Participants
Initially, 447 participants responded to the survey link. Inclusion criteria required that participants be aged 18 or over, and were a university student. Participants were excluded if they did not complete the experimental session, provided contradictory responses to COVID-19 diagnosis questions (i.e. a positive test, but reporting not ever having COVID-19), or showed evidence of repetitive responding. Repetitive responders were identified by recording the number of times a participants’ response was identical to a previous item worded in the opposing direction. Those who scored over 2 SDs on this measure were excluded. This data screening protocol is based on the longstring data-screening method (DeSimone et al.,
2015); however, this adaptation (designed for the current study) focused more on repeated implausible responses rather than all repeated responses, and was designed to exclude clear cases of unmotivated responding without relying on subjective judgment. The final sample consisted of 382 participants with the mean age of 21.24 (
SD = 4.95, Range 18–72).
1 In total there were 313 female participants, 61 male participants, 5 who reported identifying as neither male/female or non-binary, and 3 who preferred not to report. Participants all provided informed consent prior to completion of the study.
Within the sample, 88 participants reported being certain or almost certain they had been infected with COVID-19, of which 58 reported a positive test.
2 Twelve reported asymptomatic cases, 70 mild illness, 4 a moderate illness requiring hospitalisation, and 1 reporting a severe illness requiring hospitalisation.
Participants were recruited across three different universities (Roehampton, Sussex, KCL) by volunteer mailing list or in exchange for course credit. All Study 2 data was collected between March and May 2021. The recruitment strategy was to recruit as many participants as possible in a single academic term. The final sample was consistent with guidelines for factor analysis which recommend a minimum sample of 300 (Comrey & Lee,
1992), as well as those recommending a minimum number of participants per item (e.g. 10:1 ratio; Osborne & Costello,
2004). Additionally, the average item loading on each factor from Study 1 was .71 (
SD = .13), and the average number of items per factor was 4 across the five factors; therefore, according to recommendations for Confirmatory Factor Analysis (CFA) by Wolf et al. (
2013), the final sample size would be sufficient to validate the factor structure found in our initial sample, based on expected loadings.
Results and Discussion
To validate the five-factor model of the PRWQ (see Study 1 above), we conducted CFA which was conducted using a maximum likelihood estimator using JASP software. This analysis found an adequate fit to the data with the previous five-factor model (Hu & Bentler,
1999),
χ2(df = 160) = 568.58,
p < .001 CFI = .908. RMSEA = .082 (90CI[.075, .089]), SRMR = .051 (see Table
2).
Table 2
Confirmatory factor analysis results, with parameter estimates, p-values, and 95% Confidence Intervals (95CI) for the five-factor model
Decline in QoL | 24 | .772 | .045 | 17.175 | < .001 | .684 | .86 |
30 | .593 | .049 | 12.104 | < .001 | .497 | .689 |
26 | .685 | .047 | 14.558 | < .001 | .593 | .777 |
9 | .728 | .046 | 15.817 | < .001 | .638 | .819 |
19 | .619 | .048 | 12.778 | < .001 | .524 | .714 |
28 | .749 | .046 | 16.442 | < .001 | .66 | .838 |
18 | .705 | .047 | 15.127 | < .001 | .614 | .796 |
COVID infection severity | 20 | .893 | .044 | 20.329 | < .001 | .806 | .979 |
22 | .87 | .044 | 19.643 | < .001 | .783 | .957 |
Risk to loved ones | 21 | .865 | .042 | 20.718 | < .001 | .783 | .947 |
5 | .737 | .045 | 16.327 | < .001 | .648 | .825 |
11 | .795 | .044 | 18.215 | < .001 | .709 | .88 |
23 | .895 | .041 | 21.892 | < .001 | .815 | .975 |
COVID infection probability | 3 | .837 | .043 | 19.589 | < .001 | .753 | .921 |
1 | .866 | .042 | 20.65 | < .001 | .784 | .948 |
2 | .798 | .044 | 18.24 | < .001 | .713 | .884 |
4 | .785 | .044 | 17.774 | < .001 | .698 | .871 |
Financial concerns | 8 | .824 | .047 | 17.514 | < .001 | .732 | .917 |
7 | .704 | .049 | 14.386 | < .001 | .608 | .8 |
12 | .656 | .05 | 13.175 | < .001 | .558 | .753 |
Conversely, CFA on the alternative four-factor solution identified in Study 1, which combined worries about infection severity and infection probability within a single factor, revealed a poorer fit to the data, χ2(df = 183) = 836.42, p < .001; CFI = .867; RMSEA = .097 (90CI[.09, .103]), SRMR = .06. Indeed, a nested model comparison suggested that the five factor model had significantly better fit compared to the four factor model, Δχ2(23) = 267.84, p < .001.
