More Website Templates @ - September 9, 2013!

    Brice D. L. Acree

To contact me, send an e-mail to brice [dot] acree [at] unc [dot] edu.

Brief Bio

I’m a graduate student in statistics and political science. I grew up in rural Kentucky and attended LaRue County schools, where I spent my later years as a debater and extemporaneous speaker on the speech and debate team. I graduated from Dartmouth College in 2009 under the supervision of Prof. Dean Lacy. During my undergraduate years in New Hampshire, I worked on several local, state and national campaigns. I spent two stints in Washington, D.C., working at Greenberg Quinlan Rosner and Lake Research polling firms. I joined the Carolina political science department as a Ph.D. student in 2012 and the statistics department as a M.S. student in 2013. My advisor is Jim Stimson.


Download my research statement.
My primary research focus centers on quantitative methodology, machine learning and computational text analysis. In particular, I’m interested in methods for extracting ideological cues and other important signals from political texts. My dissertation examines advanced methods for analyzing political corpora. Part I builds on the logistic-normal structural topic modeling approach. Part II develops and reviews deep learning neural network approaches to extracting patterns from political texts. I evaluate the performance of various neural network models, and compare these to structural models more commonly found in political science research. Part III evaluates the ideological content of presidential candidate stump speeches in the 2008, 2012 and 2016 elections, and to presidential rhetoric from 2009-2015.

I also have a broad interest in Bayesian statistics and Monte Carlo simulation, inferring networks from overlapping language, and in tracing the diffusion of ideas in political media. For more on these, see my projects below.

My dissertation, "Ideological Rhetoric," advances methods for extracting ideological signals in political texts.
I am a member of the Carolina Text as Data Lab, a collaborative group of scholars working on computational text methods for social science research.
I mostly program in R, but I also use C++, Python, Julia, and Matlab, and teach Stata when useful. For many Bayesian models, I recommend Stan.
I’m currently working on several exciting projects with a diverse and interdisciplinary group of collaborators.
My dissertation seeks to make three contributions to the field: to explore the content of ideological schools of thought in the United States; to collect and employ a novel source of data; and to advance methods for modeling large-scale text corpora.

Chapter 1 builds on my past work and contributes to the structural topic modeling approach made popular by Roberts, Stewart, Tingley & Airoldi (2013). With a theoretically-motivated model of ideological rhetoric, I show the advantages of adding additional structure to the traditional logistic-normal topic model, especially for measuring complex language features like ideological perspective. This additional structure comes from three sources: variation in ideological language over time; differences in ideological attention to various topics, and differences in the tone with which ideologues discuss political issues. The first two components are employed widely in political methodology, though almost always separately; and tone---i.e., sentiment, or positivity/negativity of language---is not often discussed in political research. This chapter resolves that gap by providing a straightforward model and inference procedure to account for all three sources of information simultaneously.

Chapter 2 explores the use of deep learning techniques for the classification of political language. ANNs are models that attempt to learn patterns in data by passing information through layers of nodes, each of which applies a regression-like weighting function to the input. The structure of multi-layer ANNs allow them to account for high-order interaction and nonlinearities when mapping inputs (e.g., word or phrase use), to outputs (e.g., document labels, speaker ideology, sentiment). This, in turn, can improve classification accuracy in some natural language processing contexts. Deep-learning neural networks have not been discussed broadly in political science methodology, and this article shows how ANNs can serve as an important classification tool in computational text modeling.

The chapter has three methodological sections: 1. constructing multilayer neural networks and estimating model parameters; 2. making inferences on new data; and 3. assessing performance and avoiding common problems like overfitting. I then describe additional possible structures for ANNs, paying particular attention to how word embedding vectors can capture even complex lexical features in text. Word embeddings, made popular by Google's word2vec or Stanford NLP's GloVe, reduce the dimensionality of text data by projecting words into a lower-dimensional space. Distances between words in this space serve as a measure of how similar words or phrases are to each other. In both Monte Carlo and real-data applications, I show that ANNs, achieve competitive classification accuracy compared to other methods. This performance is often improved by including word embedding vectors in the ANN specification.

The third chapter of my dissertation explores the dynamics of political rhetoric. Previous work on elite responsiveness to public opinion suggests that leaders should tailor their political messages to prevailing public opinion. Using the methods described in the first two chapters, I measure the ideological content of political rhetoric over time. First, I explore the ideological content of campaign speeches by major candidates in the 2008, 2012 and 2016 elections, and measure the relationship between their language and the dynamics of the campaign. Second, I measure the variation in presidential rhetoric using major speeches and statements over the course of Obama's presidential term, and compare the ideological content of his statements against aggregate public opinion.

I show that over the course of presidential elections, candidates tend to adapt to prevailing opinion of the electorate. During the primaries, candidates tend to maintain fairly partisan language, but they move strategically even within this electoral stage. Barack Obama began his primary campaign using predominantly liberal language. As Obama's popularity and delegate count grew, he transitioned slowly to a more generalist, balanced rhetoric we would expect of a major-party nominee. Meanwhile, Clinton followed an opposite trajectory. At the start of the primary, when Clinton was the prohibitive favorite for the nomination, she employed fairly moderate ideological language. As she lost ground to Obama, however, she began adopting an increasingly liberal voice.

