Abstract
To adjust to stressful environments, people seek information. Here, we show that in response to stressful public and private events the high-level features of information people seek online alter, reflecting their motives for seeking knowledge. We first show that when people want information to guide action they selectively ask “How” questions. Next, we reveal that “How” searches submitted to Google increased dramatically during the pandemic (controlling for search volume). Strikingly, the proportion of these searches predicted weekly self-reported stress of ~ 17K individuals. To rule out third factors we manipulate stress and find that “How” searches increase in response to stressful, personal, events. The findings suggest that under stress people ask questions to guide action, and mental state is reflected in features that tap into why people seek information rather than the topics they search for. Tracking such features may provide clues regrading population stress levels.
Similar content being viewed by others
Explore related subjects
Discover the latest articles, news and stories from top researchers in related subjects.Introduction
Every person will face unexpected adversities during their lifetime. These stressful events may be global (e.g., war, pandemic) or unique to the individual (e.g., being diagnosed with cancer, losing one’s job, divorce). Such events often lead to stress, anxiety, confusion, and a reduced sense of control1,2,3,4,5,6.
One available adaptive reaction is to seek information that can help guide action to promote adaptation7,8,9. Such actions can be directly related to the event experienced (e.g., during wartime people may search for information on how to secure windows from being shuttered by rockets) or indirectly related (e.g., searching for information on activities that can distract oneself from the adversity).
To date, research on the relationship between information-seeking and stress has mostly focused on the frequency of information-seeking. Some studies propose that stress is associated with greater information seeking10,11,12, which may decrease the sense of uncertainty that is heightened under stress. Others, however, suggest that stress leads to an avoidance reaction which is characterised by less information-seeking about the stressor13,14.
Here, we take a different theoretical viewpoint. Rather than focusing on whether stress generally enhances or reduces information seeking, we test the hypothesis that when experiencing abrupt stressful life events people are more likely to search for information that can direct action—a reaction which may be adaptive. People may seek information that can direct action to deal with the stressful event directly (e.g., during a pandemic a person may ask “how do I make a mask?”) and/or information to direct action that distracts from the stressor (e.g., “how do I make a Margarita?”). In other words, stress may alter the type, rather than frequency, of information people seek. Across multiple studies, we test whether such changes can be detected and quantified by simple analysis of web searches. We propose that by examining the features of the information that people seek, we can gain insight into their external state and their internal reaction to that state.
To that end, we also examine whether the negative context people find themselves in is associated with a change in the valence of web searches. That is, whether a negative state may lead to more negatively valanced searches due to that state, or alternatively to more positive searches in an attempt perhaps to counter the negative state. The former possibility is supported by studies showing that anxious individuals have a bias towards negative stimuli15,16,17, which suggests that anxiety may increase the search for negative information.
We test our proposal both in the context of a global stressful event (i.e., the COVID-19 pandemic) and of personal aversive life events. Our approach differs dramatically from past attempts to relate web searches to mental state. Current research on web searches has focused on relating web searches for specific content terms (e.g., “suicide, “anxiety”, “Prozac”) with mental health indicators of a population, a method that has resulted in mixed findings18,19,20,21,22,23,24,25,26,27,28,29,30,31. In contrast, we assess whether changes in the high-level features of searches reflect stress levels. This approach, which is based on a recent theory of information-seeking motives8, may be more sensitive as it does not make an assumption about which topics people are searching for, but rather the characteristic of information they are searching for (i.e., information that may guide action).
First, we conducted a control study to identify which question-words people would use when seeking information to guide action (Study 1). The results clearly show that participants selectively use “How” for this purpose. Next in Study 2, we calculated the percentage of Google searches containing the “How” question-word submitted in the UK and US out of all searches submitted in that region, every week for a year from the date a “National Emergency” was declared and compared to the years before. Changes to the proportion of “How” searches cannot be explained by changes in the volume of Google searches during the pandemic, as we examined the change in the percentage of “How” searches out of all searches at that time. We also quantified the valence of the most frequent questions submitted to the Google search engine each week in the UK and US using a machine learning approach (HuggingFace, 2022). We then examined how these features related to weekly stress reports of approximately 17K individuals in the UK. Importantly, we dissociate the effects of stress on information-seeking from the effect of COVID-19 related confinement. Together, these analyses enable us to examine how features of web searches alter under a global stressor.
To be able to generalize our findings to private stressful events and to rule out potential third factors, we then tested our hypotheses in a controlled environment in Study 3. Here, we manipulate stress levels and examined if the manipulated stress impacts the likelihood of asking “How” questions in relation to private events.
Results
Study 1
“How” web searches are selectively associated with the need to guide action
To determine which question-words are associated with guiding action, we asked 100 participants (Age = 39.29 (SD = 13.86), Females = 56%, Males = 44%, Other = 0%) to think about a goal they were trying to achieve. They were then instructed to select from a list of eight question words (“What”, “Which”, “Who”, “Where”, “Why”, “When”, “Whose”, and “How”), the word that they would use in a Google query to help guide their actions to achieve the goal (i.e., test question). Participants were also asked to think about a topic they are interested in learning more about. They were then instructed to select a question-word from the same list of eight words, that they would use in Google query to increase their knowledge about the topic (i.e., control question). This design allowed us to identify which words people use to seek information to guide action and test whether such words were used generally for information-seeking, or more specifically to guide action.
The likelihood of selecting the word “How” to ask a question to guide action was equal to 87%, which was significantly greater than the likelihood of selecting “How” to increase understanding, which was 29%. (p = 0.009, Fisher exact test see Fig. 1). Moreover, the likelihood of selecting “How” to guide action was significantly greater than the likelihood of selecting all other question words put together (p < 0.001, Fisher exact test; see Fig. 1), while for increasing knowledge, the likelihood of selecting “How” was significantly less than the likelihood of selecting all other question words put together (p < 0.001, Fisher exact test; see Fig. 1).
