Mental Health and Behavior of College Students During the Early Phases of the COVID-19 PandemicLongitudinal Smartphone and Ecological Momentary Assessment Study

Imagine you’re already inside a live experiment when a global shock hits. That’s what happened at Dartmouth. Huckins and colleagues had been following a cohort of undergraduates for two years, quietly collecting a dense stream of data about their daily lives and moods. Then Winter 2020 arrived. COVID-19 started to rewrite campus life. Because the study was already in progress, they could watch—almost in real time—how behavior and mental health shifted within the same people. Here’s the setup. Two hundred nineteen students agreed to smartphone sensing; after one incompatible device and one early withdrawal, 217 were in the analyses. The group skewed female—147 students, or about sixty-seven point eight percent—and they were 18 to 22 years old at enrollment. Dartmouth runs on a four-term calendar, roughly 10 weeks per term with a break in between. Winter 2020 began on January sixth. That timing created a natural hinge: earlier terms and breaks form each student’s baseline, and Winter 2020 becomes the stress test. What makes this powerful is the texture of the data. Phones quietly logged GPS, accelerometer, and lock and unlock status through the StudentLife app. That turned raw signals into daily behaviors: how long someone stayed stationary, a proxy for sedentary time; how far they traveled; how many distinct places they visited; how long their phone stayed unlocked; and inferred sleep duration from light, movement, and screen cues. Location traces were clustered with a method called Density-Based Spatial Clustering of Applications with Noise, or DBSCAN, to count meaningful places. On the mood side, there were weekly ecological momentary assessments using the Patient Health Questionnaire-4, or PHQ-4, which is a brief mental health screen. It bundles the Patient Health Questionnaire-2, or PHQ-2, for depressive symptoms and the Generalized Anxiety Disorder 2-item, or GAD-2, for anxiety, each scored from zero to six. Alongside all that, the team pulled a daily index of COVID-19 media coverage from Media Cloud—the ratio of coronavirus stories to all articles—so they could ask whether information in the environment tracked what students were doing and feeling. Coverage was solid, which matters when you’re trying to see within-person changes. Location data were available for a mean of nineteen point seven hours out of 24, other sensing indices for twenty-two point three hours out of 24, and students completed eighty point one percent of their weekly mood check-ins. It’s not perfect, but it’s enough to see patterns rise above noise. Now, how do you separate the jolt of a pandemic from the ordinary heartbeat of a term, with its midterms, finals, and breaks? The team leaned on mixed-effects linear models. In plain language, they modeled weekly outcomes—sedentary time, anxiety, depression—as functions of where you are in the term and whether that term is COVID-affected, while giving each student their own baseline. Term week was included as a linear slope and, when needed, a bend, or a quadratic, to capture the curve of stress rising and falling across the quarter. A binary COVIDTerm flag marked Winter 2020 as different, and they allowed that difference to change the weekly trajectory by including interactions between COVIDTerm and those term-week terms. They compared model structures by how well they fit the data and used standard methods, like Satterthwaite’s approach, to compute p-values. They also stepped down to a daily lens to link behavior with the media environment. That model asked: on days when COVID-19 dominated the news a bit more, how did students’ mobility, screen use, and sleep look? They included fixed effects for unlock duration, number of unlocks, sedentary time, sleep duration, and locations visited, plus the linear and quadratic term-week controls and random intercepts for each person. Then they layered in the weekly anxiety and depression scores to see whether the association between news and behavior also showed up in self-reported mood. All continuous predictors were scaled so coefficients could be compared on the same footing. So what happened when COVID-19 hit? Things moved, and they moved together. Of the 217 students in the sample, 178—about eighty-two percent—contributed data in Winter 2020. Compared with earlier terms and the usual post-finals lull, Winter 2020 brought clear increases in sedentary time and in both anxiety and depression. Those weren’t just the familiar finals jitters. In the mixed models, the shifts during the COVID-affected weeks exceeded the typical term rhythm, with p-values below 0.001 for sedentary time, anxiety, and depression when compared to prior terms and to the following break. On the ground, that looked like this: students stayed put more often, visited fewer locations as inferred from GPS clusters, and kept their phones unlocked longer. Interestingly, the raw number of unlocks initially dipped, which sounds like less phone use at first glance. But Huckins and colleagues point out a simple explanation—if you’re keeping the screen on for longer stretches, you need fewer unlocks to rack up more screen time. When they added the media index to the mix, the pattern sharpened. Days with more COVID-19 coverage saw greater sedentary behavior, fewer places visited, longer phone-use durations, and altered sleep. When self-reported mood entered the model, anxiety tracked the