The President, Polarization and the Party Platforms, 1944-2012

Scholars generally agree that political elites in the US are polarized. Yet most of our evidence, especially longitudinal evidence, is built on proxy measures of elite ideology that fail to identify the unique dimensions that drive the cleavages between the parties. And our understanding of when elite polarization reemerged is also unclear. This study leverages the party platforms, along with the tools of content analysis, to shed new light on elite polarization. We ﬁnd that, consistent with the literature, elite polarization is an asymmetric phenomenon driven by Republicans primarily motivated by economic issues. Further, we show that modern elite polarization emerged starting with the 1980 election..


Introduction
We focus in this article on the timing of and issues driving elite polarization in the US. Polarization at both the elite and mass level has drawn increasing attention from political scientists in recent years. Broad consensus exists that elite actors -particularly members of Congress -are ideologically divided by party identification (Fleisher and Bond 2004;McCarty, Poole, and Rosenthal 2008). Evidence for mass polarization is more contentious, with some claiming that the electorate is "sorting" ideological views based on party identification (Levendusky 2009;Fiorina, Abrams, and Pope 2011) and others noting the increasing ideological divergence, particularly of politically active partisan identifiers (Abramowitz 2010). Still others claim that the electorate only "appears" polarized because of the biased menu of choices they are provided by party elites (Sniderman 2000;Sniderman and Bullock 2004;Fiorina, Abrams, and Pope 2011).
Currently, when scholars discuss elite polarization, they are referring to polarization in Congress. Scholars typically measure polarization in Congress using the divergence of aggregate roll-call based measures of congressional ideology, such as inflation-adjusted ADA scores or NOMINATE scores. We use a different approach to studying elite polarization by examining changes in the party platforms through time.
Released every 4 years, party platforms are an explicit construction of party ideology by a broader set of party elites. Accordingly, they offer unique leverage and insight into the changing ideology of elites (and perhaps ideological divergence as well). Yet surprisingly few studies have systematically analyzed the content of the party platforms, and no study of which we are aware specifically attempts to leverage the platforms to examine elite ideological divergence over time. We undertake such a study here, performing a systematic content analysis of the party platforms over time. We use the basic techniques of text mining, supplemented by qualitative evaluation of the content of the platforms, to examine how they cluster (or diverge) over time. We pay careful attention to the role of the president in the process, highlighting the particular influence that presidential nominees hold in constructing the platforms. Case studies of the platform writing process have found that presidential nominees exert considerable control over the content of the party platform (Fine 1994;Borrelli 2001). Accordingly, we discuss the role that particular presidential nominees might have played in establishing the current ideological divergence between the parties.

Literature Review Polarization
The study elite and mass polarization -the increasing ideological division between the major political parties -has obtained growing interest in American politics in recent years (for a thorough review, see Layman, Carsey, and Horowitz 2006, see also Levendusky 2014). We offer a brief review here.
If there is one point of universal agreement, it is that elites in Congress are polarized. The evidence is persuasive across multiple dimensions. Across studies (Hetherington 2001;Fleisher and Bond 2004;McCarty, Poole, and Rosenthal 2008;Theriault 2008), across candidates (Ansolabehere, Snyder, and Stewart 2001a), across measures (NOMINATE: Poole and Rosenthal 2007;inflation-adjusted ADA scores: Groseclose, Levitt, and Snyder 1999;Anderson and Habel 2009;common space Bayesian scores: Clinton, Jackman, and Rivers 2004;NPAT survey scores: Ansolabehere, Snyder, and Stewart 2001b), and across time, members of Congress (MCs) have grown more ideologically extreme and separated from one another on the basis of party identification.
Evidence is more mixed at the mass level. Some (such as Abramowitz and Saunders 1998;Abramowitz and Saunders 2008;Abramowitz 2010) point to a growing divergence of the mean response to the seven-point ideology scale by party identification, paired with similarly more extreme responses on issue specific questions (such as gay marriage or abortion, see Hetherington and Weiler 2009), as evidence of polarization in the electorate. Others (like Fiorina, Abrams, and Pope 2011;see also, Sniderman 2000;Sniderman and Bullock 2004) claim that mass polarization is just an artifact of the polarized choices provided by. They note that the same number of respondents (very few) place themselves in the most extreme survey response categories. Others contend that, even if ideological change is happening in the electorate, it is limited in scope (DiMaggio, Evans, and Bryson 1996). At most, individuals realign their ideological beliefs to fit their party identifications through a process called "sorting" (Levendusky 2009).
