, the most common words for the PPD are centered around government and party (e.g. gobierno, gobernador, PNP, PPD, Senado, Presidente). On the other hand, while the PNP also commonly uses such words, its members seems to use more specific government language (e.g. distrito, comision, municipio), do not mention specific parties commonly, and use more active and value laden language (e.g. participando, amigo, vivo, familia, reunido, trabajando). “Estadidad” actually appears 1365 times (PPD = 319 and PNP = 949), “estado” appears 0 times, and “statehood” appears 63 times (PPD = 10 and PNP = 51). So, while the PNP tweets close to twice as much as the PPD, they mention statehood words at even a higher proportion. The Most Common Words appearing in Legislator Tweets by Major Party. Note: Included are words that appear at least 500 times for the PPD and 1000 times for the PNP. Translations are in the Online Appendix.
Next, we examined the top words within tweets about statehood and across the PPD and the PNP. After excluding tweets that do not contain one of the statehood words and filtering out those state word, the distribution of the top words for the PPD appearing at least 10 times for the former and 20 times. We lowered the thresholds here because, obviously, as indicated above, there are much fewer tweets in this subsample than in the full corpus of tweets. We maintained the same threshold proportionality. The results are clear in Figure 2. The top words change from those in the full corpus, also, many of the words are different from those in the full corpus, and the words across parties significantly vary. The Most Common Words appearing in Tweets about Statehood by Major Party. Note: Included are words that appear at least 500 times for the PPD and 1000 times for the PNP. Translations are in the Online Appendix.
The first difference from above and across parties is that the PPD’s top word here is the PNP while the PNP rarely, if ever, mentions the PDP. The PPD’s next top word is plebiscito (this refers to the regular plebiscite to determine whether Puerto Ricans support statehood). This word does not come up in the full corpus as a top word (see Figure 1) and does not come up for the PPD. There are many other words that show up for each party, respectively, in Figure 2 that do not come up in Figure 1 (e.g. PPD - rosselló, proyecto, económico, etc., PNP - junio, igualdad, vota). It is worth noting that these examples for the PNP are their top 3 words used in those tweets talking about statehood. This suggests, not surprisingly, that the PNP uses words that are supportive of statehood, and second, there are likely different topical discussions across party in their Twitter communication.
This descriptive analysis foreshadows our most important result here; that topic modeling may require some data preparation, trimming and subsetting, to adequately capture those topics that are important parts of the discourse, but not necessarily mentioned frequently. It is unlikely that a the process by which topic modeling works would pick up on those topics coming up less frequently in elite communication - like statehood discussion in this instance. Topic modeling is perhaps most easily understood as akin to latent variable modeling methods, such as factor analysis (Ostrowski 2015). Similar to factor analysis or latent class analysis, topic modeling treats observed words as a function of a much smaller number of latent topic groups. The process groups them in vectors by maximizing the correlation matrix. It does so through a series of simulations wherein the words are organized in a preset number of possible topics (vectors). Say the preset was 10 possible topics, the simulation would organize vectors by the maximum likelihood that words showed up in tweets together if there was only one topic, two topic, three topic, all the way to 10 possible topics. Then we can calculate which number of topics maximizes the correlation across the matrix. This gives us what’s called a coherence score. The number of topics with the highest coherence score is the one where the organization of words in that number of vectors results in the highest correlation across the matrix.
Finally, we can extract the words for each vector/topic and order them by how strongly they correlate with that vector. This makes it relatively easy to infer what the topic for each vector is. For instance, if the words “plebiscito,” “junio,” and “vota” are some of the top words in one of the vectors, then obviously the topic is likely centered on the upcoming vote on statehood. This process of topic modeling as opposed to simply modeling the frequency of particular words makes better theoretical and empirical sense because it allows the data to tell us how the communication is structured rather than us making guesses about it; even if these guesses are informed (Enders et al. 2019). Indeed, capturing the general topics from which particular words and phrases arise allows for a more parsimonious and powerful examination of the structure of elite communication.