Given the results of Study 1, we used “How” questions in Study 2 and 3 as a proxy for the desire to gain knowledge that could help guide action. Note that we are not suggesting that all questions that begin with “How” are intended to guide action. Rather that, on average, if people want to ask a question to guide action, they will be likely to use “How”. In contrast, they are not especially likely to use “How” simply to learn more about a topic.
Study 2
The pandemic resulted in a significant change to the high-level features of web searches
To assess the high-level feature of web searches associated with guiding action, motivated from Study 1, we extracted the Google Search Volume Index of “How” questions. A Google Search Volume Index is equal to the number of searches for the specific term of interest in a given week and region divided by the total number of searches for that same week and region. These percentages are then normalized to represent search interest relative to the highest percent for that region for the entire time frame (i.e., January 1st, 2017–March 21st, 2021; see Methods for details). Note, that weekly changes to the Google Search Volume Index cannot be explained by weekly changes in the total volume of Google searches, as the index reflects the percent of specific searches out of all searches that week (i.e., on a scale from 0 to 100). We also calculated a second feature—a Valence Index—which indicates the valence of the most frequent questions submitted to Google search engine. To calculate this index, we first extracted the most frequent web searches each week that included question-words (i.e., “What” “Which”, “Who”, “Where”, “Why”, “When”, “Whose”, and “How”). Next, we implemented a machine learning approach to assess the valence of searches, by applying the pre-trained model distilbert-base-uncased-finetuned-sst-2-english (HuggingFace, 2022) to the extracted data. The model’s output contains a label: positive and negative, along with a score between 1 and 0 (1 = absolute confidence in the model output label, 0 = zero confidence in the model output label). As the sum of both values is equal to 1, we used only the positive confidence score as our measurement of valence and then transformed this number to be on a scale from 0 to 100 (i.e., a score of 100 denotes the most positively valenced score and a score of 0 the most negatively valenced score), such that it would be easily comparable to the Google Search Volume Index for “How”, described above. We validated the algorithm score in this context by asking a naïve human subject to categorize 192 randomly sampled searches as either ‘more positive’ or ‘more negative’. We found a significant positive association between the algorithmic score and the human score (r(190) = 0.463, p < 0.001). Note, that this correlation is only of medium size, which may be because when humans rate the valence of a question they consider the predicted answer while the algorithm does not.
Analyses were conducted separately in the UK and the US. In each country we quantified the measures above for every week from the date the “National Emergency” was declared (UK: March 23rd, 2020; and US: March 13th, 2020) through March 21st, 2021, as well as every week from January 1st, 2017 until “National Emergency” was declared. There was a significant increase in the Google Search Volume Index of “How” questions following the declaration of a “National Emergency” relative to the three years previous (“How” UK: before “National Emergency” declared: M = 65.37, SD = 3.27, after “National Emergency” declared: M = 84.19, SD = 8.52, t(55.71) = − 15.585, p < 0.001, Cohen’s d = − 3.754; “How” US: before “National Emergency” declared: M = 71.65, SD = 2.87, after “National Emergency” declared: M = 85.62, SD = 4.90, t(63.67) = − 19.682, p < 0.001, Cohen’s d = − 4.029; Fig. 2a–d).
Moreover, the Valence Index for question queries after the “National Emergency” was declared was found to be more negative compared to before (Valence Index UK: before “National Emergency” declared: M = 59.37, SD = 16.71, after the “National Emergency” declared: M = 49.77, SD = 13.26, t(218) = 3.788, p < 0.001, Cohen’s d = 0.601; Valence Index US: before “National Emergency” declared: M = 64.09, SD = 14.52, after “National Emergency” declared: M = 57.07, SD = 17.41, t(218) = 2.917, p = 0.004, Cohen’s d = 0.460; Fig. 2e–h).
Population stress-levels are selectively associated with asking “How”
Thus far, we have shown that there is an increase in the proportion of “How” and negatively valenced searches submitted to the Google search engine during the pandemic relative to before. We next examined whether these were related to population stress levels. We had access to self-report stress levels collected every week in the UK between March 21st, 2020 and March 21st, 2021. Approximately 70K unique individuals completed the survey, on average 17,468 individuals a week in the UK32. Specifically, participants were asked to indicate if over the previous week they felt worried and/or stressed about any of the following factors: (i) catching COVID-19, (ii) becoming seriously ill from COVID-19, (iii) finance, (iv) unemployment and (v) getting food. We computed the mean proportion of individuals who reported stress or worry over these factors. We conducted separate linear models, each predicting on a weekly basis either the Google Search Volume Index in the UK of “How” searches or the Valence Index, from stress and from UK COVID-19 confinement scores. The inclusion of the latter enabled us to disentangle the effects of stress on web searches from the effects of confinement due to restrictions placed by the Government. COVID-19 related confinement data for each week in the UK was obtained from the Oxford University COVID-19 Government Response Tracker33. The data includes ordinal variables coded by severity/intensity of confinement, on a daily basis (from January 1st, 2020 to March 21st, 2021), due to the following: (i) school and university closures, (ii) workplace closures, (iii) public event cancelations, (iv) restrictions on gatherings, (v) public transport restrictions, (vi) stay at home requirements, (vii) restrictions on domestic travel, and (viii) restrictions on international travel; see Table 4 for coding. To obtain weekly values, we computed weekly averages of the daily ratings, where all eight variables were normalised to range between 0 and 1 and averaged together.
Importantly, to account for simple temporal trends we removed linear trend34,35 from the dependent variables and predictor variables (stress scores and COVID-19 related confinement), using the detrend function in the ‘pracma’ R package. The detrended dependent and predictor variables were then Z-scored before being entered in the linear models.
The linear model predicting “How” questions from stress levels and COVID-19 related confinement scores, revealed that both high stress (β = 0.182 ± 0.074 (SE), t(49) = 2.464, p = 0.017) and greater COVID-19 related confinement (β = 0.797 ± 0.074 (SE), t(49) = 10.812, p < 0.001; Fig. 3a) predicted proportion of “How” searches. In other words, the relationship between stress levels and “How” searches cannot be solely explained by increased restrictions during the pandemic, as our model controls for COVID-19 related confinement.