This study is less focused on the particular contentions of the polarization debate. Rather, we accept that elites are polarized -a point of agreement across studies -and remain agnostic on whether or not mass partisan identifiers are polarized. Instead, we focus on the nature of the polarized choices elites offer to the electorate through party platforms. Zaller (1992) posits that individuals form opinions on issues as a result of the elite choices offered to them, particularly through the mass media. This is consistent with earlier studies of issue evolution (Carmines and Stimson 1989), whereby issues change in their polarizing nature as a result of elite discourse through time. Recent theories of elite polarization, such as conflict extension (Layman and Carsey 2002;Carsey and Layman 2006) posit that elites polarize iteratively on new issues while retaining partisan cleavages on past issues. For each new issue that the parties adopt opposing positions, some new subset of the electorate might be drawn to the party.
The overarching point is that elites are polarized and send polarizing messages to the electorate in an attempt to garner support through party-based signals on ideological issues. Most of our knowledge of this phenomenon, however, relies on on proxy measurement of elite ideology in Congress. Even the few direct measures that are available, such as the National Political Awareness Test (NPAT) (see Ansolabehere, Snyder, and Stewart 2001b) are rarely, if ever, repeated over time, probably because of elite resistance to being surveyed for fear of being audited by members of their constituency. In large part, then, our understanding of elite polarization has been built on indirect expressions of preferences of members of Congress through roll-call votes, and may not reflect the views of the broader ideological elite.
Important signals transmitted to the electorate exist in the national party platforms. The party platforms are direct measures of partisan ideology, because they are plainly a construction of party elites defined more broadly. Moreover, and particularly useful for the study of polarization, they are repeated regularly over time. As a result, they offer an opportunity to examine the changing nature (and divergence) of elite ideology over time through a direct, repeated measure of elite ideology. In this process, there is no single more influencing elite actor than the president (Fine 1994). Even when just nominees, presidents serve as a rallying point under which to consolidate the party ideology (Borrelli 2001). Accordingly, we pay special attention to the role the president (or nominee) has in constructing the party platform.

Party Platforms
Platforms spell out the issues and programs defining the political parties. They are manifestations of the policy preferences of those most central to the party. These documents are meant to appeal to the party faithful, as well as attract new supporters. Platforms provide members with blueprints for campaigning and for governing after the election (Pomper and Lederman 1980, pp. 152-153;Budge and Hofferbert 1990;Sanbonmatsu 2006, p. 98; but see King and Laver 1993). They also constitute a communication mechanism for party elites with those who would represent the party. As written documents, platforms are a mechanism for holding party members accountable, thereby facilitating party responsibility. Since the 1970s and the rise of candidate-centered, rather than party-centered elections, they also manifest the preferences of the presidential candidates, as well as all others seeking office under the party banner (Maisel 1993).
There is debate surrounding the relevance of party platforms. While scholars studying platforms argue that elites use platforms to communicate their policy goals to the public and that the public uses platforms to make decisions about candidates (Fine 1994;Kidd 2008;Simas and Evans 2011), others argue that it is unlikely that a generally uniformed electorate would take the time to process all of the information in lengthy party platforms. At the very least, voters might become informally aware of platform content. According to Paddock (2011), "while most voters do not read party platforms, there is empirical evidence that platform pledges reach voters through indirect means (e.g., news accounts)" (713).
