What Best Predicts Depression and Anxiety in Teenagers?
A survey of 1,088 teens put eight suspects side by side. Struggling with emotions and loneliness beat social media, sleep and exercise in all four outcomes.
In one paragraph
Researchers surveyed 1,088 secondary school students in Wuhan, China, and used machine learning to rank eight factors linked to depression, anxiety, stress and well-being. The best model explained 49% to 60% of the variation. Difficulty managing emotions and loneliness ranked highest across all four outcomes, academic stress led for stress, and social media mattered mainly at high use. It is a one-time survey in one city, so it shows associations, not causes.
How did researchers rank the causes of teen depression and anxiety?
The team surveyed 1,088 students aged 13 to 18 at six secondary schools in Wuhan, China. Students answered standard questionnaires on depression, anxiety, perceived stress and well-being, and on eight possible factors: difficulty managing emotions (called emotional dysregulation), loneliness, social media addiction, self-esteem, sleep quality, academic stress, family support and physical activity.
Four kinds of model, from simple linear regression to a method called XGBoost, learned from 80% of the students and were then tested on the remaining 20%, so the scores reflect students the models had never seen. A technique called SHAP then estimated how much each factor pushed each prediction up or down.
Where a factor's questions overlapped with an outcome's questions, it was left out of that model. For example, the depression questionnaire asks about sleep, so sleep quality was excluded from the depression model.
Which factors mattered most for teenagers' mental health?
XGBoost predicted best for every outcome. On the test students it explained 49% of the variation in depression, 55% in anxiety, 60% in perceived stress and 50% in well-being.
Two factors stood out everywhere: emotional dysregulation and loneliness. They stayed among the top three in each of the five ways the data were split. Beyond those, each outcome had its own pattern. Academic stress ranked first for perceived stress, and well-being went with higher self-esteem and stronger family support.
Social media addiction had a nonlinear effect: its contribution grew mainly at high levels of use. Physical activity, age and gender added little. Another model type, Random Forest, fitted the training students closely but dropped much more on new ones, a sign of overfitting.
What does this mean for parents and schools?
The two strongest factors are not tied to one disorder. Difficulty handling feelings and feeling alone went along with depression, anxiety, stress and lower well-being alike, which suggests they reflect broad parts of how teenagers are doing.
The authors see models like these as tools for research and for spotting patterns across a school population, not as a replacement for a clinical assessment of any individual teenager. They also note that general exercise may have looked weak because it was crowded out by stronger factors, and because their measure captured exertion rather than team sport, which brings friends and structure.
The social media result adds nuance to a heated debate: in this sample, moderate use was not where most of the association sat.
What can't this teen mental health study tell us?
Read before you quote it
It is a single survey at one point in time, so it cannot say whether loneliness leads to depression or depression to loneliness. Everything was self-reported, and all students came from one large city, with no rural students and no information on family income or parents' education.
The family support and exercise questionnaires were somewhat less reliable than usual, which may make those factors look weaker than they are. The models were not tested on teenagers from another population, and around half of the variation in symptoms remained unexplained.
Because sleep was deliberately left out of the depression model, the study says nothing about how sleep relates to depression. It is not a diagnostic tool and not medical advice.
- It does not show that reducing loneliness or improving emotion regulation would lower depression.
- It does not estimate how sleep quality relates to depression, because sleep was left out of that model.
- It is a one-time survey, so it cannot show that any factor causes depression or anxiety.
- All measures were self-reported, which can add bias.
- All students came from one city, with no rural students and no data on family income or parental education.
- The models were not tested on an independent population, and much of the variation stayed unexplained.
Takeaways
- saysAcross depression, anxiety, stress and well-being, difficulty handling emotions and loneliness ranked as the strongest factors.
- saysSocial media only mattered at high levels, and physical activity, age and gender added little.
- saysIt's a one-time survey in one city, so it shows associations, not causes.
- explainsTesting a model on people it has never seen shows whether it learned real patterns or just memorised its training data.
- interpretsA fair reading is that emotional skills and social connection deserve as much attention as screen time in teen mental health.
Ask the paper
In PaperSays you can ask this paper anything, by text or by saying “Hey Aiden”. Three questions listeners asked:
What is the strongest predictor of depression in teenagers?
In this study, difficulty managing emotions and loneliness ranked highest, ahead of social media, sleep and exercise. Hear the full answer in the app →
Does social media cause depression in teens?
This survey cannot show causes; social media addiction mattered mainly at high levels of use. Ask in the app →
Does exercise matter for teen mental health?
General physical activity added little in these models, possibly because stronger factors crowded it out. Ask in the app →
Briefing reviewed by an editor before publication. The paper belongs to its authors; this page summarises it and links to the original. Paper licence: CC BY 4.0. Corrections: support@papersays.com.