Background: Anecdotal reports suggest a rise in anti-Asian racial attitudes and discrimination in response to COVID-19. Racism can have significant social, economic, and health impacts, but there has been little systematic investigation of increases in anti-Asian prejudice. Methods: We utilized Twitter’s Streaming Application Programming Interface (API) to collect 3,377,295 U.S. race-related tweets from November 2019–June 2020. Sentiment analysis was performed using support vector machine (SVM), a supervised machine learning model. Accuracy for identifying negative sentiments, comparing the machine learning model to manually labeled tweets was 91%. We investigated changes in racial sentiment before and following the emergence of COVID-19. Results: The proportion of negative tweets referencing Asians increased by 68.4% (from 9.79% in November to 16.49% in March). In contrast, the proportion of negative tweets referencing other racial/ethnic minorities (Blacks and Latinx) remained relatively stable during this time period, declining less than 1% for tweets referencing Blacks and increasing by 2% for tweets referencing Latinx. Common themes that emerged during the content analysis of a random subsample of 3300 tweets included: racism and blame (20%), anti-racism (20%), and daily life impact (27%). Conclusion: Social media data can be used to provide timely information to investigate shifts in area-level racial sentiment.

This study examined whether killings of George Floyd, Ahmaud Arbery, and Breonna Taylor by current or former law enforcement officers in 2020 were followed by shifts in public sentiment toward Black people. Methods: Google searches for the names “Ahmaud Arbery,” “Breonna Taylor,” and “George Floyd” were obtained from the Google Health Application Programming Interface (API). Using the Twitter API, we collected a 1% random sample of publicly available U.S. race-related tweets from November 2019–September 2020 (N = 3,380,616). Sentiment analysis was performed using Support Vector Machines, a supervised machine learning model. A qualitative content analysis was conducted on a random sample of 3,000 tweets to understand themes in discussions of race and racism and inform interpretation of the quantitative trends. Results: The highest rate of Google searches for any of the three names was for George Floyd during the week of May 31 to June 6, the week after his murder. The percent of tweets referencing Black people that were negative decreased by 32% (from 49.33% in November 4–9 to 33.66% in June 1–7) (p < 0.001), but this decline was temporary, lasting just a few weeks. Themes that emerged during the content analysis included discussion of race or racism in positive (14%) or negative (38%) tones, call for action related to racism (18%), and counter movement/arguments against racism-related changes (6%). Conclusion: Although there was a sharp decline in negative Black sentiment and increased public awareness of structural racism and desire for long-lasting social change, these shifts were transitory and returned to baseline after several weeks. Findings suggest that negative attitudes towards Black people remain deeply entrenched.

Background: In the United States, racial disparities in birth outcomes persist and have been widening. Interpersonal and structural racism are leading explanations for the continuing racial disparities in birth outcomes, but research to confirm the role of racism and evaluate trends in the impact of racism on health outcomes has been hampered by the challenge of measuring racism. Most research on discrimination relies on self-reported experiences of discrimination, and few studies have examined racial attitudes and bias at the US national level. Objective: This study aimed to investigate the associations between state-level Twitter-derived sentiments related to racial or ethnic minorities and birth outcomes. Methods: We utilized Twitter’s Streaming application programming interface to collect 26,027,740 tweets from June 2015 to December 2017, containing at least one race-related term. Sentiment analysis was performed using support vector machine, a supervised machine learning model. We constructed overall indicators of sentiment toward minorities and sentiment toward race-specific groups. For each year, state-level Twitter-derived sentiment data were merged with birth data for that year. The study participants were women who had singleton births with no congenital abnormalities from 2015 to 2017 and for whom data were available on gestational age (n=9,988,030) or birth weight (n=9,985,402). The main outcomes were low birth weight (birth weight ≤2499 g) and preterm birth (gestational age <37 weeks). We estimated the incidence ratios controlling for individual-level maternal characteristics (sociodemographics, prenatal care, and health behaviors) and state-level demographics, using log binomial regression models. Results: The accuracy for identifying negative sentiments on comparing the machine learning model to manually labeled tweets was 91%. Mothers living in states in the highest tertile for negative sentiment tweets referencing racial or ethnic minorities had greater incidences of low birth weight (8% greater, 95% CI 4%-13%) and preterm birth (8% greater, 95% CI 0%-14%) compared with mothers living in states in the lowest tertile. More negative tweets referencing minorities were associated with adverse birth outcomes in the total population, including non-Hispanic white people and racial or ethnic minorities. In stratified subgroup analyses, more negative tweets referencing specific racial or ethnic minority groups (black people, Middle Eastern people, and Muslims) were associated with poor birth outcomes for black people and minorities. Conclusions: A negative social context related to race was associated with poor birth outcomes for racial or ethnic minorities, as well as non-Hispanic white people.

The objective of the current study is to investigate whether an area-level measure of racial sentiment derived from Twitter data is associated with state-level hate crimes and existing measures of racial prejudice at the individual-level. We collected 30,977,757 tweets from June 2015–July 2018 containing at least one keyword pertaining to specific groups (Asians, Arabs, Blacks, Latinos, Whites). We characterized sentiment of each tweet (negative vs all other) and averaged at the state-level. These racial sentiment measures were merged with other measures based on: hate crime data from the FBI Uniform Crime Reporting Program; implicit and explicit racial bias indicators from Project Implicit; and racial attitudes questions from General Social Survey (GSS). Living in a state with 10% higher negative sentiment in tweets referencing Blacks was associated with 0.57 times the odds of endorsing a GSS question that Black-White disparities in jobs, income, and housing were due to discrimination (95% CI: 0.40, 0.83); 1.64 times the odds of endorsing the belief that disparities were due to lack to will (95% CI: 0.95, 2.84); higher explicit racial bias (β: 0.11; 95% CI: 0.04, 0.18); and higher implicit racial bias (β: 0.09; 95% CI: 0.04, 0.14). Twitter-expressed racial sentiment was not statistically-significantly associated with incidence of state-level hate crimes against Blacks (IRR: 0.99; 95% CI: 0.52, 1.90), but this analysis was likely underpowered due to rarity of reported hate crimes. Leveraging timely data sources for measuring area-level racial sentiment can provide new opportunities for investigating the impact of racial bias on society and health.

Deep learning and, specifically, convoltional neural networks (CNN) represent a class of powerful models that facilitate the understanding of many problems in computer vision. When combined with a reasonable amount of data, CNNs can outperform traditional models for many tasks, including image classification. In this work, we utilize these powerful tools with imagery data collected through Google Street View images to perform virtual audits of neighborhood characteristics. We further investigate different architectures for chronic disease prevalence regression through networks that are applied to sets of images rather than single images. We show quantitative results and demonstrate that our proposed architectures outperform the traditional regression approaches.