Others argue that party platforms may not be made for public consumption at all. Instead, party elites simply use the platforms to ensure the support of "established constituencies" (Monroe 1983). Elites author the platform to lock in the support of some interest groups and sway the support of others. Reinhardt and Victor (2012) argue that interest groups use party platforms to articulate their interests and parties incorporate group preferences to reward groups that can mobilize voters. By grouping enough interests together, the platforms are engineered to give the party the best chance of achieving its electoral outcome. Platforms are not intended to mobilize support for the party directly. Even if platforms might not be directly influential on policy formation, there is some evidence that policy makers attempt to adhere to the goals established in their parties' platforms. Budge and Hofferbert (1990) demonstrate strong links between election platforms and government outputs -specifically, federal spending priorities. Pomper and Lederman (1980, p. 164) find that parties follow through on the pledges they make in platforms more than 50% of the time.
From our perspective, none of these questions are particularly relevant. Our intention is to use the party platforms as a direct measure of party ideology to investigate ideological polarization among party elites, broadly defined, over time. Accordingly, to the extent that the platforms are a valid construction of elite ideology, we can use them to track elite polarization, regardless of whether they are consumed by the public or correlate with policy outcomes.
Only a handful of scholars have used party platforms to study polarization. Ginsberg (1972Ginsberg ( , 1976 analyzed the text of party platforms in his work studying the link between electoral choices and the policy positions taken by major parties. He found that changes in "party cleavages" are linked to electoral realignments associated with "critical conflicts" where one or more parties changes positions in hopes of mobilizing a new segments of the electorate. Ginsberg (1972Ginsberg ( , 1976 sought to understand how parties' policy preferences changed over time, but the "conflicts" and "party cleavages" he described are akin to polarization. In recent years, polarization scholars have studied state level party platforms (Coffey 2011;Paddock 2011). Coffey (2011, p. 311) credits the renewed interest in party platforms with the advancement of content analysis software that makes it easier to wade through lengthy political texts. One study has applied these new techniques to national party platforms. Kidd (2008) used the "wordscores" coding program to analyze the content of national party platforms between 1996 and 2004. He found that American political parties were not ideologically polarized over this time period âȂŞ especially on domestic issues. Coffey (2011) argued that the unbelievable finding reported by Kidd (2008) was a consequence of his over-reliance on automated text analysis software and the lack of attention paid to the actual content of the platforms.
The data presented in the next section attempts to combine new content analysis techniques with some of the systematic content analysis strategies of the past. Our approach leverages the platforms to examine elite ideology change from 1944 through 2012. In the next sections, we describe the data and the specific analytic procedures applied.

Data
The data for this project come from the thirty-six platforms for the Democratic and Republican parties during presidential election years from 1944 through 2012. The platforms were retrieved as electronic text from the website of The American Presidency Project (Peters and Wooley 2012). The text lengths varied across years from a minimum of 1375 words in 1944 to a maximum of 38,195 words in 1980 for the Democrats, and from a minimum of 2755 words in 1948 to a maximum of 35,847 words in 1988 for Republicans. Because the text lengths vary so much, we standardized the counts for quantitative analysis to provide comparability across platforms. Our goal in this project was to extract meaning and quantitative measures from the Democratic and Republican platforms that would reflect the timing and degree of polarization between the parties through time. Fine (1994) describes in detail how the political party platforms are constructed. Prior to the presidential nominating conventions, each party's national committee designates a platform writing committee. The platform writing committee conducts hearings prior to the convention. Testimony is typically received from interest groups, trade associations, labor unions, corporations, educational institutions, research institutes, campaign organizations, party committee members, elected officials, as well as private individuals. Fine argues that presidential candidates are also "important in the platform writing process in the post-reform era" (Fine 1994, p. 856). Especially as the primary season continues to determine a nominee relatively early in the election cycle, presidential nominees are able to consolidate their influence in the platform writing process (Borrelli 2001). For example, recall that President Obama used a heavy hand during the 2012 Democratic convention to get the word "God" back into the Democratic platform, as well as to acknowledge Jerusalem as the capital of Israel (Jackson 2012). After hearing testimony, the platform writing committee constructs a draft document to be presented to the delegates at the convention. The convention then deliberates over the draft document, possibly amending it, and then approves the final document.
The platform writing committee is typically comprised of party leaders from various segments of the political system. The final results from these platform-writing efforts for each major political party and presidential election year yielded the data analyzed by this project. The next section details the specific methodologies used to analyze the content of the party platforms.

