Infant Screen Time Tracked Teen Anxiety in 168 Singapore Children

It is eleven at night and a one-year-old is finally asleep. The parent who spent forty minutes of the afternoon holding a phone up so that dinner could be cooked is now reading a headline that says early screen exposure has been linked to slower decision-making and higher anxiety in children. The arithmetic that follows is the familiar kind, done in the dark: how many afternoons like that one there have been, and whether anything has already been decided.

The paper underneath is narrower than the headline and considerably more interesting. It appeared in eBioMedicine on 29 December 2025, from Peng Huang, Ai Peng Tan and colleagues working with a long-running Singapore birth cohort. It does not ask whether screens are bad, but something harder: whether a measurable chain can be traced from screen time in infancy, through the physical development of brain networks in early childhood, through how long a child takes to make a choice in mid-childhood, to how anxious that child reports feeling at thirteen. The answer it returns is a qualified yes, and the qualifications are where the story is.

The cohort is GUSTO, or Growing Up in Singapore Towards healthy Outcomes, which has followed a group of Singaporean children from before birth. This analysis used the 168 of them who had all four of the things it needed: a record of screen time in infancy, diffusion MRI scans at ages 4.5, 6.0 and 7.5, a decision-making assessment at 8.5, and an anxiety questionnaire at 13. That is twelve years of measurement per child, and the sample is small because very few studies anywhere hold all four.

The exposure was collected from parents. At the child's first and second birthdays, parents were asked to estimate the average screen time on a typical weekday and a typical weekend in the past week; those were combined into a daily average for each year, and the two years were averaged into the single figure the analysis used. Across the wider cohort, that came to roughly 1.6 hours a day at age one and 2.5 at age two. It is worth naming that measure, because everything downstream inherits it: a parent's recollection of a typical week, not a log, not a device readout, and no record of what was on the screen or who else was in the room.

The brain measure is less familiar. From the three MRI scans, the authors modelled how the integration of seven major cortical networks changed over those three years, fitting each child a personal developmental slope. Decision-making was the Cambridge Gambling Task at 8.5, from which the relevant output is deliberation time, the seconds a child spends before committing to a bet. Anxiety was the Multidimensional Anxiety Scale for Children, fifty self-report items across six subscales, completed by the thirteen-year-olds themselves. All of it was then fitted into a single structural equation model, which is the statistical machinery for asking whether a chain of associations holds end to end.

The first link is the strong one. Higher infant screen time was associated with a steeper decline in the integration of the visual and cognitive-control networks between ages 4.5 and 7.5 (b = -1.03; 95% CI, -1.61 to -0.46; p < 0.0003 under bootstrapped permutation testing), and it survived Bonferroni correction for testing seven networks. Some decline in integration across childhood is ordinary, as networks specialise and separate; the authors read a steeper one as maturation arriving early, and float sensory overstimulation as the reason. That reading is an interpretation laid over the measurement, and the paper presents it as one.

The second link held too, and more loosely. The network slope fully mediated the association between infant screen time and deliberation time (b = 0.20; 95% CI, 0.01 to 0.91), with children on steeper trajectories taking longer over the gambling task. For scale, deliberation time in the cohort averaged 3.87 seconds. What is described is a shift of fractions of a second in a laboratory task, not a child visibly labouring over a decision at the breakfast table.

The third link is the headline and the frailest thing in the paper. The complete serial pathway, infant screen time to network development to deliberation time to anxiety at thirteen, was statistically significant with a coefficient of 0.033 and a 95% confidence interval running from 0.002 to 0.160. A lower bound of two thousandths is a finding that clears zero and very little else. The interval is also eighty times wider at one end than the other, which is what an estimate looks like when 168 children are asked to support four linked inferences at once.

One number in wide circulation is not in the paper at all. The coverage that carried this study reported roughly 25 percent slower decision-making per additional hour of screen time. No such figure appears in the abstract, the results, or the tables of the published article, which reports standardised path coefficients rather than effects per hour. A reader doing sums against a remembered afternoon is working from a number the study does not contain.

A serial mediation model is a chain, and every link in this one was estimated from the same 168 children, so confidence in the whole is bounded by confidence in its weakest part. The paper is honest about this: it calls itself exploratory, describes what it found as a plausible pathway, and says plainly that the result does not preclude alternative pathways or mechanisms.

The adjustments were also modest. The models controlled for maternal age, maternal education and household income, combined into a single latent measure of the child's environment. They did not control for family psychiatric history, for the quality of parent-child interaction, or for sleep, because those measures were either not collected or incomplete. Each of the three is a plausible common cause of both ends of this chain: a household where anxiety runs in the family, or where nobody sleeps, is also a household where the television goes on.

The authors themselves supply the sharpest version of this worry. Discussing prior work, they note a study in which the association between screen use and outcome disappeared once time spent alone was included as a covariate, and draw the moral for their own field: displaced social interaction is a confound when assessing the effects of screen use. What a screen occupies is time not spent doing something else, and this analysis measured only the hours, never the content, the context, or what they replaced.

Two further constraints come from the design. Because brain topology, decision-making and anxiety were each measured at only one stage, the direction of travel between them cannot be pinned down, and the ordering rests on where the measures happen to fall. And with imaging at three timepoints, only straight-line trajectories could be fitted; a development that speeds up and then settles would be invisible here.

The value here is not the effect size, which is small and unstable, but the shape of the question. Most of the screen-time literature correlates an hours figure with a symptom score collected at one sitting, with nothing in between. This study put twelve years and a mechanism in between, and specified the mechanism precisely enough that it could have failed. That is a claim with working parts, and working parts can be checked by somebody else.

That is also the fairest way to hold it. The finding that a parent can rely on is the first link, where the interval is comfortably clear of zero: in this cohort, infants with more screen time showed a measurably different developmental trajectory in one set of brain networks. The chain from there to a thirteen-year-old's anxiety score is a hypothesis that survived one test in a single small sample, and it needs a larger cohort before it means anything about a particular child. What would settle it is a replication with more children, more timepoints, and a record of what was on the screen.

None of which reaches the parent doing arithmetic at eleven at night, so it is worth saying what the evidence does not support in that direction. Nothing here identifies a threshold, a safe number of hours, or a point past which something has been decided. The study describes average trajectories across a group of children over twelve years; it contains no fact about any individual one of them, and was never designed to.

This is a better paper than its coverage and a weaker one than its headline, and both gaps run the same direction. What the authors have built is a genuine mechanistic proposal in a literature that mostly traffics in bare correlations, and the first link in it is solid work: infant screen time tracked a distinct developmental trajectory in the visual and cognitive-control networks, with an interval well clear of zero and a correction for multiple testing already applied. The chain from there to adolescent anxiety is another matter. A coefficient of 0.033 whose interval bottoms out at 0.002, estimated across four linked steps in 168 children, with no adjustment for family psychiatric history, parent-child interaction or sleep, is a hypothesis that has passed a first test, not a finding that has arrived. The authors say as much in their own limitations, in plainer language than the wire copy allowed. The number that reached most readers, a 25 percent slowdown per additional hour, is not in the paper anywhere, and a parent counting up afternoons against it is measuring against something that was never published. On the evidence as it stands, the honest summary is that a plausible route from infancy to adolescence has been drawn for the first time and now needs a bigger cohort to survive; anyone converting that into a number of minutes is well ahead of what 168 children can support.