Whilst the five-factor model was adequate, examination of the modification indices (MI) and expected parameter change (EPC) values revealed it was possible that a six-factor model could have been missed. Allowing the errors of the item pair with the highest modification index (MI = 90.56, EPC = .24; items 5 and 11) to covary in the model did somewhat improve the model fit, χ2(df = 159) = 478.55, p < .001; CFI = .928; RMSEA = .073 (90CI[.065, .08]), SRMR = .053, and this improvement was significant relative to the unadjusted five-factor model, Δχ2(1) = 90.03, p < .001. This pair of items reflected the worries about the probability that a loved one may be infected with COVID-19; whilst the remaining 2 items within the ‘risk to loved ones’ factor reflected infection severity. It may be, therefore, that the severity risk worries and infection risk worries for loved ones reflect distinct factors which are missed due to their under-representation in the initial pool of items in the PRWQ. Any future versions of the PRWQ may therefore require additional items when exploring worries about the risk to loved ones, specifically. As noted below, however, worries about risk to loved ones were more weakly correlated with cognitive outcomes relative to other factors within the scale, thus the five-factor model is valid in the current context.
One possible confounding factor to be considered could be the development of vaccines, which could alter the level of different pandemic-related worries. In the UK, the vaccination programme began in December 2020, after Study 1 but prior to Study 2. This is however unlikely to be a substantial influence on perceptions of personal risk from COVID-19, as many of the participants were ineligible for vaccination during the March–May 2021 recruitment period, due to those below the age of 30 (96.6%; N = 369) only becoming eligible for vaccination in June 2021. Although it is possible that worries about risks to older loved ones could have been reduced, our data revealed that there was no difference between Study 1, when no vaccine had been approved, and Study 2, when it was approved but not widely available, on any of the subscales related to the risk of infection from COVID-19 to self or loved ones, Cohen’s
d < .05,
t < .64,
p > .521. We would, however, expect that widespread vaccination would likely have reduced worries about severity and probability of infection due to its efficacy (Sadoff et al.,
2021), but also resulted in novel worries about side-effects, encountering unvaccinated individuals, or worries about mandatory vaccination (Bendau et al.,
2021).
Though worries about the direct threat of COVID-19 did not differ between Study 1 and 2, worries about QoL were significantly lower in Study 2, d = − .52, t(515.28) = 6.38, p < .001. With worries about financial concerns also showing a similar pattern, albeit weaker and failing to reach significance, d = − .16, t(486.97) = 2.01, p = .052. This difference is consistent with Study 1 occurring during full lockdown, and Study 2 occurring after lockdown was lifted (with restrictions in place) when QoL concerns would be less salient. It could however also reflect a general increase in optimism about the course of the pandemic caused by vaccine development. Importantly, the factor structure of PRW was replicated across both Study 1 and Study 2, suggesting a limited overall change in the relational structure between factors between the two timepoints.
The Relationship Between PRW and Cognitive Functioning
As predicted, zero-order correlations (see Table
3) revealed a significant positive correlation between both the MFS (assessing memory) and ARCES (assessing attention) and all subscales of the PRWQ, which were in line with the probable effect size based on meta-analytic estimates of the relationship between anxiety and cognitive function (
r = .28; Shi et al.,
2019). All other variables including all DASS subscales, and pre-pandemic trait anxiety and worry, measured with GAD-7 and PSWQ, also positively correlated with all PRWQ subscales.
