A graph of second-language test scores slopes downward with age of first exposure. A fitted curve bends near adolescence. It is tempting to place a vertical line at that bend and call everything before it possible and everything after it too late. Yet a critical period is a stronger proposition than an age trend. The investigator must show a particular developmental process has an interval in which a particular experience produces a qualitatively different outcome, and explain why later exposure cannot produce the same change under comparable conditions. A visually sharp graph can conceal gradual change, different learning histories and the assumptions built into a statistical model.
This Methods & Measures entry examines how researchers test the strong window claim. It is distinct from measuring sensitive periods, where an experience may simply have a larger effect at one age than another. It is also distinct from measuring experience-expectant plasticity, a claim about how expected environmental input shapes developing circuits. Critical-period research spans animal visual systems, human vision, audition and language; a result from one cannot be transferred wholesale to the others. The outcome, exposure and species must stay visible.
Say what would count as a critical period
In classic visual-system experiments, the question was whether deprivation during a restricted age interval altered cortical responses more strongly than deprivation at other ages. That is a comparatively precise target: one eye is occluded for a known duration, at a known age, and the response of visual neurons is measured. Hubel and Wiesel’s kitten experiments established an age-limited susceptibility in that preparation. They did not define a universal window for human vision, let alone for learning any skill.
A human critical-period hypothesis needs an equally explicit statement. One could propose that the effect of early patterned visual input on later acuity changes markedly after a particular age. Another could propose a transition in the capacity to acquire native-like phonetic contrasts. Those are different hypotheses. “The brain is less plastic after childhood” names neither an exposure nor an outcome and is too broad to falsify. Researchers should write down the contrast and what pattern would make them reject it.
The word “critical” is used inconsistently. Sometimes it refers to a period of greatest susceptibility to deprivation; sometimes to a period when successful input is most effective; sometimes to an apparent asymptote in later performance. These are not equivalent. A person can become less vulnerable to one disruption while remaining able to benefit from training in another task. A study should specify whether it concerns vulnerability, opportunity, closure or recovery.
Hensch’s circuit review describes mechanisms that regulate critical-period plasticity in the visual cortex, including maturation of inhibition. This mechanistic literature makes the phenomenon biologically plausible in some systems. It does not allow a human behavioral age cutoff to be inferred from a fitted score curve. The strongest designs connect a manipulation, circuit response and later behavior within the same domain; where such evidence is missing, the behavioral claim should remain behavioral.
Measure exposure and response without moving the goalposts
Timing can refer to onset, offset, duration, dose or age at recovery from deprivation. A child treated for cataracts at age three might have had visual impairment since birth, but the exact quality of input before treatment can be uncertain. A person beginning a second language at age twelve may encounter it daily or only in a weekly class. Coding both with a single “age of acquisition” variable compresses important differences. A strong test records how much relevant experience each person actually received.
The endpoint must be chosen before seeing the most striking curve. Visual acuity, contrast sensitivity, face recognition and motion perception need not share one developmental window. Speech perception, grammar and conversation likewise need not align. If researchers test many outcomes and report the one with the sharpest age break, the resulting “critical period” may be a selection artifact. Report the full outcome family and any corrections or replication plans.
Human visual research reviewed by Maurer and colleagues illustrates another difficulty: medical records may not date the onset of congenital or developmental cataracts precisely. The difference between deprivation from birth and later deterioration can be decisive, yet often rests on retrospective report. Researchers should distinguish documented from inferred onset and test whether their conclusions survive uncertain cases. A precise estimate from imprecise histories is not truly precise.
Age at outcome assessment matters as much as age at exposure. Earlier starters may simply have had more years to practice. Testing groups after equal years of experience controls duration but compares different chronological ages. Testing them at equal age controls maturation at test but leaves exposure duration unequal. A useful program of research uses multiple comparisons and models the several time scales rather than pretending one design holds them all fixed.
