Study design and participants
Data came from a larger panel study with six repeated assessments of social interaction and emotional wellbeing conducted monthly from May to October 2020. The purpose of the study was to evaluate the mental health impact of social distancing guidelines imposed as part of the public health measures to limit the spread of COVID-19. Therefore, participant recruitment was limited to fourteen US states with active stay-at-home orders in May 2020, as residents in these areas were more likely to experience social isolation due to the restrictions. The states represented in our study were: Alabama, California, Oregon, Kentucky, Louisiana, Maine, Maryland, New Jersey, New York, Pennsylvania, Tennessee, Virginia, Washington, and Washington DC. Blacks, Hispanics, and individuals who had low household income (defined as less than $30,000 a year) were oversampled, since the pandemic disproportionately affected racial and ethnic minorities and the poor [39,40,41].
For the current report, we drew from the original study a subsample of adults aged 50 and over. As shown in Fig. 1, from May 5 through May 26, 2020, a total of 380 participants aged 50 and over were enrolled in the study. They completed an online baseline (Month 0) survey via Qualtrics. The original study by design introduced panel attrition due to funding restrictions. At each follow-up, only the participants who had responded to the previous survey were reinvited and recruitment was closed upon reaching enrollment. As a result, the number of older adult participants in the current analytic sample was 300 at Month 1, 272 at Month 2, 138 at Month 3, 145 at Month 4, and 159 at Month 5. Compared to their younger counterparts, older participants in the study were less likely to have an incomplete panel (i.e., missing one or more follow ups, 81% versus 88%, p < . 001) and on average completed more follow ups (3.67 versus 2.97, p < 0.001). These participants contributed a total of 1,394 repeated observations over the course of the study for analysis. The IRB at University of Pennsylvania determined that the study was exempt from review. All research was performed in accordance with relevant guidelines and regulations. Online written consent was obtained from all participants prior to beginning the survey.
Measurements of social interaction
We included both pandemic-specific (in-person interaction) and general (living arrangement, relationship quality and social support) metrics of social interaction. For living arrangements, we asked participants to first indicate how many people currently live in their home (excluding self) and then to specify their relationship to each of the persons in their home. Two dichotomized variables were created to indicate living arrangement—living alone (= 1, otherwise = 0) and living with spouse or partner (= 1, otherwise = 0). To measure in-person interaction, we asked them to indicate how many people from outside the household they had an in-person conversation with during the past month. Response categories include none (= 0), 1–2 per week (= 1), 3–4 per week (= 2), 5–6 per week (= 3) and 7 or more per week (= 4).
Family relationships and friendships were assessed separately. The participants were first asked how the quality of relationships with family members had changed during the past month and second, how the quality of relationships with friends changed during the past month. For each question, response categories included a lot worse (= -2), a little worse (= -1), about the same (= 0), a little better (= 1), and a lot better (= 2). Perceived social support was measured with an adapted version of the six-item brief version of the Perceived Social Support Questionnaire [42]: (1) I experience a lot of understanding and security from others; (2) I know a very close person whose help I can always count on; (3) If necessary, I can easily borrow something I might need from neighbors or friends; (4) I know several people with whom I like to do things; (5) When I am sick, I can without hesitation ask friends and family to take care of important matters for me; (6) If I am down, I know to whom I can go without hesitation. The original response categories for each item included strongly disagree, disagree, neutral, agree, and strongly agree and they were dichotomized (strongly agree = 1, otherwise = 0) due to a ceiling effect. Internal consistency of the recoded items ranged from 0.78 to 0.87 across months and a summed score of six dichotomized items were created.
Living arrangement, in-person interaction, and relationship quality were time-varying and measured consistently across six monthly assessments. Perceived social support was not measured at Month 2 due to limits in survey length. We used estimated trajectory of social support to impute values at Month 2 when examining its relationship with emotional wellbeing (see Analysis Plan below).
Measurements of emotional wellbeing
Emotional wellbeing during the pandemic was measured by isolation stress, COVID worry, and sadness. All three variables were time-varying and measured consistently across six monthly assessments. Isolation stress was measured by four items asking participants to evaluate the impact of pandemic-related physical restrictions [43]. Specifically, they were asked to indicate on a 5-point rating scale ranging from not at all (= 0) to extremely (= 4) for the following: During the past month. (1) how stressful have the restrictions on leaving home been for you? (2) how stressful have changes in family contacts been for you? (3) how stressful have changes in social contacts been for you? (4) how much has cancellation of important events (such as graduation, prom, vacation, etc.) in your life been difficult for you? Given good internal consistency of the items in the current analytic sample across monthly assessments (α’s = 0.77 to 0.82), a summed score was created to indicate overall level of isolation stress the participant experienced during the past month.
