Whom We Trust, What We Fear: COVID-19 Fear and the Politics of Information

TL;DR

This study examines how different information sources influence COVID-19 fear levels, revealing that source type significantly impacts emotional responses and varies with demographics, political orientation, and pandemic trends.

Contribution

It introduces a comprehensive analysis linking information sources to COVID-19 fear, applying causal inference to quantify their effects and demonstrating the influence of demographics and political factors.

Findings

Fear levels correlate strongly with infection trends.

Information sources significantly influence fear levels.

Source preferences align with political orientations.

Abstract

The COVID-19 pandemic triggered not only a global health crisis but also an infodemic, an overload of information from diverse sources influencing public perception and emotional responses. In this context, fear emerged as a central emotional reaction, shaped by both media exposure and demographic factors. In this study, we analyzed the relationship between individuals' self-reported levels of fear about COVID-19 and the information sources they rely on, across nine source categories, including medical experts, government institutions, media, and personal networks. In particular, we defined a score that ranks fear levels based on self-reported concerns about the pandemic, collected through the Delphi CTIS survey in the United States between May 2021 and June 2022. We found that both fear levels and information source usage closely follow COVID-19 infection trends, exhibit strong correlations within each group (fear levels across sources are strongly correlated, as are patterns of source usage), and vary significantly across demographic groups, particularly by age and education. Applying causal inference methods, we found that among age, education, and information source, the latter is the most influential factor affecting individuals' fear levels. We further quantified the impact of different information sources on fear by estimating the average treatment effect, indicating how each source alters fear relative to a control. Furthermore, we demonstrated that information source preferences can reliably match the political orientation of U.S. states. These findings highlight the importance of information ecosystem dynamics in shaping emotional and behavioral responses during large-scale crises.

Tables

Table 1. Table 1. Selected questions and responses from the Delphi US CTIS survey, shown with original survey labels and the corresponding variable names used throughout the paper. Only the responses included in the analysis are presented, aggregated as indicated. All variables are binary or categorical.

Question	Responses	Name	Measured
(D1) What is your gender?	(1) Male, (2) Female	Gender	$✓$
(D2) What is your age?	(1) 18-34 years, (3) 35-64 years, (6) 65+ years	Age	$✓$
(D7) What is your race? Please select all that apply.	(1) American Indian or Alaska Native, (2) Asian, (3) Black or African American, (4) Native Hawaiian or other Pacific Islander, (5) White, (6) Some other race	Race	$\times$
(D8) What is the highest degree or level of school you have completed?	(1) High school or less, (2) Bachelor / Master degree, (3) Postgraduate	Education	$✓$
(D9) In the past 4 weeks, did you do any kind of work for pay?	(1) Yes, (2) No	Occupation	$✓$
(D12) What language do you speak most often at home?	(1) English, (2) Other	Language	$✓$
(C1) Have you ever been told by a doctor, nurse, or other health professional that you have any of the following medical conditions? Please select all that apply.	(1) Cancer (other than skin cancer), (2) Heart attack, heart disease, or other heart condition, (3) High blood pressure, (4) Asthma, (5) Chronic lung disease such as COPD, chronic bronchitis, or emphysema, (6) Kidney disease, (7) Type 1 diabetes, (8) Type 2 diabetes, (9) Weakened or compromised immune system, (10) Obesity, (11) None of these	Health	$\times$
(C15) How worried are you about your household’s finances for the next month?	(1) Not worried at all, (2) Not too worried, (3) Somewhat worried, (4) Very worried	Economic	$\times$
(A3) What is your current ZIP code?	Open answer (we map ZIP codes with U.S. states and U.S. states into regions, i.e. Northeast, Midwest, South, West)	Region	$✓$
(A4) How many additional people in your local community do you personally know who are sick with a fever, along with at least one other symptom from the above list?	Open answer (we keep only values less than or equal to 100, a threshold which is conservative and filters out unrealistic responses, like those reporting millions)	#Sick	$✓$
(V1) Have you had a COVID-19 vaccination?	(1) Yes, (2) No	Vaccinated	$✓$
(I5) In the past 7 days, from which of the following sources have you received news and information about COVID-19?	(1) Doctors and other health professionals you go to for medical care, (2) Scientists and other health experts, (3) Centers for Disease Control (CDC), (4) Government health authorities or officials, (5) Politicians, (6) Journalists, (7) Family and friends, (8) Religious leaders, (9) None of the above (other sources not listed here)	Source	$✓$
(I6) How much do you trust the following sources to provide accurate news and information about COVID-19?	For each I5: (1) Do not trust, (2) Somewhat trust, (3) Trust	Trust	$✓$
(G1) How much do you worry about catching COVID-19?	(1) Not at all, (2) A little, (3) A moderate amount, (4) A great deal	Fear	$✓$