We then conducted the same linear model as above, but this time predicting the Valence Index. The Valence of searches was not predicted by either variables (stress: β = − 0.144 ± 0.147 (SE), t(49) = − 0.980, p = 0.332, COVID-19 related confinement: β = 0.216 ± 0.147 (SE), t(49) = 1.470, p = 0.148; Fig. 3b).
Next, to examine if the relationship between stress and “How” searches was specific, or rather reflected a general tendency to ask more questions when stressed, we conducted a third linear model relating question asking to stress levels controlling for confinement. In particular, we extracted the Google Search Volume Index for all other common question-words (i.e., What, Which, Who, Where, Why, When, and Whose; Z-scored and detrended) and then averaged these together. We then predicted the average Google Search Volume Index of all other common question-words from stress scores and COVID-19 related confinement. Importantly, the proportion of other common questions asked was selectively predicted by COVID-19 related confinement (β = 0.701 ± 0.104 (SE), t(49) = 6.744, p < 0.0001), but not stress levels (β = 0.067 ± 0.104 (SE), t(49) = 0.645, p = 0.522; Fig. 3c).
Thus far, we have shown that the relative volume of searches that can direct action is related to stress levels. Next, we wanted to test the predictive validity of this simple model. Specifically, we used our stress score to predict the proportion of “How” searches using a leave one out analysis. To account for a simple temporal trend, we first removed the linear trend from the dependent variable (“How” questions) and the predictor variable (stress levels). The detrended predictor variables were then Z-scored before being entered in the simple linear model. The simple model was then run on all the data saved for one time point which was held out from the analysis. We then used the regression beta to predict the proportion of “How” searches of the left-out time point. This process was repeated so that each week’s proportion of “How” searches was estimated from the simple model parameters generated without using that week to fit the data. The actual proportion of “How” searches of a week (data) and the predicted proportion of “How” searches (estimation) were then correlated and also compared using a paired sample t-test (see Methods for details). We observed a correlation between the predicted proportion of “How” searches (estimate) and the actual proportion of “How” searches (data) (r(50) = 0.366, p = 0.007). The means of the two sets of values were not significantly different from one another (t = − 0.041, p = 0.967). This analysis suggests that stress levels in a population is a good predictor of the proportion of “How” searches during the pandemic.
Next, we tested whether stress was better predicted by “How” Google searches than searches for specific content terms (i.e., “stress”, “anxiety”, “mental health” and “psychiatrist”), which are often used in attempt to predict population mental state. Thus, we ran multiple linear models to predict stress from each term separately. Once again, the dependent and predictor variables were first detrended and then Z-scored. The strongest association was seen by “How” question volume and self-reported stress scores (β = 0.439 ± 0.127 (SE), t(50) = 3.435, p = 0.001, R2 = 0.437), followed by the Google Search Volume Index “stress” (β = 0.322 ± 0.134 (SE), t(50) = 2.403, p = 0.020, R2 = 0.322) and “psychiatrist” (β = − 0.348 ± 0.131 (SE), t(50) = − 2.887, p = 0.006, R2 = − 0.378); all other predictors were not significant (p’s > = 0.594).
We repeated the above analysis using the average Google Search Volume Index for “How to” and “How do,” which may be an even more explicit measure of action-related searches. Our findings remained the same (see Supplementary Analysis).
In the US, we did not have access to measurements of population stress levels. However, we did have access to COVID-19 related confinement data which enabled us to examine the relationship between web searches and residential confinement in the US, when stress levels are not controlled for. We observed that increased COVID-19 related confinement was related to greater Google Search Volume Index “How” (β = 0.384 ± 0.129 (SE), t(50) = 2.940, p = 0.005, R2 = 0.384) and to more negatively valenced searches (β = − 0.316 ± 0.134 (SE), t(50) = − 2.354, p = 0.023, R2 = − 0.316). We did not observe a relationship between COVID-19 related confinement and the average Google Search Volume Index of the other question-words (β = − 0.142 ± 0.140 (SE), t(50) = − 1.015, p = 0.315, R2 = − 0.142). Note, all results presented above remain when not detrending.
In addition, we conducted a content analysis to explore the topics of commonly queried “How” searches before and after the declaration of a “National Emergency” in the UK and US, using the LIWC lexicon. We observed that queries regarding work and health were more frequent after a “National Emergency” was declared relative to before in the UK and US (see Supplementary Analysis).
Study 3
Personal stressful events are associated with an increase in “How” and negative valenced searches
To assess whether stressful events influence the propensity to ask “How” and the valence of questions in other situations, we ran a third study. First, we asked participants to recall in detail, and write about, either a stressful past event (e.g., “I had a deadline at work and didn't know if I was going to meet it.”; stress condition; n = 99, age = 39.45, SD = 14.74; Females = 78.8%, Males = 19.2%, Other = 2.0%) or a relaxing and happy past event (e.g., “My holiday in [retracted] with my aunt, cousin and her children. The weather was great, nice and warm. We were staying at a resort on the beach”); control condition; n = 93, age = 38.46, SD = 11.91; Females = 76.3%, Males = 22.6%, Other = 1.1%). Participants reported their stress level on a scale ranging from very calm (− 50) to very stressed (+ 50) before and after recalling the event (see Methods for details).
A 2 (condition: stress, control) by 2 (time: pre-induction, post induction) ANOVA on self-reported stress revealed a significant interaction (F(1, 190) = 42.074, p < 0.001, partial eta squared = 0.181). Follow up t-tests revealed that the interaction was characterised by participants in the stress condition reporting higher stress post induction (M = − 3.59, SD = 24.30) compared to pre-induction (M = − 12.57, SD = 24.94, t(98) = 5.158, p < 0.001, Cohen’s d = 0.518), in contrast, in the control condition, participants mood significantly increased post induction compared to pre-induction (Post induction: M = − 18.80, SD = 27.34, pre-induction: M = − 13.38, SD = 27.79, t(92) = − 4.028, p < 0.001, Cohen’s d = − 0.418). Importantly, participants in the stress condition reported greater stress post induction relative to controls (Stress condition: M = − 3.59, SD = 24.30; Control condition: M = − 18.80, SD = 27.34, t(190) = 2.581, p < 0.001, Cohen’s d = 0.589), with no difference pre-induction (Stress condition: M = − 12.57, SD = 24.94; Control condition: M = − 13.38, SD = 27.79, t(190) = 1.712, p = 0.832, Cohen’s d = 0.031; see Fig. 4a).
