Mental health, financial, and social outcomes among older adults with probable COVID-19 infection: A longitudinal cohort study

Significance Longitudinal evidence on the impact that contracting COVID-19 may have on the individual’s mental health, personal finances, and social relationships is scarce. Using longitudinal data from the English Longitudinal Study of Aging, this study shows that older adults with probable COVID-19 infection experienced higher levels of depression and anxiety, poorer quality of life, elevated feelings of loneliness, and greater financial difficulties compared with those without probable infection. The associations were independent of prepandemic mental health and financial circumstances, and they were evident both in the acute phase of the infection and up to 6 months later. These results suggest that the adverse psychosocial impact of COVID-19 infection is long-lasting and more broadly present across the population.


sMethods Sample
The English Longitudinal Study of Ageing (ELSA) is a longitudinal cohort study of men and women aged 50 years and older living in England. The study commenced in 2002/03 (wave 1) with a nationally representative sample of 12,099 participants and their partners, drawn from the Health Survey for England (HSE). To record changes in their health, social, and economic circumstances, since the commencement of the study these participants have been repeatedly interviewed at biennial interviews (i.e. waves) using face-to-face computer-assisted personal interviews and self-completion questionnaires. The most recent full wave of data collection was carried out in 2018/19 (wave 9). The study sample is periodically refreshed with new participants to maintain its original size and representativeness.

Mental health outcomes
Depression was ascertained using the 8-item Centre for Epidemiological Studies Depression (CESD-8) scale(1). We used a cut-off point of four or more symptoms to identify likely cases of clinical depression (2). Anxiety was ascertained using the 7-item Generalised Anxiety Disorder scale (GAD-7)(3), with a threshold score of 10 or greater to identify likely cases of generalised anxiety disorder (GAD). Quality of life (QoL) was evaluated through the 12-item Control, Autonomy, Self-realisation, and Pleasure (CASP) scale (4). The resulting item scores were summed to create a QoL index where higher scores indicate poorer wellbeing (range: 1-48). Loneliness was measured using the 3-item revised University of California (UCLA) Loneliness scale (5) and an additional item asking participants how often they feel lonely. The item scores were summed to derive a total score, with higher values indicating greater loneliness (range: 1-12).

Covariates
Pre-pandemic mental health, social contact, and financial difficulties were included as covariates to account for differences between people with and without probable COVID-19 infection before the onset of the pandemic. Pre-pandemic mental health and social interactions were measured at the most recent pre-pandemic assessment available for each participant (i.e. wave 9). Assessments of depression, QoL, and loneliness in earlier waves of ELSA were conducted using the same scales as those included in the COVID-19 substudy. The GAD-7 was not administered in previous waves; to adjust the analyses for anxiety levels pre-pandemic we used the Office for National Statistics (ONS) anxiety scale administered in wave 9. The anxiety scale is one of the four wellbeing measures that have been developed by the Office for National Statistics (ONS) to assess various domains of personal wellbeing, including evaluative, eudemonic, and affective experience (6). Participants were asked to respond to the following question: "On a scale where 0 is 'not at all anxious' and 10 is 'completely anxious', overall, how anxious did you feel yesterday?". The resulting scores range between 0 and 10, with higher scores indicating greater anxiety levels. Further, as the social contact measures of the COVID-19 substudy differed from those included in earlier waves, the analyses of social interactions were adjusted for an index of social isolation combining information on whether participants had infrequent contact with family and friends and were not involved in local organisations at wave 9. Social isolation before the onset of the pandemic was measured using four indicators that were obtained from the wave 9 survey. These included having less than monthly contact (face-to-face, telephone, or writing/email) with (i) children, (ii) other family, and (iii) friends, and (iv) no membership of nine types of organisation (i.e. a political party, trade union, or environmental group; a tenants groups, resident groups, neighbourhood watch; a church or other religious group; a charitable association; an education, arts or music groups or evening class; a social club; a sports clubs, gym, or exercise class; any other organisation, club or society). This measure of social isolation has been previously used in earlier ELSA studies (7,8). Pre-pandemic financial hardship was measured based on the participants' ratings of their financial situation in the three months prior to the COVID-19 outbreak (i.e. "living comfortably", "doing all right", "just about getting by", "finding it quite difficult", or "finding it very difficult"). Covariates obtained from the first COVID-19 assessment included sex, age (52-59, 60-74, and 75+ years), whether living alone, work status (employed, retired, other not working), and whether vulnerable to COVID-19 infection (i.e. participants advised by the NHS or their general practitioner that they are at risk of severe illness if they catch COVID-19 and encouraged to stay at home and avoid face-to-face contact). The analyses of the outcomes at the second COVID-19 assessment also accounted for whether participants reported testing positive or being hospitalised for COVID-19 in Nov-Dec 2020 (information on COVID-19 symptoms was not collected in the second COVID-19 assessment). Additional covariates taken from the wave 9 survey were wealth and limiting long-standing illness. Wealth was used as an indicator of the participant's socioeconomic resources, including financial, housing, and physical wealth (such as land, business wealth, and jewellery) but excluding pension wealth, and it was divided into tertiles for the purpose of the analysis (i.e. low, medium, and high wealth). Participants were classified as having a limiting long-standing illness if they reported having a long-standing illness, disability, or infirmity that limited their activities in any way. The variables sex, age, living alone, work status, and wealth were also considered as possible effect modifiers in the analysis in order to explore whether the psychosocial impact of COVID-19 infection might vary across distinct sociodemographic groups.