  • Text as Data Lab
    • In 2015, I established the Carolina Text as Data Lab. The lab brings together graduate students and faculty with interests in using text as data in their research. We meet regularly to discuss research projects, data collection and for workshops on how to process and model text. In summer 2015, three research groups developed from the TaDL: (1) a Congressional bills project, scraping the timelines of changes made to draft U.S. legislation; (2) a U.S. Supreme Court project, scraping the content of Supreme Court oral arguments and opinions; and (3) a U.S. political media project, collecting thousands of stories from hundreds of television shows about politics and current events.
  • Generalized Exponential Random Graph Models
    • I'm currently working with a team of researchers on the application and implementation of generalized ERGMs. As part of my MS in statistics, I'm exploring how to use network models, particularly GERGMs, to map the lexical similarities in political writings and speeches. My current project uses GERGMs to explore the structure political news coverage of the European migrant crisis. The eventual goal is to develop a GERGM approach to sampling correlation matrices for logistic-normal topic models.
  • Bill Diffusion through Text
    • Legislative and state politics scholars care deeply about how states share legislative ideas---that is, how legislation diffuses through states, which Justice Brandeis described as the "laboratories of democracy." It's unclear, however, how scholars should best measure legislative diffusion. We examine various methods for measuring text similarity, and most importantly, the statistical properties of these measures. We also offer a simple but powerful estimator that combines the best components of two approaches: cosine similarity and substring matching.

Curriculum Vitæ

   University of North Carolina at Chapel Hill
  • Ph.D., Political Science, Expected 2016
  • M.S., Statistics, Expected 2016
  • M.A., Political Science, 2014
    • Thesis: Testing the Post-Primary Moderation Hypothesis     abstract | pdf
      • Pundits and political observers accept nearly as a matter of course that candidates will present themselves as strong partisans in primary elections, and then move `toward the center' upon advancing to the general election. Yet candidates also face incentives to not ``flip-flop’’ on their policy positions. I argue that candidates will not shift their policy positions after winning primary elections, but that they will change how they present their policies to the public. Using a supervised Bayesian model for text, I find evidence that presidential candidates in 2008 and 2012 used more ideologically extreme language during primary campaigns, and then moderate their tone when shifting to the general election.
   Dartmouth College, Hanover, N.H.
  • A.B., Government and French Language & Literature

"Emotional Responses to Disturbing Political News: The Role of Personality" with Timothy J. Ryan and Matthew S. Wells. Journal of Experimental Political Science. Forthcoming.     abstract

Recent scholarship in political science identifies emotions as an important antecedent to political behavior. Existing work, however, has focused much more on the political effects of emotions than on their causes. Here, we begin to examine how personality moderates emotional responses to political events. We hypothesized that the personality trait need for affect would moderate the emotions evoked by disturbing political news. Drawing data from a survey experiment conducted on a national sample, we find that individuals high in need for affect have an especially vivid emotional response to disturbing news—a moderating relationship that has the potential to surpass those associated with symbolic attachments.

"Measuring Ideological Proportions in Political Speeches" with Yanchuan Sim, Justin H. Gross & Noah A. Smith. In Empirical Methods in Natural Language Processing. 2013.     abstract | pdf | supplemental materials | replication

We seek to measure political candidates' ideological positioning from their speeches. To accomplish this, we infer ideological cues from a corpus of political writings annotated with known ideologies. We then represent the speeches of U.S. presidential candidates as sequences of cues and lags (filler distinguished only by its length in words). We apply a domain-informed Bayesian HMM to infer the proportions of ideologies each candidate uses in each campaign. The results are validated against a set of preregistered, domain expert authored hypotheses.

"Evaluating Post-Primary Moderation: The Ideological Etch-a-Sketch" with Justin H. Gross, Amber Boydstun, Noah A. Smith and Yanchuan Sim. Under Review.     abstract | slides

Formal models of two-stage elections usually predict post-primary ideological moderation. According to the hypothesis, candidates face pressure to appeal to more extreme partisans during primary elections, and after securing the nomination should moderate their politics to appeal to the general electorate median voter. Motivated by this largely untested hypothesis, we test the theory using candidates' campaign speeches as data. We employ a two-stage text analysis model with presidential candidates' speeches from 2008 and 2012, in addition to qualitative evaluation for robustness. The results show that Barack Obama, John McCain and Mitt Romney did indeed make substantively significant shifts away from the ideological extremes after securing their parties' presidential nominations.