Next, participants were asked to enter two searches they could have entered to Google during the time. Our hypothesis is that a stressful event leads to “How” searches as an adaptive mechanism to deal with the event, either by acquiring information related to the event directly and/or as a distraction. For each participant, we counted the number of questions that began with “How” (i.e., 0, 1, 2) and also assessed the average valence of participants’ two searches, by applying the pre-trained model distilbert-base-uncased-finetuned-sst-2-english36. Results show that participants in the stress condition asked significantly more “How” questions (M = 0.82, SD = 0.77) than those in the control condition (M = 0.16, SD = 0.28, t(124.58) = 94.154, p < 0.001; Cohen’s d = 1.110; see Fig. 4b). In addition, participants in the stress condition also submitted significantly more negative searches (M = 40.55, SD = 36.51) than those in the control condition (M = 53.97, SD = 36.11, t(190) = − 2.558, p = 0.011; Cohen’s d = − 0.369; see Fig. 4c). This study strengthens the conclusion of Study 2, that stress inducing events are associated with asking more “How” questions and negatively valenced question by generalising the finding of Study 2 to other stressful events and providing support for the conclusion in a controlled setting. As shown in Table 1, there was no increase in the use of other question words under stress relative to control.
To examine if the increased proportion of asking “How” questions in the stress condition was associated specifically with stress rather than with the novelty of the stressful events, we assessed the novelty of events recalled by participants and repeated the analysis above controlling for novelty scores. All results remain the same (see Supplementary Analysis).
In addition, we assessed the semantic similarity between participants’ queries and their recalled events. The analysis revealed a moderate to high level of similarity (see Supplementary Analysis), which suggests that some of the participants questions shared semantic similarity with their recalled event.
Discussion
The global pandemic generated a new set of practical and mental challenges. To overcome some of these challenges, people turned to technology. On average, people spent almost 7-h a day online in 2020, up 7.3% from the previous year37. A large fraction of this time was dedicated to searching for and consuming information37. Here, we examined how the high-level features of searches submitted to Google changed in response to the pandemic and private aversive life events and how such changes relate to stress.
In particular, we were interested if under stress people would seek more information that could guide their actions—a behaviour that could facilitate the process of adapting to a stressful event (according to the APA a stressful event can be an occurrence or circumstance that individuals perceive as threatening, challenging, or demanding38). Study 1 revealed that participants specifically use the word “How” when they want information to guide their action. Thus, in Study 2 we examined the frequency of Google searches that included the word “How”. We found that both in the UK and US, the proportion of searches that included the word “How” were greater during the year following the declaration of “National Emergency” than in the years prior. It is important to emphasize that any change in the proportion of “How” searches cannot simply be explained by a general increase in number of Google searches, as the former are calculated as proportion of the latter. Neither can it be explained by temporal linear trends, as the data was detrended. The rise in “How” searches may reflect an adaptive human tendency to ask questions that can facilitate rapid adjustment to new and potentially aversive environments. This interpretation is in accord with a recent study suggesting that changes to search during the pandemic reflect a change in people’s needs31. We were also interested in changes to the valence of Google searches during the pandemic relative to before. We observed that the most popular searches in the UK and US were more negatively valenced during the year following the declaration of “National Emergency” than in the years prior. This aligns with the notion that people may search for information that aligns with their emotional state.
Weekly fluctuations in the proportion of “How” questions submitted to Google during the pandemic was positively associated with weekly fluctuations in the proportion of individuals who reported experiencing COVID-related stress in the UK in a sample of over 17K residences. This association could not be attributed to COVID-related confinement, as this factor was controlled for in the model. Furthermore, the relationship was specific to “How” searches, and did not generalize to general question asking.
Markedly, we show that the frequency of “How” searches predicted COVID-related stress better than the frequency of searches that include stress related content (i.e., “anxiety”, “stress”, “mental health”, and “psychiatrist”).
Consistent with the above, in Study 3 we show that people are also more likely to ask “How” questions, and more negatively valenced questions, in response to a personal stressful life event compared to a non-stressful event. In particular, participants were instructed to recall in detail a stressful life event or a calming event. The former, but not the latter, increased self-reported stress. They were then asked which questions they could have submitted to a search engine at the time. Recalling stressful events was associated with significantly more “How” questions and negative questions than control events. This suggests that the relationship between asking “How” and stress is not specific to the pandemic or other national events such as national elections but extends to stressful events in general.
This raises the novel idea that tracking the frequency of “How” searches may provide clues regarding population-level stress beyond the time of a pandemic and perhaps provide clues regarding an individual’s stress level. Of course, while we show (in Study 3) that stress causes a rise in “How” questions, a rise in “How” questions does not necessitate a rise in stress, as there may be other reasons for why people ask more “How” questions. Thus, tracking “How” queries cannot on its own be an indicator of stress but may provide an indicator of possible rising stress levels which can be then followed up using other indicators.
Our investigation was guided by previously identified factors that motivate information-seeking (for review see8). In particular, studies show that people seek information more when it is useful in guiding action9,39,40,41. The current results suggest that this motive is ‘up weighted’ when experiencing stress, perhaps because the need to select adaptive actions is heightened under such circumstances. While past studies also show people prefer to seek good news over bad8,39,41,42,43,44,45,46,47,48 we find that under stress web searches were in fact more negative. This could simply be due to a significant proportion of searches relating to the stress itself (i.e., illness in Study 2 and a range of personal aversive events in Study 3). We did not, however, observe a significant association between valence of web searches and self-reported stress levels. It is interesting to note that studies examining information-sharing (e.g., tweets) rather than information-seeking have revealed small but significant associations between mental state and the valence of information shared49,50,51. Because a significantly smaller slice of the population regularly shares than seeks information, understanding how information-seeking relates to mental health is crucial and may diverge from patterns observed for information sharing.