Multiple imputation of missing data
The percentage of missing data in the variables ranged between 0 and 6%. In addition, due to a coding error, for around 75% of the sample the last item of the CESD-8 questionnaire was not administered at the first COVID-19 wave. This type of missing data is classified as missing completely at random (MCAR), and can be dealt efficiently with multiple imputation (9). We used multiple imputation by chained equations with all variables of the analysis included as predictors of the imputation models as well as auxiliary variables. We created twenty imputed datasets. The distribution of the variables in the imputed and observed data was similar, suggesting that the MI procedure produced accurate model estimates. In regression analyses, we pooled the regression estimates across the imputed datasets using Rubin's rules (10). For the IPTW analyses, propensity score estimation and weighting were performed within each imputed dataset, and the regression estimates across the weighted datasets were then pooled together -i.e. 'within' approach (11). We favoured this approach over the 'across' approach (i.e. propensity scores are averaged across the imputed datasets, and weighting within each dataset is performed using this averaged measure) because the latter has been shown to have inferior statistical performance compared with the 'within' approach across different scenarios (12,13).

Sensitivity analyses
The following analyses were conducted: • We applied inverse probability of treatment weighting (IPTW) to account for measured confounding factors in the associations between probable COVID-19 infection and the outcomes. • To account for multiple testing, we corrected the p-values of all associations tested in the main analyses using the false discovery rate (FDR) method. • We tested the associations between probable COVID-19 infection and the outcomes using two alternative definitions of probable COVID-19 symptoms (see Manuscript, 'Measures'). • We evaluated the associations of probable infection with social interactions considering only the amount of real-time contact that participants had with their family and friends. • The regression models of the main imputed data analysis were tested again in a subsample of participants with complete data on all variables. • We further adjusted the associations of probable COVID-19 infection with the outcomes for the presence of a mental health condition (either at the first COVID-19 assessment or at wave 9) and for the development of new physical health conditions at the first COVID-19 assessment, including cardiovascular conditions (e.g. hypertension, stroke), diabetes, respiratory illnesses (e.g. chronic lung disease, asthma), arthritis, cancer, and diseases of the visual, digestive, and nervous systems. In addition, we examined the interactions between these two variables and probable COVID-19 infection on all outcomes. • Lastly, we tested the associations of pre-pandemic levels of depression, anxiety, poor QoL, and loneliness (2018-19) with probable COVID-19 infection (June-July 2020) in order to clarify whether the nature of their relationship might be bidirectional.