Table 3
Zero-order correlations between the Pandemic-Related Worries Questionnaire (PRWQ) and identified subscales, Attention-Related Cognitive Errors Scale (ARCES), Memory Failures Scale (MFS), Depression, Anxiety, and Stress Scale (DASS) scores
1 | PRWQ decline in QoL | – | | | | | | | | | | | | |
2 | PRWQ infection probability | .31*** | – | | | | | | | | | | | |
3 | PRWQ infection severity | .26*** | .59*** | – | | | | | | | | | | |
4 | PRWQ risk to loved ones | .44*** | .54*** | .50*** | – | | | | | | | | | |
5 | PRWQ financial concerns | .58*** | .34*** | .38*** | .46*** | – | | | | | | | | |
6 | PRWQ total | .66*** | .77*** | .77*** | .80*** | .73*** | – | | | | | | | |
7 | Memory-related errors (MFS) | .25*** | .31*** | .23*** | .21*** | .20*** | .32*** | – | | | | | | |
8 | Attention-related errors (ARCES) | .37*** | .29*** | .25*** | .25*** | .25*** | .37*** | .75*** | – | | | | | |
9 | Depression (DASS) | .47*** | .15* | .28* | .22*** | .28*** | .32*** | .36*** | .43*** | – | | | | |
10 | Anxiety (DASS) | .37*** | .29*** | .31*** | .32*** | .33*** | .43*** | .37*** | .44*** | .59*** | – | | | |
11 | Stress (DASS) | .46*** | .28*** | .21*** | .29*** | .31*** | .40*** | .38*** | .47*** | .72*** | .65*** | – | | |
12 | Negative affect (DASS total) | .50*** | .27*** | .30*** | .31*** | .34*** | .43*** | .42*** | .51*** | .90*** | .83*** | .90*** | – | |
13 | Trait anxiety (pre-pandemic recall) | .32*** | .29*** | .24*** | .23*** | .23*** | .37*** | .43*** | .47*** | .47*** | .55*** | .59*** | .59*** | – |
14 | Trait worry (pre-pandemic recall) | .36*** | .32*** | .24*** | .28*** | .21*** | .37*** | .28*** | .34*** | .43*** | .47*** | .55*** | .55*** | .69*** |
To explore which subscales of the PRWQ were most strongly associated with attention and memory functioning, we entered all five PRWQ subscales into two separate regression models with ARCES scores and MFS scores as separate outcome variables (see Table
4). Interestingly, two of the PRWs were independently associated with MFS and ARCES scores, these were the worries about the decline in QoL and the worries about infection risk. All other subscales became non-significant when simultaneously entered into the model.
Table 4
Linear regression analyses for Attention-related errors and Memory-related errors with all Pandemic-Related Worries Questionnaire (PRWQ) subscales as predictor variables
Step 1—PRWQ subscales | Decline in QoL | .30 | 5.11 | < .001 | .19 | .42 | .082 | .17 | 2.81 | .005 | .05 | .29 | .032 |
Infection probability | .15 | 2.36 | .019 | .03 | .27 | .035 | .22 | 3.45 | .001 | .10 | .35 | .049 |
Infection severity | .10 | 1.56 | .119 | − .03 | .22 | .022 | .06 | .98 | .298 | − .06 | .19 | .018 |
Risk to loved ones | − .01 | .17 | .868 | − .13 | .11 | .02 | − .02 | − .32 | .727 | − .15 | .11 | .012 |
Financial concerns | < .01 | < .01 | .999 | − .12 | .12 | .017 | .01 | .11 | .918 | − .12 | .13 | .011 |
| | R2 = .176, F(5,376) = 16.07, p < .001 | | R2 = .121, F(5,376) = 10.36, p < .001 | |
Step 2—trait covariates | Decline in QoL | .22 | 3.76 | < .001 | .10 | .33 | – | .10 | 1.59 | .114 | − .02 | .21 | – |
Infection probability | .10 | 1.70 | .090 | − .02 | .21 | – | .18 | 2.95 | .003 | .06 | .30 | – |
Infection severity | .04 | .63 | .532 | − .08 | .15 | – | < .01 | .06 | .951 | − .11 | .12 | – |
Risk to loved ones | < .01 | < .01 | .940 | − .11 | .12 | – | < .01 | .06 | .956 | − .12 | .12 | – |
Financial concerns | < .01 | .03 | .979 | − .11 | .11 | – | < .01 | .09 | .932 | − .11 | .12 | – |
Pre-pandemic trait anxiety | .39 | 6.29 | < .001 | .27 | .51 | – | .40 | 6.19 | < .001 | .27 | .52 | – |
Pre-pandemic trait worry | − .04 | .69 | .491 | − .17 | .08 | – | − .08 | − 1.28 | .201 | − .21 | .05 | – |