"The Cue-Lag Model for Ideological Proportions" with Justin H. Gross, Noah A. Smith and Yanchuan Sim. In Preparation.     abstract | slides

Over the past decade, political researchers have increasingly recognized the promise of computational text analysis. Political drama often plays out via the written and spoken word, and the discipline has sought out methods for leveraging text to inform our empirical studies. Many richly nuanced constructs of direct interest to researchers, like political ideology, can be advanced by using data from political writings and speeches. Focusing on individual ideological expression in political rhetoric, we develop a model that integrates scholarly expertise with computational efficiency to extract ideological cues from text. Our approach involves two stages: First, we identify a vocabulary of n-grams that distinguish between sets of authors grouped by their ideological persuasion. Then we tailor a hidden Markov model to represent speakers as transitioning between latent ideological states manifested through the emission of cue-terms from the learned vocabulary. We conclude with several validation results and a discussion of possible applications of the model across discipline subfields and research questions.

"Citizen Evaluation of Ideological Text" with Michael B. MacKuen. In Preparation. Presented at MPSA 2015.     abstract | slides

Prior work shows that elites use rhetoric strategically, sending ideological signals to voters through language. Do citizens understand those signals? Can subtle changes in language affect how citizens perceive candidate ideology? Using a battery of survey experiments, we present respondents with political messages, making only subtle changes to language. Results show that participants receiving more ideological language placed the source quite far from the ideological center. Using open-ended responses, we also assess the process by which citizens perceive the signals. While participants give heterogenous explanations for their assessments, different criteria do not seem to predict different placements of the source.
Invited Talks

"Ideological Rhetoric" PoliInformatics Retreat. Depts. of Political Science, Computer Science, Statistics and Journalism. Iowa State University. May 2015.     abstract | slides

The PoliInformatics group at Iowa State University is exploring new ways to marry machine learning to substantive questions about politics. They invited me, along with researchers from Harvard, Arizona State and the University of California at Irvine, to present our work and broader lessons about methods for computational social science. My talk focused on assessing features like 'ideology' in political texts.

"IdEaS: The Ideological Etch-a-Sketch model: Inferring Ideology from Key Phrases in Text" with Justin H. Gross. Political Methodology Series. Dept. of Politics, Princeton University. February 2012.    


Download my teaching statement.

My primary teaching interests center on methodological training for graduate and undergraduate students; particularly: probability theory, linear and generalized linear models, Bayesian statistics, latent variable and scaling models, text-as-data models and introductory machine learning. In American politics, my teaching interests include ideology, political behavior, campaign and elections, and political communication.

At Carolina, I have served as TA for multiple graduate-level methods courses. As TA, I've taught a weekly one-credit-hour lab in addition to holding office hours, contributing questions to assignments and grading problem sets. Labs cover computational statistics in R, as well as supplemental material, background mathematics and typesetting in LaTeX.

In Summer 2015, I served as a Teaching Assistant for two courses at the Inter-University Consortium for Social and Political Research (ICPSR) summer program: advanced Bayesian methods, and multi-level modeling.

{UNC} POLI 281: Introduction to Quantitative Political Methods
  • (Fall 2015) Instructor: A first course in quantitative research methodology for undergraduates. We cover basic probability theory, summarizing variables, evaluating covariation, principles of measurement and basic modeling techniques. Throughout, we discuss how statistics fits into the broader context of research design, causal inference and the philosophy of science.
    • Resources: Forthcoming

{UNC} POLI 783: Introduction to Statistics & Mathematics for Political Research
  • (Fall 2013, 2014) Teaching Assistant: This class is the first in the graduate methods sequence, which orients students toward probability theory, sampling theory, Frequentist and Bayesian inference and hypothesis testing. Lab sessions cover computational statistics in R, as well as supplemental topics like Monte Carlo sampling, optimization, simulation and background calculus.
{UNC} POLI 784: Intermediate Statistics & Linear Models
  • (Spring 2014, 2015) Teaching Assistant: This class is the second in the graduate methods sequence, which focuses mostly on the linear model, its assumptions and modeling deviations from those assumptions. Lab sessions cover applications R, supplemental topics like Bayesian linear regression, and typesetting in LaTeX.
{ICPSR}: Advanced Bayesian Methods
  • (2015) Teaching Assistant: As TA, I wrote and graded assignments, and led one lecture on Bayesian finite mixture modeling. The course covered the theoretical and applied foundations of Bayesian statistical analysis. Topics include: Bayesian stochastic simulation (MCMC), with an orientation towards deriving important properties of the Gibbs sampler and the Metropolis Hastings algorithm; model checking, model assessment, and model comparison, with an emphasis on computational approaches; Bayesian variants of "workhorse" political science models, such as generalized linear models; and advanced Bayesian models, such as hierarchical/multilevel models, models for panel and time-series cross-section data, latent factor and item response theory (IRT) models, as well as instrumental variable models.
{ICPSR}: Multilevel Models: Pooled & Clustered Data
  • (2015) Teaching Assistant: As TA, I led a one-day intensive session on Bayesian hierarchical modeling, from the foundations of Bayesian inference through hierarchical regression models in Stan. The course provided general introduction to a variety of applications of multilevel modeling in the social sciences. Equal emphasis was placed on the underlying statistical model and on the estimation and interpretation of empirical data. The course explored methods for so-called robust standard errors in the face of clustered data, along with traditional methods for pooled time series and multi-level models.