news signal strongly, with a p-value below 0.001. Depression still moved with it, but more modestly—significant at a p-value of 0.03. Timing matters here, too. The local pandemic arrived during week nine of the term. By week ten, as campus, local, and national policies changed almost overnight, the trajectories bent. Anxiety rose, depressive symptoms climbed, sedentary time ticked up, and mobility dropped, not for a day or two but across the closing weeks of the term and into the first two weeks of break. Normally, mood spikes around finals and then eases when the academic pressure lifts. This time, those elevations didn’t simply melt away. That persistence is the tell. Why all this fuss about modeling term week with lines and curves? Because the undergraduate year has its own seasonality. Everyone feels the squeeze before exams; everyone exhales on break. If you ignore that, you risk calling the ordinary extraordinary. By explicitly modeling both the level shift in COVID-affected terms and the way the week-to-week slope changed under COVID-19, the authors could say, with some confidence, that Winter 2020 deviated from the usual script. Let’s pause on what the media index contributes, because causality is tricky. The media ratio—coronavirus stories divided by total stories—isn’t a measure of what any one student read. It’s a barometer of the information weather. In the daily models, that barometer correlated with behavior even after accounting for where you were in the term and how much you were sleeping, moving, and using your phone. When weekly PHQ-2 and GAD-2 scores were added, anxiety still stood out as closely linked to the news signal. That doesn’t prove the news made anyone anxious. It does show that as the information environment grew more COVID-saturated, students’ behavior and reported anxiety moved in tandem. Of course, smartphones are imperfect mirrors of life. If you leave your phone on a desk while you pace your room, the app might read stillness as sedentary time. If you switched from your phone to a laptop during lockdown, the system undercounts your screen hours. And this is a particular group: Dartmouth undergrads with compatible iOS and Android devices, mostly 18 to 22 years old, roughly two-thirds female. The numbers are not small—217 students total, with strong daily coverage and eighty point one percent completion of weekly mood checks—but they’re not a census of young adults everywhere. The sensing itself has limits. GPS clustering with DBSCAN is a smart way to define "places" from thousands of points, but it smooths nuance. Two coffee runs on opposite ends of the same block might be one location in the data. Weekly EMAs provide clinically meaningful snapshots of anxiety and depression, but because they’re weekly, the daily news models and the mood scores don’t always line up in time. The authors are frank about these caveats, and they matter for generalizing beyond a campus in New Hampshire in early 2020. Still, the through-line is hard to miss. As COVID-19 encroached, students moved less, went fewer places, kept their phones on longer, and reported worse anxiety and depression compared with their own prior terms. Those changes bent in concert with the media environment and did not snap back immediately after finals ended. That convergence—behavioral sensing, self-reported mood, and the information landscape all pointing in the same direction—makes the story compelling. There’s also a methodological lesson tucked in here. Dense, longitudinal, within-person designs let you ask sharper questions in messy real life. By anchoring each student to their own baseline across multiple terms and by modeling the normal ebb and flow of a quarter, Huckins and colleagues could detect a pandemic-specific deviation rather than mistake it for the usual crunch time blues. By pairing that with an external, daily measure of COVID-19 coverage, they could situate those deviations in the broader information context without pretending headlines equal personal exposure. What should we take away, beyond the particulars of Winter 2020 at Dartmouth? First, that passive sensing plus brief mood check-ins can serve as early indicators of population mental health during fast-moving crises. You don’t need perfect measures to see a reliable bend in the curve. Second, that anxiety may be especially sensitive to the drumbeat of crisis news, showing a tighter link than depression in this dataset. And third, that behavior—how much we move, where we go, how long we keep a screen on—can be an unobtrusive proxy for how we’re coping. Closing on a modest note is the right thing here. The sample is specific, the measures are proxies, and causality remains complicated. But the design shows what’s possible: real-time windows into mental health that respect the rhythms of everyday life. If you were to build on this, you’d want richer content signals—distinguish doomscrolling from a video chat with friends—and wearable data that better capture movement when the phone is on the nightstand. You’d also want broader, more diverse samples. In the meantime, the headline holds. During the first visible term of the pandemic, students became more sedentary, went fewer places, used their phones longer, and felt more anxious and depressed than in their own prior terms, with changes that rose and fell with the pandemic’s presence in the news. That’s not the whole story of COVID-19 and mental health. But it’s a clear, measured chapter, written in the rhythm of an academic term and the glow of a lock screen.