| | R2 = .286, F(7,374) = 21.44, p < .001 | | R2 = .225, F(7,374) = 12.49, p < .001 | |
We note however that the PRWs were highly correlated, meaning that variation in one PRW may be dependent upon changes in another PRW subscale, making it difficult to determine the relative associative strength with the outcome variable. To address this, we supplemented the results with dominance analysis, which assesses the relative explanatory power of one variable over another when every combination of predictor variable is analysed (Budescu,
1993). This was conducted using the R
dominanceanalysis package (Navarrete & Soares,
2020). We first calculated the average unique contribution to the
R2 of attention and memory-related errors by each predictor variable, relative to competing variables in the model (see Table
4, ‘unique
R2’ columns). This replicated the pattern of importance found in the linear regression. To assess the generalisability of the dominance structure, we then assessed the proportion of times the dominance structure was reproduced across 5000 bootstrapped samples (Azen & Budescu,
2003).
When assessing the strength of the unique associative contributions to attention-related errors (ARCES score), we found that worries about decline in QoL completely dominated all other PRWs in 78.4% to 99.3% of bootstrap samples; and worries about COVID infection probability completely dominated all other PRWs in 47.4% to 67.5% of bootstrap samples. No other variable completely dominated another variable in the model. For memory-related errors (MFS score), we found worries about COVID-19 infection probability completely dominated decline in QoL in 53.1% of bootstrapped samples and all other variables in 90.1% to 94.6% of bootstrap samples. Decline in QoL then dominated all other PRWQ subscales in 50.9% to 88.7% of bootstrap samples. Though worries about COVID-19 severity dominated risk to loved ones in 39.4% of samples, there was no clear pattern of dominance across the other comparisons (see Supplementary materials 3 for bootstrapped dominance analysis tables). The dominance analysis therefore confirmed findings from the linear regression suggesting that decline in QoL and COVID-19 infection probability were the factors most strongly correlated with cognitive function.
The significant relationship between worry about declining QoL and cognitive functioning is consistent with evidence that young adults are highly concerned about the indirect impact of the pandemic on their mental health and social life, more than about the direct threat from the virus itself (Groake et al.,
2021; Ranta et al.,
2020). The reason that worries about infection probability was correlated with cognitive functioning more than the infection severity worries may be because the sample were predominantly young adults, and would be less at risk from severe infection, but no less likely to be infected (Davies et al.,
2020). On the other hand, it could be that probability of infection may be an indirect measure of the current state of the pandemic at the time of report. As cases increase in the population, so would the probability of infection, meaning that probability of infection could also reflect concerns about the overall progression of the pandemic and its subsequent impacts.
To explore whether the relationship between PRW and attention and memory-related errors was independent of pre-existing trait anxiety and worry, these measures were entered into the regression model (Table
4). When added to the model, recalled trait anxiety was the primary predictor of both attention and memory-related errors, above recalled trait worry or any of the PRWQ subscales, consistent with anxiety linked impairments in executive functioning (Eysenck et al.,
2007; Moran,
2016).
