Information seeking and the expected utility of information about COVID-19 can be associated with uncertainty and related attitudes

Abstract

Understanding how people decide when to seek out information can offer important insights into best practices for scientific communication, which may be critical in the face of global challenges like the COVID-19 pandemic. We examined how expected information utility, affective characteristics, and attitudes predicted COVID-19 information seeking behavior in a sample of 191 midwestern undergraduate students in late 2020. Participants completed five rounds of an information seeking task in which they read about a potential gap in their knowledge about COVID-19 and chose whether to read an excerpt that could fill that information gap. We found that information seeking in a given round (i.e. “round-wise information seeking”) was best predicted by expected cognitive utility (i.e., expected reduced uncertainty). When collapsed across rounds, information seeking was positively correlated with COVID-19 preventive behaviors and trust in science, which also correlated with each other. Additionally, exploratory analyses regressing round-wise utility ratings on personality variables revealed that intolerance of uncertainty was associated with higher ratings of all three information utilities. Together, these results suggest that pandemic-related information seeking may have been especially driven by how individuals relate to and manage uncertainty. We discuss how these findings relate to extant literature on information utility and seeking behaviors and highlight the potential for work in this area to improve scientific communication.

Introduction

Predictors of COVID-19 information seeking: Expected information utility and seeking decisions may be particularly associated with uncertainty and related attitudes.

The modern media landscape affords access to an intractably large cache of information–far beyond what any individual could consume. When an exhaustive search is intractable, an information consumer must decide which information is worth the time and effort required to access it. This kind of value-based decision making is thought to direct information seeking (i.e., the process by which individuals acquire information), much like it does other choice behaviors¹. Consequently, examining how state, trait, and contextual features influence individuals’ information evaluations and seeking behaviors has become an area of active research, especially in recent years^2,3. In the present study, we examined the contributions of trait and state features to information-seeking and evaluation in the context of the COVID-19 pandemic—a context that particularly affords the examination of mental health, trust in science, and health-related attitudes and stands to benefit from a more nuanced understanding of what drives information seeking.

The seemingly sudden onset of the COVID-19 pandemic, combined with its major implications for health and daily living, generated wide-scale interest in – and availability of – information about COVID-19. Intuitively, the information an individual is willing and able to access may influence their thoughts, feelings, and actions. This had potentially major consequences in the context of the pandemic, which was characterized by striking political disagreements on public health policies⁴, notable negative impacts on mental health^5,6,7,8,9, and calls for highly coordinated actions (e.g., mask-wearing, hand washing, social distancing, stay-at-home orders¹⁰). Understanding what factors may have directed information-seeking during the COVID-19 pandemic could thus help us make sense of how people responded to this devastating shift in the health status quo and give us cues for similar future events.

Traditionally, information-seeking behaviors have been examined from a variety of perspectives. Early work sought to quantify how a buyer should search for information about the best price for a good in a non-centralized market (i.e., one where there are many sellers whose prices for comparable goods may differ¹¹). From this perspective, information-seeking behaviors are conceptualized as a problem of expected utility: individuals select the information that they expect to be most applicable to reaching their goals. Alternatively, other approaches center on internal mental or emotional states as drivers of information-seeking behaviors, with a particular emphasis on curiosity. Loewenstein¹², characterized curiosity as a feeling of deprivation arising from an individual’s awareness of a gap in their knowledge. This information-gap theory posits that curiosity motivates information seeking as a way to fill an information gap, thus alleviating the feeling of deprivation. As the field of information-seeking has advanced, these perspectives have grown to complement each other and contemporary work often integrates them, affording a more nuanced model of information evaluation¹³.

We extend this literature by examining how expected utility predicts search behaviors after participants were alerted to a potential gap in their knowledge about the COVID-19 pandemic. We focused on the information-utility framework proposed by Sharot and Sunstein¹⁴, which posits that individuals consider three types of information utility in their information-seeking decisions: instrumental, hedonic, and cognitive. Instrumental utility follows from traditional conceptualizations of information utility, where information is evaluated based on whether it will increase the chances of selecting actions that maximize reward and/or minimize harm. For example, information about the effectiveness of masks for reducing the spread of COVID-19 could help one decide if and when they should wear a mask. Hedonic utility describes how information can itself be rewarding or harmful, producing positive or negative feelings. Reading about hopeful doctors describing promising developments in vaccine production could offer high hedonic utility, whereas reading rising cases or death tolls would have negative hedonic utility. Finally, cognitive utility refers to the capacity for information to reduce uncertainty and build on the information-seeker’s existing concepts and schemas. Because this valuation relies on what one already knows, it is expected that people are more likely to seek out information that pertains to topics they are interested in or think about frequently. For example, lay individuals with an interest in microbiology may be particularly inclined to seek information about the SARS-CoV-2 and how infection impacts the immune system, even if this information does not appear to have strong hedonic or instrumental value.

Although the information utility framework is relatively new, early work on the subject supports the tripartite categorization of information utility. Cogliati Dezza et al. found that individuals were able to accurately predict how information will increase their task performance, affect their mood, and reduce their uncertainty and use these predictions to guide their information-seeking decisions¹⁵. Similarly, across four studies Kelly & Sharot found that the tripartite model outperformed other models in predicting information seeking across two domains of information content¹⁶. Interestingly, the weight individuals assigned to cognitive utility in these decisions was inversely related to their score on a general index of psychopathology that included factors labeled “Anxious-Depression”, “Compulsive Behavior and Intrusive Thought” and “Social-Withdrawal”. Together, these results suggest that there are stable individual differences in how individuals calculate information utility and that these calculations may be affected by psychopathology. Applying Sharot and Sunstein’s tripartite utility framework to COVID-19 information seeking thus builds on prior work by examining how psychological distress interacts with individual differences in utility calculation to influence pandemic-related information seeking.

Although existing work has examined potential motivators of prior search behavior, less is known about how different types of information utility drive individuals’ real-time, pandemic-related information seeking. For example, in one study of 1,760 Chinese participants, Zhao and Liu found that self-reported information-seeking had an inverted U-shaped relationship to expected information insufficiency¹⁷. Upon reflection, individuals reported seeking information when a gap in their knowledge about the pandemic existed – suggesting that individuals may seek out information with high cognitive utility to reduce uncertainty about the pandemic. Furthermore, certain types of both positive and negative affective responses to the pandemic (e.g., anger, excitement) predicted information-seeking, suggesting that individuals may have sought information to modulate their emotional states about the pandemic (i.e., information with high hedonic utility). However, it is worth noting that these findings were based on reports of participants’ general affective reactions and self-reported information-seeking behaviors during the pandemic. Much less is known about how expected information utilities predict observed pandemic-related information seeking or whether particular types of information utility differentially predict these choices. Thus, comparing results from observational and self-report methods is an important step towards understanding how observed information seeking aligns with perceived information seeking.

Moreover, information seeking can be driven by more than the expected characteristics of the available information: it can also be driven by characteristics of the decision-maker¹⁸. For example, Ramanadhan and Viswanath found that cancer patients who distrust media reporting on health-related issues are more likely to avoid information-seeking, as were individuals who reported less engagement in preventive behaviors (e.g., not smoking, diet, exercise)¹⁹. Likewise, recent work on COVID-19 information-seeking behaviors showed that higher engagement in preventive behaviors was associated with increased information-seeking²⁰. Thus, we might expect that those who trust health and science communicators or are already performing preventive behaviors would be more likely to engage in COVID-19 information-seeking.

Information-seeking behaviors may also be affected by an individual’s emotional state, mood, and affective traits. There is a particularly rich, but conflicting, literature on the relationship between anxiety and information-seeking, where anxiety has been associated with both increased information-seeking^21,22,23, decreased exploration and information-seeking^24,25,26, or has been shown to be unrelated to information-seeking²⁷. It is unclear whether these contradicting findings arise from the way that anxiety alters exploration and information-seeking processes (e.g., alterations in expected information utility, increased information-seeking when it is unhelpful or decreased information-seeking when it is helpful²⁸) or from qualitative differences in the nature of anxiety individuals experience (e.g., health-related versus general, state versus trait, somatic versus cognitive, clinical versus nonclinical). Interestingly, despite its close association with anxiety, depression has not been as consistently associated with information-seeking^27,29. Furthermore, when such relationships have been found, depression appears to be negatively correlated with information-seeking³⁰. Thus, compared to depression, anxiety is unique in that, under certain circumstances, it may have a positive relationship with information-seeking.

