Design and study population
A cohort study of university students in Stockholm, Sweden before the outbreak and during COVID-19 was conducted. The study is nested within a larger on-going dynamic cohort study of university students: the Sustainable University Life (SUN-study) (http://clinicaltrials.gov/ ID: NCT04465435).
Participants came from six selected universities in Stockholm, covering medical, economic technical and sport science educations (Table 1). Full-time undergraduate students with at least one year left to complete their degree were eligible for inclusion in the study. Data collection in the SUN-study started in August 2019 and will be ongoing until November 2021. Only participants completing their baseline questionnaire before March 13, 2020 were included in this study.
Students received information about the study through in-class presentations by study staff and were provided with access links to the study questionnaire via e-mail. Information about the study was also given in relevant social media channels (e.g., student union social media channels), and at on-campus information sites. Included students were followed with web surveys every three months over one year. Participants not responding to the follow-up surveys received reminders by email, phone text-message and one phone call over the following month. The study was approved by the Swedish Ethical Review Authority (reference number: 2019-03276, 2020-01449), and informed consent was provided by all participants electronically.
The collected data was divided into three time-periods: before the pandemic (August 19, 2019 to March 13, 2020), follow-up period 1 (FU1; March 14 to June 15, 2020) and follow-up period 2 (FU2; June 16 to September 10, 2020) (Figure 1, eFigure 1). The date marking the start of the pandemic, March 13, is the date that the Public Health Agency of Sweden declared that Sweden entered a new phase in the pandemic (9). Four days later, March 17, Sweden closed all on-campus education on universities switching to on-line education. The cut-point for FU2, June 15, was made since most university semesters had ended by then (eFigure 1).
Symptoms of depression, anxiety, and stress were measured with the short-form Depression, Anxiety and Stress Scale (DASS-21) (23) at each time-period. DASS-21 consists of 21 items rated on a 4-point scale ranging from 0 (Did not apply to at all) to 3 (Applied to me very much, or most of the time). Scores for the depression, anxiety and stress subscales are the sum of the 7 items in each subscale, ranging from 0-21 with higher scores indicating more severe symptoms. DASS-21 has good psychometric properties, with convergent validity for all subscales, and Cronbach’s α of 0.77-0.92 for the three subscales (24).
Pre-pandemic mental health problems was classified as scoring above the cut-offs for moderate symptom levels on any of the three subscales (≥ 7 on the depression subscale or ≥ 6 on the anxiety subscale or ≥ 10 on the stress scale) (25) on DASS-21 at baseline. These cut-offs are slightly above the cut-off point found to separate normal population from an outpatient psychiatric population (26)
Loneliness before the pandemic was measured using the UCLA Three-Item Loneliness Scale (27), consisting of three items rated from 1 (Hardly ever) to 3 (Often). The score of the scale is the sum of the three items. In accordance with previous research, we used a cut-off of ≥ 6/9 to define loneliness (28). The scale has acceptable internal consistency (Cronbach’s α= 0.72) and high correlation (r=0.82) with the 20 item Revised UCLA Loneliness Scale (27).
Sleep quality before the pandemic was measured using the Pittsburgh Sleep Quality Index (PSQI) (29). The scale consists of 19 items covering seven components of sleep with scores from 0-3. The global score is a sum of the components scores, ranging from 0-21. A global score of > 5/21 was used to classify poor sleep quality. This cut-off has shown a sensitivity of 89.6 % and a specificity of 86.5 % for differentiating between good and poor sleepers (30). The PSQI has adequate internal consistency (Cronbach’s α =0.82) and test-retest reliability (r=0 .82) over one month (30).
The statistical models including the exposures were adjusted for gender and age, which were selected a priori based on previous literature.
We used Generalized Estimating Equations (GEE) to model mental health symptoms during three time periods. GEE models treat correlation between observations from the same individual as nuisance parameters and provide estimates of the marginal population mean of the outcome. Our data was not normally distributed, which was one reason for choosing GEE since the model do not rely on the assumption of normally distributed outcome measures or the normality of residuals. We built three separate models, one each for symptoms of depression, anxiety, and stress, to assess overall mean differences in symptoms over the three time periods. These models included only time-period as the predictor. Since the models evaluated mean differences between time-periods for the full group, no covariates were used in these models. All GEE models, including the models described below, were specified with exchangeable working correlation structures, robust sandwich variance estimators and gaussian link functions.
Subsequently nine separate models were fitted to assess the differences in the trajectories over time by exposure levels. We dichotomized our exposures (loneliness, poor sleep quality or pre-pandemic mental health problems), into exposed and non-exposed groups. An interaction term between exposure and time-period was included, letting the differences in mean symptom level between exposed and non-exposed vary over time, giving an estimate of the difference-in-difference between mean symptom levels among exposed/non-exposed over time. These models were adjusted for age and gender.
Sensitivity analyses were conducted to examine the possible attrition bias by performing complete case-analyses (using data from participants who provided complete follow-up). We further conducted a sensitivity analyses in a sample of 496 participants followed before the pandemic, from August-September 2019 to November 2019-January 2020. This was done to compare the trajectories of exposed and unexposed groups during the pandemic to those of an earlier time-period (eFigure 2). Graphs (Figure 3 and eFigure 2) were compared by ocular inspection.
All analyses were performed using RStudio version 1.2.5001, the packages ‘geepack’ and ‘emmeans’ were used to perform GEE analyses, and to derive estimated marginal means from the models.
Mean imputation by the individual scale means was used to handle missing data on the PSQI that arised due to initial technical problems with the web survey. Three items of the scale were missing (5b, 5f and 5j) for the first 512 included participants. No other variables on any scales had missing values.