Lifestyle and mental health disruptions during COVID-19

We enrolled three cohorts of students from the University of Pittsburgh in the study: spring 2019, fall 2019, and spring 2020. The study was approved by the University of Pittsburgh Institutional Review Board and was preregistered in the American Economic Association Randomized Controlled Trials (AEA RCT) Registry (RCT ID AEARCTR-0003235). Data and materials can be accessed at Open Science Framework (https://osf.io/f85e3/) (33). A detailed description of the methods and measures can be found in SI Appendix.

At the beginning of each semester, we invited college students at the University of Pittsburgh to participate in a semester-long experiment on wellness. Eligible participants signed a consent form in the laboratory at the beginning of the study. They then filled out a baseline survey, received a wearable tracker (a Fitbit Alta HR device), and installed a custom-made smartphone application on their phone which allowed us to track their Fitbit data.

Throughout the semester we continuously collected daily Fitbit data, which measures steps, physical activity, and sleep based on heart rate and movement. We also collected weekly measures of time use through a diary survey following the structure of the American Time Use Survey (34).^†

We measured mental health at the beginning, middle (spring 2020 only), and end of the semester. Our primary measure of mental health is depression, which we assessed using the Center for Epidemiologic Studies Depression Scale (CES-D) (35). The CES-D is a validated self-report instrument designed to assess the frequency of symptoms of depression, such as helplessness or loneliness, on a scale from 0 (rarely or none of the time) to 3 (most or all of the time) and has a total score between 0 and 60. Our primary benchmark for depression is a CES-D score of 16 or above, which is considered the cutoff for clinical concern, implying high levels of depressive symptoms (36). We additionally assessed anxiety, resilience, and life satisfaction.

In the spring 2020 cohort we continued to track a subsample of participants who agreed to continue their participation after the semester ended in April 2020. In June 2020 we randomized half of the participants to an intervention group to increase physical activity and half to a control group for a 2-wk intervention period. This RCT was preregistered in the AEA RCT Registry (RCT ID AEARCTR-0005949). We measured mental health in May just before the intervention, mid-June at the end of the intervention, and mid-July a month after the intervention ended.

Our sample includes all participants for whom we have a baseline survey, including baseline mental health measures: spring 2019 (n = 150), fall 2019 (n = 315), and spring 2020 (n = 316). ^$‡$The combined cohorts include $N=682$ unique participants. In SI Appendix, Fig. S.1 we report the sample sizes at each stage of the study.

We present descriptive statistics for our sample in SI Appendix, Table S.1. While our sample is not nationally representative, as we noted above, our measures of baseline mental health are in line with estimates from representative surveys. We present estimates reweighted to match a nationally representative sample on gender, age, and race/ethnicity and the results do not change (SI Appendix, Table S.11).

The main analyses include all participants who have at least one observation for the relevant outcome. We examine attrition directly and also conduct several sensitivity checks to address attrition concerns (SI Appendix, section 4B). Unless noted otherwise below, our results are robust to these sensitivity analyses.

Our biometric and time use measures reveal that the pandemic led to major disruptions in daily behavior. Fig. 1 plots average daily physical activity and sleep across the semester for the spring 2019 and spring 2020 cohorts.

In the spring 2019 cohort, daily steps are fairly constant, with an average of about 10,300 to 10,400 steps throughout the term (Fig. 1A). At the beginning of the semester (February), the spring 2020 cohort is statistically indistinguishable from the spring 2019 cohort. In March, there is a sharp drop in the average number of steps from 10,000 to 4,600, a more than 50% decline (P $\hspace{0.17em}<\hspace{0.17em}$0.001, from a regression of difference-in-differences across cohorts, SI Appendix, Table S.2, Panel B). We observe a similar pattern for physical activity (Fig. 1B), which is measured as minutes in which a person is nonsedentary for at least 10 continuous minutes, where nonsedentary minutes are defined as activity that raises heart rate enough to burn at least 1.5 times as many calories as at rest. Time spent in active (nonsedentary) activities dropped by about 1.5 h from 4.4 h to 2.9 h per day. The decline in active hours throughout the term is 10 times larger than in 2019 ( $P<0.001$; SI Appendix, Table S.2, Panel B). These results are robust to corrections for attrition and incomplete syncing (SI Appendix, section 4B).