Inverse probability of treatment weighting (IPTW)
This technique provides unbiased estimates of the average treatment effect of an exposure on the outcome independent of measured baseline confounding factors, and it enabled us to understand how the results compare across different strategies to account for confounding in observational studies (14). The propensity scores were estimated with logistic regression models, including all covariates, the pre-pandemic outcome scores, and the longitudinal survey weights as predictors of the probability of COVID-19 infection. The balance of the covariates across the exposure groups in the unweighted and weighted datasets was assessed by computing the standardized mean differences (SMDs) for each covariate. Although no threshold is universally accepted, covariates with SMDs lower than 0.1 are typically considered well balanced (15).

Descriptive statistics
The distribution of the covariates and outcomes in participants with and without probable COVID-19 infection at the first COVID-19 assessment is shown in Table 1. Participants with probable COVID-19 infection were younger on average, were more likely to be employed than to be retired, and had lower wealth compared with those without probable infection. People with probable COVID-19 infection were also more likely to experience elevated symptoms of depression and anxiety; they reported poorer quality of life and loneliness; and they were more worried about their future financial situation and more likely to experience greater financial difficulties than before COVID-19. The levels of infrequent contact with family and friends were slightly lower in people with probable COVID-19 infection than in those without, but these differences were not statistically significant (Table 1). Similar differences in mental health and financial hardship between participants with and without COVID-19 infection were also observed at the second COVID-19 assessment (SI -sTable2).

Sensitivity analyses (i) IPTW analyses
A comparison of the associations between COVID-19 infection and mental health, financial hardship, and social connections found in the IPTW regression analyses versus standard logistic/linear regression analyses is presented in sFigure2 (SI). The full statistical results of the IPTW regression models are reported in sTable13 (SI). The balance of the covariates across the exposure groups in the propensity score weighted and unweighted datasets is shown in sFigure1 (SI). SMDs for all covariates were lower than 0.1 in the weighted sample. .039) at both COVID-19 assessments, and they also were more likely to experience greater financial difficulties in June-July 2020 than before the start of the pandemic (ORassess1 1.40 [95%CI: 1.11;1.77] p=0.005). In contrast, no significant differences in the levels of financial worries and social contact with family and friends were observed between people with and without probable infection ( Figure 2; SI -sTable13).  , and limiting long-standing illness and including interaction effects between COVID-19 infection and sociodemographic factors; June-July 2020 = COVID-19 assessment 1; Nov-Dec 2020 = COVID-19 assessment 2; OR = odds ratio; CI = 95% confidence interval; FDR = false discovery rate.  , and limiting long-standing illness and including interaction effects between COVID-19 infection and sociodemographic factors; June-July 2020 = COVID-19 assessment 1; Nov-Dec 2020 = COVID-19 assessment 2; OR = odds ratio; CI = 95% confidence interval; FDR = false discovery rate.  , and limiting long-standing illness and including interaction effects between COVID-19 infection and sociodemographic factors; June-July 2020 = COVID-19 assessment 1; Nov-Dec 2020 = COVID-19 assessment 2; OR = odds ratio; CI = 95% confidence interval; FDR = false discovery rate.  , and limiting long-standing illness and including interaction effects between COVID-19 infection and sociodemographic factors; June-July 2020 = COVID-19 assessment 1 (June-July 2020); Nov-Dec 2020 = COVID-19 assessment 2 (Nov-Dec 2020); CI = 95% confidence interval; FDR = false discovery rate. sFigure1. Balance of the covariates in the treatment and control groups across the imputed datasets before and after propensity score weighting.  (N=5,146); pooled estimates from logistic regression models across 20 imputed datasets, weighted using survey weights; Models adjusted for sex, age groups, pre-COVID-19 outcome, Nov-Dec 2020 COVID-19 infection (Nov-Dec 2020 outcome only), whether living alone, employment status, wealth, whether vulnerable to COVID-19, limiting long-standing illness; June-July 2020 = COVID-19 assessment 1; Nov-Dec 2020 = COVID-19 assessment 2; OR = odds ratio; CI = 95% confidence interval.