Using Twitter to Capture Racism in the Air: Integrating Macro and Micro-Level Factors to Predict Asian American Well-Being during the COVID-19 Pandemic

Background and Purpose: Anti-Asian violence and harassment have escalated during the Covid-19 pandemic, catalyzed by the racial framing of the virus, and converging with a national awakening to systematic racism. Whereas the negative association between perceived racism and mental health has been well-established in the literature, studies rarely consider the effect of macro-contextual factors on racialized individuals’ lived experiences. We conducted two large scale surveys of Asian Americans in 2020 and 2022, integrating geocoded indices of sociocultural climate and structural inequalities (e.g., racial attitudes, residential segregation, economic inequities) to examine how macro-contextual factors interact with individual-level psychological variables to predict Asian Americans’ experiences of and responses to racism during the pandemic. As an example, we examined how geographic variations in racial climate (as indicated by geocoded Twitter analyses of anti-Asian sentiment) are associated with perceived racial discrimination and substance use outcomes.

Methods: Geocoded Twitter racial sentiment data from 2020, the first year of the COVID-19 pandemic, were combined with individual-level responses from our December 2020 survey of 684 Asian Americans from 43 states. We performed a multilevel analysis. First, regression analyses were used to examine associations between online/in-person discrimination and substance use, with Asian subgroup, education, age, sex, and nativity (U.S.- vs. foreign-born) as individual-level covariates. The second model added an index of state-level Twitter anti-Asian racial sentiment developed using a machine learning algorithm, support vector machine. Of 183,837 geocoded tweets, 42,817 tweets were identified as Asian-related, and 9,807 tweets indicated positive (n=9,367) or negative sentiments (n=440). State-level anti-Asian racial sentiment was operationalized as the ratio of negatively valenced tweets about Asians to the total number of Asian-related tweets produced in the state.

Results: The intraclass correlation coefficient was 3.73% for the Twitter racial sentiment variables. The mean proportion of anti-Asian tweets across states was 1.04%, ranging from 0% to 2.97%. For the individual-level only model, in-person discrimination (β=.25, CI=[.14, .37]), online discrimination (β=.16, CI=[.07, .26]), and male gender (β=.11, CI=[.00, .22]) were positively associated with substance use as coping (p<.05). In the second model, state-level anti-Asian racial sentiment also was significantly associated with substance use (β=-0.23, CI=[-0.36, -0.10], p=0.001) and Likelihood ratio test with model 1 was statistically significant (p<.05), taking into account individual difference factors, including experiences of discrimination.

Conclusions and Implications: Consistent with an ecological approach, these preliminary results suggest that beyond direct experiences of racial discrimination, macro-level conditions such as racist public discourse on social media, can affect minoritized individuals’ health risk behaviors, such as substance use. In other words, ambient racist sentiment as captured by Twitter may both mirror and perpetuate interpersonal forms of racism to impact community well-being. Besides accounting for how macro-level variables trickle down to affect individual experiences and behaviors, future studies will also explore individual-difference risk and protective factors that mitigate the impact of adverse macro-level conditions on the individual.