Cognition varies across the calendar year in multiple large-scale datasets Authors: Arielle S. Keller, Alisha Shetty, Ran Barzilay, Monica E. Calkins, Yap-Seng Chong, Niyati Dave, Damien A. Fair, Peter D. Gluckman, Raquel E. Gur, Ruben C. Gur, Allyson P. Mackey, Michael J. Meaney, Lucille A. Moore, Tyler M. Moore, Theodore D. Satterthwaite, Ai Peng Tan, Brenden Tervo-Clemmens, Bart S. Larsen Published: 2025 ABSTRACT SignificanceHow well we perform on challenging cognitive tests may depend on when tests are administered. We explore one potential source of variation in test performance—the “summer slide”—referring to worse academic test scores following school vacation. To investigate whether seasonal variation in cognition is consistent across cognitive assessments, we gathered data from thousands of individuals from varied geographic regions and sociodemographic backgrounds. We found a small but consistent minimum in cognitive performance during the summer months among school-age youth but not young adults, aligning with school vacation across datasets. These results provide useful insight for those invested in supporting children’s development and for researchers across disciplines who rely on large-scale population studies conducted across the calendar year. Children’s cognitive abilities vary across both short and long timescales, from circadian fluctuations to year-by-year developmental changes. The “summer slide” is a commonly described example of temporal variation in youth cognition across the calendar year, with better performance on academic assessments during the school year and lower performance during or after school vacation, most often occurring during the summer months (1, 2). This phenomenon goes by many names, including the “summer setback,” “summer slump,” or “summer learning loss.” While many previous studies have investigated seasonal variation in cognition by comparing academic test scores before and after periods of school vacation, concerns have been raised about the replicability of these effects, even among recent studies with large samples (3). Moreover, it remains contested whether seasonal variation in cognitive performance reflects a “loss” or “forgetting” of specific types of learned academic material (e.g., math skills) during periods without practice, or whether it represents a broader difference in cognitive skills (e.g., executive functioning) during the summer months. Understanding time-of-year variation in youth cognition is critical not just for caregivers and educators invested in supporting children’s development, but also broadly for understanding cyclical patterns in behavior, particularly as researchers increasingly rely on large-scale population studies with measurements taken across the calendar year. The most common approach to identifying the summer slide is through academic performance measures, typically quantified as absolute or relative change in academic test scores before and after school vacation, to measure learning loss from the end of one school year to the start of the next (2). Several studies have suggested that student achievement scores do tend to decline following school vacation periods (4). The use of longitudinal assessments in prior study designs has been an important advantage for identifying changes in performance within individuals. However, results have been mixed and often observed effects are quite small (5, 6). Some studies have reported variability by subject area, with larger declines in math skills compared to reading skills, though these findings are also mixed (2, 4, 7). More broadly, recent work has called into question whether summer slide effects are in fact replicable across studies, pointing to inconsistent measurement approaches and small, local samples in past work (3). While a long history of past longitudinal studies has laid important groundwork for understanding changes in individual children’s academic test scores with school vacation, we note four critical gaps in knowledge worth addressing. First, given the focus on scholastic performance, past studies have typically used academic test scores rather than standardized cognitive assessment batteries to measure decreases in performance during school vacation. It therefore remains unclear whether these declines pertain to specific knowledge learned in school (e.g., forgetting new material that was taught in a classroom several months prior but not practiced since) or whether they reflect a more general difference in cognitive performance. Furthermore, it has been postulated that the magnitude of observed decreases in academic performance after summer vacation may be exaggerated by measurement choices such as grade-scaling