Methodology
We used Text Mining methodologies to extract meaningful quantitative measures from the party platforms. Information on the details of text mining, as well as our specific methodological choices, can be found in the Appendix. Instructive exam-ples in both Perl and R for implementing the technology can be found in Francis and Flynn (2010). A more comprehensive introductory text is Weiss, Indurkhya, and Zhang (2010).
Our text mining was performed with the unique platforms as the document units, with their component terms reduced to stemwords, stop words removed (e.g., demonstration/demonstrated becomes demonstrat; national/nationhood becomes nation, etc.), and sparse terms eliminated. We used the R package tm to pre-process the documents, and as an engine for the initial Text Mining analysis (Feinerer 2008;Feinerer and Meyer 2008). This package was used to create the document term matrix for use with other R functions and packages. It was also used to develop term frequencies for each term within documents, as well as to find word associations. Analyses were conducted separately on the Democratic and Republican platforms, as well as combining them within a single corpus.
We first conducted hierarchical cluster analyses on the document term matrix using the built-in R function hclust. Information on the details of cluster analysis can be found in the Appendix. We used k-means clustering to identify a bend in the plot of within groups sum of squares for extracted clusters from two to ten. This approach is recommended to determine the proper number of clusters to initiate the algorithm (similar to interpreting a scree plot in factor analysis) (Everitt and Hothorn 2009, p. 251). Based on this plot, and the theory that the two parties should form two distinct clusters, we report below the two-cluster dendrogram tree defined by using Ward's (1963) minimum variance method. At the broadest, the entire set of data points (i.e., all platforms) may be considered a single cluster, while at the finest, each data point (i.e., each platform) can be considered a unique cluster. Ward's method incrementally adds data points that are closest to all members of an existing cluster, thereby producing compact, spherical clusters. Note that the results reported below do not deviate significantly when we set the number of clusters at three or more. The dendrogram tree depicts which data platforms are nearest to each other in content, thereby showing the relative similarity of the party platforms.
Following the cluster analysis, we evaluated the platform content to determine what was driving the unique clusters. Thirty-six platforms produced 8000 unique stemwords for Democratic platforms and 9000 for Republican platforms. Accordingly, we used four different approaches to help identify polarizing words. The first was a human coding of all stemwords as polarizing or not using coder judgment. The second was a human coding of only frequently used stemwords as polarizing using coder judgment. Frequently used stemwords were identified as those at least one standard deviation above the mean number of uses, the idea being that frequently used words might be more important to polarization. Two coders performed the human coding, both cross-validating one another. Finally, knowing that elite polarization has increased over time, we regressed the counts for each of the stemwords, by party, on a simple time trend and a second time trend that began in 1980. We were left with four potential indicators of "polarizing" words: the human coding of all terms, the human coding of only frequent terms, and the significance of two trend regressions. Polarizing words were defined by stemwords appearing on at least three of these four measures. The result was 150 polarizing stemwords for the Democrats and 220 for the Republicans.
To evaluate what it was about the content of the party platforms that produced particular clusters, we classified polarizing terms into five dimensions: economics (stemwords like compani, economi, manufactur), regulation/size of government (stemwords like ban, climat, pollut, bureaucrat), morality (stemwords like abort, faith, gay), social welfare (stemwords like crime, health, povert), and defense/ foreign policy (stemwords like arsenal, homeland, iran, nuclear). The full list of polarizing stemwords, by dimension, are shown in Appendix Tables 1 and 2. Later, we report time series graphs of the frequencies of the polarizing words, aggregated by dimension for each party and standardized by platform length, to gain a sense of the timing and magnitude of the divergence between the party platforms. Eisenhower-Nixon years, perhaps suggesting the presidency's distinct imprint on party platforms. The separate twig for Republican platforms through this era shows they were distinct from both earlier and later eras of the Republican Party, as well as from Democrats. The second twig of the second branch contains Democrat and Republican platforms for 1944 and 1948, along with the 1988 and 1992 Democratic platforms. During the early period, there was very little separation between the two parties. And, the Democratic platform in 1988 was not all that distinct from those of the early party. However, it is interesting that the 1992 Democratic platform, with its own twig, is distinct from the early platforms. Interestingly, this corresponds to case-study evidence that Clinton specifically attempted to eschew traditional liberal priorities in favor of a more centrist platform (Borrelli 2001). Again we reiterate the importance of the nominee in the platform writing process: Clinton's influence over the platform was such that the 1992 document was "not just compatible with but thoroughly imbued with the Clinton agenda" (Borrelli 2001, p. 455).