Table 2. Table 2. Top-20 pairwise associations with Cramér’s V V .

Variable 1	Variable 2	Cramér’s $V$
Fear	Vaccinated	0.345
Scientists / health experts Source usage	Scientists / health experts Source trust	0.320
Journalists Source usage	Journalists Source trust	0.311
CDC Source usage	CDC Source trust	0.303
Government authorities Source usage	Government authorities Source trust	0.260
Fear	Other / none of the above Source usage	0.242
Religious leaders Source usage	Religious leaders Source trust	0.207
Fear	CDC Source usage	0.204
Education	Scientists / health experts Source usage	0.187
Fear	Gender	0.178
Education	CDC Source usage	0.169
Fear	Scientists / health experts Source usage	0.167
Family and friends Source usage	Family and friends Source trust	0.161
Fear	Language	0.159
Gender	CDC Source trust	0.153
Fear	Government authorities Source usage	0.152
Age	Vaccinated	0.149
Doctors Source usage	Doctors Source trust	0.143
Education	Other / none of the above Source	0.133
Education	Government authorities Source usage	0.131

Table 3. Table 3. Contributors to the variance explained in Fear .

Parent Variable $\to$ Fear	Variance Explained	Variance Explained (%)
Gender	0.083	8.22%
Vaccinated	0.083	8.22%
#Sick	0.065	6.44%
Other / none of the above Source sources	0.064	6.34%
Occupation	0.055	5.45%
CDC Source usage	0.049	4.85%
Scientists / health experts Source usage	0.041	4.06%
Politicians Source usage	0.030	2.97%
Government authorities Source usage	0.028	2.77%
Education	0.027	2.68%
Region	0.026	2.58%
Age	0.024	2.38%
Language	0.021	2.08%
Doctors Source usage	0.019	1.88%
Family and friends Source usage	0.015	1.49%
Journalists Source usage	0.009	0.89%
Religious leaders Source usage	0.005	0.50%
Total	0.644	63.80%

Table 4. Table 4. ATE estimates of Source on Fear .

Treatment Variable	ATE
CDC	0.227
Scientists / health experts	0.147
Government authorities	0.114
Doctors / health professionals	0.089
Family and friends	0.027
Journalists	-0.029
Religious leaders	-0.141
Politicians	-0.284
Other / none of the above sources	-0.464

Full text

COVID-19, Fear, Causal inference, Sources of information, Survey-based analysis

1. Introduction

The COVID-19 pandemic represented a global health crisis, while simultaneously it generated an infodemic: a rapid surge of information from heterogeneous sources (Cinelli and others, 2020). Individuals relied on news media, social platforms, government announcements, and public health organizations for guidance. The credibility, consistency, and framing of these sources shaped both public understanding of the threat and emotional responses across populations (Jones et al., 2021; Lu and others, 2021).

In this context, fear emerged as one of the main psychological responses to the pandemic (Coelho et al., 2020). This emotion was driven not only by the objective health threat, but also by the volume, tone, and perceived trustworthiness of consumed information. The impact of information sources varied across demographic groups: age, education, gender, and socioeconomic status shaped how individuals selected, interpreted, and trusted information (Nino et al., 2021). These disparities highlight the need to account for both information characteristics and recipient profiles when evaluating the psychological effects of information exposure.