and differences in test forms between prevacation and postvacation assessments (3, 8). If differences in performance are found not just on academic assessments but also on well-validated standardized cognitive assessments conducted in a controlled laboratory setting, this would have far-reaching implications for many psychological studies of youth cognition, which generally do not consider or control for time-of-year effects during study design, participant enrollment, data collection, or analysis. Second, temporal changes in children’s cognitive performance have most often been assessed by treating time as a continuous linear variable. Although this approach has yielded important insights into continuous changes in cognition (e.g., over many years), linear methods might fail to capture smaller magnitude cyclical fluctuations, such as time-of-day, day-of-week, or time-of-year effects. Statistical approaches that can account for cyclical variation therefore have the potential to uncover important and underdescribed sources of variation in children’s cognitive performance. Generalized additive models (GAMs) (9), for example, do not assume strictly linear relationships (e.g., between age and cognitive performance) and can allow one to model complex associations using a basis set of flexible splines which can take many forms. Cyclic cubic splines in particular can be used as a basis set to model cyclical effects like time-of-year, where each day of the year falls along a circle with December 31st and January 1st positioned side by side. Third, previous studies have investigated whether seasonal variation in cognitive performance exacerbates or alleviates existing social inequities by examining whether factors such as socioeconomic status (SES), which could affect involvement in or access to various activities during school vacation, moderate the size of the effect (2, 4–6, 10, 11). Some studies have reported disproportionate declines in performance on academic tests during school vacation for children from low-income households, widening the gaps in cumulative academic achievement between high- and low-SES groups (2, 10). However, other studies have shown the opposite effect, reporting that the gaps between low- and high-SES groups shrank in both reading and math (8). Another study examining participation in reading and math programs during school vacation found that achievement gaps in math tended to expand over time whereas reading achievement gaps did not, potentially due to higher participation of low-income students in reading programs over math programs during school vacation (5). Reconciling these disparate findings will require large datasets with more comprehensive, multidimensional measures of each child’s environmental context, including physical/chemical, psychosocial, and socioeconomic factors that may moderate the effect size. Fourth, there remain important open questions regarding specificity and generalizability. Although it has been presumed that the summer slide is caused by school vacation itself, it would not be feasible to conduct a randomized controlled trial to demonstrate causality. As most prior studies have assessed young children (e.g., 5 to 8 y old) from western countries such as those in North America, it remains unclear whether seasonal variation in cognition is specific to school-age children and adolescents but not young adults, and whether the effect generalizes to youth in other locations. Geographic diversity is particularly important in this case given that the timing of school vacation during the calendar year can vary substantially from place to place, along with differences in climates and seasonal weather patterns. Here, we seek to build on prior work and address these four challenges by quantifying seasonal variation in children’s cognition across four independent datasets of youth (total n = 23,251) with a wide range of measures, participant demographics, and geographic locations. To determine whether declines in cognitive performance generalize beyond scholastic assessments, we leverage multiple well-validated standardized cognitive assessment batteries each assessing multiple cognitive subdomains. To model time-of-year as a cyclical rather than unidirectional measurement, we fit GAMs with cyclic cubic splines. To investigate whether declines in performance are specifically related to school vacation, we compare school-age children to young adults and compare school-age children across countries that exhibit different seasonal patterns in average temperature, weather patterns, and the timing of school vacation. Finally, we evaluate potential moderators by leveraging deep phenotyping of each child’s environmental context (“exposome”) (12–14), comprising a multitude of physical/chemical exposures (e.g., air pollution), psychosocial settings (e.g., family dynamics), and socioeconomic resources (e.g., household income). Together, this set of approaches allow us to expand upon previous work by systematically testing the replicability and generalizability of patterned seasonal variation in youth cognition. Results Children’s Cognitive Performance Is Lowest Following the Summer Months in the United States. To model cyclical variation in cognitive performance by time-of-year, we fit GAMs with cyclic cubic splines to the baseline cohort of the Adolescent Brain Cognitive Development (ABCD) Study (n = 11,040, 9 to 11 y old) (15). We observed a significant relationship between time-of-year and cognitive performance across three domains capturing the majority of variance in a standardized battery of laboratory-based cognitive assessments (16) (Fig. 1A and Table 1 and SI Appendix, Fig. S1). These models revealed a characteristic drop in cognitive performance during the summer months that reaches a minimum during September (General Cognition: min = September 18th, CI = [September 2nd–October 7th], R2partial = 0.026, PFDR < 0.001; Executive Functioning: min = September 6th, CI = [August 18th–September 24rd], R2partial = 0.014, PFDR < 0.001; Learning/Memory: min = September 4th, CI = [August 11th–September 26th], R2partial = 0.013, PFDR < 0.001). These cyclical variations in cognitive performance were consistent across demographically-matched discovery (n = 5,366) and replication (n = 5,320) subsamples derived from the ABCD Reproducible Matched Samples (17, 18) (ARMS; SI Appendix, Fig. S2 and Table 1). These end-of-summer minima were also consistently observed when we computed GAMs for each of nine cognitive tasks independently (Fig. 1B and SI Appendix, Table S1) with consistent minima following the summer months (Table 2) and were also largely consistent across ABCD sites (SI Appendix, Fig. S3). Cyclical variation in youth cognition across the calendar year was also observed irrespective of participant demographic features such as household income, parental education, self-reported biological sex, and racial identification (no significant interaction with time-of-year, Ps > 0.19; Fig. 1C). Follow-up analyses using two simplified approaches confirmed the robustness of these results: comparing month-by-month variation across the calendar year and comparing participants who were assessed in the months just before vs. after summer (SI Appendix, Fig. S1), we find the same pattern of lowest performance following summer months across all three cognitive domains. Cognitive functioning consistently varies by time-of-year. (A) GAMs with cyclical smoothing splines reveal that cognitive performance [principal component scores derived previously (13), y-axis] varies systematically across the calendar year in the ABCD Study dataset (n = 11,040). Plots depict partial effects of time-of-year (intercept removed). This effect is consistent across three broad cognitive domains (General Cognition, Executive Function, and Learning/Memory) as well as across random split-half subsamples of the data. Radial plots depict model-estimated minima in cognitive performance across domains (Table 2). Shaded regions represent 95% credible intervals. (B) Associations between time-of-year and cognitive performance are consistent across individual cognitive tasks and composite scores from the NIH Toolbox, with consistent minima following the summer months. (C) Associations between time-of-year and cognitive functioning reveal declines during summer months that are consistent across sociodemographic categories including parent education, participant sex assigned at birth, and household income (no significant interactions with time-of-year, Ps > 0.19). Box plots and histograms depict the distribution of participants’ study visit days across the calendar year. Table 1.Cyclical GAMs reveal consistent variation in cognitive performance across the calendar year in school-age children across multiple datasets and cognitive measuresSampleCognitive MeasureModel TermEDFPartial R2 P PFDR ABCD TotalGeneral CognitionMain Effect: Time-of-Year8.5850.026<2 × 10−16<2 × 10−16Executive FunctionMain Effect: Time-of-Year5.6050.014<2 × 10−16<2 × 10−16Learning/MemoryMain Effect: Time-of-Year4.4340.013<2 × 10−16<2 × 10−16ABCD Split 1General CognitionMain Effect: Time-of-Year6.0060.0261.010 × 10−52.020 × 10−5Executive FunctionMain Effect: Time-of-Year2.2210.0160.0170.017Learning/MemoryMain Effect: Time-of-Year3.8980.0231.679 × 10−75.037 × 10−7ABCD Split 2General CognitionMain Effect: Time-of-Year3.2010.0302.512 × 10−55.023 × 10−5Executive