Anxiety, gender, and social media consumption predict COVID-19 emotional distress

The study addresses which socio-psychological factors independently predict COVID-19-related emotional distress in the general U.S. population during the early pandemic. Although many variables (e.g., demographics, political ideology, media use, mental health, personality, emotion regulation, knowledge, behaviors) have been associated with pandemic responses, their unique predictive power remains unclear due to interdependencies among factors. The authors aim to build a data-driven psychological profile of COVID-19 emotional distress by simultaneously assessing 30 variables across seven domains and using cross-validated modeling to determine which factors most strongly and uniquely predict distress. Understanding these predictors is important for targeting mental health interventions and identifying vulnerable groups as the pandemic persists.

Prior research has documented increased anxiety and depression early in the pandemic and linked various factors to pandemic behaviors and emotions. Demographic factors such as gender and age relate to preventive behaviors and worry, with women often reporting higher distress and engaging more in protective measures, while older adults report less worry but more compliance. Political ideology has been tied to risk perceptions and distancing behaviors, particularly in the U.S., where COVID-19 became politicized. Media consumption patterns influence compliance and attitudes in pandemics (e.g., H1N1), and the type and amount of media exposure can shape responses. Personality traits (e.g., neuroticism), intolerance of uncertainty, and emotion regulation capabilities are associated with stress responses to crises and viral outbreaks. However, findings across studies are sometimes inconsistent (e.g., mixed links between age and COVID-related fear), and single-variable analyses make it difficult to assess unique contributions when factors are interrelated. This study situates itself within this literature by testing a broad set of variables simultaneously with cross-validation to evaluate unique predictive power for COVID-19 emotional distress.

Design and participants: Cross-sectional online survey administered March 24–26, 2020, via Prolific using a U.S.-representative sampling approach stratified by age, sex, and ethnicity. Of 1000 recruited (506 females; mean age ≈ 44.75–44.85, SD ≈ 15.9), 45 were excluded by preregistered checks and 7 removed for missing gender, yielding N = 948. Study preregistration: https://osf.io/2y2jg/. Compensation: $3.50 for ~20 minutes. Ethics approval and informed consent obtained.

Outcome measure (COVID-19 emotional distress): A composite index from two scales assessing current worry/stress (6-point agreement scale) and time spent thinking/ruminating about coronavirus. High internal consistency reported (Cronbach’s alpha = 0.93).

Predictor variables (30 total across seven categories):

• Mental health: Depression (CES-D), anxiety (GAD-7), alexithymia (TAS-20).
• Personality/traits: Extraversion and neuroticism (BFI-2-S), intolerance of uncertainty (IUS).
• Emotion regulation: ERQ (suppression/reappraisal) and Interpersonal Regulation Questionnaire (IRQ; reliance on social support to regulate emotions).
• COVID-19 knowledge: Custom quiz on symptoms, transmission, quarantine duration, etc.
• Media consumption: Sources and frequency for COVID-19 information (e.g., Facebook, Twitter, government sources, television, internet, U.S.-specific media), rated 1–5.
• COVID-19 behaviors: Preventive behaviors (e.g., handwashing frequency) and related practices.
• Demographics: Age, gender, political ideology (100-point liberal–conservative scale), socioeconomic status, education.

Statistical analysis: Data split into training (75%) and held-out test (25%). Within training, 10-fold cross-validation used with a hybrid stepwise selection (forward/backward) based on Akaike Information Criterion (AIC) to select the optimal model minimizing cross-validated error. Partial correlation network visualized relationships among variables in the selected model (edges shown for |r| > 0.20). Model performance evaluated on the held-out test set, reporting explained variance and calibration. Data and R markdown code available: https://github.com/jpfeffer/covid_models.

• Cross-validated model performance on held-out data: r^2 = 0.46; calibration slope g = 0.95 ± 0.07, t(234) = 14.02, p < 0.001.
• Strongest predictors of higher COVID-19 emotional distress: Anxiety (GAD-7) was the most predictive; gender (women > men); social media consumption (Twitter, Facebook) and U.S.-specific media; intolerance of uncertainty; interpersonal emotion regulation (greater reliance on social support); alexithymia. Preventive behaviors and greater COVID-19 knowledge had small positive associations with distress. Age showed a small negative association (younger more distressed). Liberal political orientation showed a small negative beta (liberal-leaning associated with higher distress in narrative; table shows negative coefficient for Liberalism variable after coding).
• Variables commonly linked to distress in bivariate analyses (e.g., neuroticism, depression) did not retain unique predictive power once anxiety and other variables were included.
• Government sources of information did not uniquely contribute to distress when other media sources were considered.
• Table 2 (standardized betas, significance): Anxiety b = 0.49 (p < 0.001); Gender b = 0.17 (p < 0.001); Twitter b = 0.16 (p < 0.001); USA media b = 0.13 (p < 0.001); Interpersonal (IRQ) b = 0.10 (p < 0.001); COVID behavior b = 0.10 (p = 0.016); LUS/IUS b = 0.09 (p = 0.002); Facebook b = 0.07 (p = 0.013); COVID knowledge b = 0.07 (p = 0.010); Internet media b = 0.05 (p = 0.035); Age b = -0.06 (p = 0.016); Liberalism b = -0.09 (p = 0.006). Extremism and social events were not significant.
• Many simple correlations overestimated effects compared to cross-validated model coefficients, highlighting the importance of multivariable, cross-validated modeling.

Findings indicate that individuals who are more anxious, female, liberal-leaning, higher in intolerance of uncertainty, and who consume substantial COVID-19 content via social network platforms (Twitter, Facebook) were more vulnerable to emotional distress early in the pandemic. The type of media matters: social platforms and U.S.-specific media predicted higher distress, whereas government sources did not uniquely contribute when accounting for other media. Small positive links between accurate knowledge and preventive behaviors with distress may reflect heightened engagement or vigilance accompanying worry. The cross-validated framework helps contextualize prior reports by showing that depression and neuroticism, although correlated with distress, offer little unique explanatory power once anxiety is modeled—consistent with theoretical accounts where anxiety precedes and centers on future-oriented worry. Overall, the model clarifies which factors uniquely explain variance in distress and can guide targeted public health messaging and mental health interventions, including attention to social media exposure and support for high-anxiety and female populations.

The study contributes a data-driven, cross-validated model identifying key predictors of COVID-19 emotional distress in a U.S. sample during the pandemic’s early phase. Anxiety emerged as the dominant predictor, followed by gender and specific media consumption patterns (notably social media). Other factors, such as depression and neuroticism, showed reduced unique importance when considered alongside anxiety. These insights can help public health officials and clinicians identify and support vulnerable groups and refine communication strategies. Future research should examine causal pathways, longitudinal dynamics across pandemic waves, the role of platform-specific information quality, and generalizability across countries and cultural-political contexts.

The study uses cross-sectional, self-report data from an online U.S. sample collected in the pandemic’s early stage, limiting causal inference and potentially affecting generalizability over time and to other countries. The liberal–conservative scale is U.S.-specific and may not map onto political spectra elsewhere, constraining cross-national interpretation. Although the variable set was broad (30 predictors), it was not exhaustive, and measurement/recall biases may influence self-reports of media use, behaviors, and distress.