Notably, most of the extant work regarding exploration and information-seeking in anxious or depressed individuals examines behavior in simple behavioral tasks (e.g., bandit task) or relies on post-decisional measures (e.g., self-report, Google history). Much less work has examined whether these aspects of negative affectivity are associated with high-level information-seeking behaviors (e.g., selection of articles of interest) in controlled settings. Additionally, the extent to which psychiatric traits such as depression or anxiety may alter the way that individuals perceive or weigh different kinds of information-utility is still unclear. The present work seeks to examine these queries in the context of COVID-19 information-seeking.

In the present study, we examined several potential predictors of COVID-19 information-seeking in a sample of college students from the Midwestern United States (n = 190), during a time where the COVID-19 pandemic was ongoing (Fall 2020). We employed a within-subjects behavioral design in which participants opt whether to seek information about particular COVID-19 topics after reporting the cognitive, hedonic, and instrumental utility of that information. We examined whether individuals’ ratings of information utility were related to individual differences in anxiety, depression, self-reported history of COVID-19 preventive behaviors, and trust in science. We then fit a mixed effects logistic regression model to examine whether individuals’ ratings of information utility and their scores on the questionnaire measures predicted COVID-19 information-seeking. Our study design and results thus offer new insights into how the qualities of both the available information and an individuals’ characteristics shape information-seeking decisions. In particular, our findings suggest a prominent role of cognitive utility in guiding pandemic-related information seeking and identify intolerance of uncertainty, trust in science, and engagement in preventive behaviors as personal characteristics relevant to information seeking.

Methods

Preregistration

This study was preregistered with the Open Science Framework (OSF; https://osf.io/k7z4b/) prior to data collection. The analyses and hypotheses reported here were preregistered, unless otherwise specified. We treated any analyses that were not preregistered as exploratory and have labeled them accordingly.

We preregistered our experimental design and analytical plan for the model described above. Additionally, we preregistered the following confirmatory hypotheses about our individual differences measures’ relationships to information seeking.

• 1. Information-seeking decisions will have a positive relationship with anxiety, intolerance of uncertainty, trust in science, and COVID-19 preventive behaviors.

• 2. Information-seeking decisions will have a negative relationship with depression.

We further hypothesized that some traits may be associated with a bias in how individuals perceive the three types of information utility:

• 3. We expected higher anxiety scores to be associated with higher expected cognitive utility, increased weighting of cognitive utility, or both.

• 4. We expected depression to be associated with lower ratings and/or weights of instrumental and hedonic utility.

Additionally, we tested the following hypotheses:

• 5. Participants with higher levels of trust in science will be more likely to report engaging in COVID-19 preventive behaviors (e.g., handwashing, social-distancing).

• 6. Greater perceived risk of infection or fatality from COVID-19 will be associated with increased information seeking.

• 7. Individuals may use information seeking to regulate their mood – thus we may expect a significant change in mood ratings from before and after information-seeking decisions.

Participants

Participants were University of Minnesota undergraduate students enrolled in Psychology coursework during Fall 2020 which did the experiment in exchange for course credit. We conducted an a priori power analysis using G*Power 3.1³¹ to determine the appropriate sample size for our study given a medium effect size for correlations in individual differences (r = .20³²) and a desired power of 0.80. This power analysis indicated that a minimum of 150 participants would be required to reach a power of 0.80. To participate in the study and be included in any of the analyses, we required participants to be at least 18 years old, have normal or corrected-to-normal vision, and pass at least three of four attention check items. Additionally, because mixed-effects analyses require complete data, some participants were dropped from these analyses, as indicated in the results. We anticipated a potentially high exclusion rate after applying our inclusion criteria, so we allowed open enrollment in our online study throughout the Fall 2020 semester.

In total, we recruited 223 undergraduate students from the population of students enrolled in introductory psychology courses. After removing participants who failed more than one of our four attention checks, our final sample size was 191. One participant was dropped from the mixed effects modeling due to incomplete data (N = 190), but all 191 participants were included in analyses that did not require complete data. These participants were predominantly young (M = 19 years, SD = 2.2), white (61% White, 27% Asian, 6.0% African American or Black, and less than 5% in each of the following: Native American and White, Asian and White, Other, or Prefer not to respond), and identified as female (73% female, 25% male, 2% non-binary or genderqueer; see Supplementary Fig. 3). Participants completed the experiment online.

Research ethics statement

We believe our study protocol and its execution followed the principles of ethical human subjects research set forth by the Declaration of Helsinki and the Belmont Report. Furthermore, our study protocol was approved for human subjects research by the University of Minnesota’s Institutional Review Board (ID: STUDY00005811).

Materials

Information seeking task

Participants completed five rounds of an information-seeking task. In each round of this task participants had the opportunity to choose whether to seek out information in the form of an article excerpt about topics related to COVID-19. We conducted a search for articles about COVID-19 from a variety of popular news media sources (e.g., National Public Radio, Science Magazine). We also attempted to choose a set of topics that would fully represent the three different types of information utilities. Our final set of topics included herd immunity, efficacy of antibody tests, Vitamin D as a potential treatment and preventative measure, a predicted “baby bust” and its potential economic consequences, and the function of killer T cells in the immune response against COVID-19. The number of excerpts and their content were chosen to balance the representativeness of the included information excerpts against the task length limitations imposed by the threat of participant exhaustion. The order of these topics was counterbalanced across participants (see Supplementary Table 7).

At the beginning of each round, participants saw a prompt that defined an information gap related to that round’s topic, indicating what information the participant could learn about the topic (see Table 1, left). For example, the prompt for the herd immunity topic is given here:

“Herd immunity describes the phenomenon in which a person without immunity to a particular infectious disease is protected from the disease by a large percent of the population who are immune to it. In the following section, you can learn more about what factors will determine when and if herd immunity will take effect for COVID-19.”

After reading the prompt, participants used seven-point Likert scales to rate the hedonic utility, instrumental utility, and cognitive utility they associated with this information (see Table 1, right). They also rated the likelihood that they felt the article would actually contain this level of utility (“How likely is it that your feelings/actions/certainty would change in this way”; for hedonic/instrumental utility/cognitive, respectively). Finally, before making the decision to view or skip the excerpt, participants indicated whether they were familiar with the topic by answering the yes-or-no question “Have you looked into this topic before?”.

Table 1 Sample information gap prompts.

After deciding whether to seek the information, participants rated how much weight they gave to each utility (seven-point Likert scale from “Not at all influential” to “Very influential”) and their post-choice mood (seven-point Likert scale from “Strongly negative” to “Strongly positive”) before reading the article excerpt or moving onto the next round. Each round included exactly one information seeking decision for a single article excerpt. Only one excerpt was used for each topic across all participants. If participants chose to read the article excerpt, they also gave post-excerpt ratings of hedonic, instrumental, and cognitive utility based on how useful the article actually was. To avoid adding potential motivational confounds that might influence information-seeking behaviors, we did not include any extrinsic rewards for engaging in information seeking (e.g., points, money).

COVID-19 risk perception items

To index participants’ perceptions of COVID-19 related risks, we asked participants to rate the probability of four different COVID-19-related scenarios: (1) “I will get sick with COVID-19”, (2) “Someone I know will get sick with COVID-19”, (3) “I will die due to COVID-19”, and (4) “Someone I know will die due to COVID-19”. Participants provided probability ratings ranging from 1 to 100 on a sliding scale. In accordance with the data privacy protocol approved by our IRB, we did not ask participants to directly report information about their general medical or COVID-19 diagnostic history. Consequently, our data cannot distinguish whether participants that provided a risk rating at or near 100% for any of these items represent reports of diagnosed or confirmed cases of COVID-19.