FunctionMain Effect: Time-of-Year3.1320.0173.833 × 10−71.150 × 10−6Learning/MemoryMain Effect: Time-of-Year2.8850.0120.0010.001PNCSocial CognitionMain Effect: Time-of-Year0.0590.0060.4290.429Social CognitionInteraction: Age × Time-of-Year13.0760.0063.353 × 10−63.018 × 10−5Complex CognitionMain Effect: Time-of-Year1.1390.0050.1300.259Complex CognitionInteraction: Age × Time-of-Year10.9280.0051.054 × 10−58.428 × 10−5Executive FunctionMain Effect: Time-of-Year3.9540.0053.807 × 10−40.002Executive FunctionInteraction: Age × Time-of-Year7.5070.0050.0010.004MemoryMain Effect: Time-of-Year2.3580.0050.0010.004MemoryInteraction: Age × Time-of-Year7.5420.0051.837 × 10−40.001Overall CognitionMain Effect: Time-of-Year2.5570.0070.0100.029Overall CognitionInteraction: Age × Time-of-Year11.9040.0077.665 × 10−77.665 × 10−6ADHD-1000Overall CognitionMain Effect: Time-of-Year2.3350.0340.0130.013Overall CognitionInteraction: Age × Time-of-Year9.4600.0340.0210.021GUSTOMatrix ReasoningMain Effect: Time-of-Year2.9240.0330.0180.037Block DesignMain Effect: Time-of-Year0.0115.912 × 10−50.4110.411GAMs with cyclical smoothing splines reveal that cognitive performance on standardized laboratory-based assessments varies systematically across the calendar year in school-age children. For interaction models, the partial R2 reflects the combined time-of-year effect across the main effect and interaction terms. Abbreviations: ABCD = Adolescent Brain Cognitive Development Study; PNC = Philadelphia Neurodevelopmental Cohort; ADHD-1000 = Attention-Deficit Hyperactivity Disorder-1000 dataset; GUSTO = Growing Up in Singapore Toward healthy Outcomes study; EDF = effective degrees of freedom. Cyclical GAMs reveal consistent variation in cognitive performance across the calendar year in school-age children across multiple datasets and cognitive measures GAMs with cyclical smoothing splines reveal that cognitive performance on standardized laboratory-based assessments varies systematically across the calendar year in school-age children. For interaction models, the partial R2 reflects the combined time-of-year effect across the main effect and interaction terms. Abbreviations: ABCD = Adolescent Brain Cognitive Development Study; PNC = Philadelphia Neurodevelopmental Cohort; ADHD-1000 = Attention-Deficit Hyperactivity Disorder-1000 dataset; GUSTO = Growing Up in Singapore Toward healthy Outcomes study; EDF = effective degrees of freedom. Table 2.Minima in cognitive performance across the calendar year systematically align with the end of school vacation in school-age childrenSampleCognitive MeasureMinimumCIABCD TotalGeneral CognitionSeptember 18[September 02–October 07]ABCD TotalExecutive FunctionSeptember 6[August 18–September 24]ABCD TotalLearning/MemorySeptember 4[August 11–September 26]ABCD Split 1General CognitionSeptember 26[August 31–October 14]ABCD Split 1Executive FunctionAugust 29[July 04–October 27]ABCD Split 1Learning/MemorySeptember 17[August 22–October 12]ABCD Split 2General CognitionOctober 21[September 07–December 19]ABCD Split 2Executive FunctionSeptember 6[August 06–October 07]ABCD Split 2Learning/MemoryAugust 17[June 12–September 22]PNC (age 9.9 y)Social CognitionJuly 20[June 06–September 27]PNC (age 9.9 y)Complex ReasoningAugust 17[June 28–October 04]PNC (age 9.9 y)Executive FunctionSeptember 5[July 25–September 30]PNC (age 9.9 y)MemoryAugust 16[July 13–September 21]PNC (age 9.9 y)Overall CognitionAugust 18[July 11–September 20]PNC (age 20 y)Social CognitionJanuary 25[December 02–February 22]PNC (age 20 y)Complex ReasoningJanuary 24[October 11–March 15]PNC (age 20 y)Executive FunctionOctober 2[July 20–March 22]PNC (age 20 y)MemoryOctober 10[May 04–March 11]PNC (age 20 y)Overall CognitionJanuary 22[September 25–March 05]ADHD-1000 (age 9.9 y)Overall CognitionAugust 16[June 20–September 20]ADHD-1000 (age 20 y)Overall CognitionJuly 30[March 09–January 03]GUSTOBlock DesignMay 12[November 21–November 03]GUSTOMatrix ReasoningFebruary 10[December 21–April 12]Average minima in cognitive performance are reported for each sample and each cognitive domain. Bootstrapped 95% CI are also reported. Minima in cognitive performance across the calendar year systematically align with the end of school vacation in school-age children Average minima in cognitive performance are reported for each sample and each cognitive domain. Bootstrapped 95% CI are also reported. Time-of-year variation in cognitive performance among school-age youth also generalized to independent datasets across the United States with different sampling schemes, demographic features, and cognitive assessments, but was not found in young adults. We again fit GAMs with cyclic cubic splines to data from the Philadelphia Neurodevelopmental