Screen time, too coarse a variable

Screen time adds up unrelated activities: reading, playing together, calling a relative, passive viewing, schoolwork. Studies that break down those uses find different associations depending on content and context.

That aggregation is a measurement convention, adopted because it is easy to report. It produces a variable that is robust to collect and uninformative to interpret. It also mixes ages: an hour of screen time does not denote the same use in a three-year-old and in a fifteen-year-old, yet public debate often files them under the same heading.

Amy Orben and Andrew Przybylski published in 2019 an analysis of large adolescent samples, systematically testing the possible combinations of variables rather than a single one.

The method, known as specification curve analysis, consists in computing the result for each reasonable way of handling the data, then examining their distribution rather than a single choice. The article's most widely quoted comparison likened the observed association to that of eating potatoes regularly.

Their result is that the association between digital use and reported well-being is very weak, comparable in size to associations nobody finds worrying. They also show that by choosing the variables, almost any result can be produced.

Jean Twenge and others argue that averages mask effects concentrated on certain uses and certain groups, notably adolescent girls and social networks, and that self-reported measures underestimate real exposure.

The debate therefore turns less on the raw data than on how to cut them. Both camps often use the same datasets, which is unusual and makes the discussion technical rather than doctrinal.

Comparisons between reported time and time recorded by devices find poor agreement. People underestimate or overestimate according to their relationship with the object, and the gap is not random.

Part of the debate therefore rests on a variable of low reliability, which authors on both sides acknowledge and which is steering research towards directly collected usage data.

The few longitudinal studies and experiments find weak and often bidirectional effects: a degraded state increases passive use as much as the reverse.

The experiments most often involve asking volunteers to cut their use for a few days or weeks. Their results are mixed, and the impossibility of blinding plays a part, as in every trial where participants know what they are doing.

Public recommendations have shifted accordingly. Several professional bodies have dropped uniform hourly thresholds in favour of criteria bearing on content, timing and what the use displaces.