Measurements must also be comparable across ages. An infant’s looking-time task, a child’s verbal response and an adult’s button press carry different motor and language demands. A floor effect in younger participants or a ceiling effect in older ones can mimic closure or exaggerate its sharpness. Neural markers offer another view but introduce age-dependent physiology and recording issues. Convergence across carefully defined measures is more persuasive than one dramatic test.
Test a boundary instead of drawing one afterward
A threshold is a model, not a visible biological fence. Analysts can fit a step, a piecewise slope, a smooth decline or several domain-specific curves to the same observations. The chosen function may move the estimated break. Report what alternatives were considered, the uncertainty around any transition and the fit outside the observed age range. A model that extrapolates beyond the youngest or oldest participants should not be used to announce what happens at those unobserved ages.
Hartshorne, Tenenbaum and Pinker used a very large online sample of English grammar judgments to estimate age patterns in second-language learning. The scale of the data helped reveal structure that small studies might miss, but the design remains observational and depends on an inferred learning trajectory. Van der Slik and colleagues’ reanalysis questioned the sharply defined transition and found learner-type differences. The disagreement is scientifically useful: it shows that a critical-period claim depends on how the learning process and sample are modeled, not just on the number of responses.
Pre-registering a hypothesized age range and analysis can reduce the chance that a striking breakpoint is discovered by trying many specifications. An independent cohort, different test items and a different recruitment method can test whether it generalizes. Even replication of a statistical break would not alone identify why it occurs. School schedules, migration, motivation, language immersion and prior language knowledge can all change with age. The biological claim requires evidence that these alternatives do not explain the transition.
Age groups should have overlapping exposure opportunities. If nearly all early learners were immersed and most later learners attended classes, a threshold comparison confounds age with input intensity. A model can adjust for measured immersion, but crude categories may leave residual differences. Comparing siblings, immigrants with similar language environments or policy changes can reduce some confounding, although each introduces its own selection and context problems. No one design licenses a universal conclusion.
Differentiate susceptibility from irreversibility
In animal vision work, closing one eye during development changed ocular-dominance responses, and later reopening did not fully restore the measured physiology under the original conditions. This supports age-limited vulnerability and incomplete recovery in that preparation. But “incomplete under this recovery protocol” is narrower than “irreversible under every possible intervention.” Researchers should specify the recovery opportunity, its duration and the level at which recovery was assessed.
Modern visual-plasticity work has examined whether some adult responses can change after training or manipulations of circuit constraints. Hensch’s review treats critical-period closure as regulated, not a metaphysical end to learning. The existence of later plasticity does not erase evidence of earlier high susceptibility; it narrows the absolute version of the claim. An outcome can be much harder to obtain later without becoming impossible.
Human intervention studies face ethical and practical limits. Researchers cannot assign untreated deprivation to establish a cutoff. Cataract treatment timing is shaped by diagnosis, access and health. Delaying care for the sake of a clean experiment is not an acceptable measurement strategy. Observational follow-up and natural variation can be informative when investigators describe selection, treatment differences and uncertainty rather than treating age as randomly assigned.
In speech perception, Werker and Hensch review developmental tuning and the relation between experience and plasticity. Their evidence helps identify mechanisms and time courses for phonetic learning. It does not imply that every aspect of language shuts on the same date, or that adults cannot learn new contrasts. The broad word “language” masks several capacities, each with different input, task and endpoint. A critical-period study should be small enough in scope to be meaningfully tested.
Protect the analysis from common confounds
Selection can be severe. People who begin a new skill late may differ in education, health, resources and reason for learning from those who begin early. Those who persist to a high-level test are not a random sample of all starters. A cross-sectional sample that includes only successful adult learners can understate the difficulty of later learning; a sample recruited from remedial services can overstate it. Recruitment and attrition should be reported alongside age curves.
Cohort effects can masquerade as development. Older participants may have grown up with different schools, technology or cultural exposure. Testing different birth cohorts at one time confounds age of first experience with historical conditions. Longitudinal designs help but can still suffer attrition and repeated-test practice. A convincing program uses several designs whose biases differ, rather than assuming one corrected regression settles the matter.
