« Back to Results

Analyzing Bias with Text, Images, And Sound

Paper Session

Saturday, Jan. 8, 2022 3:45 PM - 5:45 PM (EST)

Hosted By: American Economic Association
  • Chair: Carlo Schwarz, Bocconi University

“A Picture is Worth a Thousand Words”: Implicit Visual Bias in News Media

Elliott Ash
ETH Zurich
Ruben Durante
Pompeu Fabra University
Carlo Schwarz
Bocconi University


Virtually all measures of media bias used in the political economy literature are based on analysis of the text of news reports either to identify the topics covered in them or to track the use of ideologically charged expressions. None of the existing measures uses information contained in the images featured in news articles, and, most crucially, on the subtle links between such images and the content of the article. To make progress on this issue, we propose a novel image-based measure of implicit gender and racial biases of media outlets. The measure exploits how often media outlets use a picture of specific gender, race or ethnic group to accompany an article. For the creation of this measure we have trained a deep-learning machine learning classifier that is able to accurately classify the gender and ethnicity of faces in pictures. Using this measure, we show that the pictures of media outlets exhibit significant degrees of bias. Women are underrepresented in images and so are several ethnic minorities. We further analyse the development of the bias across outlets, over time and by occupations that get mentioned in an article. We show that the occupation shares by group shown in news images diverge from the occupation shares present in labor-market data.

Emotions and Violence: The Effect of Biased News on Hate Crimes

Jacob Miller
University of Zurich


Can what the media discusses, and the way it discusses it, have consequences for hateful behavior? To examine this question, I measure the effect of biased news about immigration on violent crimes towards immigrants. Motivated by a recent literature establishing that unexpected sports losses can stimulate emotional reactions, I utilize unexpected sports losses as a trigger for emotionally driven violence. When the local National Hockey League (NHL) team unexpectedly loses, I observe heightened hate crimes against immigrants if cable news had been intensely discussing immigration. To dig deeper into the intensive margin of this effect, I develop and utilize state-of-the-art machine learning algorithms that can quantify immigration bias in two different ways: from the text of the broadcasts, the degree of slant, and from the audio, the degree of anger. I find that the effects are further magnified when the media's coverage is especially angry, and in the case of Fox News, if the coverage is slanted against immigrants. These novel findings establish that, in the context of emotionally driven violence, the topics and tone of the news cycle can play a consequential role.

The (great) Persuasion Divide? Gender Disparities in Debate Speeches & Evaluations

Huyen Nguyen
University of Hamburg


Do men and women persuade differently? Are they evaluated differently? Using a novel data set of 1517 speech transcripts, evaluation scores and demographic data from highest-profile intervarsity debate tournaments, this research investigates spoken verbal tactics across genders and any ensuing impacts on their performance evaluations. The well-defined competitive rules of debate tournaments along with transparent merit-based evaluation system enables me to disentangle whether gender disparities in outcomes is due to differences in speech behavior or gender-specific evaluation standards (i.e. discrimination). To extract persuasion-relevant linguistic variables, I use widely validated word-and-phrase-based methods from persuasiveness studies. This method provides easily replicable linguistic variables to analyze unstructured, high-dimensional impromptu spoken text data, with limited observations. I find significant variation in speech patterns of male and female speakers. Specifically, female speakers use more personal and disclosing speaking styles, with more hedging phrases and disfluencies in their speeches. In their answers to questions from opponents, they negate less while having notably longer and more vague answers. Evaluation-wise, within debates, except for disfluencies, there is no robust evidence of gender-specific evaluation standards. These findings suggest that women receive lower scores than men because their speeches contain more score-reducing and fewer score-enhancing features, rather than discrimination.

Identifying Physician Practice Style for Mental Health Conditions

Kelli Marquardt
University of Arizona


This paper proposes new methods for identifying and estimating physician practice style in the context of mental health diagnoses. Empirical identification in these settings can be difficult as there is no formal biological/medical test to determine the presence of a mental health condition. Instead, diagnosis depends on physician-patient interviews to extract the existence of behavioral symptoms and match to documented diagnostic guidelines. To address these identification concerns, I propose a two step estimation procedure. First, I describe unique text-analysis methods applied to digitized clinical doctor notes as a way to measure how closely the patient interview matches mental health diagnostic guidelines according to The Diagnostic and Statistical Manual of Mental Disorders (DSM-V). I then use this measure as a control in a reduced-form model to identify two components of physician practice style: diagnostic intensity (the mean propensity to diagnose) and diagnostic compliance (the weight that physicians place on official medical guidelines). As an application, I use electronic health record data to estimate physician practice style for a widely prevalent mental health condition, Attention Deficit Hyperactivity Disorder (ADHD). I find significant variation in physician practice style, with physician gender and experience being the strongest predictors of this variation. Finally, I discuss how mental health practice style estimates can be used to guide potential healthcare policies, and I provide a list of extensions and suggestions on how these methods can be used in future mental health care research.
Ashesh Rambachan
Harvard University
Oda Nedregard
BI Norwegian Business School
JEL Classifications
  • J7 - Labor Discrimination
  • L8 - Industry Studies: Services