Contributions. This work provides three main contributions. First, we introduce a causal inference framework combining epidemiological indicators, demographic information, and information-source exposure to study fear dynamics during the COVID-19 pandemic. Second, we investigate how these demographic, epidemiological, and informational factors contribute to fear, emphasizing the need to distinguish causal effects from statistical associations. Third, by fitting a Probabilistic Causal Model (PCM) and using variance decomposition of arrow strengths and Average Treatment Effect (ATE) estimation, we quantify the role of information sources in shaping fear, showing how communication channels can amplify or mitigate emotional responses during public health crises.

2. Related work

Fear emerged as a central psychological response during the COVID-19 pandemic, shaped more by subjective perceptions and social transmission than by epidemiological indicators. Cross-national longitudinal research found that roughly 23% of fear variance was explained by micro-level psychological factors (including worries about shortages, perceived vulnerability, germ aversion, and infections in close social circles), while macro-level indicators such as case counts, mortality, and restrictions contributed minimally (Eder and others, 2021).

Social media analyses showed that fear peaked during early outbreak phases, and that females, healthcare workers, lower-educated, and married individuals showed higher fear levels (Chai and others, 2022; Oliveira and others, 2023; Mertens et al., 2022; Doshi and others, 2021). This had impact in the evolution of the pandemic. Agent-based models showed that declining fear can trigger new disease waves by reducing protective behavior (Retzlaff and others, 2022). On the other end, early population-level fear modestly predicted later vaccination willingness but was insufficient to explain vaccine hesitancy.

Causal inference studies revealed bidirectional relationships between media exposure, interpersonal communication, and fear. Media consumption and perceived danger increased fear, while elevated fear motivated additional information seeking (Li, 2022). Despite extensive research, the influence of different information sources on fear across demographic groups remains poorly characterized. Our study addresses this gap through causal inference methods evaluating how demographics, information sources, vaccination, and other variables shape fear responses.

3. Data

We used two COVID-19 data sources: official data from the COVID-19 Data Hub (Guidotti and Ardia, 2020; Guidotti, 2022) and survey data from the Delphi US CTIS (Delphi Group, CMU, 2020). From the Data Hub, we extracted daily fatalities and confirmed cases to estimate infections via Sybil (Baccega et al., 2024), reconstructing a Susceptible-Infected-Recovered-Deceased-Susceptible (SIRDS) model over the study period assuming a 14-day average recovery time.

The Delphi US CTIS survey collected approximately 50,000 daily responses from Facebook users in the U.S. and was administered by Carnegie Mellon University under Facebook’s Data for Good program (Facebook, 2020; Astley and others, 2021; Fan and others, 2020). It includes self-reported information on symptoms, testing, vaccination, behaviors, trust in information sources, and fear. These data were used to estimate prevalence trends across U.S. states while accounting for direct and indirect reporting (Srivastava et al., 2024). Access was obtained through an agreement with CMU, UMD, and Facebook. Table 1 reports the subset of Delphi US CTIS questions considered in this study (the full questionnaire is available at (Delphi Group, Carnegie Mellon University, 2020)) Data for Race, Health, and Economic variables are not available due to dataset limitations and inconsistencies (see Section 6 for details).

Our analysis covers May 2021 to June 2022 across all U.S. states, combining epidemiological indicators with self-reported behavioral and psychological information. Although these sources are biased toward individuals with stable internet access, particularly active Facebook users, such biases are less pronounced in U.S. states with widespread connectivity, still providing valuable insights into large-scale behavioral trends during the COVID-19 pandemic.

4. Model and Methods

To define the causal inference model shown in Figure 1, we use the DoWhy package (Sharma and Kiciman, 2020; Blöbaum and others, 2024), a Python library integrating graphical causal models with statistical methods to specify assumptions, identify estimands, estimate causal effects, and perform robustness checks through refutation tests. The Directed Acyclic Graph (DAG) encodes the assumed causal relationships among demographic and contextual confounders, trust in information sources, source usage, epidemiological variables, and fear outcomes.