Cohort (PNC) (19) (n = 9,416, 8 to 22 y old) and the Oregon ADHD-1000 dataset (20) (n = 843 participants, 2,454 observations, 7 to 21 y). Because of the wide age-ranges of the PNC and the Oregon ADHD-1000 samples, we tested whether age moderated the association between time-of-year and cognitive performance. We observed significant age-by-time-of-year interactions in both the PNC and Oregon ADHD-1000 datasets across all cognitive domains (Fig. 2A and Table 1). For younger individuals in both datasets, we observed the same pattern as in the ABCD Study dataset: performance fell during the summer months across all five cognitive domains in the PNC dataset as well as in a composite cognitive score (Materials and Methods) from the Oregon ADHD-1000 (Fig. 2B). Predicted values from the GAM at the median age of the ABCD Study dataset (9.9 y old) showed minima between July and September in the PNC and Oregon ADHD-1000 (Fig. 2B and Table 2). This effect was consistent across all tasks that made up the composite score from the Oregon ADHD-1000 with the exception of 2-back working memory performance (SI Appendix, Fig. S4 and Table S2). In contrast, young adults from the PNC tended to have lowest performance during the winter months across all cognitive domains, and young adults from the Oregon ADHD-1000 had a less pronounced local minimum (Fig. 2). Furthermore, time-of-year variation in cognition was present in both neurotypical youth and youth with a diagnosis of ADHD in the Oregon ADHD-1000 sample (time-of-year by clinical diagnosis interaction P = 0.27; SI Appendix, Fig. S5). Associations between time-of-year and cognition replicate across datasets of youth with and without ADHD, but not young adults. (A) GAMs with cyclical smoothing splines reveal that cognitive functioning in the PNC dataset (zero-centered factor scores, y-axis) and the Oregon ADHD-1000 dataset (PC scores, y-axis) varies systematically across the calendar year. We also observed a significant interaction between age and time-of-year across all cognitive factors and in both datasets (Table 1). Predicted values from the GAM are plotted at the median age of the ABCD Study dataset (9.9 y old) and at 20 y of age (dotted vs. solid lines). (B) Radial plots depict model-estimated minima in cognitive performance (Table 2), which are found during the summer months for youth from these datasets collected in the United States. Shaded regions represent 95% credible intervals. Cognitive Performance Minima Consistently Align with the Timing of School Vacation Across Countries. To investigate the generalizability of the observed minima in cognitive performance following the summer months, we leveraged data from the Growing Up in Singapore Towards healthy Outcomes (GUSTO) study (21) collected in Singapore (n = 342, 7-y-olds). Notably, while the typical school vacation time in the United States occurs approximately between June and August, the typical school vacation time in Singapore occurs approximately between November and January. Seasonal variation in temperature and weather conditions also differs between the United States and Singapore, with much less seasonal variation in Singapore compared with most climates in the United States. GAMs with cyclic cubic smoothing splines fit to data from the GUSTO study revealed significant nonuniform associations between time-of-year and performance on the Wechsler Abbreviated Scale of Intelligence (WASI) Matrix Reasoning task (min = February 10, CI = [December 12–April 12]; R2partial = 0.033, PFDR = 0.037; Fig. 3) but no significant association for the WASI Block Design task (PFDR = 0.411). In summary, school-aged youth from three US-based samples had lowest cognitive performance during the end of summer, just following the most typical school vacation months, whereas youth in Singapore had lowest cognitive performance in February, following the typical school vacation time in Singapore (Fig. 4). Associations between time-of-year and cognition follow the timing of school vacation across countries. (A) GAMs with cyclical smoothing splines reveal that performance on the WASI Matrix Reasoning task (t score) varies systematically across the calendar year in the GUSTO dataset of youth in Singapore (Table 1). This effect was not observed for the WASI Block Design task. (B) Radial plots depict model-estimated minima in cognitive performance (Table 2). Minimum performance on the WASI Matrix Reasoning task was found on February 10th, aligning with the end of the longest school vacation time in Singapore. Shaded regions represent 95% credible intervals. The timing of minimum performance across cognitive tasks consistently aligns with the timing of school break. Datasets of US-based