key: cord-0766964-kaharu3x authors: Shahar, Golan; Aharonson-Daniel, Limor; Greenberg, David; Shalev, Hadar; Malone, Patrick S; Tendler, Avichai; Grotto, Itamar; Davidovitch, Nadav title: Changes in General and Virus-Specific Anxiety during the Spread of COVID-19 In Israel: A Seven-Wave Longitudinal Study date: 2021-08-11 journal: Am J Epidemiol DOI: 10.1093/aje/kwab214 sha: 02ce0c5fb23ec0a10af485b8327317138cff23b5 doc_id: 766964 cord_uid: kaharu3x We compared three hypothetical trajectories of change in both general and COVID-19-specific anxiety during the 1(st) wave of the spread in the state of Israel: panic (very high anxiety, either from the outset or rapidly increasing), complacency (stable and low anxiety), and threat-sensitive (a moderate, linear increase compatible with the increase in threat). A representative sample of 1018 Jewish-Israeli adults was recruited online. A baseline assessment commenced two days prior to the identification of the first case, followed by six weekly assessments. Latent Mixture Modeling analyses revealed the presence of the three trajectories: (1) "threat-sensitivity" (29% and 66%, for general and virus-specific anxiety, respectively), (2) Panic (12% and 25%), and (3) Complacency (29% and 9%). Only for general anxiety, a fourth class representing a stable mid-level anxiety was identified ("balanced": 30%). For general anxiety, females and the initially anxious - both generally and specifically from the spread of the virus - were more likely to belong to the panic class. Men and older participants were more likely to belong to the complacency class. Findings indicate a marked heterogeneity in anxiety responses to the first wave of the spread of COVID-19, including a large group evincing a "balanced" response. Wuhan, China. 1 The outbreak was declared a Public Health Emergency of International Concern on 30 January and as a global pandemic on March 11, 2020 . COVID-19 commonly present as cold symptoms (cough ,fever, malaise, myalgias), gastrointestinal symptoms, and anosmia, but may exacerbated into shortness of breath, severe pneumonia, respiratory failure and death. Men, the elderly and people with pre-existing medical conditions are most vulnerable to severe illness. 2 As of writing these words, during July, 2021, effective vaccines are used to prevent the spread of COVID-19, and Israeli is leading in terms of offering them to the public. Up until the emergence of these vaccines, however, the only preventive treatment was behavioral, and includes hygiene, physical-distancing, masks-wearing, and large-scale societal measures ranging from targeted quarantine to general lockdowns. 3, 4 . Because some variants of COVID-19 appear This is the 1 st report from the COVID-19-Israeli Public Behavior Project (COVID-19-IPBP), a project aimed at documenting the Israeli public's reaction to the pandemic, with a focus on compliance, public trust, resilience, and emotional distress. An integrated conceptual framework guiding this study was utilized, drawing from behavioral medicine, psychology and psychiatry, and public health. In this first report from COVID-19-IPBP, we utilized data from the 1 st wave of the spread (although we have been, and still are, collecting data). Herein we report on the Israeli public's anxiety, a key factor in populations' behavior during medical crises such as mass-vaccination, plagues, and global hostilities 6 . Anxiety is largely considered a negative force, usually a clinical outcome of the crisis 7 . It is construed as an alarming outcome among medical teams 8 , as well as a factor derailing the public's compliance with instructions issued by policy makers 7 . Importantly, some levels of fear which are compatible with the medical threat are expected in the public, and some levels of anxiety can act to positively encourage people to adhere to guidelines. Nevertheless, extreme levels of fear and anxiety, tantamount to "panic", are likely to exacerbate the spread of epidemics and pandemics 9 . At the same time, low levels of anxiety may be associated with an underestimation of the medical threat, and often represent "complacency " 10,11 . The latter might hinder preparedness 10 , and compliance with governments' instructions, such as those concerning social distancing 11 . Because most pandemic-anxiety research is cross-sectional, not enough is known about trajectories of public anxiety during pandemics. However, one longitudinal study revealed that, in response to the H1N1 pandemic, anxiety increased rapidly, but then rapidly decreased 12 . Moreover, responses to the Ebola epidemic (distinguishable from H1N1 and COVID-19, which are pandemics) appear to reflect immediately high anxiety that was increasing further 13 . We investigated a large cohort of Jewish-Israeli adults over six weeks, using seven weekly assessments. The first assessment (Wave 0) transpired two days prior to the identification of the first COVID-19 carrier in Israel. We assessed general and virus-specific anxiety, demographics, and other variables not pertinent to the present report. Charting changes in general vs. virus-specific anxiety, we compared three hypothetical models: (1) A "Panic Model", evincing extreme levels of anxiety, either starting at a low level and increasing very fast or as high anxiety already appearing from the outset. (2) A "complacency model", in which general and virus-specific anxiety start and stay relatively low, i.e., not exceeding the mid-level of the scales. (3) A "threat-sensitivity model", a title developed specifically for this study. This model reflects a linear increase in anxiety which is commensurate with the increase in the threat (see 9 ). It should be noted that all three possibilitiesand of course additional onesmay occur concurrently at any single population, reflecting a heterogeneous public response to COVID-19. We allowed for this possibility in our analyses (see below). To summarize, the three goals of the present study are: (1) To identify various trajectories of anxiety in the investigated population during the 1 st wave of the spread of COVID-19 in Israel. (2) To characterize individuals belonging to the various trajectories using baseline anxiety and demographic variables. (3) To examine the direction of relationships between general and virus-specific anxiety. We were particularly interested in examining the predictive effect of baseline virus-specific anxiety on trajectories of general anxiety. Such an effect would suggest that at least some of the change in general anxiety during the 1 st wave of the spread is attributable to the spread itself, rather than to preexisting conditions. Participants and procedure. Tables 1-4 ). The recruitment procedure is depicted in Figure 1 . people were prohibited, and people younger than 65 are instructed to refrain from visiting the elderly. The fifth assessment wave, i.e., Week 4, took place on March the 18 th , 2020, after 524 -cases‖ were identified. All Israelis arriving from other countries were instructed to be selfquarantined. The education system was inoperative. Gathering of >10 people was prohibited. Lockdowns was placed on targeted areas were spread was high. The sixth assessment wave, i.e., Week 5, took place on March the 25 th , 2020, after 2436 -cases‖ were identified. One hospitalized patient died from the virus. People were instructed to minimize outings to strictly crucial activities, otherwise not to leave home for more than 100 meters. Group praying ("Mynian") was prohibited. The seventh assessment wave, i.e., Week 6, took place on April 1 st , 2020, several days prior to the Passover holiday and after 6168 -cases‖ were identified. Prohibitions and guidelines were dramatically increased, and were enforced thereafter (e.g., fines for not wearing masks). All measures were self-report questionnaire items. Following Israeli et al 5 not), employment (transformed into employed or not), and income (transformed into below average, above average, refused to answer). Income was ultimately not used in the present analyses because of the substantial number of refusals (n = 111, 11%). All analyses were conducted in Mplus v8 respondents for whom any data were available for the analysis (e.g., even if only one week of anxiety data). ML estimation from raw data accommodates data missing at random. The missing data mechanism is that not every panel member responded to every week's survey. The -robust‖ maximum likelihood (MLR) estimates use a sandwich estimator for robust standard errors of parameter estimates and, for the single-class models, yield a scaled test statistic. Data analysis was conducted in the following three phases (with Phase 3 being divided into six steps): This was done via two types of outlier checks: Time to complete survey and Mahalanobis's distance, as well as via a patterned response check (psychological synonyms; see 16 , and examples in 17, 18 ). Checks were applied to the first assessment wave, as subsequent assessment included only three questions each-too few for us to confidently identify outliers or careless responding. This was done on the restricted sample resulting from Phase 1. The study variables were characterized using frequencies for binary variables and means, standard deviations, and ranges for the continuous variables. As well, means, standard deviations, and ranges were also calculated for general and virus-specific anxiety across the seven assessment waves. Missing value rates (numbers and proportions) were computed for each variable, and summary statistics were estimated in Mplus with maximum likelihood. 