In particular, the model represents how demographic characteristics (Gender, Age, Education, Occupation, Language, and Region), together with pandemic-specific variables such as the COVID-19 Wave  (in the period considered, we have two waves:Delta and Omicron) and epidemiological Trend  (which maps the derivative of the number of infected individuals: -1 for values $\leq 0$ and 1 for values $>0$), jointly influence usage of information sources and fear levels.

The treatment variables consist of nine binary Source  indicators obtained through one-hot encoding of the original source-usage variable, while the outcome variable is COVID-19 Fear. The DAG also includes other covariates: trust in information sources (Trust), and two epidemiological ones which directly contribute to Fear: the number of infected individuals known locally (#Sick) and vaccination status (Vaccinated). We additionally acknowledge potential unmeasured confounders such as health status (Health), economic conditions (Economic), and race (Race)—see Section 6 for details.

Since demographic and contextual variables affect both Source and Fear, naive associations between source usage and fear may be confounded. The DAG, therefore, defines the adjustment set required to identify the causal effect of information-source usage on fear. Rather than modeling calendar time explicitly, temporal dynamics are captured through the categorical COVID-19 Wave indicator and the epidemiological Trend variable, which reflects whether infection rates are increasing or decreasing. This formulation distinguishes periods of pandemic expansion and contraction.

DoWhy was primarily used to estimate the causal relationships represented in the DAG by fitting a PCM to the observed data. The PCM enabled estimation of arrow strengths while preserving the conditional dependencies specified by the graph. We additionally computed the ATE for each source variable to quantify the average causal effect of using a specific source on reported fear levels, after adjustment for DAG-identified confounders. Whereas we use variance to quantify causal arrow strength (i.e., how much a parent node influences its child node), the ATE measures, under the assumptions encoded in the graph, the expected change in Fear associated with exposure to a given information source relative to non-exposure.

5. Results

To preliminarily assess relationships among the variables, we computed pairwise associations using Cramér’s V. Table 2 reveals weak to moderate associations between several variable pairs. Notably, Trust and the corresponding Source variables show consistent positive associations, indicating that individuals who trust a given source also tend to rely on it. Fear exhibits moderate associations with Vaccinated and, to a lesser extent, with Gender, Language, and Source from the CDC, scientists and health experts, government authorities, the journalists, and other unspecified sources. However, these associations are not causal because demographic and epidemiological variables may jointly affect both information-source usage and fear. This motivates the need for causal inference methods.

To address these limitations, we estimated the PCM described in Section 4. Table 3 reports the variance explained by the main causal arrows targeting Fear. Overall, the variables included in the model account for about 64% of the variance in fear (which is equals to 1.0093), indicating that the considered demographic, epidemiological, and informational variables collectively capture a substantial portion of fear variability.

Information-related variables emerge as the dominant explanatory component. Combined, information-source variables explain more than 25% of the total variance in Fear, exceeding the contribution of any other variable group in the model. In particular, receiving information from other unspecified sources, the CDC, scientists and health experts, politicians, and government authorities explains a substantial portion of fear variance. Among individual variables, Gender and Vaccinated contribute the most (a result supported by (Chai and others, 2022) and (Mertens et al., 2022)). #Sick also contributes strongly, confirming that epidemiological conditions shape fear levels. Other demographic variables, such as Occupation, Education, Age, and Language, further explain part of the variance, suggesting heterogeneous psychological responses across population groups.