youth (red: ABCD, blue: PNC, purple: Oregon ADHD-1000) reveal minima in cognitive performance near the end of summer while youth in Singapore from the GUSTO dataset (green) have minimum performance on the WASI Matrix Reasoning test in February following the typical period of school vacation (WASI Bock Design scores did not vary significantly by calendar year). Each point reflects the model-estimated day of minimum performance for a cognitive domain with 95% credible intervals indicated by the shaded region (see Table 2 for a list of all minima by cognitive domain). For the PNC and Oregon ADHD-1000, model-estimated minima are plotted at the median age of the ABCD Study (9.9 y). Environmental Features Associated with Seasonal Variation in Cognitive Performance. We next sought to determine whether various features of childhood environments might be associated with the extent of cyclical time-of-year variation in youth cognitive performance. To investigate this question, we leveraged the deep phenotyping of many environmental features that were available in the ABCD and PNC datasets. To quantify the potential effects of many interconnected and co-occurring environmental exposures at once, including physical/chemical exposures (e.g., air pollution), psychosocial settings (e.g., family dynamics), and socioeconomic resources (e.g., household income), we leveraged previously validated measures of the multidimensional exposome from bifactor analyses conducted in our prior work in these datasets (22, 23). This bifactor analysis approach yields both a general exposome score (hereafter referred to as exposome), which primarily captures co-occurring household-level and neighborhood-level features of socioeconomic inequality, as well as orthogonal subfactors capturing specific aspects of familial, social, and physical environments. Using GAMs with cyclic cubic smoothing splines as before, we investigated whether participants’ exposome was related to the time of year that they participated in the ABCD and PNC studies. We then investigated whether the exposome moderated the cyclical variation in cognitive performance by time-of-year by including interaction terms for the exposome score. We uncovered a significant association between the exposome and the time of year that individuals participated in the study, whereby participants with lower exposome scores (associated with measures of lower SES) were more likely to participate in the ABCD Study during the summer (R2partial = 0.021, P < 0.00001). In the PNC, age significantly moderated this effect such that this participation effect was more pronounced in school-aged youth (time-of-year: P < 0.0035; age-by-time-of-year interaction: P = 0.0145; combined R2partial = 0.005; Fig. 5A). Covarying for exposome score reduced the effect size of the relationship between time-of-year and each domain of cognitive performance, but it did not fully account for these effects in either dataset (ABCD: time-of-year P = 0.0005, R2partial = 0.007; PNC: time-of-year-by-age interaction P < 0.0001, combined effect of time-of-year R2partial = 0.006; Fig. 5B). Furthermore, the exposome score did not significantly moderate the association between time of year and cognitive performance in the ABCD Study dataset (P = 0.12) or the age-by-time-of-year interaction in the PNC dataset (P = 0.07). The exposome is associated with the date of study participation but does not fully account for associations between time-of-year and cognitive performance. (A) Cyclical GAMs reveal significant variation in exposome scores across the calendar year in both the ABCD Study and PNC datasets. This analysis reveals that individuals with lower exposome scores (associated with measures of lower SES) disproportionately participated in these studies during the summer months. (B) This time-of-year variation in exposome scores does not fully account for time-of-year variation in cognitive performance. Minima in cognitive performance during the summer months are still observed across datasets collected in North America in cyclical GAMs that account for exposome scores. Shaded regions represent bootstrapped 95% CI. To examine whether associations between time-of-year and cognitive performance were explained by residual confounding of other socioenvironmental variables, we also fit a model that accounted for the general exposome and all six orthogonal exposome subfactors as covariates. This analysis revealed that the association between time-of-year and cognitive performance remained significant (P = 0.0007, R2partial = 0.007). To further confirm the robustness of our results, we conducted a sensitivity analysis using propensity matching to determine whether individuals who are