Individual differences questionnaires

Depression, anxiety, stress scale (short-form; DASS-21)

The DASS-21 is a self-report inventory commonly used to measure general distress over the past week (S. H. Lovibond et al., 1995). The inventory contains 21 items which participants answer using a 4-point Likert scale. Higher scores indicate higher levels of distress. It is particularly notable for its ability to distinguish between depression, anxiety, and stress – a task which other measures of these constructs often struggle to achieve³³. The depression scale (M = 2.12, SD = 3.64) is characterized by low self-esteem and a general dysphoric mood, the anxiety scale (M = 1.22, SD = 1.77) is characterized by enduring state anxiety and fear, and the stress scale (M = 3.51, SD = 3.78) is characterized by bodily arousal, tension, and irritability³⁴. The DASS-21 has demonstrated acceptable to excellent levels of convergent and divergent validity, as well as strong internal reliability³⁵.

Intolerance of uncertainty scales, short form (IUS-12)

The IUS-12 is a 12-item self-report measure of intolerance of uncertainty³⁶. It is the short-form version of the IUS-27, an inventory designed to assess individuals’ emotional, cognitive, and behavioral reaction to uncertainty³⁷. The short form is defined by two subscales: seven items related to prospective anxiety (fear regarding the future, e.g. “Unforeseen events upset me greatly”) and five items related to inhibitory anxiety (the extent to which fear of uncertainty prevents action, e.g. “When it’s time to act, uncertainty paralyzes”). The sum of these two subscales can be interpreted as a general intolerance of uncertainty. The total IUS-12 has an expected mean score of 25.85 (SD = 9.45) and excellent internal reliability (α = 0.96³⁶).

Trust in science and scientists inventory (TSSI)

While some researchers have included items to examine trust in science and scientists, almost none have formally developed an inventory for such a purpose. The TSSI was developed to fill this niche³⁸. Participants rate the extent to which they agree with various statements about scientists and their work (e.g., “We can trust science to find the answers that explain the natural world”) using a five-point Likert scale. The initial item formation for the TSSI relied on an interdisciplinary research team of scientists and educators and was later tested on two rounds of undergraduate college students to test and restructure the inventory. The final TSSI was reported to have good internal reliability (α = 0.86), but there are little to no studies beyond the TSSI’s original publication that examine or report its psychometric properties.

COVID-19 preventive behaviors questionnaire

To score participants’ self-reported engagement in preventive health behaviors, i.e., behaviors aimed at reducing the spread of COVID-19, we wrote a set of nine items based on the United States’ Center for disease prevention and control (CDC) COVID-19 guidelines³⁹ (see Supplementary Materials, Appendix A for items). Each item is rated on a seven-point Likert scale ranging from “strongly disagree” to “strongly agree” and include statements like “Even if I do not feel sick, I still follow social distancing guidelines (e.g., staying 6 feet apart, wearing a mask in public)” and “I frequently wash my hands thoroughly, for at least 20 seconds.” Participants’ total scores on this inventory were intended to provide an estimate of the extent to which they had been following CDC guidelines. The total set of nine items demonstrated good internal reliability (Cronbach’s α = 0.85).

Procedure

During the study, participants completed three Qualtrics forms: a consent form, a response form, and a compensation information form. Participants’ identifying information (e.g., name, email) was always kept separate from their study responses. Prior to beginning the study, all participants provided informed consent – those who did not were redirected to an end of survey message explaining that informed consent was required for participation. The response form contained sections of the study that were presented in a fixed order across participants, as follows: demographic item, risk perception items, the information-seeking task, and the individual difference questionnaires (e.g., DASS-21, TSSI, COVID-19 preventive behaviors). Although the order of these sections was fixed, the order of the topics in the information-seeking task and the order of the questionnaires in the questionnaire block were both counterbalanced across participants.

Analyses

Identifying the most predictive calculation of utility

The best method for calculating information utility scores has not yet been clearly established. Previous work has asked participants to provide ratings of information utility^14,15 and have sometimes asked participants to estimate a weight of how much each type of utility contributed to their decision¹⁶. Furthermore, some methodologies have incorporated likelihood or certainty into participants’ estimates of each type of expected information utility¹⁵.

In our study, participants rated the utility associated with the decision to seek out the information in each article, as well as how likely they felt the article would actually offer the expected utility and the weight they gave to each utility in their decision-making. We calculated the expected utility of information seeking I given both the expected state s experienced after acquiring the information and the observed information gap G. State s is a variable that corresponds to each type of utility: s_h is the expected state associated with hedonic utility, s_i is the expected state associated with instrumental utility, and s_c is the expected state associated with cognitive utility.

We identified four different calculations of expected utility. Our first measure of expected utility only involves the utility ratings (index 1):

$$\:EU\left(I|s,G\right)=\:U\left({s}_{h},G\right)+\:U\left({s}_{i},G\right)+\:U\left({s}_{c},G\right)$$

(1)

Where U(s, G) is the utility of the expected state s after acquiring the information missing from gap G. Our second measure of expected utility takes into consideration both the utility ratings and the likelihood (index 2):

$$\:EU\left(I|s,G\right)=\:U\left({s}_{h},G\right)P\left({s}_{h}|I,G\right)+\:U\left({s}_{i},G\right)P\left({s}_{i}|I,G\right)+U\left({s}_{c},G\right)P\left({s}_{c}|I,G\right)$$

(2)

Where P(s|I, G) represents the probability that the state s will occur given that the participant seeks out the information (I) for a specific information gap G. Our third measure of expected utility included the utility ratings and the weighted value of the utility ratings (index 3):

$$\:EU\left(I|s,G\right)=\:U\left({s}_{h},G\right)W\left(U({s}_{h},G)\right)+\:U\left({s}_{i},G\right)W\left(U({s}_{i},G)\right)+U\left({s}_{c},G\right)W\left(U({s}_{c},G)\right)$$

(3)

Where W(U(s, G)) indicates the weighting associated with the corresponding utility U(s, G) as rated by participants on a seven-point Likert scale. Our fourth measure of expected utility included all of the above-mentioned measures: the utility ratings, the likelihoods, and the utility weightings (index 4):

$$\:EU\left(I|s,G\right)=\:U\left({s}_{h},G\right)P\left({s}_{h}|I,G\right)W\left(U({s}_{h},G)\right)+\:U\left({s}_{i},G\right)P\left({s}_{i}|I,G\right)W\left(U({s}_{i},G)\right)+U\left({s}_{c},G\right)P\left({s}_{c}|I,G\right)W\left(U({s}_{c},G)\right)$$

(4)

To identify the best utility index, we constructed three different logistic regression models with an index for each type of utility (i.e., hedonic, instrumental, and cognitive) included as separate predictors: (1) Model A included the three expected utility measures as fixed effects and participant as a random effect, (2) Model B included the three expected utility measures and dummy-coded predictors for each of the five topics as fixed effects with participant as a random effect, and (3) Model C was a fixed-effects model using the utility measures and dummy-coded topic predictors. We compared model fit using Bayesian Information Criterion (BIC) for each model using the full data set. The k-fold cross-validation method is designed to identify the model with the best out-of-sample predictive ability, whereas the BIC method tries to identify the ideal tradeoff between model fit and complexity⁴⁰. For a more comprehensive evaluation, we also examined Akaike Information Criteria (AIC) for each model as an exploratory analysis, as AIC offers a more liberal allowance for additional predictors than BIC. Finally, to examine out-of-sample prediction error performance more directly we ran 100 five-fold cross-validations on the Brier Scores for each of the models and constructed a kernel density plot comparing the distributions for each model (see Fig. 2, described later in Results). Per our preregistration, in the event of a disagreement between model comparison metrics, we opted for the model with the lowest BIC value.

Identifying predictors of round-wise information seeking

One of our primary research goals was to model how participants made information-seeking decisions over each round of our information-seeking task. To capture the repeated-measures design of our study, we fit a mixed-effects logistic regression model to our data. Mixed-effects models require complete observations (i.e., no missing values), so our sample size for the model (N = 190) was slightly smaller than our overall sample size of eligible participants in the study (N = 191). We then randomly assigned participant data to either the training (n = 152) or test (n = 38) subset, approximately corresponding to an 80/20 split, and scaled the predictors by training set responses to ensure numerical comparability of variances. Before engaging in model selection, we first examined the full model to afford a broad understanding of the relationship between each variable and information-seeking. We fit a mixed-effects logistic regression model using glmer()⁴¹ to predict information-seeking from round-wise utility ratings, participants’ self-reported scores for depression, anxiety, trust in science, and COVID-19 preventive behaviors, and random intercepts by participant, topic, and topic order.