Finally, we quantified the causal effect of each Source through ATE estimation using propensity score stratification. Table 4 reveals substantial heterogeneity across sources. Institutional and expert-driven sources, including the CDC, scientists and health experts, government authorities, and doctors and health professionals, are associated with increased fear levels. Conversely, relying on alternative or unspecified sources (“None of the above”), politicians, and religious leaders is associated with reduced fear after adjustment for confounding. Importantly, the ATE estimates differ substantially from the pairwise associations observed in the Cramér’s V analysis. Variables strongly associated with fear in the correlation analysis may exhibit weaker causal effects after adjustment for demographic and epidemiological confounding, while others emerge only after causal adjustment. This discrepancy highlights the importance of causal inference approaches when studying fear determinants during the COVID-19 pandemic.

6. Limitations

This work has several limitations related to data availability and survey structure. First, we could not include potentially important confounders such as race (Race), previous health conditions (Health), or detailed economic status (Economic). Race was not available for ethical reasons, while Health and Economic variables could not be reliably integrated because of dataset limitations and inconsistencies. Consequently, unmeasured confounding may still bias part of the estimated causal effects despite DAG-based adjustment.

A second limitation is the absence of longitudinal user-level information. The Delphi US CTIS survey does not guarantee repeated observations of the same individuals and provides no persistent identifiers. Consequently, we could not construct a panel dataset or include lagged variables capturing delayed effects of information exposure or prior fear levels, preventing explicit modeling of temporal causal dependencies. We explored aggregating observations into daily averages to introduce a temporal dimension. However, conditioning on the large set of demographic and contextual confounders produced subgroup sample sizes too small for reliable estimation. Directly modeling time was therefore not feasible within the adopted framework.

To partially address this limitation, we introduced two variables capturing epidemic evolution without explicit temporal indexing: the categorical Wave variable and the epidemiological Trend variable. The former distinguishes major pandemic phases, while the latter captures epidemic expansion or contraction through the derivative of estimated infections. Together, these variables provide a compact representation of epidemic dynamics despite the absence of longitudinal observations. Within the DAG, Wave and Trend influence trust in information sources, information-source usage, vaccination behavior, and the perceived number of sick individuals, but not Fear directly. The assumption is that individuals do not directly observe epidemiological indicators; rather, information reaches them through media, institutions, social interactions, and other channels represented in the model. Although simplified, this assumption allows the model to isolate the indirect role of epidemic evolution through the information ecosystem.

7. Conclusion

This study investigated the determinants of fear of COVID-19 infection using epidemiological, behavioral, and demographic data from the Delphi US CTIS survey during the Delta and Omicron waves.

Using a Probabilistic Causal Model and the Average Treatment Effect estimation, we quantified the contribution of demographic, epidemiological, and informational factors to fear dynamics. Information sources emerged as the largest explanatory component of fear variance, exceeding the contribution of any other variable group. Institutional and expert-driven sources were associated with increased fear, whereas politicians, religious leaders, and alternative information channels were associated with reduced fear after causal adjustment.

Our findings suggest that fear was shaped not only by epidemiological conditions, but also by how the pandemic was communicated through the information ecosystem. The framing, credibility, and emphasis of information sources could amplify or mitigate emotional responses independently of underlying epidemiological conditions.

Methodologically, this work highlights the importance of causal inference approaches for studying behavioral responses during large-scale crises. By accounting for confounding variables and interdependencies among information channels, causal models provide a more robust framework for analyzing fear determinants than correlation analyses.

8. Ethical declaration

This work received ethical approval from the Ethics Board (IRB) of IMDEA Networks Institute on 2021/07/05. IMDEA Networks signed Data Use Agreements with Facebook and Carnegie Mellon University (CMU) to access data from CMU project STUDY2020_00000162 (ILI Community-Surveillance Study). The data were collected by CMU through the Global COVID-19 Trends and Impact Survey in partnership with Facebook. Informed consent was obtained from all participants. All methods were carried out in accordance with relevant ethical and privacy regulations.

9. Availability of materials and data

The COVID-19 Data Hub (Guidotti and Ardia, 2020; Guidotti, 2022) is open-source and publicly available. The microdata of the Delphi US CTIS survey cannot be shared due to the Data Use Agreements signed with Facebook and Carnegie Mellon University (CMU).