19 . Correlations among the study variables were computed concurrently, enabling future replications and meta-analyses. We also calculated the proportions of respondents meeting a binary criterion for the general anxiety scale. We set the cutoff at >=4 on the 1 to 5 scale, which correspond to the "very much" and "very strongly" verbal anchors of the scale. Because the item measuring virus-specific anxiety had anchors only for the extreme scores (i.e., 1 and 7), we did not calculate proportions for this item. Two outcomes were considered: general and virus-specific anxiety, assessed at the first post-detection assessment (Week 1), for the subsequent six assessment waves. Because we anticipated heterogeneous responses to the coronavirus, we estimated latent class growth analyses (LCGAs) and growth mixture models (GMMs) in a structured search, separately for each of the two outcomes. Analyses were conducted following recommendations by Jung and Wickrama 20 . In addition, we also followed van de Schoot, Sijbrandij, Winter, Depaoli, and Vermunt, 21 , who provided guidelines for reporting on latent trajectory studies (i.e., GRoLTS-Checklist). In Figure 2 we present a flowchart of the various stages employed, linking these stages to Jung and Wickrama 20 . In Web Table 5 we detail our compliance with the GRoLTS-Checklist. Before addressing the Jung and Wickrama's 20 recommendations, we employed a visual inspection of individual trajectory plots. Many plots showed a conspicuous inflection point at the fifth assessment (-Week 4‖) for both outcomes, with individual trajectories tending to show a leveling off after increasing to that point. Given the above-mentioned visual inspection, we based our models on a piecewise functional form, with linear change before and after a spline knot at the fifth assessment. In all models, growth was modeled as linear from Week 1 through Week 4, then allowed a different linear slope from Week 4 through Week 6. The latent intercept was parameterized as a factor with loadings of 1 for each assessment. The first latent slope was parameterized with loadings [-3, -2, -1, 0, 0, 0] and the second slope with loadings [0, 0, 0, 0, 1, 2]. As a result, the model-implied value at Week 4 (when both slope loadings equal 0) is the estimate of the intercept factor. Disturbance variances were independent and constrained to invariance across time within latent class, in order to improve estimability. This is done in order to assess whether a parsimonious latent trajectory model fits the data well. We employed LCGAs, which impose a within-class homogeneity in the growth factors, namely, the variance-covariance estimates (VCVs) are fixed to zero. We then employed GMMs, for which the homogeneity of the within-class VCVs is relaxed. Because we hypothesized trajectories near the floor and/or ceiling of the scale, we decided a priori to allow the variance-covariance (VCV) matrix of the growth factors to differ between classes, as variances would be expected to be lower the nearer the means are to the scale endpoints. This was then followed by estimating models with the VCV matrices constrained to equality, both to improve estimability and to serve as a sensitivity analysis. Information Criterion (SABIC) to compare fit between types of models. We predicted class membership via respondent age (grouped), gender, religiosity (binary), and education (binary: degree completed), as well as both general and virus-specific anxiety from the first assessment. Employment was not entered into this analysis because there was little variability: 92% were employed. Prediction was tested using the 3-step method incorporated into the R3STEP option in Mplus. This procedure automates the prediction of individual class membership based on the original mixture models incorporating individual-level uncertainty in class assignment 22 . We computed odds ratios and confidence intervals for the effect of each predictor in the multiple multinomial regression on likelihood of membership in each class, contrasted pairwise with every other class. Because this pairwise examination resulted in more contrasts than degrees of freedom, we applied Holm's family-wise error rate correction, separately for each predictor 23 Our first assessment of careless responding was to examine the time taken to complete the assessment for left-side (i.e., suspiciously brief) outliers. Visual inspection of plotted completion times indicated none. Our second assessment of careless responding was a psychometric synonyms measure, In total, 27 (2.7%) of the 1,018 respondents were screened out of the sample following these procedures: 14 for careless responding, 12 as multivariate outliers, and one who met both criteria. All results below are based on the remaining sample of 991 participants. In Web Table 6 we present the intercorrelations among the study variables. In Tables For clarity of presentation, means of the two outcomes across time are plotted in Figures 