Nick Beauchamp

Debate and deliberation play essential roles in politics and government, but most models of debate presume that debates are won mainly via superior style or agenda control. Ideally, however, debates would be won on the merits, as a function of which side has the stronger arguments. We propose a predictive model of debate that estimates both the effects of linguistic features and the latent persuasive strengths of different topics, as well as the interactions between the two. Using a dataset of 118 Oxford-style debates, our model's combination of content (as latent topics) and style (as linguistic features) allows us to predict audience-adjudicated winners with 74% accuracy, significantly outperforming linguistic features alone (66%). Our model finds that winning sides employ stronger arguments, and allows us to identify the linguistic features associated with strong or weak arguments.

Spatially or temporally dense polling remains both difficult and expensive using existing survey methods. In response, there have been increasing efforts to approximate various survey measures using social media, but most of these approaches remain methodologically flawed. To remedy these flaws, this paper combines 1200 state-level polls during the 2012 presidential campaign with over 100 million state-located political Tweets; models the polls as a function of the Twitter text using a new linear regularization feature-selection method; and shows via out-of-sample testing that when properly modeled, the Twitter-based measures track and to some degree predict opinion polls, and can be extended to unpolled states and potentially sub-state regions and sub-day timescales. An examination of the most predictive textual features reveals the topics and events associated with opinion shifts, sheds light on more general theories of partisan difference in attention and information processing, and may be of use for real-time campaign strategy.

Advertisements, talking points, and online and social media content often take the form of short chunks of spoken or written text, yet crafting these short documents remains more art than science due to the extremely high-dimensional nature of textual content. Focus groups, A/B testing, and substantive theories of political opinion can suggest the general themes for a persuasive text, but do little to help shape it on the word or sentence level. Recent progress has been made in moving beyond binary experimental testing into higher dimensions \cite{hainmueller2014causal}, harkening back to earlier work outside of political science in fractional factorial design, but such approaches remain insufficient for the extreme case of free-form textual optimization. This paper instead develops a new approach and machine learning optimizer to craft short chunks of text in order to maximize their persuasive impact. First, a large collection of sentences in support of Obamacare are scraped from obamacarefacts.com and parameterized via a simple 7-topic LDA. Each text treatment comprises three sentences from this pool, parameterized in 21D-space. The persuasive effects of each treatment is assessed using Mechanical Turk subjects, and these treatments iteratively improved using nonlinear optimization over the 21D parameter space to suggest progressively more persuasive three-sentence combinations. A new optimization algorithm is designed specifically for this class of problem and shown though extensive Monte Carlo testing to out-perform the dominant existing method. Furthermore, lasso techniques allow us to assess the response surface using the samples collected during the maximization procedure, allowing more general inferences about the individual and interactive effects of the topics on opinion. Together, these procedures constitute a new approach to designing and testing new persuasive text, and not just assessing the effects of existing treatments.

This paper develops a new tool for the visualization and analysis of individual documents, representing the progression of a document as a trajectory through its own word-space. We present two models for estimating this document trajectory, the first a slower probabilistic generative model, and the second a faster approach using principal component analysis, which produce similar results. Document trajectories are then analyzed for a large corpus of important political speeches, with the goal of identifying characteristic topological patterns than may reflect hidden structural patterns underlying substantively very different speeches. Speech trajectories are clustered into topologically similar patters using affine transformations, revealing both the most common rhetorical patterns in these speeches, and differences in core patterns between speakers (such as Bush vs Obama) and over time. Modeling temporal semantic structures in single documents opens the way for a new analysis of rhetorical structure in political speech, and ultimately understanding the effects of these structures on listeners.

This paper shows how legislators' written and spoken text can be used to ideologically scale individuals even in the absence of informative votes, by positioning members according to their similarity to two reference texts constructed from the aggregated speech of every member of each of two major parties. The paper develops a new Bayesian scaling that is more theoretically sound that the related Wordscores approach, and a new vector-based scaling that works better than either at matching the vote-based scaling DW-Nominate in the US Senate. Unsupervised methods such as Wordfish or principal component analysis are found to do less well. Once validated in the US context, this approach is then tested in a legislature without informative voting, the UK House of Commons. There the scalings successfully separate members of different parties, order parties correctly, match expert and rebellion-based scalings reasonably well, and work across different years and even changes in leadership. The text-based scaling developed here both matches existent scalings fairly well, and may be a much more accurate window into the true ideological positions of political actors in legislatures and the many other domains where textual data are plentiful.

This paper presents a new, bottom-up approach to modeling the effects of television advertising on vote intention. Because ads are so numerous and varied, existing studies generally begin with a specific theory of persuasion, or must simplify the data down to a few latent dimensions or effective ads. Instead, this new approach first develops a one-at-a-time regression technique to estimate the effects of hundreds of different ads on vote intention during the 2004 presidential campaign. The aggregate effect of advertising is found to be significant, though many individual ads have small or backfiring effects. To explain these varying effects, new automated text analysis procedures are developed which can predict the effects of ads based only on their text, and reveal complex and asymmetric strategies that mix affect, policies, issue ownership, negativity, and targeting. This bottom-up procedure constitutes a new method for understanding persuasion in campaigns and politics more broadly.