To further explore the proportion of unique variance accounted for by PRW, independent of trait anxiety and worry, the order that the trait covariates and PRWQ subscales were entered into the model was reversed, to assess the change in R2. The trait measures accounted for 22.1% of the variance in attention-related errors, and 18.1% of the variance in memory-related errors when entered in step 1 (p’s < .001). The entry of the PRWQ subscales in the model did however result in a significant increase in R2, revealing that PRW accounted for an additional 6.6% of the variance in attention-related errors, R2change = .066, F(5, 374) = 6.89, p < .001, and 4.4% of the variance in memory-related errors, R2change = .044, F(5, 374) = 4.23, p = .001, independent of trait anxiety and worry. The independent relationships are consistent with the hypothesis that individuals without pre-existing high levels of trait anxiety or worry may have experienced interference from PRW on cognitive functioning.
In order to determine whether PRW correlated with both attention and memory function independently, or whether their relationships reflect the same underlying process, we conducted an exploratory follow-up partial correlation analysis. The relationship between total PRWQ score and memory-related errors, whilst controlling for attention-related errors, was found to be non-significant,
r(379) = .06,
p = .26; conversely, the relationship between attention-related errors and total PRWQ score remained significant when controlling for memory-related errors,
r(379) = .22,
p < .001. Thus, attention-related errors were more strongly related to PRW than memory-related errors, with the MFS likely correlating with PRW due to shared variance with attention-related functioning; either due to a common relationship with general cognitive ability, or attention-related functioning’s partial role in long-term memory retrieval (Kane & Engle,
2000; Unsworth,
2010). The current findings are consistent with anxiety disrupting attention and executive function, rather than the recall of information from long-term memory (Eysenck et al.,
2007).
The ARCES measures the frequency of cognitive errors in real-world situations, such as losing concentration whilst reading, zoning out during conversations, increased interference whilst multitasking, and distractibility. Increased inattention in these situations can result in a wide range of negative outcomes, for instance in occupational settings the elevated inattention would likely result in poorer work performance which could contribute to job insecurity and stress; and more broadly, constitute a hidden economic cost caused by the pandemic (Cutler & Summers,
2020).
For students, who made up our second sample and most of our first sample, PRW may pose a direct threat to their ability to perform academically. For instance, worry has been linked to reduced academic performance over time, with earlier levels of worry predicting subsequent lower academic achievement (Owens et al.,
2012). PRW may reduce academic achievement both through the ability to learn, by reducing the ability to focus in lectures and when reading (Risko et al.,
2012; Unsworth et al.,
2013), as well as performance in assessment situations, where current concerns increase off-task thoughts (Jordano & Touron,
2017; Mrazek et al.,
2011). The novel PRW may therefore result in poorer academic performance during the pandemic due to its disruption of attentional mechanisms.
To explore whether attention-related errors mediated the relationship between PRW and negative affect, we conducted a mediation analysis using the SPSS PROCESS macro model 4 (Hayes,
2017). We entered attention-related errors (ARCES scores) as a mediator variable in the relationship between the total PRWQ score and overall negative affect (total DASS score). The rationale being that the attention-related errors would be indicative of reduced attentional resources required to regulate emotion. In order to preserve statistical power we collapsed the subscales of the DASS and the PRWQ to their total scores, however, conducting the analysis with any subscales from these measures produced the same pattern of results (Fig.
1).
We found a significant overall regression model, R2 = .33, F(1,380) = 93.07, p < .001, where the PRWQ score positively correlated with attention-related errors, β = .38, SE = .05, p < .001, 95CI[.28, .47], and attention-related errors correlated with total negative affect, β = .41, SE = .05, p < .001, 95CI[.32, .50]. The standardised indirect path from PRW to negative affect through attention-related errors was also significant, β = .16, SE = .03, 95CI[.11, .21]. The model showed evidence of partial mediation, as the direct relationship between PRW and negative affect remained significant even with the inclusion of the mediator, β = .27, SE = .05, p < .001, 95CI[.18, .36]. The total relationship between PRW and negative affect without ARCES score in the model was significant, β = .43, SE = .05, p < .001, 95CI[.34, .52].
In line with cognitive models of worry and anxiety (Eysenck et al.,
2007; Hirsch & Mathews,
2012), the current mediation relationship could be interpreted as PRW occupying attentional capacity required to disengage from negative thoughts. Without an effective mechanism to disengage from specific PRWs, these negative thoughts would have a persistent effect on mood, elevating levels of depression, anxiety, and stress.