3 and 4 . This enables the reader to understand why we examined trajectory models that allow for a plateau transpiring after weeks of a linear growth. As shown in the figure, the plateau started at Week 4 (i.e., the 5 th assessment). Strongly corresponding to the above pattern is our findings from calculating proportions of general anxiety (scores >4), which were 13.5%, 18.4%, 23.4%, 33.5%, 46.2%, 43.7%, and 42.6%, for weeks 0-6, respectively. However, the RMSEA estimate and its lower confidence bound were both higher than common standards for -close fit‖ (.050; e.g., 24 ), In Table 3 , we also present the LGCAs and GMMs that were tested. For reasons of clarity and brevity, we relegate the description of each of the model specification and fit to Web Appendix 2. By way of a summary, we note that two LCGAs were examined (with 2 and 3 classes, respectively), and the more parsimonious 2-class model was preferred to the 3-class model. As well, four GMMs were examined, and the GMM fitting best was a 4-class model, which was preferred over the 2-class LCGA based on the fit indices. Table 5 . In suggests that too little and too high anxiety are detrimental, we title this class "Balanced" (see Table 4 ). Detailed results of the R3STEP prediction of class membership are shown in Web Table 7 . Listwise deletion resulted in N = 958 for this analysis. Older respondents were more likely to be in the complacency class than the balanced Estimated descriptive statistics from the BCH method for the covariates within class are shown in the right-hand panel of Web Table 7 . Virus-Specific Anxiety. Again, the visual inspection identified a knot, or a spline, at Week 5. The estimated models and their relative fit are summarized in Table 6 . but not -closely‖ by RMSEA and its confidence interval. In Table 6 , we also present the LGCAs and GMMs that were tested. An elaborated description of these models appears in Web Appendix 2). To summarize, 2-and-3 classes LCGAs were examined, and fit indices identified the 3-class LCGA was preferred. As well, three GMMs were examined, culminating in a 3-class, equated variance-covariance as the best fitting model. Characteristics of the final model are presented in Tables 7 and 8 Detailed results of the R3STEP prediction of class membership are shown in Table 9 . Listwise deletion resulted in N = 956 for this analysis. As expected, higher baseline virus-specific anxiety clearly predicted class membership. Respondents with higher baseline virus-specific anxiety were more likely to be in panic class The BCH procedure to estimate descriptive statistics within class failed, with Mplus reporting an error in computations. After multiple attempts to resolve the error, we concluded the problem was likely to be related to the zero-constrained variance. given studies showing that the media can bias our perceptions of disease 28, 29 . Documenting the unfolding of general anxiety, as well as specific anxiety concerning the virus, was therefore essential. For both general and virus-specific anxiety, our hypothesized three classes emerged: (1) a threat-sensitive class evincing a linear increase till Week 4, then plateauing (29% and 66%, for each outcome, respectively), (2) a complacency model, characterized by low (general and virusspecific) anxiety slightly increasing over the study period (29% and 9%, respectively), and (3) a third, panic class starting very high and increasing in a linear fashion close to the ceiling (12% and 25%, respectively). In addition, for general anxiety, a non-hypothesized, albeit large (30%) and highly intriguing, fourth class emerged, evincing stable, mid-levels of general anxiety throughout the study period. We now discuss the implications of each of these classes in turn. The two extreme classes in each outcomelow and high anxiety --can be said to reflect the hypothesized complacency and panic trajectories. As for the low anxiety class, although it did not pertain to a straight line, the linear increase in this class was so modest (and culminating in a low anxiety level at the last wave) to be deemed as reflecting complacency: Despite the apparent threat, 29% of our sample were unperturbed in terms of general anxiety, and 9% were not worried about the spread of the virus. As shown in previous research 10, 11 , complacency can be identifiable at times of medical crises, and it can be quite harmful. As for the high-anxiety class, it is consistent with a "panic" trajectory where levels of anxiety are high already for the start and are even increasing with time 9 . This interpretation, however, should be tempered given the possibility that a subset of the population (12% and 25%, for each outcome, respectively), may constantly suffer from high anxiety regardless of the circumstances. That this class is relatively small may explain previous research reporting