The Platforms Cluster Analysis
The second cluster on the left side of Figure 1 contains the most interesting result from the analysis. It consists of the nine Republican platforms from 1980 through 2012, along with two Democratic platforms in 1980 and 1984. The right branch of the second cluster consists of Republican and Democratic platforms during the Reagan era. Democrats moved to the right along with Republicans during the Reagan administration, but only temporarily. The left branch of the second cluster consists solely of Republican platforms starting in 1992. More generally, these results show that Republican platforms after 1980 were very distinct from Republican platforms prior to 1980. They were also distinct from Democratic platforms prior to 1980 and after 1988. In this regard, note especially that on the dissimilarity scale (left side), these platforms form a cluster that is dissimilar relative to every other platform before 1980, Republican or Democratic. Republicans seemingly became more cohesive in their differences.
Particularly interesting in this process is the influence of the polarizing Ronald Reagan. It is well known that, even though they are more extreme ideologically than other members of the party (Stone and Abramowitz 1983), convention delegates are focused on winning elections, rather than preserving ideological purity. In 1976, traditional party activists (who backed Ford for the nomination) outweighed ideological purists (who supported Reagan) (Sullivan 1977). By 1980, the much more conservative Reagan had engineered a legitimate sense of electability: 70% of Republican convention delegates perceived Reagan as having a better chance against Carter than Bush (Stone and Abramowitz 1983, p. 949). For the first time, then, the presidential nominee of the Republican Party was viewed as widely electable without sacrificing ideological purity to centrists for electability. As the nominee, Reagan and the conservative intellectual forces behind him (e.g., the Heritage Foundation, the American Enterprise Institute) exerted strong presidential influence over the platform writing process, setting in motion a movement by the Republican Party to the right.
Thus, modern Republicans have clearly separated ideologically from the old Republican Party. The old Republican Party was distinct from the Democratic Party between 1956 and 1968, but was not all that distinct during the 1940s, early 1950s, and 1970s. Since the 1980s, Republicans seemingly cater to their median partisans, while the old Republican Party seemingly catered to the median voter, as did Democrats. These results are consistent with past work suggesting that polarization of the American system in the modern era is an asymmetrical phenomenon (Wood and Jordan 2011). Modern Republicans are the primary source of polarization in the American system.
Of course, the cluster analysis alone cannot tell us what was distinct about the party platforms, or what in them caused the change through time. We must, therefore, move to an analysis of what is in the platforms to make these determinations. We turn in the next section to an analysis of the platform content to shed light on these topics.

Analysis of the Platform Terms
Recall that we have counts of polarizing words, aggregated by dimension, and standardized by platform length for each platform-year. We present these data two ways: grouped by issue dimension and grouped by party. Figure 2 presents the trends by issue dimension. Note especially the economic dimension from the first panel. The usage of polarizing economic words, particularly by Republicans, jumps considerably in 1980, suggesting that the impetus for the divergence observed in the cluster analysis was the Republican focus on economic issues. In fact, this is the only dimension that jumps considerably in 1980 for either party. Figure 3 presents the trends by party. Again, the only dimension to grow considerably in 1980 is economics for Republicans. In fact, the consistently lowest dimension for Republicans is the moral dimension, which includes such issues as abortion and gay marriage. This analysis suggests the primacy of economics in driving polarization, relegating other issues to a place of secondary importance. Note as well that economics and regulation/size of government dimensions are always the most important of the polarizing dimensions for Republicans. Figure 3 also suggests that Democrats are responsive to Republican polarization, and that polarization is driven asymmetrically by Republicans (echoing the findings of the cluster analysis). Polarizing words account for very few of the total words in the Democratic platform until the mid-1980s, after which Republicans had begun to polarize along the economic dimension. Reiterating the importance of economics, however, when Democrats did polarize, they tended to do so along the social welfare and economic dimensions.