Again the descriptive analysis presented above suggests that there may be some data preparation, trimming and subsetting, for topic models to capture statehood discussion. Below we first estimate a topic model of the full corpus of data allowing for 10 possible topics including the correlations between both single words and two word combinations. This allows us to test whether, generally speaking, statehood was one of the primary topics in elite Puerto Rican Twitter communication. After the descriptive analysis above, we suspected that it was not. Then we use the trimmed data from Figure 2 (only tweets discussing statehood), and we estimate subset models based on tweets from the PPD and PNP. Despite the power of the tool, topic modeling may require some subjective interpretive choices with priors and/or post hoc inductive reasoning for it to identify those less pervasive topics in the Twitterverse.
Topic Models of Puerto Rican Legislators’ Twitter Communication
To begin, we fit a topic model on the corpus of Twitter data. Before presenting the results, we needed to determine the optimal number of topics, so we know how many vectors of words to extract from the estimation. While one might initially think that politicians’ digital communication covers are hundreds of topics, the reality is that there are actually a limited range of topics that get any significant and widespread discussion (see Enders et al. 2019). Of course, individual politicians may communicate about topics that do not get broad coverage, but topic models will not pick up on these because the word patterns/usage are largely uncorrelated to other usage. The individualized topics are basically noise in the model. The model will identify those topics with broad attention. As such, we set the maximum number of topics to 10. That was clearly sufficient; the relative coherence scores are plotted in Figure 3. Putting the words into four vectors maximizes the correlation matrix, hence the model indicates that the coherence is clearest when the words are organized into four topics. Estimating the Best Fitting Number of Topics using Coherence Scores. Note: We did not exclude the statehood words here as we did in Figure 2 because we were examining whether they were central to the topics.
Top 20 words representing each topic extracted via latent topic modeling with the full corpus of twitter data.
Note: Translations are in the Online Appendix.
We decided to isolate those tweets that we know are communicating about statehood, and fit the model to see patterns of topics within statehood communication. Further, because we know that the two major parties are divided on this issue, it stands to reason that the topics within statehood communication may vary across party. As such, we also subsample the tweets by party. First, again, before extracting words from each of the topics from the model we utilize coherence scoring to determine the number of topics that maximizes correlation. The results for both the PPD and PNP are presented in Figure 4. Both the model for the the PPD and the PNP indicate that the highest coherence is at nine topics each. Estimating the Best Fitting Number of Topics by Party using Coherence Scores. Note: We did not exclude the statehood words here as we did in Figure 2 because they may be central to the topics.
Top 20 PPD words representing each topic extracted via latent topic modeling for tweets about statehood.
Note: Translations are in the Online Appendix.
Nonetheless, the model does identify a set of statehood subtopics in PPD tweets. Those topics are presented in Table 2. The PPD tends to be negative in their communication about statehood. This is not surprising given that they are opposed to statehood. Notice that the first subtopic in Table 2 is centered on Representative Rob Bishop’s support for statehood. He is a Republican representative from Utah. Members of the PPD are expressing concern because Republicans are typically opposed to statehood. Following that subtopic is discussion about statehood status options and the effect on economic recovery, and then a general discussion about statehood and government including Congress. The next topic also gets at economics claiming that citizen taxes will be adversely affected under statehood. There is also a subtopic here that is simply a clear expression against statehood. The PPD claims that the plebiscite vote creates the appearance of a majority but it is artificial because not everyone participated.
Further, other subtopics express even more negativity - communication about the promise of economic recovery and statehood (a sample of tweets from this subtopic shows that this is critical of argument that statehood would stimulate recovery), critique of the use of public funds for the plebiscite on statehood, tweets about corruption and lies from the PNP, and finally claims that the whole idea of statehood is an illusion.
Top 20 PNP words representing each topic extracted via latent topic modeling for tweets about statehood.
Note: Translations are in the Online Appendix.
