Self-selected youth screen-time and wellbeing survey
A national youth wellbeing charity runs a campaign on Instagram and TikTok asking 14- to 18-year-olds to complete a 10-minute survey. It receives 1,842 responses. 78% are female, 62% are 16–18, and 41% report more than six hours of recreational screen time on a typical day. The charity computes a correlation of -0.21 between self-reported screen time and WHO-5 wellbeing. After adjusting for sleep duration, the correlation is -0.08. A media release says 'screen time damages teen wellbeing'. The charity has no sampling frame, but it does have 12,000 email subscribers from schools who opted in. It wants to know what it can responsibly claim and whether its subscriber list can help.
1. Evaluate the inference from the -0.21 correlation to 'screen time damages teen wellbeing'. Address selection, measurement, confounding, and reverse causation.
Read the worked answer
The correlation comes from a self-selected online sample, not a probability sample. People who respond to a social media campaign differ in age, gender, wellbeing, and screen habits; 78% female and 62% aged 16–18 shows the sample is not representative. Self-reported screen time is prone to recall and social-desirability error, and error in the predictor attenuates correlations. Wellbeing and screen time are measured at one time, so reverse causation is possible: low wellbeing may increase screen use. Sleep, family circumstances, school stress, and device access are confounders, and sleep may also be a mediator. A correlation of -0.21 is weak and does not establish damage. The causal media claim goes well beyond the design.
2. Propose a sampling and analysis strategy using the 12,000 subscriber emails that would support a more defensible claim. Include how you would handle nonresponse and weighting.
Read the worked answer
Use the 12,000 subscriber emails as a sampling frame if they are reasonably representative of the target schools. Draw a stratified random sample by school, age, and gender; invite with reminders; record nonresponse. Weight for selection and nonresponse, and post-stratify to known school rolls. Measure screen time with a diary or device log rather than a single self-report, and measure wellbeing with a validated scale at two time points. Collect sleep, family, and school covariates. Pre-register the primary analysis. Even then, observational associations need cautious language; a stronger design would follow a cohort or use a natural experiment. Report response rates, weights, and sensitivity to unmeasured confounding.
3. The adjusted correlation falls to -0.08. Explain what this attenuation does and does not tell you, and how you would communicate it.
Read the worked answer
Attenuation from -0.21 to -0.08 means sleep explains part of the association. It does not tell you whether sleep is a confounder (e.g., poor sleep causes both) or a mediator (screen time displaces sleep, which affects wellbeing). Adjusting for a mediator can remove a real indirect effect; adjusting for a confounder is appropriate. Measurement error could also attenuate. Communicate that the association is weak and sensitive to adjustment, that the study cannot establish direction or cause, and that the difference between the raw and adjusted estimates is itself an important uncertainty. Avoid 'no effect' as well as 'damages'.