An alternative interpretation could be that the higher attention-related errors may have resulted in the disruption of everyday functioning leading to more stressful outcomes (e.g. missing bill payments, poorer exam grades), and that this indirectly increased negative affect. Both interpretations of the mediation relationship are not mutually exclusive and would likely interact (Moran,
2016).
As with the relationship between PRW and attention and memory errors, all paths in the mediation model remained significant when controlling for recalled trait worry and anxiety prior to the pandemic, R2 = .48, F(4, 377) = 85.69, p < .001. PRW still significantly correlated with attention-related errors, β = .24, SE = .05, p < .001, 95CI[.14, .33], and attention-related errors correlated with total negative affect, β = .25, SE = .04, p < .001, 95CI[.17, .34]. The standardised indirect path from PRW to negative affect through attention-related errors also remained significant when controlling for trait worry and anxiety, β = .06, SE = .02, 95CI[.03, .10], as was the direct path (c′) between PRWQ score and negative affect, β = .15, SE = .04, p < .001, 95CI[.07, .24].
The reason that this independent relationship is so important is because it could be indicative of an elevated risk of developing more prolonged anxious symptoms (and the development of anxiety disorders), even in individuals who previously did not experience high levels of anxiety. One of the key moderating factors which determines whether an individual goes on to develop anxious symptoms is the ability to effectively disengage from negative thoughts, which prevents the persistence of anxious states (Fox et al.,
2021). For instance, in children, poorer performance on executive function tasks predicts anxiety at later timepoints, beyond their initial levels of anxiety (White et al.,
2011; Zainal & Newman,
2018). Without this ability to disengage from negative automatic thoughts, these thoughts can become more habitual and more easily activated resulting in chronic anxiety and mood disorders (Watkins & Nolen-Hoeksema,
2014; Wells,
1995). Further, the general negative mood induced by these habitual thought patterns can increase negative interpretation of other ongoing problems, allowing worry to persist for even longer (Davey & Meeten,
2016; Startup & Davey,
2001). Therefore, rather than reflecting a transient increase in worry and cognitive interference, which will be alleviated once the threat of the pandemic dissipates, initially flexible concerns about the pandemic may develop into habitual worries and more entrenched anxiety disorders if not identified and addressed.
Overall, the current results are consistent with the predictions of cognitive models of anxiety/worry, whereby the attention required for the control of emotion and behaviour is disrupted by the pre occupation with task-irrelevant negative thoughts, and the impaired ability to control attention is a risk factor for the development of anxiety (Derakshan,
2020; Eysenck et al.,
2007; Hirsch & Mathews,
2012; Wells,
1995). Existing cognitive models focus mainly on individuals with dispositional anxiety or worry, with limited focus on how similar mechanisms can explain the transition to more severe levels of negative affectivity when exposed to stressful situations external to the individual, independent of predetermined traits. The current investigation therefore extends these models by highlighting how the same attentional processes could plausibly also underpin the general transition to higher levels of negative affect when exposed to increased levels of worry about external situations, in this case the COVID-19 pandemic (see Songco et al.,
2020 for further discussion of cognitive-emotional theories and the development of anxiety).
Limitations
One limitation of the current investigation was that trait anxiety and worry measures were based on recalled baseline scores. Previous evidence suggests that anxious individuals often over-estimate their previous negative emotions, consistent with a mood-congruent recall bias (Cutler et al.,
1996; Safer & Keuler,
2002). The current measure of trait anxiety could therefore reflect, or be contaminated by, current levels of anxiety. If true, however, we would have expected that controlling for recalled trait anxiety would have resulted in the relationship between PRW and negative affect becoming non-significant due to shared variance, which was not the case. Additionally, when pre-pandemic baseline data was controlled for in a previous study, it was found that the relationship between poorer cognitive functioning and a single item measure of COVID-19-related anxiety remained significant, replicating our pattern of results with related non-recalled measures (Fellman et al.,
2020).