We initially included topic and topic order random effects in our model formula because we wanted to treat our chosen article excerpts and the orders in which they were presented to participants as random variables (i.e., samples drawn from a larger population of potential excerpt topics and orders). However, it is likely that including these predictors produced some issues with model fit due to redundancy c. Indeed, including these terms in our model design produced a singular fit (i.e., it was impossible to fit the model using linear optimization). To circumvent this issue, we dropped the topic and topic order random intercepts that were initially preregistered, resulting in the following full model:

$$\:{y}_{ij}={\beta\:}_{0}+{\beta\:}_{1}{x}_{cog}\:+{{\beta\:}_{1}{x}_{hed}+\:{\beta\:}_{1}{x}_{instr}+\:{\beta\:}_{1}{x}_{dep}+\:{\beta\:}_{1}{x}_{anx}+\:{\beta\:}_{1}{x}_{tssi}+\:{\beta\:}_{1}{x}_{covid}+\:\mu\:}_{i}+{\epsilon\:}_{ij}$$

(5)

To select our final information-seeking model formula, we fit a model for each possible combination of predictors and selected the model with the lowest BIC using the dredge() function from the MuMIn package for multi-model inference⁴². Finally, we used the selected model to predict information-seeking in our test set and examined its classification accuracy. Because this model includes information that changed across round (i.e., participants’ information utility ratings), we refer to the model’s predictions a “round-wise” predictions. We use this term to distinguish these results from those that collapse information across rounds.

Modeling the effect of information seeking on mood

To examine whether information seeking modulated round-wise mood ratings, we fit a mixed-effects linear regression model regressing the difference in pre- and post-decision mood scores on participants’ information seeking decisions and anxiety scores and added random intercepts by participant, topic, and topic order. Once again, including topic and topic order produced a singular model matrix, making it impossible to proceed with model fitting procedure using all of our parameters. Thus, to be consistent with the model we fit to predict information-seeking decisions (see Eq. (5)), we decided to drop the topic and topic order random intercepts to resolve the of singular fit.

Exploratory analyses

Correlations between weighted utility ratings and individual differences

To further explore relationships between our questionnaire measures and information utility, we conducted Spearman correlations between all the scaled questionnaire and cross-topic mean utility scores in our sample using the corrplot() function in R⁴³ (see Fig. 5, described later in results). When interpreting these results, we applied a Bonferroni adjustment for multiple comparisons, which reduced our significance threshold from α = 0.05 to α = 0.0045. We then used hierarchical clustering functionality of corrplot() to help visualize which variables were most closely related to one another.

Fitting a model to predict round-wise utility ratings

To probe predictors of round-wise information utility ratings, we fit an exploratory mixed-effects linear regression model with weighted utility ratings as the outcome variable. We included utility type (a factor with three levels: cognitive, hedonic, instrumental) and scaled scores of participants’ total intolerance of uncertainty, anxiety, depression, mean TSSI scores, and COVID-19 preventive behaviors as fixed effects and added a random intercept by participant. We also examined fixed-effect interactions between the utility types and each of the scaled scores listed above. We compared all possible combinations of these fixed effects and ranked the models by their BIC and AIC scores (see Supplementary Results section titled “Model Comparisons for the Exploratory Model Predicting Round-Wise Weighted Utility Ratings”). Performing model selection using BIC is more likely to produce the true model, assuming that the true model is defined by a subset of the predictors included in the full mode. However, AIC is more tolerant of more complex models, which may be desirable for hypothesis generation despite its drawbacks for hypothesis testing. Moreover, AIC can outperform BIC when the true model is not a subset of the available predictors (i.e., when potentially relevant predictors were not measured in the study or included in the model). Given this model’s role as an exploratory analysis, we report both the AIC and BIC model selection results.

Results

Descriptive statistics

Information seeking

On average, participants sought information on just under half of the five rounds (M = 2.35, sd = 2.03; see Fig. 1 for detailed rates of seeking different proportions of rounds). Each topic elicited similar amounts of information seeking across participants, despite subtle variation in cognitive, hedonic, and instrumental utilities and dramatic variation in the topic familiarity (i.e., whether a participant had searched for information on this topic in the past; see Table 2). Exploratory paired-samples Wilcoxon signed-rank tests revealed that none of the topics elicited a significant change in mood after the information seeking decision (see Supplementary Fig. 4 for details), with the exception of the article about killer T cells, which produced a significant increase in post-decision mood (W = 834, p = 5.20*10 − 6).

Table 2 Descriptive statistics for information seeking, familiarity, and information utility for each topic.

Fig. 1

Rates of Total Information Seeking. Bar plot depicting the proportion of participants that sought information at different rates. Notably, 30% of participants never sought information on any of the rounds, and 26% of participants sought information on every round.

Questionnaire measures

We examined the distribution of participants’ questionnaire scores using box plots (see Supplementary Fig. 5). Given that our sample was drawn from a student population and that depression and anxiety scores on the DASS-21 have potential clinical relevance, it is possible that our sample would lack sufficient variance in the severity of depression or anxiety scores to capture the relationship between these traits and our outcomes of interest. To address this concern, we compared scores on depression and anxiety from our sample compared to mean values from a relatively small norming study sample of nonclinical participants³⁵ (see Supplementary Fig. 6, top). On average, participants in our sample scored much higher on both anxiety and depression relative to the nonclinical norming sample. Notably, participants in the norming sample were screened such that individuals with a previous history of psychopathology were excluded, whereas our study did not remove any participants based on their clinical symptom scores. Further supporting the notion that our sample had adequate variability in depression and anxiety scores, we compared scores from our against the recommended clinical severity cutoffs provided by the DASS-21 scoring instructions³⁴ (see Supplementary Fig. 6, bottom). Our sample contained participants from all five categories of clinical severity (normal, mild, moderate, severe, and extreme), suggesting that our sample contained a fairly wide range of clinical severity scores.

Best-fitting calculation of information utility

Our model fitting procedure identified weighted utility ratings as the best-fitting method of calculating information-utility, with converging evidence across BIC, AIC, and Brier score values (index 3; see Eq. (3), Fig. 2 and Supplementary Table 8).

Fig. 2

Visualizing model fit statistics for different calculations of utility. Index 1 incorporates ratings only, index 2 ratings and likelihood, index 3 ratings and weightings, and index 4 ratings, weightings, and likelihood. Lower Brier scores (bottom) indicate better out-of-sample predictive ability. Likewise, lower BIC (top left) and AIC (top right) values indicate better model fit. AIC and BIC use maximum likelihood estimation rather than relying directly on prediction accuracy and consequently tend to penalize model complexity more harshly. BIC punishes model complexity more than AIC. All three model indices provide convergent support for the use of index 3 (ratings x weightings) as our utility measure, although differences in model fit between these four models appear relatively small.

Cognitive utility and self-reported COVID-19 preventive behaviors predicted information-seeking decisions

We fit the full logistic mixed effects model Eq. (5) to the training data, predicting information-seeking using weighted utility ratings from each round (separate predictors for cognitive, hedonic, and instrumental utility), participants’ self-reported scores for depression, anxiety, trust in science, and COVID-19 preventive behaviors, and random intercepts by participant. We did not detect any violations of the homogeneity of variance or normal distribution of error variance assumptions (see Supplementary Fig. 7). Our predictors did not show concerning levels of multicollinearity – our Variance Inflation Factor (VIF) analysis revealed relatively small VIFs (all VIFs < 1.68) for all variables compared to standard cutoffs suggested in the literature (VIF < 5^40,44. Examination of the full model fit to the training data revealed that both cognitive utility (β = 1.03, 95% CI [0.60, 1.46], p = 2.83*10 − 6) and COVID-19 preventive behaviors (β = 0.87, 95% CI = [0.26, 1.47], p = .005) positively predicted information-seeking decisions. Exploratory analysis revealed that, together, these fixed-effects alone explained approximately 18% of the variation in information-seeking decisions ($\:{R}_{m}^{2}=\:0.19$), whereas the full mixed-effects model explained the majority of the variance in responses ($\:{R}_{c}^{2}=\:0.76$). None of the other predictors demonstrated a significant association with information-seeking behaviors in the model (see Table 3), although it is worth noting that the use of significance testing for interpreting the results of mixed-effects models has been criticized in the past⁴⁵.