comparable in their environmental exposures but differ in their testing occasion have different cognitive scores. Specifically, we used propensity scores to match an equally sized group evaluated in the Spring months prior to school vacation (February to May) to those tested in the months following school vacation (August 27 to September 25, defined as those within the credible interval of our estimate of the day of minimum performance for general cognition; Table 2) in the ABCD Study dataset, with participants matched according to all seven exposome factor scores (general exposome and six subfactors) as well as age and sex. Repeating our analyses using this subset of data (n = 1,700), we found a significant effect of testing time (estimate = −0.07, SE = 0.03, Cohen’s f2 = 0.004, P = 0.013), providing further evidence for the robustness of our results (SI Appendix, Fig. S6). Contextualizing Seasonal Variation in Cognitive Performance in the Frame of Cognitive Development. It is important to note that while the seasonal variation in cognitive performance we observed in school-aged youth is statistically significant and consistent across datasets, time-of-year accounts for only a small proportion of the overall variance in cognitive performance (R2partial range = 0.005 to 0.034). To contextualize this, we find that in the ABCD Study (ages 9 to 11), age explains a moderate amount of variance in overall cognitive performance (R2partial = 0.063), while household income accounts for a much larger proportion (R2partial = 0.104) and the exposome accounts for even more (R2partial = 0.260). This comparison highlights that the small but reliable variation in cognition across the calendar year is much smaller than developmental improvements with age and is overshadowed by the cognitive variability attributable to environmental factors like SES. Without accounting for environmental factors, during the summer months, when cognitive scores are lowest, the total difference in general cognitive performance in the ABCD sample appears equivalent to nearly 1 y of development (General Cognition: 0.93 y, 95% CI = [0.68, 1.20]; Executive Function: 0.54 y, [0.29, 0.87]; Learning/Memory: 0.99 y, [0.54, 1.50]). This difference would translate to a loss of nearly 5 d of cognitive development for every 1 d of summer between the highest peak and lowest trough of the calendar year (General Cognition: 4.94 d; Executive Function: 1.89 d; Learning/Memory: 4.41 d). However, this difference is substantially attenuated after adjusting for the exposome. After accounting for the exposome, summertime differences in cognitive performance are equivalent to only 0.22 y of development (95% CI = [0.08, 0.40]) for General Cognition, 0.29 y (95% CI = [0.12, 0.47]) for Executive Function, and 0.43 y (95% CI = [0.15, 0.76]) for Learning/Memory. Correspondingly, the estimate of days of development lost per day of summer also decreases to 0.94 d for General Cognition, 1.05 d for Executive Function, and 1.48 d for Learning/Memory. Thus, while seasonal variation in cognitive performance may appear comparable in size to other educational effects, we note that the largest effect size is still seven times smaller than the effect size for associations between the exposome and cognition in youth. Discussion Youth cognition is dynamic across multiple timescales. While many studies of youth cognitive performance have investigated long-term variation from year to year during development or short-term variation from moment to moment during learning, far fewer studies have investigated cyclical variation from month to month across the calendar year. Here, we investigated seasonal variation youth cognition using standardized cognitive assessments in thousands of children. By leveraging GAMs with cyclical cubic smoothing splines to model cyclical time-of-year variation, we found a systematic, generalizable, small-magnitude decrease in cognitive performance in youth that aligned with the timing of school vacation across datasets. This variation appears to be specific to school-age youth, as it is not observed in young adults, and specifically aligns with the end of school vacation, as it is independent of when school vacation occurs during the calendar year. Importantly, although seasonal variation in youth cognition appears to be replicable across samples and consistent across many cognitive assessments, the dip in cognitive performance around the timing of school vacation is much smaller in magnitude than has been previously appreciated, particularly when situated in the broader context of much larger magnitude increases in cognition with age and even larger magnitude differences in cognition by SES. These results extend