Political opinion, and hence political behavior, is shaped largely via talk: with family, friends, and increasingly, online. But such discussions are often taken to be unstructured ideological posturing with little persuasive effect. This paper instead proposes a more deliberative model, where argument consists of the strategic exchange of topics, frames and ideas that are interconnected in a complex conceptual network. This network of ideas is inferred using new Bayesian topic modeling methods applied to a new dataset of millions of political discussions from the largest political forum online. By modeling arguments as a Markov process with the network as transition matrix, we can predict what topics arguers will deploy in response to each other: contrary to framing or expressive models of speech which predict that speakers will echo or ignore their interlocutor, this new model shows discussion to be more deliberative, where speakers offer ideas, facts, and topics relevant to, but missing from, what their interlocutor has said. In the longer term, panel vector autoregression methods reveal that a significant subset of users appear to change their views in response to what they hear, although listeners are biased against speech too unlike their own. Finally, because users can vote to recommend posts, factor analysis of this voting data reveals a strong underlying ideological dimension, largely centering (on this mainly Democratic forum) around criticism or praise of Obama. This ideological behavior is illuminated by the conceptual network: we see criticism of Obama largely based on left-wing policy issues, whereas praise is largely emotional and personal. This asymmetry is consistent with numerous psychological models of ideology, and also reveals which discussed topics may influence voting behavior in the long term. This text-based network of ideas allows us to model both the short-term dynamics of political argument and long-term opinion change, using a framework that is as rich, complex, and substantively interpretable as people have always claimed their arguments were.

This paper finds that the decisions of the Supreme Court over the last 10 years can be systematically predicted using the text of the briefs and oral arguments that precede those decisions. This automated text analysis also sheds new light on the variety of mechanisms underlying decisions, including the tradeoffs between political and procedural decision-making. Support vector machines and ensembles of univariate regressions are first used to predict decisions of the court as a whole and of individual judges, with up to 62\% (out-of-sample) accuracy -- better than many experts. These ensemble methods are then used to extract a small subset of words that are significantly associated with tendencies in the court and individual judges to vote more liberally or more conservatively. These terms reveal a large array of different decision-making strategies at work, with implications for the ongoing debates between legal reasoning and precedent on the one hand, and policy preferences and ideological attitudes on the other. The decisions of the conservative justices appear especially predictable, revealing a tendency to vote more conservatively on constitutional and criminal topics, but more liberally when briefs emphasize the intentions of Congress and statutory language. Analysis of oral transcripts reveals how the qualities of debate affect decisions, in particular a tendency for some conservative judges to vote more liberally when the conversation is intense but marked with laughter. Together, the text of briefs and oral arguments provide distinct but complementary insights into the varied decision-making of different justices. More broadly, this approach both constitutes a useful predictive tool, and sheds light on numerous existing theories of judicial decision-making while suggesting many new avenues of exploration.

This paper introduces a new maximization and sampling algorithm, "Blossom," along with an associated R package, which is especially well suited to rugged functions where even approximate gradient methods are unfeasible, such as those encountered in highly nonlinear likelihoods and in matching for causal inference. The Blossom algorithm is an evolutionary strategy related to the Estimation of Multivariate Normal Algorithm (EMNA) or Covariance Matrix Adaptation (CMA), within the general family of Estimation of Distribution Algorithms (EDA). It works by successive iterations of sampling, selecting the highest-scoring subsample, and using the variance-covariance matrix of that subsample to generate a new sample, with various self-adapting parameters. Compared against a benchmark suite of challenging functions introduced in Yao, Liu, and Lin (1999), it finds equal or better maxima to those found by the genetic algorithm Genoud introduced in Mebane and Sekhon (2011). The algorithm is then applied to two real-world problems from political science: maximizing a difficult likelihood function combining both utility and spatial metric parameters, and two high-dimensional matching problems, where it produces better results than many existing packages in R such as GenMatch (Sekhon, 2011). Finally, to utilize the samples generated in the process of maximization to accurately sample from the posterior likelihood, approximate voronoi cells around sample points are used to approximate numerical integrals. This sampling method is shown to produce better results than a simple metropolis MCMC sampler using benchmark distributions such as the "banana" function.

Most spatial models of preference assume that the spaces in question are Euclidean, and that utility functions are quadratic. Although increasing work has recently been done in estimating utility functions from empirical data, and some theoretical work has been done with non-Euclidean spatial metrics, relatively little has been done to estimate spatial metrics from empirical data in the political context. This paper employs maximum likelihood techniques to directly estimate both spatial metrics and utility functions from ANES survey data. A simulation is also conducted to confirm that these methods can indeed accurately recover spatial and utility parameters. The results show that in the most general case, the spatial metric appears close to Euclidean, but the utility function is much less "risk-adverse" than generally assumed. Furthermore, different combinations of issues produce different estimates for both spatial metrics and utility functions, although in all cases the utility functions are far from quadratic. Of practical importance, coefficients on policy variables appear to vary with different spatial metrics and utility functions, indicating that assumptions made about the metric of a space may be biasing empirical results.

Research

Research Interests

Publications

Working Papers

Research in the news

Teaching

Education and Employment

Contact