The most burnt-out residents were the lightest users, not the heaviest

The intuition is that the resident in trouble is the one who cannot put the phone down at three in the morning. In this survey of 3,621 Chinese residency trainees, she was not. The heaviest users did screen positive most often, but the worst depression, anxiety, insomnia and burnout scores in the entire sample sat with the lightest users, who also crossed the screening threshold more often than the group just above them.

Who cleared the threshold

The instrument is the Bergen Social Media Addiction Scale. Six items, one for each element of the components model: salience, mood modification, tolerance, withdrawal, conflict, relapse. Each is rated 1 to 5, so the total runs 6 to 30, and the authors used the published cut-off of 24. That is a demanding threshold. To reach it a respondent has to endorse nearly every component at "often" or "very often". The vocabulary was built for substance dependence and has been carried over wholesale to an app.

At that threshold, 211 of 3,621 residents screened positive, 5.8%. The authors set this against a pooled global BSMAS estimate of roughly 8% (95% CI 4%-12%) and against 3.49% reported in Chinese adolescents under identical criteria. The width of that global interval deserves a pause: 4% to 12% is a threefold spread for one scale at one cut-off, before anyone moves the cut-off. The group means say the same thing from the other side. Mean BSMAS was 13.07 (SD 5.57) among those using social media an hour or less per day and 16.90 (SD 5.36) among those using it more than four hours. Both sit far below 24. The average heavy user is not near the threshold at all.

Two things in this paper do not agree with each other, and both matter. The composition first: Table 1 reports 1,567 men (43.2%) and 2,054 women (56.8%), but the sex cells of Table 2 sum to 1,694 men and 1,927 women. Table 2's column totals are internally correct and reconcile to 3,621, yet its percentages are computed on Table 1's marginals, so the male row percentages sum to 108.2% and the female row to 93.9%. The narrative sex comparison, 1,233 of 2,054 women against 728 of 1,567 men using two hours or more, inherits the same mismatch; the direction survives either denominator, the exact percentages do not. A smaller version of the same problem affects only-child status: 2,173 in the text, 2,139 in Table 1, and 2,144 when Table 2's cells are added. Where they disagree I have followed the tables, and where the two tables disagree I have used the cells that reconcile to the column totals.

The group nobody screens

Then the shape of the dose relation, which is the more consequential disagreement. Screen-positive rates run 4.7%, 3.7%, 5.9%, 6.8%, 12.2% across the five usage bands. The dip at the low end is not a rounding artefact to be waved past, and it is instructive that the paper's own Discussion quotes the ascending range as 3.7% to 6.8% up to four hours against 12.2% above it, quietly leaving the lightest band out of the sequence. The lightest users were also the most unwell people in the sample: PHQ-9 9.54 against 6.4 in the three-to-four-hour group, GAD-7 7.07 against 4.39, Insomnia Severity Index 9.99 against 6.84, burnout 2.64 of 5 against 2.21. A doctor too depleted to open an app is not a low-risk doctor. She carries 6.47 night shifts a month, and the paper's own account of this group is digital disengagement under acute stress.

So stop treating hours as the clinical variable. The continuous score did climb smoothly with duration here, 13.07 to 16.90, but the proportion crossing the threshold did not. Why it did not is something the paper never addresses, because it denies the non-monotonicity outright. My own reading is that the lightest-use band also carries the widest spread of scores (SD 5.57), and at a fixed cut-off a wider spread pushes more people past it for the same mean. That is a hypothesis and not a result: a group mean and a standard deviation do not determine a tail proportion. Duration moves the middle of the distribution; it does not tell you who sits in the tail. A patient who reports six hours a day has told you a fact about their week, not a diagnosis. The six components carry the diagnostic load, and asking about them takes about a minute.

What held up was the workload variable. Each point of self-rated burnout on a five-point scale corresponded to 1.409 times the odds of screening positive (95% CI 1.226-1.617, p < 0.001), and that association was stable across all three residency years (year 1 OR 1.327, year 2 OR 1.442, year 3 OR 1.391). A psychiatric history went with twice the odds (OR 2.000, 95% CI 1.406-2.838). Daily use duration was significant in the pooled model (OR 1.388, 95% CI 1.233-1.562) but fell out of every year-stratified model (all p > 0.100). None of this establishes direction: burnout may drive compulsive use, compulsive use may deepen burnout, or a third state may produce both. What the design does support is a screening habit. In an exhausted professional with a psychiatric history, ask about the components rather than the clock, and treat unusually low engagement as worth one question rather than assuming that only the high end signals trouble.