We also find disruptions to sleep habits (Fig. 1C), as students started to sleep about 25 to 30 min more per night throughout the pandemic ( $P<0.001$; SI Appendix, Table S.2). As shown in Fig. 1D, the increase in sleep is driven by later wake-up times: Average wake times shift by 30 to 40 min from about a quarter after 9 AM to a little before 10 AM ( $P<0.001$; SI Appendix, Table S.2). There is little change in bedtimes (SI Appendix, Fig. S.5). Previous studies document that misalignment of sleep timing with respect to the natural dark–light cycle may have detrimental effects on sleep quality, health, and depression (37–40). Thus, the later timing of sleep during the pandemic may have contributed to mood disorders or exacerbated depression symptoms in individuals predisposed for mental health disorders. We note, however, that the estimated impacts of the pandemic on sleep are less robust than those for active hours and steps. There are baseline differences in the sleep habits of our cohorts and so they may not be as comparable, and the estimates are sensitive to corrections for attrition (SI Appendix, Table S.7). The imbalance on sleep across cohorts may also bias our estimates of the impact of the pandemic on physical activity.

We next examine shifts in self-reported time use. Fig. 2 shows average daily social interactions and screen time in February compared to April for the spring 2019 and spring 2020 cohorts. In the 2020 cohort, screen time—time spent watching TV, playing video games, or surfing the internet outside of work or studying—more than doubled after the announcement that classes would be moved remotely, reaching an average of 5 h per day at the end of the term. By contrast, screen time averaged around 2 h per day throughout the semester in 2019 with only a moderate decrease at the end of the term ( $P<0.001$; SI Appendix, Table S.2, Panel C). We also observe a substantial drop in the number of hours spent interacting with friends, from approximately 1.5 h per day at the beginning of 2020 to less than 30 min per day at the end of April, a more than 50% decline ( $P<0.001$; SI Appendix, Table S.2, Panel C). This drop is consistent with self-reported declines in face-to-face interactions (SI Appendix, Fig. S.6). In SI Appendix we also document a drop in the number of work hours—driven by a subset of our participants who lost their jobs as a result of the campus closure—and a significant drop in the number of hours spent studying in the second half of the semester (Table S.2, Panel C).

Our primary measures of mental health are assessments of depression using the CES-D scale. Fig. 3 shows average CES-D scores for the spring 2019 and spring 2020 cohorts. We present the baseline measures taken at the beginning of the semester (February), the midsemester measures taken in March (spring 2020 cohort only), and the end-of-semester measures taken in April.

Our results show large increases in depression during the pandemic. In the spring 2019 cohort, we estimate a 1.5-point increase in average CES-D scores over the term from 12.4 at the beginning of the semester to 13.9 at the end of the semester. The spring 2020 cohort has very similar scores at baseline, averaging 12.1 at the beginning of the semester. However, the estimated increase in scores across the term is over four times larger than in spring 2019. We estimate that average CES-D scores increase from 12.1 to 19.5, a more than 60% increase ( $P<0.001$ from a difference-in-differences regression across cohorts; SI Appendix, Table S.2, Panel A).^§ .

As shown in SI Appendix, we find directionally similar results when we look at anxiety using the Generalized Anxiety Disorder (GAD-7) scale (SI Appendix, Table S.2 and Table S.3, Panel A).

The pandemic shifted the distribution of CES-D scores with a substantial increase in the share of subjects with a CES-D score above 15, the threshold commonly used to identify clinical depression (SI Appendix, Fig. S.4). By the end of the semester in April 2020 we estimate that 61% of our participants were at risk for depression, about a 90% increase over the baseline rate of 32% just 2 mo earlier prior to COVID-19 ( $P<0.001$). By comparison, these same rates increase by only 6 percentage points over the course of the spring 2019 semester from 31% to 37% ( $P=0.30$). Using Lee bounds (41), we estimate a difference-in-difference increase in depression rates of 15 to 24 percentage points in spring 2020 compared to spring 2019 ( $P<0.001$; SI Appendix, Table S.7).^¶

About 80% of the increase in CES-D scores occurs by the time of the midline survey in March 2020 (20 March). This pattern aligns with the timing of lifestyle disruptions shown in Fig. 1, in which, for example, average daily steps largely decline within a 2-wk period in the middle of March and then plateau.

Building on the analysis above, we examine the link between the large disruptions to behaviors during the pandemic and endline depression in 2020. We measure lifestyle changes as the difference between average behavior during the onset of the pandemic in March and April minus average baseline behavior in February. Fig. 4 compares rates of depression among participants with smaller and larger disruptions using below-vs. above-median changes in physical activity, sleep, and time use.^#

Larger declines in physical activity are associated with significantly higher rates of depression. There is a 15 to 18 percentage point gap in depression rates between participants who experience large disruptions and those who maintain their baseline habits ( $P=0.012$; SI Appendix, Table S.17). These results are robust to Anderson (2008) corrections (42) for multiple hypothesis testing ( $P=0.036$). We additionally examine the relationship between students’ location during lockdown and endline depression (by the end of the term approximately 79% of students were not in Pittsburgh). We find no evidence of an association between depression and locations with higher COVID-19 cases or deaths (SI Appendix, Table S.19).