To further explore whether the recalled trait anxiety and worry levels were likely to be accurate, we searched for existing published and unpublished data which contained measures of trait anxiety and worry from shortly before the pandemic, and were from similar samples to Study 2: Students from the University of Roehampton, Sussex, or KCL, recruited through random opportunity sampling. Four samples were found which matched this criteria, two were from KCL students who completed the GAD-7 and PSWQ as part of a screening process and were collected in 2017 (N = 165; Feng et al.,
2019) and 2019 (N = 274; Feng et al., unpublished data). The mean levels of trait anxiety and worry from these samples were comparable with our recalled levels from this period on the same measures: The GAD-7 scores from 2017 (M = 8.07, SD = 5.01) and 2019 (M = 7.59, SD = 4.99), as well as the PSWQ scores from 2017 (M = 55.87, SD = 12.84) and 2019 (M = 56.56, SD = 14.21), were nearly identical to our mean scores on our recalled measures (i.e. recalled mean GAD-7 = 8.06; recalled mean PSWQ = 56.31).
In other published data, a sample of University of Sussex students (N = 216) recruited by Davey et al. (
2022) in 2019 also reported a similar level of trait worry to our sample (M = 57.39, SD = 9.87). Further, a sample of University of Roehampton students (N = 546) recruited between 2017 and 2018 by Norbury and Evans (
2019) reported levels of anxiety on the Trait Anxiety Inventory (Speilberger et al.,
1983) in the upper levels of the mild to moderate range, i.e. 40–50 (47.37; SD = 10.9; Van Dam et al.,
2013). Consistent with our measure of trait anxiety reflecting accurate recall of pre-pandemic anxiety, our sample also reported anxiety in the higher end of the mild to moderate range on the GAD-7 (moderate anxiety score criteria = 5–9; Spitzer et al.,
2006). Thus, from across four samples drawn from the identical population we recruited from, our measures of recalled trait anxiety and worry were equivalent to actual scores from the pre-pandemic period. Our measures of pre-pandemic trait anxiety and worry are therefore likely valid, despite potential recall bias.
When interpreting the current findings, we must account for the cross-sectional design, which limits the ability to draw strong conclusions about causal relationships between factors. Indeed, it is possible that negative affect results in an increase PRW, rather than the opposite relationship in the current model. Though difficult to reliably separate the true direction of effects out statistically, we conducted a reverse mediation analysis, whereby the variables were reversed in the mediation model (PRW → Negative affect → attention-related errors). This revealed that the opposite pattern of results was also found, as the indirect effect of PRW on attention-related errors mediated through negative affect was significant (whilst controlling for recalled trait anxiety and worry),
β = .08,
SE = .02, 95CI[.05, .12]. The results therefore suggest that the opposite interpretation remains plausible. Indeed, the data could reflect a bidirectional relationship, where worry’s impact on negative affect could reduce attentional control, concurrent with worry’s deleterious effect on attentional control increasing negative affect through the persistence of worry episodes – consistent with a self-perpetuating relationship (Eysenck et al.,
2007; Hotton et al.,
2018; Moran,
2016). We note that the reverse mediation method does not always reliably detect the ‘true’ direction of causality (Lemmer & Gollwitzer,
2017; Thoemmes,
2015). Future investigations exploring the lasting effects of PRW on cognition and mental health should therefore utilise a longitudinal design, as this would allow the inference about directional relationships. Here, however, we present a plausible theoretically grounded model of how PRW may disrupt cognitive function, especially attentional capacity, and how this may result in elevated negative affect.
Additionally, when investigating the enduring effect of PRW on cognitive function, more objective task-based behavioural measures could be used. These could identify specific executive functions which are most susceptible to interference from PRW, such as the impact on inhibition or shifting (Mennies et al.,
2021). However, using more sophisticated task-based measures would require larger sample sizes to account for potentially lower statistical power in such studies (Hedge et al.,
2018). Specific moderating risk factors not explored in the current investigation could also be explored in relation to persistent PRW in later stages of the pandemic, for instance, the death of a loved one due to COVID-19 or redundancy due to the pandemic (Blustein & Guarino,
2020; Torrens-Burton et al.,
2022).