Outcome scaling can also create false transitions. A test with many easy questions compresses high performers near a ceiling, hiding gains among early starters and changing the apparent slope. A small number of hard items can make lower scores noisy. Nonlinear transformations and categorical “native-like” labels impose further thresholds. Show the raw score distribution, item difficulty, reliability and alternative scoring models before interpreting a change-point parameter as a developmental boundary.
Researchers should inspect within-age variation. A mean decline with age may coexist with many high-performing later starters and some low-performing early starters. That overlap is not merely an exception to be explained away: it tests whether age alone is an adequate account. Report prediction intervals or distributions, not just a line through means. If the hypothesis concerns the probability of reaching a criterion, define the criterion and justify why it matters.
Model comparison should include a sensitive-period alternative, in which effects decline gradually, and a critical-period alternative, in which a qualitative change is predicted. It should also consider cumulative experience and recency. A better-fitting critical model does not prove a cellular mechanism; a better-fitting smooth model does not show that timing is irrelevant. The inference remains tied to the tested outcomes and the data’s ability to distinguish the models.
What the evidence does not show
The classic kitten experiments do not set human clinical deadlines. Human visual treatment studies do not establish the timing of language acquisition. Large language datasets do not directly measure inhibitory circuit maturation. Park and Fine propose a model of cross-modal plasticity and skill acquisition; their comparison of periods is not a validation study of a general critical-period measure. Each source has a domain and a level of analysis.
The cited research does not support the popular rule that meaningful learning ends at a particular birthday. A later learner may progress substantially even if earlier learning yields a different probability of an eventual outcome. Nor does an age-related average prove that an individual has missed an opportunity. Treatment and educational decisions require the person’s circumstances and domain-specific evidence, not a headline cutoff extracted from a group curve.
There is no single assay that reads “critical-period status” in a person. A test score, evoked response or fitted breakpoint is evidence about a specified process under a specified history. A strong claim of closure needs more than a significant age coefficient: it needs discriminating evidence about timing, exposure, alternative curves and the ability to change later. Where those pieces remain uncertain, “age-varying sensitivity” is the more defensible description.
Return to the age graph
Before interpreting the bend in the second-language graph, ask how age of onset was reported, whether exposure was comparable, how long each group had learned, and whether the break survives a different curve and sample. Ask which part of language the test measures. If a sharp transition remains, it is a serious result to explain; if it softens, the developmental question remains worthwhile. The answer is not decided by how compelling a line looks.
The Institute’s measurement questions here synthesize the cited visual, speech and statistical literature; they are not a validated diagnostic instrument. Critical-period research is most revealing when it makes a narrow prediction that competing timing accounts could falsify. The result should retain its domain, species and outcome all the way to the final sentence. A boundary that is truly critical will survive that precision; one that does not should be redrawn.
Research references
- Hubel, D. H., & Wiesel, T. N. (1970). The period of susceptibility to the physiological effects of unilateral eye closure in kittens. The Journal of Physiology. View source
- Hensch, T. K. (2005). Critical period plasticity in local cortical circuits. Nature Reviews Neuroscience, 6, 877–888. View source
- Mitchell, D. E., & Maurer, D. (2022). Critical periods in vision revisited. Annual Review of Vision Science, 8, 291–321. View source
- Werker, J. F., & Hensch, T. K. (2015). Critical periods in speech perception: New directions. Annual Review of Psychology, 66, 173–196. View source
- Hartshorne, J. K., Tenenbaum, J. B., & Pinker, S. (2018). A critical period for second language acquisition: Evidence from 2/3 million English speakers. Cognition. View source
- van der Slik, F., Schepens, J., Bongaerts, T., & van Hout, R. (2022). Critical period claim revisited: Reanalysis of Hartshorne, Tenenbaum, and Pinker (2018) suggests steady decline and learner-type differences. Language Learning, 72, 87–112. View source