To better understand the dramatic rise in depression during COVID-19, we combine our rich data to identify risk factors for depression in the spring 2020 cohort. We then compare those predictors to prior cohorts (pooling the spring 2019 and fall 2019 cohorts). We focus on risk factors for having an end of semester CES-D score that meets or exceeds 16, the threshold for clinical depression (36).

Building on the above results suggesting the importance of lifestyle disruptions, we focus on a “differences” model that uses changes in lifestyle measures (physical activity, sleep, and time use) along with baseline measures of mental health and demographics. We feed these variables as potential features into the XGBoost machine-learning algorithm (43), a flexible and robust decision tree-based classification method. The pooled 2019 and spring 2020 models achieve 89% and 91% predictive accuracy, respectively (i.e., the overall percentage of observations that are correctly predicted by the model). See SI Appendix, section 5 for the full list of variables and a detailed description of the prediction exercise.

Before pooling the 2019 cohorts we first estimated the model for each cohort: spring 2019, fall 2019, and spring 2020. We then examined the accuracy of the spring 2019 model for predicting endline depression in fall 2019 compared to predicting endline depression in spring 2020. We find that the spring 2019 model is significantly more accurate for fall 2019 than for spring 2020 ( $P<0.001$). These results suggest an overall shift in risk factors for depression during the pandemic compared to prior cohorts.

We find further support for this hypothesis when comparing the specific risk factors for depression across cohorts. Fig. 5A reports the cumulative importance—adding up to 1—of the different features for the pooled 2019 and spring 2020 cohorts, grouped by category. Below the figure we list the three most important features for each model along with their relative importance, which approximates the average gain in predictive accuracy from using that feature in the model.

There are substantial shifts in the importance of each category across cohorts. For the 2019 cohorts, who participated prior to the pandemic, baseline measures of mental health at the beginning of the semester largely explain depression rates at the end of the semester, with baseline depression by far the leading factor, accounting for an estimated 36% of the predictive accuracy of the model. The cumulative importance of the baseline mental health measures declines during the pandemic from 0.57 in 2019 to 0.31 in 2020, though not all of the measures move in the same direction. Baseline measures of depression and life satisfaction decline in relative importance in 2020, while baseline measures of anxiety and resilience increase in relative importance compared to prepandemic cohorts.^∥ Our results suggest that those most resilient to stress and least susceptible to anxiety may be especially protected against depression during the pandemic (SI Appendix, Fig. S.15). This is in line with work suggesting that resilience protects individuals against stressful events (27) and helps them preserve their health despite adversity (44).

Differences between endline and baseline lifestyle behaviors (physical activity, sleep, and time use) become more important in 2020: Their cumulative importance increases from 0.34 in 2019 to 0.58 in 2020. The increased importance of these measures suggests that the pandemic has tightened the link between between lifestyle behaviors and mental health. In particular, disruptions in physical activity emerge as a critical predictor of depression during COVID-19, increasing in estimated relative importance from 0.026 to 0.114.

The importance to well-being of maintaining physical activity is illustrated in Fig. 5B. The figure displays estimated Shapley Additive Explanations (SHAP) values by differences in physical activity. A SHAP value approximates the marginal contribution of a feature to a particular observation’s predicted risk, where a higher SHAP value indicates a higher risk of endline depression. In the 2020 cohort, depression risk increases substantially with larger declines in daily active hours, while in 2019 the relationship is largely flat. The high-risk group in 2020 experiences a decline of about one to three fewer daily active hours (around the 2020 average of 1.5 h), with disruptions of such magnitude largely absent in the 2019 cohort. Importantly for the 2020 cohort, those participants who maintain daily active hours similar to baseline (i.e., differences near zero) demonstrate strikingly lower risk of endline depression. These results suggest that sustaining healthy physical habits is strongly associated with well-being during the pandemic.

We find similar results if we exclude baseline mental health measures as potential risk factors: The importance of lifestyle behaviors increases in 2020 and disruption to physical activity is the leading predictor of endline depression during COVID-19 (SI Appendix, Fig. S.18). We also report the results of “baseline” models that use baseline measures of lifestyle habits rather than differences (SI Appendix, section 5C). In line with our main results, the importance of baseline lifestyle behaviors increases in 2020 compared to 2019. Also, some of the typical relationships between baseline habits and endline depression diverge during the pandemic. For example, whereas walking the recommended 10,000 steps per day minimizes depression risk in prepandemic cohorts, these same baseline activity levels are associated with increased risk of depression during the pandemic.**

In the 2020 cohort we continued to track a subsample of our participants after the semester ended in April. These participants (n = 205) filled out a consent form, agreed to continue wearing the Fitbit, and completed weekly time-use surveys as well as mental health surveys in May, June, and July. We find that in May lifestyle behaviors—and physical activity in particular—demonstrate a “bounce-back” moving toward baseline levels. As shown in Fig. 6A, we find that daily steps increase from 4,600 in April to 6,400 in May, which closes a third of the decline from baseline ( $P<0.001$; SI Appendix, Table S.25). We also observe small decreases in sleep duration of about 10 min ( $P<0.05$) and a small increase in social interactions of about 10 min ( $P=0.10$), while screen time continues to increase ( $P<0.01$; SI Appendix, Table S.25).