These findings are highly consistent with the broader literature on polarization and elite cleavages. McCarty, Poole, and Rosenthal (2008) find evidence that elite polarization in Congress -as measured through NOMINATE scores -is related to growing economic inequality. Aldrich and Rohde (2000) also find that the first dimension of what they term "conditional party government," a measure of homogeneity of the majority party versus the heterogeneity of the minority party, is almost entirely explained through time by economic policy, relegating social and moral policy to a second dimension of conflict. Additionally, our findings fit with theories of conflict extension , rather than conflict displacement. Only after first polarizing on economic issues do elites move to secondary social issues. But they retain their established cleavages on economic issues, rather than only polarizing on a single dimension at a time.

Conclusions
Party platforms are a construction by party elites, broadly conceived, of party ideology. They spell out the issues and programs of the political parties. Since the rise of candidate-centered elections, presidential nominees have had greater influence in constructing these documents and, in turn, party ideology. Advancements in content analysis software have generated renewed interest in the study of party platforms. These programs make it easier to navigate these lengthy documents. Surprisingly, most of this work has been applied to platforms at the state level. We use political party platforms to develop a direct longitudinal indicator of changing elite ideology at the national level.
The results from the cluster analysis show that Republican platforms in the 1970s were not all that distinct from Democratic platforms. Starting in 1980, the Republican Party moved to the right, and their platforms have clustered together ever since. These results are consistent with past work suggesting that polarization of the American system in the modern era is an asymmetrical phenomenon (Wood and Jordan 2011). Modern Republicans are the primary source of polarization in the American system. This movement to the right started with the rise of conservative President Ronald Reagan. For the first time, electability and ideological purity coincided for the Republican Party, leading to the nomination and election of a conservative leader that welcomed a conservative movement within the party. This movement was directly reflected in the party platforms, in no small part because of the tremendous influence that nominees exert on the platform writing process.
We also developed a catalogue of polarizing words and aggregated the words into five issue dimensions. Our analysis suggests that economic issues began driving an important wedge between the parties starting in 1980. Economics and regulation/size of government are important issue areas for Republicans. Demo-crats have tended to polarize along the social welfare and economic dimensions. These findings are consistent with the broader literature on polarization and elite cleavages, suggesting that polarization is related to growing economic inequality.
More importantly, the party platforms examined here are a critical source for identifying elite ideology, broadly conceived, over time. Instead of relying on proxy measures rooted in Congress to investigate elite polarization, we can utilize direct measures of party ideology, as constructed by the parties themselves. Pairing these party platforms with modern tools of content analysis, we can better understand the changing nature of elite ideology: both in the aggregate (how the parties cluster together) and the specific dimensions along which these fractures are occurring.
cluster the objects into sub-groups based on their inter-object similarities. Thus, we first scaled the document term matrix using the built-in R function scale to standardize the columns for each term. Then, the Euclidian distance between each column was calculated using the built-in R function dist. The resultant distances were then clustered hierarchically using hclust and a variety of clustering algorithms.   cuba  famili  corrupt  learn  cost  defens  infant  court  medic  costeffect  global  kill  endang  medicaid  cut  immigr  life  energi  medicar  debt  iran  mainstream  environ  medicin  dollar  iraq  marriag  environment  patient  earn  israel  moral  expand  prescript  econom  korea  murder  feder  read  economi  mexico  parent  fraud  school  employe  militari  prayer  freedom  scienc  entrepreneur  missil  pregnanc  fuel  senior  financi  nuclear  punish  gas  sentenc  fund  palestinian  religi  govern  social  gain  patriot  religion  gun  student  grow  russia  sexual  illeg  teacher  growth  sovereignti  stem  individu  tuition  homeownership  taiwan  teen  infrastructur  import  terror  toler  law  invest  threat  tradit  lawsuit  job  threaten  unborn  lawyer  lowincom  troop  victim  legal  manufactur  war  violenc  liberti  market  weapon  violent  limit  marketplac  world  women  litig  middl  local  money  lower  own  mandatori  owner  oil  ownership properti