Table 3 Regression coefficients from the full model using the training data.

We compared all possible combinations of the predictors and found that the model with the lowest BIC included only cognitive utility and COVID-19 preventive behaviors as fixed-effects, with a random intercept by participant (see Supplementary Table 9).

$$\:{y}_{ij}={\beta\:}_{0}+{\beta\:}_{1}{x}_{cog}+{\beta\:}_{2}{x}_{cov\_prev}+{\mu\:}_{i}+{\epsilon\:}_{ij}$$

(6)

As shown in Eq. (6), the reduced model predicts binary information-seeking decisions for each participant on each round (y_ij) using weighted cognitive utility judgements (x_cog), COVID-19 preventive behaviors (x_{cov_prev}), a random intercept by participant (µ_i), and subject-by-round specific error terms (ε_ij). Fitting the reduced model to the training data produced significant, positive regression coefficients for both cognitive utility (β = 1.10, p < .001) and COVID-19 preventive behaviors (β = 0.95, p = .002; see Table 4), consistent with the full model. The final fitted model demonstrated strong within-sample predictive accuracy (see Fig. 3A).

We conducted an exploratory analysis to examine the prediction accuracy of the final model (see Eq. (6)), which revealed the model’s accuracy was largely dependent on the random intercept by participants term (µ_i). We fit a model with only the fixed effects terms to the training data and tested its out-of-sample prediction accuracy on the test set. The fixed-effects-only model performed above chance level on out-of-sample prediction accuracy (62.20% correct, d’ = 0.63, AUC = 0.61), but the addition of the random intercepts term (fitted to the test set) greatly improved accuracy (87.80% correct, d’ = 2.46, AUC = 0.87). Consistent with these findings, further exploratory analyses revealed that the fixed-effects alone explained a much smaller percentage of variance in information-seeking ($\:{R}_{m,\:train}^{2}=\:0.18,\:\:{R}_{m,\:\:test}^{2}=\:0.03$) compared to the mixed-effects version of the model ($\:{R}_{c,\:\:train}^{2}=\:0.76,\:{\:R}_{m,\:\:test}^{2}=\:0.80$). Although it showed a similar pattern of results, the variance explained by the fixed effects was much smaller in the test sample model fit compared to the training sample model fit.

Table 4 Regression coefficients for the final reduced model fit to the training data.

To examine out-of-sample prediction error, we fit our final reduced model Eq. (6) to the test set data using the coefficient values given in (Table 4). We then classified participants’ rounds as “seeking” if the predicted probability score was above 50% and “non-seeking” if it was below. Necessarily, the random intercepts term (µ_i) was refitted in the mixed effects model, as the test set included exclusively participants absent from the training set. Consistent with the findings in the training set, exploratory analyses revealed that out-of-sample prediction accuracy was near chance-level when fixed-effects were examined alone (52% correct, d’ = 0.40, AUC = 0.52), but the full mixed-effects model showed strong prediction accuracy (91% correct, d’ = 2.70, AUC = 0.89 see Fig. 3B; see also Supplementary Results, section “Exploratory “Fixed-Effects Only” Information-Seeking Model”, Supplementary Fig. 1, and Supplementary Tables 1, 2 for additional model comparisons). To see a similar analysis excluding participants who never chose to seek information, refer to the Supplementary Results section titled “Exploratory Information-Seeking Model Excluding Participants Who Never Sought Information”.

Fig. 3

ROC curves of the final model fit to the training and test data. ROC curves for the final model in the training data (A) and test data (B). Each plot includes a separate ROC curve for the fixed-effects only model and the mixed-effects model (i.e., including random intercepts by participant). Model performance is quite good in both the training and test set. Notably, performance appears to depend heavily on fitting random intercepts, suggesting that cognitive utility and COVID-19 preventive behaviors are parts of a greater set of participant characteristics that can together predict information-seeking.

COVID-19 preventive behaviors and trust in science correlated with information seeking and with each other

In addition to predicting round-wise information-seeking, we wanted to examine whether any of the individual differences measures we included were associated with individuals’ overall proportion of information-seeking decisions. To this end, we conducted Spearman correlations between participants’ overall proportion of information-seeking decisions across the five rounds of the study and their anxiety, depression, intolerance of uncertainty (IUS), COVID-19 preventive behaviors, and Trust in Science scores. Our results revealed that, as predicted, the overall proportion of information-seeking was significantly positively correlated with COVID-19 preventive behaviors (ρ = 0.21, p = .003) and Trust in Science (ρ = 0.18, p = .01) although surprisingly not with the other measures (see Fig. 4). We also found a positive correlation between Trust in Science and COVID-19 preventive behaviors (ρ = 0.32, p = 7.2 * 10^− 6; see Fig. 5; see supplementary Table 9 for values). Thus, consistent with our preregistered hypothesis, higher levels of trust in science were associated with greater self-reported engagement in COVID-19 preventive behaviors, and both were positively correlated with participants’ decisions to seek out COVID-19 information.

Fig. 4

Correlations between individual differences measures and overall information seeking. Scatterplot where points represent subjects’ scores on the predictor variables (anxiety, depression, intolerance of uncertainty, COVID-19 health behaviors, and trust in science) plotted against the proportion of rounds where they chose to seek information (i.e., overall seeking, collapsed across topics). Lines represent Spearman correlation values. Only COVID-19 Preventive behaviors and Trust in Science are significantly correlated with overall proportion of information-seeking.

Neither depression nor anxiety correlated with expected information utility across rounds

To examine whether depression and anxiety were associated with participants’ utility ratings across all rounds, we conducted pre-registered Spearman correlations between standardized anxiety and mean cognitive utility scores as well as standardized depression and mean hedonic and instrumental utility scores. Contrary to our hypotheses, we found neither evidence that anxiety was correlated with cognitive utility ratings (ρ = 0.07, p = .34) or weightings (ρ = − 0.06, p = .40) nor that depression was correlated with either the ratings of hedonic (ρ = 0.01, p = .88) and instrumental (ρ = 0.03, p = .67) utility or their weights (ρ_hedonic = 0.02, p = .77; ρ_instrumental = 0.03, p = .66).

Perceived risk of infection or fatality was not associated with information seeking

On average, participants indicated relatively high levels of perceived risk of infection or fatality for themselves and others (see Supplementary Table 10 for descriptive statistics). Our preregistration used Pearson correlations to examine the relationship between perceived risk and participants’ overall information seeking for each of the four risk ratings. However, examination of the risk item data revealed major departures from normality (see Supplementary Fig. 8), so we instead calculated Spearman correlation coefficients. Contrary to our expectations, we found no relationship between any of the four risk perception ratings and participants’ overall proportion of information-seeking (ρ_{self−infection} = − 0.02, p = .82; ρ_{self−fatality} = − 0.07, p = .32; ρ_{other−infection} = 0.07, p = .40; ρ_{other−fatality} = 0.002, p = .97).

Information seeking was not associated with post-decision changes in mood

As described above, we fit a linear mixed-effects model regressing difference scores in pre- and post-decision mood on information seeking decisions, anxiety scores, and a random intercept by participant. Contrary to our hypothesis, neither information-seeking (β_seek = −0.07, p = .41) nor anxiety (β_anxiety = −0.03, p = .49) was associated with a change in pre- and post-decision mood. Exploratory analyses revealed poor overall model fit, as indicated by the marginal and conditional coefficients of determination ($\:{R}_{m}^{2}=\:0.001,\:{R}_{c}^{2}=\:0.10$). This suggests that neither the fixed nor random effects predicted much of the variance in mood changes.