COVID worry was measured by a summed score of five items designed for this study assessing an individual’s disease worry, risk perception, and perceived controllability in the context of COVID-19: During the past month. (1) how worried have you been about being infected? (2) how worried have you been about friends or family being infected? (3) how worried have you been about your physical health being influenced by Coronavirus/COVID-19? (4) how worried have you been about your mental/emotional health being influenced by Coronavirus/COVID-19? (5) how much are you reading or talking about Coronavirus/COVID-19? For the first four items, response categories included not at all (= 0), slightly (= 1), moderately (= 2), very (= 3) and extremely (= 4) and for the last item response categories included never (= 0), rarely (= 1), occasionally (= 2), often (= 3) and most of the time (= 4). Internal consistency of the items was high across monthly assessments (α’s = 0.78 to 0.86).
Finally, sadness was measured by asking the participants to indicate how happy versus sad they were during the past month. Response categories include very happy/cheerful (= -2), moderately happy/cheerful (= -1), neutral (= 0), moderately sad/depressed/unhappy (= 1), and very sad/depressed/unhappy (= 2).
Measurement of covariates
We measured the participants’ age at baseline in years. Gender was measured by a dichotomized variable indicating female (= 1, otherwise = 0). Race and ethnicity were captured by three dichotomized variables indicating non-Hispanic White (= 1, otherwise = 0; reference group), non-Hispanic Black (= 1, otherwise = 0), and Hispanic (= 1, otherwise = 0). Education was measured by an ordinal variable indicating highest level of education completed, which included less than high school (= 0), high school diploma or GED (= 1), some college or 2-year degree (= 2), 4-year college graduate (= 3), and more than college (= 4). Participants reported on their annual household income by selecting one of the following categories: less than $30,000, $30,000 to $49,999, $50,000 to $74,999, $75,000 to $99,999 $100,000 to $249,999, more than $250,000. A dichotomized variable indicating low income was used in all multivariate analyses (< $30,000 = 1; otherwise = 0). We also noted if the participant was currently working for pay (= 1, otherwise = 0). Because enrollment was capped at a lower number at Month 3 through Month 5, we adjusted for potential panel attrition bias with a dichotomized indicator of incomplete panel (i.e., completed fewer than six monthly assessments) included in all models as a covariate. In preliminary analyses (not shown), we compared this method to a correction instrument calculated from a two-stage Heckman selection bias model [44] and found the two methods do not produce substantively different results. We opted to use a dichotomized indicator to adjust for panel attrition in the final models as this method did not rely on accurate specification of the selection bias, which was difficult to ascertain in a panel study spanning over a relatively short period of time.
Statistical model
To examine intra-individual change in social interaction and emotional wellbeing across repeated observations and the associations between them, we used random-intercept logistic regression models to estimate binary outcomes (living alone and living with spouse or partner), random-coefficient proportional odds models to estimate ordinal outcomes (in-person interaction, quality of relationship with family and with friends, and sadness), and random-coefficient growth curve models with a Gaussian link to estimate continuous outcomes (social support, isolation stress, and COVID worry). These random effects models explicitly model longitudinal dependency with individual-specific trajectories of change. All models used month as the time variable. Since change over time could be nonlinear, we tested in preliminary analysis linear, quadratic, cubic, and quartic specifications of time. Best fitting model specification for each outcome was determined by statistical significance of the coefficient(s) and Bayesian Information Criterion (BIC) and used for final analysis. Specifically, we modeled change in living arrangement, in-person interaction, quality of relationship with family and with friends, and sadness as linear function of time and change in perceived social support, isolation stress, COVID worry as a cubic function of time.
Analysis plan
The analysis proceeded in three stages. First, we described the sample with regards to sociodemographic characteristics and panel attrition. Second, we described change in social interaction and emotional wellbeing over six months during the pandemic, by estimating time-based trajectories of the variables without adjusting for covariates. From these models we extracted estimated probabilities (for binary or ordinal variables) or values (for continuous variables) and plotted them over time to visualize the time trend. Third, we examined the associations between social interaction variables, as time-varying covariates, and time-based trajectories of emotional wellbeing, adjusting for covariates. Estimated values of perceived social support were extracted from the trajectory model in stage 2 and used to impute social support at Month 2 in models estimating trajectories of emotional wellbeing. Due to difficulty interpreting an interaction between a continuous time-varying covariate and time, we created a dichotomized variable representing higher levels of perceived social support (sum social support score = 6) to examine how social support affected the rates of change in trajectories of emotional wellbeing. In preliminary analysis, we tested interactions between social interaction variables and time, to determine how each social interaction variable affected the trajectory of change in emotional wellbeing. Only statistically significant interactions were retained in the final models and presented in the results.