In June 2020 we implemented a randomized intervention to further stimulate physical activity among our participants. We randomly assigned participants to receive incentives for walking a minimum of 10,000 steps a day. Participants in the treatment group received a monetary transfer of $5 every day they reached the minimum number of steps. The control group received a similar distribution of payments (see SI Appendix for experimental procedures). The intervention began on 1 June and lasted 14 consecutive days. After the end of the intervention, on 16 June, we measured mental health again. Then, we continued to track the subjects through July and measured mental health on 17 July.

As shown in Fig. 6A, the intervention had a large impact on physical activity, increasing average steps by about 2,300 steps ( $P<0.001$) and active minutes by almost 40 min ( $P<0.001$; SI Appendix, Table S.26 and Fig. S.19). As a result, daily steps in the treatment group approached baseline prepandemic levels, averaging approximately 9,000 steps per day. On average, participants in the treatment group met the step goal on 50% of days vs. 16% of days in the control group ( $P<0.001$; see SI Appendix, Fig. S.20 for the distribution).

However, as shown in Fig. 6B, we find no effect on CES-D scores measured at the end of the intervention period in mid-June. We estimate a difference between treatment and control of −0.30 points ( $P=0.85$; SI Appendix, Table S.27, column 1). Our 95% confidence intervals exclude that treatment caused more than a 3.4-point change in the CES-D score, which is about 45% of the increase in average CES-D scores observed at the onset of the pandemic.^††

We continued to track participants for a month after the intervention ended. As shown in Fig. 6A, after the intervention ended physical activity in the treatment group declined within about a week to the control group’s levels. In the week after the intervention, we estimate a difference of about 1,200 steps between the treatment and control group ( $P<0.001$). When we examine the month-long postintervention period as a whole we do not find a significant difference in steps between the treatment and control groups ( $P=0.32$; SI Appendix, Table S.28, column 1). In the follow-up survey of mental health a month after the intervention ended we estimate a difference between the treatment and control group of 0.60 points ( $P=0.70$; SI Appendix, Table S.29, column 1). Our 95% confidence intervals exclude that treatment caused more than a 2.4-point decline in the CES-D score in July, less than a third of the increase in average CES-D scores observed at the onset of the pandemic.

In preregistered heterogeneity analysis we examine treatment effects within the following subgroups: preintervention depression (CES-D score above 15 vs. 15 or below) and preintervention recovery in physical activity, where we split participants by above/below median improvement in physical activity during the “bounce-back” period in May.

We find suggestive evidence that the impact on physical activity is larger and more persistent for participants who experienced smaller improvements in physical activity prior to the intervention. We estimate that during the intervention period, treatment increased their physical activity by almost an hour a day compared to about 30 min for participants who experienced larger bounce-backs in May ( $P<0.001$ compared to control, $P=0.012$ comparing effects across subgroups; SI Appendix, Table S.26, column 8).

In the postintervention period, those with limited recovery prior to the intervention continue to experience treatment impacts, with an estimated average increase of about 30 min per day over the month ( $P=0.044$ compared to control, $P<0.001$ comparing effects across subgroups; SI Appendix, Table S.28, column 8). Point estimates for the impact on steps in both the intervention and postintervention periods are consistent with the results for physical activity but are less precisely estimated.

Despite the sustained impact of the intervention on the behavior of these participants, we do not find evidence of short or longer run effects on their mental health (SI Appendix, Tables S.27 and S.29, columns 3 and 8). In our second subgroup analysis split by preintervention depression status we find no evidence of differential treatment effects on either physical activity or mental health (SI Appendix, Tables S.26–S.29, columns 4 and 9).

We note that in both the control and treatment groups, average CES-D scores improved in the period from May to July compared to April. Estimated depression rates in May, June, and July were 50%, 46%, and 48%, respectively. While this represents a significant improvement from rates of 61% in April, depression scores plateau in July and remain about 50% higher than prepandemic levels. The improvement and then plateauing in CES-D scores coincides with the bounce-back and then plateauing of physical activity from May through July 2020. Our short-term intervention successfully counteracts the plateauing of physical activity during the intervention period but has no meaningful effect on mental health.