Reflecting on why individuals tend to ask action-oriented questions during stressful times is crucial. Our theory posits that asking “How” questions could offer a sense of control, potentially mitigating stress by empowering individuals with knowledge and actionable strategies. This actionable information may be directly relevant to dealing with the stressor (e.g., if a hurricane is coming one may ask “how do I duck-tape windows?”) or aimed at actions that can distract and reduce stress (e.g., if a hurricane is coming one may ask “how do you play scrabble?”). This hypothesis aligns with the notion that acquiring control, or even the perception of it, can be a key factor in stress alleviation.
Together, the findings show that in the face of high stress people search for information that can help guide action. While in the past such information may have been sought directly from other people, with the development of the internet, individuals are now able to turn to the web for answers. This ability may have contributed to the high resilience and quick adaptation observed in response to the pandemic6,52,53.
Materials and methods
Study 1
Participants
One-hundred participants (Age = 39.29 (SD = 13.86), Females = 56%, Males = 44%, Other = 0%) completed the study on Qualtrics (http://www.qualtrics.com) and were recruited via Prolific’s online recruitment system (http://www.prolific.co). Participants received £7.50 per hour for their participation. For all studies presented in this article, ethical approval has been provided by the UCL Ethics Committee [12467/005 and 3990/003] and all participants have given their informed consent to participate. All methods were performed in accordance with UCL’s guidelines and regulations.
Procedure
Participants were asked to think about a goal they were trying to achieve. They were then instructed to select from a list of eight question-words (“What”, “Which”, “Who”, “Where”, “Why”, “When”, “Whose”, and “How”) a word that they would use to submit a query on Google to help guide their actions to achieve their goal (i.e., the test question). Participants were also asked to think about a topic they are interested in learning more about and select a question-word from the same list to ask a query on Google to increase their knowledge about that topic (i.e., the control question). The order of the two questions were counter balanced, and the order of the question-words were randomized. This design allowed us to identify which words people use to seek information to guide action and test whether such words were used generally for information-seeking and more specifically to guide action.
Analysis
To assess which question-words were associated with guiding action, we first calculated the proportion of people that selected each of the question-words. We then conducted a Fisher’s exact test to examine whether the most prevalent question-word selected (i.e., “How”) was significantly different than the proportion of all other question-words asked together for each question separately (coded ‘0’ if “How” was selected and ‘1’ if any other question-word was selected), and whether the proportion of people that selected “How” for the test question was significantly different than for the control question.
Study 2
Data extraction
Web search data
Weekly search data was extracted from Google Trends (http://www.googletrends.com) for 220 weeks (January 1st, 2017 through March 21st, 2021). This was done separately for the UK and the US. Based on the results of Study 1, to quantify action guidance, we extracted the Google Search Volume Index for the search term “How”. We also extracted the Google Search Volume Index for the search terms “What”, “Which”, “Who”, “Where”, “Why”, “When”, and “Whose” and then averaged them together to quantify general question asking (i.e., control variable). A Google Search Volume Index value is equal to the number of searches for the specific term of interest in a given week and region (for example total number of searches that include the question-word “How” in the UK the first week of 2020) divided by the total number of searches in that same time and region (for example the total number of Google searches submitted in UK the first week of 2020). These values are normalized to represent search interest relative to the highest value for that region for the entire time frame (i.e., January 1st, 2017–March 21st, 2021).
To quantify valence, we extracted the 25 most popular searches for each week and region for “What”, “Which”, “Who”, “Where”, “Why”, “When”, “Whose” and “How”, questions. That is, for each week we extract up to 25 searches per type of question (i.e., ~ 200 total web searches for each week), as this is the maximum Google Trends reports. Next, we implemented a machine learning approach to assess the valence of these searches, by applying the pre-trained model distilbert-base-uncased-finetuned-sst-2-english36 to the extracted data. The models output contains two labels: positive and negative, along with a score between 1 and 0 (1 = absolute confidence in the model output label, 0 = zero confidence in the model output label). The sum of the values is always equal to 1, for example, a potential output could be Positive = 0.9, Negative = 0.1. As the two measures are fully dependent, we used the positive confidence score as our measurement of valence and then transformed this number to be on a scale from 0 to 100, such that it would be easily comparable to the Google Search Volume Index for “How”, described above (i.e., a score of 100 denotes the most positively valenced score and a score of 0 the most negatively valenced score).
Self-reported stress
Data was extracted, with permission, from the UK COVID-19 Social Study32. The study is a panel study of over 70,000 UK citizens which aims to characterize the psychological and social experience of adults living in the UK during the COVID-19 pandemic (see Table 2 for demographics). The study commenced as a weekly survey, with participants receiving an invitation to the next wave of data collection 7 days following their last completion. All participants received up to 2 reminders (24 and 48 h following their initial weekly invitation). The link to their last reminder remained live so they could return to the study a few days later if they chose to. Following week 22 of the study, monthly follow-ups rather than weekly follow-ups were sent. To attain an equal number of responses across time, participants were randomized to receive their monthly invitation on either week 1, 2, 3 or 4 of the month, with subsequent invitations following 28 days after they completed the survey. An average of 17,468 individuals submitted data each week (see Table 3 for response frequency for each week). For full methods and demographics for the sample see http://www.COVIDSocialStudy.org. The UK COVID-19 Social Study was approved by the UCL Research Ethics Committee [12467/005 and 3990/003], and all participants gave written informed consent. All methods were performed in accordance with UCL’s guidelines and regulations.