Exploratory analysis results

Information utility ratings were strongly correlated

Given that our model selection procedure retained cognitive utility, but not hedonic or instrumental utility as predictors of round-wise information, it is important that we examine the relationship between utility ratings in the task. As shown in Fig. 5, information utility ratings were strongly, positively correlated across utility types. We also examined Spearman correlations between utility ratings of different topics at the excerpt level (i.e., de-aggregating across topic levels) and found the same pattern of results - once again, utility ratings were tightly correlated across utility types (see Supplementary Fig. 9). Thus, our information stimuli were not strongly differentiated across types of information utility.

Total information seeking differs across severity of depression, but not anxiety

It is possible that the effects of depression and anxiety symptoms on information-seeking depend on the level of symptom severity or are too small to outcompete the contributions of state-like variables like information utility ratings. This may be exacerbated by the fact that our study did not expressly control for levels of symptom severity (i.e., we did not screen out participants with high or low symptom loads or ask about diagnosis history). To address this, conducted a Kruskall-Wallis test to examine whether total information sought differed across the standard symptom severity groups defined by the DASS-21 (i.e. normal, mild, moderate, severe, and extreme³⁴). The test revealed that the severity level of depression, but not anxiety, symptoms was significantly associated with total information seeking (H_{2,n = 191} = 10.42, p = .034; see Supplementary Fig. 10), though post-hoc pairwise Wilcoxon signed-rank comparisons revealed no between-group effects (see Supplementary Table 11).

Intolerance of uncertainty and COVID-19 preventive behaviors correlated with weighted utility ratings

Exploratory Spearman correlations across questionnaires and weighted utility ratings revealed that intolerance of uncertainty was positively correlated with participants’ weighted utility ratings for all three types of information utility (ρ_cognitive = 0.24, p_cognitive = 0.001; ρ_hedonic = 0.29, p_hedonic = 5.31*10^− 5; ρ_instrumental = 0.31, p_instrumental = 1.56*10^− 5; see Fig. 5), suggesting that discomfort with uncertainty is associated with greater overall expected value of information. No other affective characteristics (e.g., depression, anxiety, stress) were correlated with information utility, however self-reported COVID-19 preventive behaviors were positively correlated with participants’ average cognitive (ρ = 0.25, p = .0004) and instrumental (ρ = 0.26, p = .0004) utility ratings. Using the hierarchical clustering functionality of corrplot(), we visualized four hierarchical clusters demonstrating strong groupings across mean utility ratings, the IUS-12 scores, the DASS-21 scores, and the COVID-19 preventive behaviors coupled with the Trust in Science and Scientists Inventory (see Supplementary Fig. 11).

Fig. 5

Spearman correlation matrix of utility scores and individual differences measures. Matrix of exploratory Spearman correlations between individual differences measures and average weighted information utility ratings across participants. Asterisks represent Bonferroni-corrected significance levels corresponding to the standard 0.05, 0.01, and 0.001 cutoffs: * = 0.0045, ** = 0.00091, *** = 9.1*10 − 5. A full table of correlation coefficients and corresponding p-values can be found in (Supplementary Table 12).

Round-wise utility ratings may be predicted by intolerance of uncertainty and depression

As described above, we fit a mixed effects linear regression predicting weighted utility ratings from each round using utility type, anxiety, depression, intolerance of uncertainty, trust in science, and COVID-19 preventive behaviors as fixed-effects and a random intercept by participants to the training data (see Supplementary Table 13 for full model summary). The full model revealed main effects of depression (β = − 0.218, p = .002), intolerance of uncertainty (β = 0.245, p = 1.78*10 − 4), and COVID-19 preventive behaviors (β = 0.176, p = .002), but no other significant main effects or interactions. Notably, depression was associated with lower weighted utility ratings, whereas intolerance of uncertainty and engagement in COVID-19 preventive behaviors were both associated with higher weighted utility ratings.

Model selection by BIC revealed that, of these affective predictors, intolerance of uncertainty alone appeared to predict information utility ratings (see Supplementary Table 5). The final reduced model thus had the following form:

$$\:{y}_{ij}=-1.18*1{0}^{-4}\:+\:0.23{x}_{ius}+{\mu\:}_{i}+{\epsilon\:}_{ij}$$

(7)

The positive coefficient on the term corresponding to participants’ IUS total scores suggests that, when random effects and grand average information utility ratings are taken into account, individuals with higher intolerance of uncertainty tend to view information as more valuable.

Discussion

We examined the role of affective traits, trust in science, COVID-19 preventive behaviors, and expected information utility in predicting college students’ COVID-19 information-seeking decisions. Our results suggest that pandemic-related information-seeking within this timeframe (i.e., Fall 2020) and population (i.e., Midwestern undergraduates) was associated with cognitive utility ratings, COVID-19 preventive behaviors, and trust in science and scientists. Contrary to our hypotheses, we found no evidence for a relationship between participants’ overall proportion of information-seeking rounds and individual differences in risk perception, depression, anxiety, or intolerance of uncertainty. Additionally, we found no evidence to support our hypothesis that information-seeking could predict changes in mood. Below we discuss these predictors and interpret their relationships to pandemic-related information-seeking. We also examine consistencies and inconsistencies between our results and the broader literature on information-seeking.

Reducing cognitive uncertainty played a major role in COVID-19 information seeking

While past research has established general principles of information seeking and value computation, the present study offers insights into the role of information utility in COVID-19 related information seeking. In our sample of undergraduate students, round-wise information-seeking behaviors were predicted by cognitive utility but not hedonic or instrumental utility. Furthermore, we found no evidence that our participants effectively used information seeking to regulate their mood, as seeking decisions were not associated with a change in participants’ mood state. Additionally, information seeking decisions were not associated with participants’ ratings of perceived risk of infection or fatality, suggesting that our participants did not use information seeking as a method of risk management. Thus, information seeking about COVID-19 in our sample was particularly characterized by a drive to reduce uncertainty, rather than a desire to influence emotional and goal-oriented outcomes.

This emphasis on cognitive utility as the strongest predictor of information seeking is somewhat unique compared to past work. Multiple studies examining the information utility framework have found that a model including all three utilities outperforms alternative models that include only a subset of these utilities or alternative predictors such as entropy¹⁵ and distinctiveness, sense-making, and recency¹⁶. Interestingly, in both of these studies, the model with the second lowest BIC rating included both hedonic utility and cognitive utility – suggesting that cognitive utility was particularly important in predicting information-seeking decisions in these studies too. It is not clear why hedonic and instrumental utility were not strong predictors of information seeking in our model. One explanation may be that the COVID-19 pandemic created a particular context wherein the importance of cognitive utility was enhanced, the importance hedonic and instrumental utility were attenuated, or both. For example, the COVID-19 pandemic was associated with a global increase in negative feelings and mental health concerns⁸, which may have masked or overshadowed participants’ expectations about the impact of information on their feelings or goals. Consistent with this notion, prior work has found a unique relationship between cognitive utility and psychopathological symptoms, wherein individuals who weighted the cognitive utility of self-referential information more highly during their information search had higher levels of psychopathological symptoms¹⁶. Thus, future work may seek to examine the extent to which different information content and contexts, especially those that elicit strong emotions, can alter perceptions of information utility and its relationship to information-seeking behaviors.

A positive relationship between expected cognitive utility and seeking behaviors has potential implications for the communication of pandemic-related information. Our results suggest that feelings of uncertainty about a topic are particularly important in guiding information seeking about this topic, thus future work may examine whether engagement with health-related information is influenced by either manipulating informational text to emphasize its ability to reduce uncertainty (e.g., via message framing) or altering the salience of feelings of uncertainty (e.g., by emphasizing information gaps). If utilized properly, this line of work may be able to help health organizations (e.g., CDC, WHO) and news media more effectively communicate, educate, and engage readers in their reporting of health-related information.