the absence of panic at times of disasters and large-scale political crises 30 , in that these studies did not allow for the possibility of various classes of trajectories included in the their sample. In subsequent analyses, we are examining the public and mental health consequences of already belonging to the high-anxiety class. For both outcomes, a large class was consistent with the threat-sensitive model, according to which levels of general and virus-specific anxiety start normatively low, but increase in a linear fashion as the threat grows, and then plateauing. We construe this plateau as reflecting habituation. As described in the landmark paper by Thompson and Spencer 31 , habituation is a form of simple, nonassociative learning in which the magnitude of the response to a specific stimulus decreases with repeated exposure to that stimulus. It has been known for several decades that the magnitude of hypothalamic-pituitary-adrenal activation occurring in response to a stressor declines with repeated exposure to that same stressor, and this decline has been referred to as -habituation‖ in the stress neurobiology literature 32 . When general, but not virus-specific, anxiety was considered as the outcome, we also identified a non-hypothesized, large (30%), fourth class, in which mid-levels of general anxiety were evinced from the outset, remaining stable across the six assessments. We title this class Week 0 predictors of the various classes for each outcome were informative. Interestingly, while both general and virus-specific anxiety predictedin expected waysmembership in the general anxiety classes, only virus-specific anxiety predicted membership in the virus-specific anxiety classes. Indeed, in developing the assessment protocol, we were explicitly influenced by literature on perceived stress 15 , whereby virus-specific anxiety is "a risk factor" for general anxiety symptoms. Nevertheless, we also allowed for a reverse, or birelational/cross-lagged prospective association, because these prospective associations are often observed in psychopathology 33 . Our actual findings are consistent with the construal of virusspecific anxiety as the stressor, and with general anxiety as the stress-reactive outcome. This pattern is important because it suggests that some of the changes in general anxiety are attributable to anxiety about COVID-19. Second, gender markedly distinguished between participants membership in the classes pertaining to general, but not virus-specific anxiety: Women were decidedly more anxious than men, and their anxiety was more likely to place the in the panic class (and men'sin the complacency class). However, because such associations were found only for general anxiety, it is possible that these gender differences, well documented in previous research, are unrelated to the COVID-19 crisis. Third, older participants were more likely to belong to a general (but not virus-specific) low anxiety (i.e. complacency) class. Overall, the moderating role of age has been documented in previous research , concerning mass disasters 34 . We are not sure why older participants in our sample reported levels of anxiety that were that low, and whether such low-end anxiety does indeed reflect complacency. Implications for public health interventions are noteworthy. That a large subset of this sample was threat-sensitive in terms of both general and virus-specific anxiety suggests that this particular public may be an intense consumer of data regarding the spread, the infected, and the deceased. To the extent that this post-hoc speculation is correct, quantitative information provided by medical leadership, as opposed to dramatic declarations, should be at the forefront of messages to the public. Relatedly, that baseline levels of virus-specific anxiety predictedin the expected direction --subsequent membership in the general anxiety trajectory, highlight the need to assess anxiety that is specific to the unfolding medical crises at the earliest stages of such crises. As for the identification of the "balanced", general anxiety class, and the predictive associations involving gender and age, we refrain from pointing out public health implications until we ascertain that this class, and such predictive associations, are specific to the COVID-19 crisis. Study This graph shows the various stages employed in the course of the complex data analytic process utilized for this study. As detailed in the text, the various stages integrates various statistical tasks, namely, data cleaning and management, preparing for modeling, numerous modeling procedures, completion of modeling checklists. This graph depicts the mean levels of general anxiety across the seven assessment waves. X axis pertain to the assessment waves (waves 0-6), whereas Y axis refers to the measurement scale (1-5). This graph