Participants were asked: “over the past week, have any of the following been worrying you at all, even if only in a minor way?” They were presented with 18 factors that may cause worry (for example internet access, boredom, neighbours) and were to pick any that they were worried about. Five of these factors were a-priori categorized by the authors of the survey32 as ones that have been impacted by COVID. These were (i) catching COVID-19 (ii) becoming seriously ill from COVID-19, (iii) finances, (iv) losing your job/unemployment and (v) getting food. Second, they were asked “have any of these things been causing you significant stress? (e.g., they have been constantly on your mind or have been keeping you awake at night)”. They were presented with the same 18 factors as above and were asked to tick any of those causing significant stress. For each week and factor, Fancourt and colleagues32 calculated the proportion of respondents that ticked that factor either in response to question 1 and/or question 2. Factors i and ii were a-priori combined by Fancourt and colleagues32 to make one factor, leaving us with four factors. For each week the proportion of people ticking 1 and/or 2 were averaged across the four factors to produce one indicator of “stress levels” for that week.
Table 2 shows the demographic of respondents to the UK COVID-19 Social Study. Importantly, data points reported by Fancourt and colleagues32 were weighted using auxiliary weights to the national census and Office for National Statistics (ONS) data. We used these weighted data points in our study. Thus, reported stress levels are representative of the UK population.
COVID-19 confinement score
To measure COVID-19 related confinement, we extracted eight confinement variables from a publicly available dataset (The Oxford University COVID-19 Government Response Tracker33). All variables are ordinal coded by severity/intensity of confinement, on a daily basis (from January 1st, 2020 to March 21st, 2021), for the following: (i) school and university closures, (ii) workplace closures, (iii) public event cancelations, (iv) restrictions on gatherings, (v) public transport restrictions, (vi) stay at home requirements, (vii) restrictions on domestic travel, and (viii) restrictions on international travel; see Table 4 for coding. To obtain weekly values, we computed weekly averages of the daily ratings. To quantify an overall COVID-19 related confinement score, we transformed all variables to range between 0 and 1 using the R function scaler from the R package, bruceR. Finally, we averaged the 8 transformed variables together.
Analysis
Analysis was conducted separately for data from the UK and the US. In each region we quantified action guidance searches (i.e., “How”) and valence of searches every week from the date the “National Emergency” was declared due to the COVID-19 pandemic, (March 23rd, 2020 in the UK and March 13th, 2020 in the US) through March 21st, 2021, as well as every week dating back to January 1st, 2017. We then compared the weekly scores before the “National Emergency” to that after using an independent samples t-test.
To assess whether our measures were related to UK stress levels, we conducted two linear models predicting on a weekly basis Google’s Search Volume Index of “How” questions and the Valence Index of questions submitted to Google in the UK, from UK stress levels. We also included weekly COVID-related confinement scores in the models to disentangle the effects of stress from the effects of confinement due to restrictions placed by the Government. To account for simple temporal trends, we removed the linear trend from the dependent and predictor variables first, using the detrend function in the pracma R package. The detrended dependent and predictor variables were then Z-scored.
Next, to examine whether the relationship between stress and “How” searches was specific or rather reflected a general tendency to ask more questions when stressed we conducted a third linear model relating question asking to stress levels controlling for COVID-related confinement scores. To do so, we extracted the Google Search Volume Index for all other common question-words (i.e., “What”, “Which”, “Who”, “Where”, “Why”, “When”, and “Whose”; Z-scored and detrended) and then averaged them together.
We were then interested in whether stress was better predicted by “How” Google searches than searches for stress related terms. To test this, we first removed the linear trend from the dependent and predictor variables, and then Z-scored the dependent and predictor variables. We then ran a model predicting the proportion of UK sample reporting COVID-related stress from the “How” Google Search Volume Index as well as Google Search Volume Index for the words “stress”, “anxiety”, “mental health”, and “psychiatrist” in the UK. In addition, we ran multiple linear models to predict UK COVID-related stress levels each time from only one of the terms above.
Finally, we tested the predictive validity of a simple model using stress levels to predict the proportion of “How” searches using a leave one out analysis. Once again, we removed the linear trend from the dependent and predictor variables first, and then the dependent and predictor variables were Z-scored. Specifically, the simple model was run on all the data saved for one time point which was held out from the analysis. We then used the regression beta to predict the proportion of “How” searches of the left-out time point. This process was repeated so that each week’s proportion of “How” searches was estimated from the simple model parameters generated without using that week to fit the data. This resulted in two values for the proportion of “How” searches for each week: the actual proportion of “How” searches (data) and the predicted value from the leave-one-out validation (estimate). The actual proportion of “How” searches of a week (data) and the predicted proportion of “How” searches (estimation) were then correlated and compared using a paired sample t-test. This analysis indicates whether the population stress levels is a good predictor of the proportion of “How” searches.
Study 3
Participants
One hundred and ninety-three participants completed the study on Qualtrics (http://www.qualtrics.com) and were recruited via Prolific’s online recruitment (http://www.prolific.co). Participants received £7.50 per hour for their participation. One participant was excluded for not providing a valid response, leaving the final participant N at 192 (stress condition: n = 99, age = 39.45, SD = 14.74; Females = 78.8%, Males = 19.2%, Other—2.0%; control condition: n = 93, age = 38.46, SD = 11.91; Females = 76.3%, Males = 22.6%, Other = 1.1%).
Procedure
Participants were asked to recall a time when they were very stressed (stress condition) or recall a time when they were happy and relaxed (control condition). They were then instructed to think about that time in as much detail as possible and describe it in a text box. They indicate their stress level on a scale ranging from very calm (− 50) to very stressed (+ 50) before and after the induction. Next, participants were asked to enter two searches they could have entered to Google during the time.
Analysis
To assess whether the manipulation was successful, a 2 (condition: stress, control) by 2 (time: pre-induction, post induction) ANOVA was run with follow up pair-wise t-tests. Next, for each participant, we counted the number of questions that began with “How” (i.e., 0, 1, 2) and for every other question word (i.e., “What”, “Which”, “Who”, “Where”, “Why”, “When”, and “Whose”). We then conducted separate independent samples t-tests to assess the difference in the number asked for each question type between conditions. Finally, we implemented a machine learning approach to assess the average valence of participants’ two searches by applying the pre-trained model distilbert-base-uncased-finetuned-sst-2-english36 to their submitted searches. Finally, we examined the difference between participants in the stress and control conditions with regard to the number of “How” searches submitted and the valence of their searches.