Anxiety and depression did not predict information seeking, except in a subsample of participants who sought information at in at least one round

We found no relationship between depression and anxiety scores and COVID-19 information seeking decisions in our primary analyses. This is particularly surprising given the similarity of our information content to that of prior work. Many studies that have found positive associations between anxiety and information-seeking examined health-related information^23,46, and several studies specifically found positive relationships between anxiety and self-reported engagement or exposure to COVID-19-related information^{17,47,48,49,50}. Notably, all of these studies, except So et al., 2019, used simple self-report measures of information-seeking aimed at understanding participants’ general decision-making tendencies. In contrast, our work focused on predicting round-wise and overall information-seeking behaviors directly. Self-report measures of information seeking may be more likely to capture a participants’ general tendencies over time but may be prone to bias (e.g., desirability bias, poor self-reflective insight), whereas observing behavior directly provides a more objective measure of behavior at the cost of scope. Thus, trait anxiety may predict participants’ perceptions about their own information seeking tendencies, but it is less clear to what extent these perceptions accurately reflect behavioral tendencies.

Notably, previous work has found that using observed rather than self-reported information-seeking measures yielded small but significant correlations between anxiety and information-seeking measures (|r| ≤ 0.10²³). However, the designed used in this study differed from ours in several important ways. First, participants were allowed to freely search for information from a set of ten topics – they could spend as little or as much time on each of these topics in any order and were allowed to skip information-seeking entirely if desired. Second, information-seeking was measured by both time spent and number of articles examined and the relationship between anxiety and information-seeking was only examined for participants who chose to seek-out information on at least one topic. And finally, anxiety was measured as a state, rather than trait-level, variable. These design differences could explain the inconsistency between our findings, in which case we might expect trait anxiety to predict the frequency and breadth of information-seeking, but not decisions of whether to seek a particular piece of information or to initiate information seeking at all. However, the previous study also had a much larger sample size (N = 916; 408 information-seekers) than the present study (N = 191). Thus, it seems likely that our study may simply have been inadequately powered to detect the small effect size of the potential relationship between anxiety and information-seeking.

Alternatively, some research suggests that anxiety may promote information avoidance rather than information seeking⁵¹. People may, for example, avoid learning about risks like diseases to minimize anxiety about uncertain, threatening events⁵². This avoidance helps maintain psychological comfort by shielding them from distressing possibilities. For example, medical patients have been shown to avoid learning about potential health risks, such as the results of genetic screening for cancer²⁴. Similarly, during the early stages of the COVID-19 pandemic, when our study was conducted, people with high anxiety may have avoided information as a coping strategy to avoid increasing their anxiety.

This perspective aligns with our supplementary analysis, which found that anxiety negatively predicted information seeking in our best fitting model, indicating that less anxious participants were more likely to seek information (see Supplementary Results section “Exploratory Information-Seeking Model Excluding Participants Who Never Sought Information”). For this analysis, we excluded participants who refrained from seeking information in all rounds. Interestingly, we also found that depression was negatively correlated with participants’ overall proportion of information-seeking rounds (ρ = − 0.18, p = .0.042) in this subsample, though this relationship did not survive correction for multiple comparisons (see Supplementary Fig. 2).

These results should be interpreted with caution, as the reasons behind the participants’ decision to refrain from seeking additional information remain unclear. It is possible that these participants were engaged with the task but were simply uninterested in seeking additional information. Alternatively, their behavior might reflect a lack of motivation to exert more effort than was necessary to complete the task. Given the current task design, it is challenging to determine which explanation best accounts for this subset of participants. However, these individuals appeared to comprehend the task requirements, as they met our inclusion criteria by answering at least three out of four attention checks correctly. This suggests they were actively processing the task material and likely understood the prompts well enough to make informed decisions.

Expected information utility may be influenced by affective individual differences

Although affective traits did not predict information-seeking, exploratory analyses suggested that they may be related to how an individual evaluates information – i.e., how they calculate information utility. We had hypothesized that depression would be associated with lower perceived hedonic and instrumental utility. While a direct correlation between depression scores and hedonic or instrumental utility did not yield any significant effects, an exploratory mixed-effects multiple linear regression predicting round-wise weighted utility ratings indicated that individuals with higher levels of depression generally perceived information to have less value across rounds. It is possible that we did not find a direct correlation between depression and information utility because co-occurring intolerance of uncertainty tends to mask this relationship. Notably, while depression was negatively associated with information utility, intolerance of uncertainty was positively associated with it. There is, however, also a significant positive correlation between depression and intolerance of uncertainty. It is thus possible that potential depression effects on decreasing perceived utility are masked by these participants also having increased intolerance of uncertainty. Future studies should evaluate both depression and intolerance of uncertainty symptoms when studying their effects on information seeking, as one may mask the effect of the other.

Also, in contrast to our expectations, depression was associated with round-wise ratings for all three information utilities (cognitive, hedonic, and instrumental) but did not show a specific or enhanced relationship to hedonic or instrumental utility (as there was no significant interaction effect). The relationship between depression and information utility was not sufficiently informative to survive penalization for model complexity using BIC as a fitness criterion but does survive model selection by AIC (see Supplementary Table 9). Overall, these results are consistent with the notion that depression is associated with a broad attenuation of value perception, tempering not only the emotional and goal-driven value of information, but also its perceived ability to help make sense of the world. However, to the extent that such an effect exists, it is likely relatively small and may be overshadowed by more prominent effects on expected information utility.

While we found no relationship between anxiety and expected information utility, our exploratory analyses indicated that the closely related construct of intolerance of uncertainty is a robust predictor of both round-wise and overall expected utility. Intolerance of uncertainty was first coined in the literature as part of an effort to understand why people worry – a phenomenon deeply intertwined with, but phenomenologically distinct from, anxiety itself (Freeston et al., 1994). Whereas anxiety purportedly incorporates both worry and negative affect, intolerance of uncertainty, instead focuses on how individuals relate emotionally, cognitively, and behaviorally to situations that are ambiguous, uncertain, or out of their control. Thus, the positive relationship between intolerance of uncertainty and information utility ratings coupled with the null relationship with anxiety, suggests that it is not negative affect, but rather an individuals’ relationship to uncertainty that affects how much they value information.

Interestingly, one might expect that intolerance of uncertainty would be associated only with cognitive utility, as conceptualized in our study, because it deals most directly with the potential for information to reduce feelings of uncertainty. Instead, intolerance of uncertainty was associated with all three types of information utility, suggesting that individuals who are more uncomfortable with uncertainty perceive all types of information as more valuable. This finding adds important context to previous work examining the relationship between negative emotions and information-seeking tendencies: while information seeking can be positively associated with negative feeling states like anger⁵³, fear, or anxiety^23,53, our work suggests that the “distress” component of this response may not be driving this relationship. Instead, related cognitive schemas (e.g., attitudes about uncertainty) may offer more robust and powerful predictions about how people perceive information utility.

Trust in science was associated with greater information seeking and engagement in self-reported COVID-19 preventive behaviors

Given the proliferation of pandemic-related misinformation, understanding how a participant’s attitude towards scientific work relates to their health-related information-seeking and behaviors is particularly important. Our results suggest that participants with greater trust in science report higher levels of engagement in COVID-19 preventive behaviors and may engage in more information-seeking, although this latter effect was only marginally significant.

Prior exposure to COVID-19 health information from scientific sources could explain the positive correlation between trust in science and engagement in COVID-19 preventive behaviors, as many scientific sources provided or directed individuals towards recommended practices for COVID-19 prevention (e.g., CDC). Moreover, individuals with higher trust in science may be more likely to trust the institutions publishing COVID-19 prevention guidelines and therefore be more likely to engage in these behaviors after learning about them. Further research is needed to identify whether a causal relationship between trust in science and engagement in COVID-19 preventive behaviors, and other preventive behaviors, exists. Additionally, future work may seek to understand the extent to which the relationship between trust in science and engagement in recommended preventive behaviors is mediated by increased health information exposure versus increased engagement intention post-exposure. Although future work will be necessary to explore the nature of the relationship between trust in science and engagement in recommended preventive behaviors, our work implies that fostering trust in science is a promising start towards the goal of increasing the prevalence of healthy pandemic-related behaviors in the population.