Discussion/Conclusion: Political Subtext in Puerto Rico, statehood and the efficacy of topic models
We used topic modeling to explore the the partisan characteristics among the two major parties in Puerto Rico - the PPD and the PNP. In particular, we were interested in seeing how these models would classify language centered on their opposing positions on statehood for Puerto Rico. The results provide a different kind of window into the political discourse. When we fit a topic model to the full corpus of data - all tweets across the period of time from which two plebiscites were held - the topic of statehood was not even classified. This is quite surprising given that we know this issue is at the very core of the difference between these parties and was key policy for the ruling party. This result suggests that an issue can be so in-bedded into the political sphere it does not get as much attention in the back and forth discourse and may be missed by methodologies that use language sorting algorithms. Politicians are not focused on this issue to persuade voters as it is likely seen as a largely static partisan position. The value of a researcher with some understanding of the political history is clear, as well as the need to tailor the models based on the nature of the data and the context.
After some deductive preliminary data trimming and subsetting, we were able to create far more useful models. Specifically, we trimmed the data by eliminating those tweets not discussing statehood and then subset the data by party, and as a result, the subtopic within this larger discussion became quite clear, and in ways that perfectly fit expectations. The PNP, the party born out of support for statehood, spoke with clarity about statehood across a range of topics, and PPD was, in fact, a little less clear as statehood, or should we say, independence, is not necessarily their primary focus. Additionally, the PNP support was evident as was the PPD opposition to the idea of statehood.
Even in a context where the political divisions are relatively clear, as they are in Puerto Rico, relying on techniques such as topic modeling alone may be insufficient and may miss some of the more fundamental political disputes that an active researcher would take into account. Language tools are useful in helping us understand part of the story, but may not be ready to stand alone in describing the data. Understanding what the audience knows or believes is relevant to measuring the impact of the words. A more nuanced understanding of context from the researcher is going to be needed to shape the model so that it captures ideas that may be below the surface conversation.
Interestingly, Puerto Ricans were not ignoring statehood, but rather forming around the split on the topic in a way that structured other issues. Thus, our results allow us to understand the current Puerto Rican political context. In the social media campaign, there were distinct differences between the parties on how they engage with propositions for economic or social reforms to address current pressing issues. However, the PNP is more effective in using the status issue to mobilize its supporters for its electoral advantage. Most of the models show that the topics in which the PNP tended to focus were either for mobilizing its followers to vote for statehood, or to support other PNP politicians at the local level. On the other hand, the PPD’s tweets show a reactionary approach to the status issue and statehood. The PPD failed to provide an alternative, not only to the issue of status, but also in terms of policy reforms. The tweets show a lack of strategy to mobilize its supporters in favor of the status quo. Hence, it is not surprising, that the PPD failed to win the plebiscite, but also the executive branch during the 2020 elections.
The absence of the statehood topic from the initial model can be explained in part by the motivations of the politicians. Where political information is imperfect and there are clear limits on the amount of information that voters are willing to receive, candidates are likely to allocate their time and messages in areas that are most effective at winning votes, while avoiding others (Page 1976). This can result in a low emphasis on important issues that are none the less less likely to be effective for voter appeal. In Puerto Rico, where the issue of statehood is baked into the partisan division, it is understandable how the issue would be a subtext rather than focus. One contribution from our research is the conventional wisdom that political elites often use the status debate as a distraction tactic from other more immediate issues is not supported. Contrary to this belief, our findings suggest that the status debate in Puerto Rico is not overtly raised as a smokescreen, but rather is more firmly embedded in the political subtext.
As for future research, clearly we need to explore different contexts as well as different general topics in Puerto Rico to best demonstrate how choices can be made before fitting topic models that will uncover things we do not already know. We must also be open to new and different approaches to topic modeling that might be better suited to capture latent topics in the discourse (Vayansky and Kumar 2020). We see this work extending beyond social media communication as well. Our next step will be to pull media coverage of the parties adopting a similar strategy of trimming and subsetting to try to produce the most nuanced machine coded classification of statehood communication across parties.
Footnotes
Declaration of conflicting interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.
