Designs you will meet here
- Cross-sectional studies — a snapshot. Association only. Direction unknown.
- Longitudinal studies — time order. Still vulnerable to confounding and to the choice of lag.
- Case-control — compare people with and without an outcome. Helpful in clinics; easy to get selection wrong.
- Clinical case series — a pattern in a specialist setting. No denominator, high signal, high bias.
- Experiments — the cleanest causal warrant, often unethical or unrepresentative for this topic.
- Natural experiments — policy, outage, pandemic, viral shock. Only as clean as the shock.
- Systematic reviews / meta-analysis — a map of a literature, including its shared flaws. Garbage in, precision out.
- Specification-curve / multiverse — shows how much a finding depends on analytic choice (Orben & Przybylski, 2019).
Confounding
Sleep, poverty, parenting, school, prior diagnosis, loneliness, and platform use travel together. If a paper “controls for some of these,” ask whether it controlled the ones that actually sit on the back door.
Selection bias
Clinic samples over-represent the unusual. Social-media samples over-represent the online. Hashtag samples over-represent the identified. Every sampling frame is an argument.
Reporting bias
Destigmatization changes what people will tick on a questionnaire. That can look like a prevalence epidemic. Foulkes and Andrews (2023) ask us to test that rather than waive it away or treat it as the whole story.
The reverse-causation problem
This question should haunt the whole project:
Does social media exposure contribute to distress? Or do distressed adolescents use social media differently? Or do both processes occur?
Heffer et al. (2019) found evidence consistent with depressive symptoms predicting later social-media use more than the reverse. Other longitudinal papers find residual associations after covariates. Bidirectional looping is a live possibility. Time-use studies still do not answer the content-and-identity questions this site is built around.
Distinctions this site treats as methodological, not rhetorical
- Correlation ≠ causation
- Temporal association ≠ causation
- Clinical observation ≠ population evidence
- Viral examples ≠ prevalence evidence
- Self-report ≠ clinical diagnosis
- Diagnostic disagreement ≠ intentional deception