Limitations, constraints on generality, and future directions

Our work offers insights into pandemic-related information seeking and the ways in which situations may shape the relationship between information value and information-seeking behaviors. However, the present work has several limitations. One of these limitations is the restricted demographic characteristics of our sample. Our sample included only participants from the university student population – specifically students participating in introductory psychology courses. This is obviously not representative of the demographics of the United States and likely produced biases in certain individual differences and demographic characteristics relevant to this work (e.g., trust in science, political affiliation, age, education). For example, many of our participants may have been majoring in psychology and related life-sciences programs (e.g., pre-medical sciences, neuroscience) – such an educational background may predispose these individuals towards certain information seeking strategies or tendencies.

The generalizability of our findings is limited by this demographic feature of our sample – it would be particularly interesting to examine the same behaviors in a sample predisposed to greater distrust in science or with less direct exposure to the perspectives of experts in the life sciences. It is possible that new relationships between our variables of interest would emerge in samples wherein the individual differences are more representative of the general population, or biased in a different direction. In particular, trust in science and engagement in COVID-19 preventive behaviors may have been inflated in our sample relative to the general population given that all students in the sample were enrolled in at least one psychology course and all students were likely aware of the university’s policies regarding COVID-19 safety. Given that trust in science and engagement in COVID-19 preventive behaviors predicted information-seeking despite this potential problem of restricted variance, these characteristics may be particularly important to examine in more representative populations.

Future research may also benefit from narrowing the demographic scope to focus on individuals with a history of mental health diagnosis or treatment. Our study did not include any data regarding mental health diagnoses or treatment (e.g., prescription medication or psychotherapeutic history), nor did we screen out participants based on their mental health symptoms or current and past treatment plans. Our sample thus includes a wide range of anxiety and depression scores, but cannot answer whether the relationship between clinical symptoms and COVID-19 information seeking changed depending on whether a participant was engaged in psychotherapeutic and/or psychopharmacological treatment plans at the time of measurement. Interestingly, the mean depression and anxiety scores in our samples far exceeded those from a norming sample of healthy control participants³⁵. This may represent a deleterious effect of the pandemic itself, but could also reflect the fact that we did not screen out participants depending on their mental health symptoms. By focusing on more specific clinical samples and accounting for individuals’ diagnoses, treatment histories, and current treatment regimens, future work may be able to offer stronger insights into the relationship between clinical symptoms and information-seeking behaviors.

Another potential limitation is our sample size. Although our sample size was decided a-priori with a power analysis, and it should give us 80% power to detect a medium effect size correlation (r = .2), it is possible that the correlations between affective characteristics and information seeking or utility are smaller, in which case our study would be underpowered. Thus, our study does suggest that, if these correlations do exist, they are likely small (r < .2).

Additionally, while our study design afforded us a great deal of control over the information content available to participants and allowed us to examine information seeking behaviors directly rather than rely on self-reports, these benefits came at a cost. First, unlike information seeking in the real world, where information options are functionally limitless, logistical demands (e.g., session times) forced us to restrict the number of topics available for participants to five stories. In a more option-rich environment, there may be more variance in information seeking options that can be explained by individual differences. Future work may examine how predictors of information-seeking are influenced by the number and kinds of information made available to participants by separately varying the maximum number of searches a participant can engage in and the number of available information topics and sources.

Additionally, when creating a set of information stimuli, other researchers might consider conducting an independent study to establish standard ratings of each utility type for each stimulus prior to collecting information seeking data. In addition to affording a sort of “ground truth” comparison, this approach allows the experimenter to sub-select information seeking stimuli that efficiently differentiate between different types of information utility from the broader stimulus set. Such an orthogonalized stimulus set would reduce concerns about multicollinearity during regression model fitting, which may be especially desirable for large stimulus sets.

Furthermore, the information sources and topics in our study were carefully chosen to reflect what we believed to be accurate, responsible reporting about the pandemic. However, information seeking in the real world requires participants to distinguish between information and misinformation. Given the “misinformation pandemic” amidst the pandemic⁵⁴, understanding how particular profiles of information utility ratings and weightings and other individual differences may be related to the ability to distinguish information from misinformation/disinformation is a particularly important line of research.

Importantly, if individuals are driven to seek a particular type of information utility, they may become more susceptible to misinformation that projects this utility and/or less amenable to true information that they perceive may reduce that utility. For example, when information seeking appears to be driven primarily by cognitive utility, as it was in our sample, individuals may be more prone to seek information that is familiar to them and consistent with their own preconceived beliefs and less likely to seek information that introduces new perspectives or challenges their beliefs. Such a bias could contribute to or enhance cognitive distortions during information processing in favor of “a priori” preferences, an effect that previous work has shown to be driven by the goal to achieve consistency (i.e., reconcile past and present information⁵⁵). Thus, future work may aim to examine how preference for and use of particular types of information utility may be related to both information seeking behaviors and information processing tendencies as a means of understanding the conditions under which different people are particularly vulnerable to misinformation.

Relatedly, our design featured all-or-nothing information-seeking decisions which allowed us to fit a relatively simple logistic regression model to predict information-seeking decisions, but real information seeking is often less black-and-white. The information seeking individuals perform in the real world involves more dimensions – like the breadth and depth of the search (Anker et al., 2011). An individual may be able to reject one information source based on its title alone, skim another before rejecting it, or examine the entire item before making a decision about whether to integrate it into their understanding of the topic.

Important individual differences in information-seeking may emerge at levels beyond binary seek/no-seek decisions or overall proportion of information-seeking. For example, even if participants with high levels of health anxiety tend to engage in a larger number of information-seeking rounds, their searches may be “shallower” compared to non-anxious participants (e.g., less time per page, smaller proportion of overall information consumed²⁶). A full picture of the relationship between any given trait and an individual’s information-seeking behaviors should examine not only the quantity of information-seeking, but its quality. Likewise, interesting relationships between expected information utility, individual differences, and information-seeking may only arise at certain levels of analysis. Thus, binary information-seeking decisions, though important and relatively easy to study, are not the only outcomes of interest in information-seeking and future work on predictors of information-seeking behaviors should consider both the quality and quantity of information-seeking.

One further limitation of our study is the absence of an objective measure of information utility. While subjective ratings provided valuable insight into participants’ perceived relevance and usefulness of COVID-19 information, the lack of an objective benchmark limits the scope of our conclusions. There is a rich literature that relies exclusively on information seeking when the objective utility is known^56,57,58,59. However, more recent work has found that a model incorporating subjective utilities can better explain information-seeking behavior than a model based solely on objective utilities¹⁵. Unfortunately, defining objective utility in the context of COVID-19 information is challenging due to the evolving nature of the pandemic and the variability in informational accuracy, expert consensus, and policy relevance over time. While the present study was not able to overcome these barriers, we encourage future research to consider how to develop paradigms that can account for objective information utility in a variety of applied contexts - preferably in ways that minimize disruption to the balance between intrinsic and extrinsic motivation that characterizes real-world behaviors.

Conclusions

We examined COVID-19 information-seeking behaviors in a sample of 190 university students in the Midwest United States during Fall of 2020. Our results revealed that the decision to seek out information on a given round was predicted by the expected capacity of the information to reduce feelings of uncertainty about COVID-19 (i.e., cognitive utility) and by self-reported engagement in COVID-19 preventive behaviors (e.g., hand washing, social distancing). Notably affective factors, including anxiety, depression, and expected hedonic value of the information failed to predict information seeking. Additionally, exploratory analyses indicated that individuals with higher levels of intolerance of uncertainty tended to perceive information as more valuable, for all types of utility. Together, this suggests that COVID-19 information seeking in our sample was driven largely by a desire to reduce uncertainty. These factors were more important in predicting information seeking than the expected emotional or instrumental value of information, implying that individuals prioritized information that would reduce uncertainty above and beyond information that could provide comfort or actionable suggestions.

Furthermore, individuals with higher trust in science reported higher engagement in COVID-19 preventive behaviors and sought out more information about COVID-19 overall, positioning trust in science as a variable of interest in understanding and increasing COVID-19 information literacy and guideline compliance. Altogether, our findings suggest that information that can help make sense of the world and reduce feelings of uncertainty may be particularly attractive during a global pandemic, and that, in terms of individual differences, trust in science and previous engagement in preventive behaviors may predict whether individuals seek out information.