depicts the mean levels of virus-specific anxiety across the seven assessment waves. X axis pertain to the assessment waves (waves 0-6), whereas Y axis refers to the measurement scale (1-7). This graph shows the various classes, or trajectories, identified for general anxiety. X axis pertains to the assessment waves (waves 1-6). Y axis refers to the measurement scale (1) (2) (3) (4) (5) . The lines in colors represent the various classes, or trajectories. The navy-blue line represents the "panic" class. The yellow line pertains to the "threat-sensitivity" class. The light blue class represents the raw data (as in Figure 3 ). The gray line corresponds to the "balanced" class. Finally, the orange line pertains to the "complacency" class. The box within the graph depicts the sample percentages of members belonging to the various classes. This graph shows the various classes, or trajectories, identified for virus-specific anxiety. X axis pertains to the assessment waves (waves 1-6). Y axis refers to the measurement scale (1-7). The lines in colors represent the various classes, or trajectories. The yellow line represents the "panic" class. The light blue class represents the raw data (as in Figure 5 ). The gray line pertains to the "threat-sensitivity" class. The orange line represents the 'complacency" class. The box within the graph depicts the sample percentages of members belonging to the various classes. Coronavirus Infections-More Than Just the Common Cold Clinical features of patients infected with 2019 novel coronavirus in Wuhan, China. The Lancet WHO Director-General's opening remarks at the media briefing on COVID-19 -11 The COVID-19 pandemic in the USA : what might we expect ? Preparedness is essential for malaria-endemic regions during the COVID-19 pandemic. The Lancet The heroic self under stress: Prospective effects on anxious mood in Israeli adults exposed to missile attacks The Tripartite Model of Risk Perception (TRIRISK): Distinguishing Deliberative, Affective, and Experiential Components of Perceived Risk Coupled Contagion Dynamics of Fear and Disease: Mathematical and Computational Explorations. Galvani AP Managing mental health challenges faced by healthcare workers during covid-19 pandemic The Role of Fear-Related Behaviors in the 2013-2016 West Africa Ebola Virus Disease Outbreak Public Complacency under Repeated Emergency Threats: Some Empirical Evidence Early assessment of anxiety and behavioral response to novel swineorigin influenza a(H1N1) Public Anxiety and Information Seeking Following the H1N1 Outbreak: Blogs, Newspaper Articles, and Wikipedia Visits Mass Media and the Contagion of Fear: The Case of Ebola in America State-Trait Anxiety Inventory: A Comprehensive Bibliography Terrorism-Related Perceived Stress, Adolescent Depression, and Social Support From Friends Identifying careless responses in survey data Using Mechanical Turk to recruit participants for internet intervention research: experience from recruitment for four trials targeting hazardous alcohol consumption Worry, avoidance, and coping during the COVID-19 pandemic: A comprehensive network analysis R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing An Introduction to Latent Class Growth Analysis and Growth Mixture Modeling: Latent Trajectory Classes. Social and Personality Psychology Compass Latent Class Modeling with Covariates: Two Improved Three-Step Approaches A Simple Sequentially Rejective Multiple Test Procedure Cutoff criteria for fit indexes in covariance structure analysis: Conventional criteria versus new alternatives. Structural Equation Modeling: A Differences in posttraumatic stress characteristics by duration of exposure to trauma Nowcasting and forecasting the potential domestic and international spread of the 2019-nCoV outbreak originating in Wuhan, China: a modelling study Euro surveillance : bulletin Europeen sur les maladies transmissibles = European communicable disease bulletin The psychological effects of quarantining a city Medicine in the popular press: The influence of the media on perceptions of disease Why the French did not choose to panic: a dynamic analysis of the public response to the influenza pandemic: The French response to the influenza pandemic Habituation: a model phenomenon for the study of neuronal substrates of behavior Habituation to repeated stress: Get used to it Erosion: The Psychopathology of Self-Criticism Community Resilience throughout the Lifespan the Potential Contribution of Healthy Elders Visual Inspection (Identification of Spline) s Phase 3: GMMs Free vs s Phase 4: Prediction of Classes GRoLTS Checklist s Phase 2: LGCAs LMR/bLRT/SABIC Arriving at Final Model The authors wish to thank Dr. Ariel Ayalon, Ph.D., Owner and CEO of the Midgam Panel, for his collaboration on this project and his comments on the Methods section of this manuscript.