Data availability
Data and code are available at a dedicated OSF repository (https://osf.io/edr74/?view_only=f2c80dcafd2447e5a4a06567cfa26b2a). The source data underlying Figs. 1, 2, 3, 4 are provided as part of this repository.
Code availability
Code supporting this study are available at a dedicated OSF repository (https://osf.io/edr74/?view_only=f2c80dcafd2447e5a4a06567cfa26b2a).
References
Finlay-Jones, R. & Brown, G. W. Types of stressful life event and the onset of anxiety and depressive disorders. Psychol. Med. 11(4), 803–815. https://doi.org/10.1017/S0033291700041301 (1981).
Francis, J. L., Moitra, E., Dyck, I. & Keller, M. B. The impact of stressful life events on relapse of generalized anxiety disorder. Depress. Anxiety 29(5), 386–391. https://doi.org/10.1002/da.20919 (2012).
McLaughlin, K. A., Conron, K. J., Koenen, K. C. & Gilman, S. E. Childhood adversity, adult stressful life events, and risk of past-year psychiatric disorder: A test of the stress sensitization hypothesis in a population-based sample of adults. Psychol. Med. 40(10), 1647–1658. https://doi.org/10.1017/S0033291709992121 (2010).
Miloyan, B., Joseph Bienvenu, O., Brilot, B. & Eaton, W. W. Adverse life events and the onset of anxiety disorders. Psychiatry Res. 259, 488–492. https://doi.org/10.1016/j.psychres.2017.11.027 (2018).
Suls, J. & Mullen, B. Life events, perceived control and illness: The role of uncertainty. J. Hum. Stress 7(2), 30–34. https://doi.org/10.1080/0097840X.1981.9936824 (1981).
Globig, L. K., Blain, B. & Sharot, T. Perceptions of personal and public risk: Dissociable effects on behavior and well-being. J. Risk Uncertain. 64, 213–234 (2022).
Hirshleifer, J. & Riley, J. G. The analytics of uncertainty and information—an expository survey. J. Econ. Lit. 17, 1375–1421 (1979).
Sharot, T. & Sunstein, C. R. How people decide what they want to know. Nat. Hum. Behav. https://doi.org/10.1038/s41562-019-0793-1 (2020).
Stigler, G. J. The economics of information. J. Polit. Econ. https://doi.org/10.1086/258464 (1961).
Drouin, M., McDaniel, B. T., Pater, J. & Toscos, T. How parents and their children used social media and technology at the beginning of the COVID-19 pandemic and associations with anxiety. Cyberpsychol. Behav. Soc. Netw. 23(11), 727–736 (2020).
Ebrahim, A. H. et al. COVID-19 information-seeking behavior and anxiety symptoms among parents. OSP J. Health Care Med. 1(1), 1–9 (2020).
Loosen, A. M., Skvortsova, V. & Hauser, T. U. Obsessive–compulsive symptoms and information seeking during the Covid-19 pandemic. Transl. Psychiatry 11(1), 309 (2021).
Kash, K. M. et al. Psychosocial aspects of cancer genetics: Women at high risk for breast and ovarian cancer. Semin. Surg. Oncol. 18(4), 333–338 (2000).
Chae, J. Who avoids cancer information? Examining a psychological process leading to cancer information avoidance. J. Health Commun. 21(7), 837–844 (2016).
Bar-Haim, Y., Lamy, D., Pergamin, L., Bakermans-Kranenburg, M. J. & Van Ijzendoorn, M. H. Threat-related attentional bias in anxious and nonanxious individuals: A meta-analytic study. Psychol. Bull. 133(1), 1 (2007).
Cisler, J. M. & Koster, E. H. Mechanisms of attentional biases towards threat in anxiety disorders: An integrative review. Clin. Psychol. Rev. 30(2), 203–216 (2010).
MacLeod, C. & Mathews, A. Anxiety and the allocation of attention to threat. Q. J. Exp. Psychol. 40(4), 653–670 (1988).
Ayers, J. W. et al. Suicide-related internet searches during the early stages of the COVID-19 pandemic in the US. JAMA Netw. Open https://doi.org/10.1001/JAMANETWORKOPEN.2020.34261 (2021).
Gunnell, D., Derges, J., Chang, S. S. & Biddle, L. Searching for suicide methods. Crisis 36(5), 325–331. https://doi.org/10.1027/0227-5910/A000326 (2015).
Sueki, H. Does the volume of Internet searches using suicide-related search terms influence the suicide death rate: Data from 2004 to 2009 in Japan. Psychiatry Clin. Neurosci. 65(4), 392–394. https://doi.org/10.1111/J.1440-1819.2011.02216.X (2011).
Hoerger, M. et al. Impact of the COVID-19 pandemic on mental health: Real-time surveillance using google trends. Psychol. Trauma Theory Res. Pract. Policy 12(6), 1–2. https://doi.org/10.1037/TRA0000872 (2020).
Barros, J. M. et al. The validity of Google Trends search volumes for behavioral forecasting of national suicide rates in Ireland. Int. J. Environ. Res. Public Health https://doi.org/10.3390/IJERPH16173201 (2019).
Ayers, J. W. et al. Novel surveillance of psychological distress during the great recession. J. Affect. Disord. 142(1–3), 323–330. https://doi.org/10.1016/J.JAD.2012.05.005 (2012).
Knipe, D., Evans, H., Marchant, A., Gunnell, D. & John, A. Mapping Population Mental Health Concerns Related to COVID-19 and the Consequences of Physical Distancing: A Google Trends Analysis (Wellcome Open, 2020).
McCarthy, M. J. Internet monitoring of suicide risk in the population. J. Affect. Disord. 122(3), 277. https://doi.org/10.1016/J.JAD.2009.08.015 (2010).
Misiak, B., Szcześniak, D., Koczanowicz, L. & Rymaszewska, J. The COVID-19 outbreak and Google searches: Is it really the time to worry about global mental health?. Brain Behav. Immun. 87, 126. https://doi.org/10.1016/J.BBI.2020.04.083 (2020).
