Does statistical learning decline from childhood to adolescence?

Learning a pattern without being told it exists

The study asks whether the ability to pick up regularities in the environment changes between childhood and adolescence. To measure that ability, participants performed an Alternating Serial Reaction Time (ASRT) task: they pressed keys matching the location of a target stimulus, and the sequence of locations alternated between a fixed eight-element pattern and random trials. Because pattern and random trials alternate, some runs of three consecutive trials, called triplets, occur more often than others.

Statistical learning is read off the difference in reaction time between high-probability triplets, which appear 62.5% of the time, and low-probability triplets, which appear 37.5% of the time. Faster responses on the more predictable triplets indicate that the participant has absorbed the regularity. This learning score is a reaction-time contrast in milliseconds, not a count of correct answers; accuracy analyses are reported separately in the paper's supplementary material.

The design is what makes the question answerable. Rather than comparing different children of different ages, the same participants were assessed at ages 7, 8, 11 and 14, so age-related change is measured within the same individuals. The main model was a linear mixed model on block-wise median reaction time, with triplet type, block and age as fixed effects and participant-level intercepts and slopes as random effects.

The learning contrast shrinks from age 7 to 14

The model found a significant main effect of triplet type: high-probability triplets were answered faster than low-probability ones (568 ms versus 585 ms), confirming that statistical learning occurred. It also found a significant age by triplet type interaction, meaning the size of that learning contrast differed across ages.

The direction was a decline. The paper reports that the amount of statistical learning decreased with development, and the individual trajectories plotted for the 97 participants show the same downward tendency. Overall reaction times also fell sharply with age, from 789 ms at age 7 to 599 ms at age 8, 509 ms at age 11 and 407 ms at age 14, with every pairwise comparison significant.

Those two findings are different quantities and should not be merged. The drop in overall reaction time is general speeding up; the shrinking triplet contrast is the change in learning itself. The authors also report that the three-way age by block by triplet type interaction was not significant, so the decline is best described as a change in the overall learning effect rather than a selective change in how learning builds within a session.

Faster responses do not explain the smaller learning scores

A natural objection is that children who respond faster have less room to improve, so their learning scores are compressed by a ceiling. The paper tests this in several ways. Across participants, correlations between overall reaction time and learning scores were not significant at any of the four ages, with Bayes factors described as anecdotal evidence for the null.

Within individuals, the authors used repeated measures correlation on the second half of the task, blocks 11 to 20, where learning is more likely to have stabilised. Only age 7 showed a significant but weak positive relationship; ages 8, 11 and 14 showed no significant association, with Bayes factors providing moderate support for the null. Age 14 matters most here because it had the lowest overall reaction times and therefore the greatest risk of a ceiling confound.

A further control analysis excluded one fifth, one fourth or one third of participants, either based on their age-14 reaction time or at random. The decreasing pattern of learning scores across ages remained present in all of these models. The authors conclude there is little credible evidence that baseline reaction time differences produced the developmental decline.

Most children decline, but a third move the other way

The average trend hides individual differences, so the authors applied latent class analysis to group participants by the shape of their learning trajectory. They selected a three-class solution over a two-class model because the two-class model was extremely unbalanced, with 96.9% of participants in one class.

The largest class, 61 of 97 participants, was labelled decrease: a relatively high starting learning score with a significant downward slope, amounting to a predicted difference of about 18 ms, or 0.89 standard deviations, between the first and last timepoint. A second class of 33 participants was labelled increase: a starting score close to zero and a slight but significant upward slope, a predicted difference of about 13 ms, or 0.65 standard deviations. A third class contained only 3 participants, with a very high start and a steep decline of about 94 ms.

The authors checked this structure against several other clustering methods, including regression-based k-means, mixed-model-based k-means, group-based trajectory modelling and growth mixture modelling. All favoured a three-class solution with qualitatively similar shapes, and three of the methods recovered the exact same classification. Agreement between methods ranged from moderate to perfect. The class solution is therefore a description of trajectory shapes in this sample, not a claim that children fall into fixed developmental types.

Higher executive function tracks a steeper decline

The study then asked whether executive function, the set of cognitive processes supporting goal-directed behaviour, predicted which trajectory a child followed. Two latent executive function factors were used as predictors: fluency and working memory. Class membership was the outcome in a logistic regression weighted by posterior probabilities, with the increase class as the reference and the three-participant class omitted because of its size.

Higher fluency scores were associated with a lower likelihood of belonging to the increase class and a higher likelihood of belonging to the decrease class, with an odds ratio of 2.03. Working memory pointed in the same direction but did not reach significance, with an odds ratio of 1.58 and a confidence interval spanning 1.

This is an association between executive function measured at age 14 and the shape of the learning trajectory from age 7 to 14, not a demonstration that executive function causes the decline. The authors also report little credible evidence that executive function at age 14 correlated with statistical learning scores at either age 7 or age 14, which is why they frame the result as a link to the trajectory rather than to the level of learning at a single point.

A gradual decline rather than a step around age 11

Earlier cross-sectional work had suggested that statistical learning improves up to about 11 to 13 years and then declines into adulthood. This longitudinal study instead found a progressive decline across the whole window from 7 to 14, with a relatively constant magnitude rather than a step change. The authors note that cross-sectional and longitudinal designs can disagree because cross-sectional comparisons mix age effects with differences between generations of participants.

The paper places its result within a broader pattern in which the direction of development depends on the type of material. Purely visual statistical learning and auditory learning with non-linguistic stimuli have been reported to increase with age, auditory learning with linguistic stimuli appears comparable across development, and visuomotor learning with probabilistic regularities, the paradigm used here, appears to decrease. The present study used probabilistic visuomotor regularities, so its declining profile is consistent with that specific strand of cross-sectional work rather than with statistical learning in general.

The authors' interpretation is that younger children may be more adept at building new predictive models from environmental input, and that the emergence of stronger top-down cognitive control by age 14 is associated with a reduced tendency to do so. That mechanism is proposed, not directly tested; the study measured executive function at a single age and cannot resolve the direction of the relationship.

Where the conclusion stops

Executive function was measured only at age 14, which limits the ability to disentangle the mechanisms behind the observed association. The authors state that future work should examine the relationship in late adolescence, when both statistical learning and executive function are closer to maturity, to determine whether the two systems compete during development.

The sample is 97 participants followed at four ages, and the smallest latent class contains only 3 people, so the steep-decline subgroup is described for completeness rather than analysed. The class solution also had modest entropy, which is why the authors invested in comparing multiple clustering algorithms before treating the three-class structure as robust.

The learning measure is a reaction-time contrast from a visuomotor probabilistic sequence task. The paper's own framing ties the declining profile to that paradigm, and it explicitly contrasts this with findings from visual and linguistic statistical learning. The result should therefore be read as evidence about probabilistic visuomotor statistical learning between ages 7 and 14, not as a general law about implicit learning across all domains.

The trajectory, not the single score, is the finding

The most useful distinction in this paper is between a child's learning score at one age and the direction in which that score moves. A high learning score at age 7 does not predict a high score at 14; in this sample it more often predicted a steeper decline, and higher executive function at 14 was associated with that decreasing trajectory rather than with a higher final level of learning.

For anyone comparing children of different ages on a statistical learning task, the paper's evidence argues against treating a lower score in an older child as a deficit. The decline appeared alongside faster overall responses, held after controls for baseline reaction time, and was concentrated in a majority decrease class rather than being a uniform shift. The subgroup split is the more informative summary: roughly two thirds of participants declined, about a third increased slightly, and executive function helped distinguish the two.

Sources used in this answer