Rana, U. Elderly suicides in India: An emerging concern during COVID-19 pandemic. Int. Psychogeriatr. 32(10), 1251–1252. https://doi.org/10.1017/S1041610220001052 (2020).
Sinyor, M., Spittal, M. J. & Niederkrotenthaler, T. Changes in suicide and resilience-related Google searches during the early stages of the COVID-19 pandemic. Can. J. Psychiatry 65(10), 741–743. https://doi.org/10.1177/0706743720933426 (2020).
Tran, U. S. et al. Low validity of Google Trends for behavioral forecasting of national suicide rates. PLoS One 12(8), e0183149. https://doi.org/10.1371/JOURNAL.PONE.0183149 (2017).
Yang, A. C., Huang, N. E., Peng, C.-K. & Tsai, S.-J. Do Seasons have an influence on the incidence of depression? The use of an internet search engine query data as a proxy of human affect. PLoS One 5(10), e13728. https://doi.org/10.1371/JOURNAL.PONE.0013728 (2010).
Suh, J., Horvitz, E., White, R. W. & Althoff, T. Population-scale study of human needs during the COVID-19 pandemic: Analysis and implications. In Proceedings of the 14th ACM International Conference on Web Search and Data Mining pp. 4–12 (2021).
Fancourt, D., Bu, F., Mak, H. W. & Steptoe, A. COVID-19 social study. Results release, 25 (2021).
Hale, T. et al. A global panel database of pandemic policies (Oxford COVID-19 Government Response Tracker). Nat. Hum. Behav. 5(4), 529–538 (2021).
Lampos, V. et al. Tracking COVID-19 using online search. npj Digit. Med. 4(1), 1–11 (2021).
Bakker, K. M., Martinez-Bakker, M. E., Helm, B. & Stevenson, T. J. Digital epidemiology reveals global childhood disease seasonality and the effects of immunization. Proc. Natl. Acad. Sci. 113(24), 6689–6694 (2016).
HuggingFace. DistilBERT base uncased finetuned SST-2. https://huggingface.co/distilbert-base-uncased-finetuned-sst-2-english. Accessed: 2022-07-13 (2022).
Kemp. S. Digital 2021: Global Overview Report—DataReportal—Global Digital Insights. https://datareportal.com/reports/digital-2021-global-overview-report (2021).
VandenBos, G. R. APA Dictionary of Psychology (American Psychological Association, 2007).
Kelly, C. A. & Sharot, T. Individual differences in information-seeking. Nat. Commun. 12, 7062. https://doi.org/10.1038/s41467-021-27046-5 (2021).
Kobayashi, K., Ravaioli, S., Baranès, A., Woodford, M. & Gottlieb, J. Diverse motives for human curiosity. Nat. Hum. Behav. 3, 587–595 (2019).
Dezza, I. C., Maher, C. & Sharot, T. People adaptively use information to improve their internal states and external outcomes. Cognition 228, 105224 (2022).
Karlsson, N., Loewenstein, G. & Seppi, D. The ostrich effect: Selective attention to information. J. Risk Uncertain. https://doi.org/10.1007/s11166-009-9060-6 (2009).
Golman, R., Hagmann, D. & Loewenstein, G. Information avoidance. J. Econ. Lit. https://doi.org/10.1257/jel.20151245 (2017).
Persoskie, A., Ferrer, R. A. & Klein, W. M. Association of cancer worry and perceived risk with doctor avoidance: An analysis of information avoidance in a nationally representative US sample. J. Behav. Med. 37, 977–987 (2014).
Vellani, V., de Vries, L. P., Gaule, A. & Sharot, T. A selective effect of dopamine on information-seeking. Elife 9, e59152 (2020).
Charpentier, C. J., Bromberg-Martin, E. S. & Sharot, T. Valuation of knowledge and ignorance in mesolimbic reward circuitry. Proc. Natl. Acad. Sci. USA https://doi.org/10.1073/pnas.1800547115 (2018).
Lerman, C. et al. What you don’t know can hurt you: Adverse psychologic effects in members of BRCA1-linked and BRCA2-linked families who decline genetic testing. J. Clin. Oncol. 16, 1650–1654 (1998).
Kobayashi, K. & Hsu, M. Common neural code for reward and information value. Proc. Natl. Acad. Sci. 116(26), 13061–13066 (2019).
De Choudhury, M., Gamon, M., Counts, S. & Horvitz, E. Predicting depression via social media. In Seventh International AAAI Conference on Weblogs and Social Media (2013).
Kelley, S. W. & Gillan, C. M. Using language in social media posts to study the network dynamics of depression longitudinally. Nat. Commun. 13(1), 1–11 (2022).
Eichstaedt, J. C. et al. Facebook language predicts depression in medical records. Proc. Natl. Acad. Sci. 115(44), 11203–11208 (2018).
Aknin, L. et al. Mental health during the first year of the COVID-19 pandemic: A review and recommendations for moving forward. Prespect. Psychol. Sci. 17, 915–936. https://doi.org/10.31234/OSF.IO/ZW93G (2021).
Daly, M. & Robinson, E. Psychological distress and adaptation to the COVID-19 crisis in the United States. J. Psychiatry Res. 136, 603–609 (2021).
Acknowledgements
We thank Liron Rozenkrantz, Moshe Glickman, Laura Globig, Valentina Vellani, Sarah Zheng, Gaia Molinaro and Christina Maher for comments on previous versions of the manuscript. The work was funded by a Wellcome Trust (No. 179378) Senior Research Fellowship to TS.
Author information
Authors and Affiliations
Contributions
C.K., B.B. and T.S. designed the study. C.K. and B.B. collected and analyzed the data with guidance from T.S. C.K. and T.S. drafted the manuscript. All authors approved the final version of the manuscript for submission.
Corresponding authors
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher's note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary Information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Kelly, C.A., Blain, B. & Sharot, T. “How” web searches change under stress. Sci Rep 14, 15147 (2024). https://doi.org/10.1038/s41598-024-65895-4
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41598-024-65895-4
- Springer Nature Limited