the extant literature in several important ways. First, our findings suggest that seasonal minima in cognitive performance are found not only on academic test scores but on a wide variety of standardized and well-validated cognitive assessments. While most scholastic assessments test specific skills that were explicitly taught during the prior school year (e.g., multiplication skills), the standardized cognitive batteries assess more general cognitive skills that are not explicitly taught in school (e.g., attention). Our results therefore suggest that time-of-year variation in youth cognition is a more general phenomenon that may extend beyond forgetting of recently learned material. Second, our results show that seasonal variation in cognitive performance consistently replicates across diverse samples and is specific to the timing of school vacation in school-aged youth. These observations point to the importance of the structure of the academic school year and vacation time rather than other factors (e.g., seasonal variation in temperature or weather patterns) that would presumably affect young adults and children similarly and exhibit different patterns across climates. Future work in adult samples may seek to investigate whether college enrollment leads to similar seasonal variation according to the academic calendar year or whether adult cognition is more stable and less influenced by time-of-year than youth cognition. Our findings also motivate future research to investigate cyclical patterns in behavior at an even more fine-grained level (e.g., shorter holiday periods occurring at other times of the year, day-of-week variation, or time-of-day variation) using denser sampling approaches. Our results also provide critical insights to inform the sampling schemes, participant enrollment strategies, analytic approaches, and interpretations of future studies of human behavior. Given our observation that cyclical time periods like the calendar year can account for structured patterns in human behavior, future studies could consider time-of-year when recruiting study participants across different segments of the calendar year and when analyzing data in light of this underlying structure. Particularly for studies investigating critical periods of cognitive neurodevelopment like adolescence, it is important to be aware of this cyclical structure in cognitive behavior around the academic calendar. Typically, studies of cognitive behavior, including large-scale consortium studies like the ABCD Study that are collected across multiple sites, do not ascertain the timing of school vacation for each participant, which may vary across school districts. As future studies continue to collect larger and more representative samples, often targeting participant recruitment during school holidays and vacation times, this challenge may become increasingly prominent. It is common practice for studies collecting data from human participants to naturally exhibit variation in participant recruitment across the calendar year for a variety of practical reasons: students who are busy taking classes during the academic year may have more time available to participate in research studies or to assist in conducting research studies during school vacation periods; caregivers who have children in multiple schools or who struggle to afford childcare may also exhibit systematic patterns in when they and their children are able to participate in research studies. Unfortunately, it is easy for potential confounds like SES to become entangled with cyclical time-of-year patterns as we have observed with both the ABCD and PNC studies, so we encourage future studies to aim to recruit representative samples as evenly as possible across the calendar year to avoid this confounding effect. With adequate sampling schemes, future studies may also apply the analytic technique of employing cyclical smoothing splines as we have outlined here to both study cyclical patterns in human behavior and to account for these patterns when studying other effects on human behavior. An additional important context for interpreting the present results comes from comparisons with two other critically relevant effect sizes for studies of development. First, age effects quantify the expected amount of improvement in cognition as children develop into adolescents and adults. Here, we find that the small-magnitude differences in cognition during school vacation are dwarfed by the large-magnitude increases in cognition during typical development (24). Although seasonal minima in cognitive performance may be consistent across children and across cognitive domains, it is important to view these relatively small variations in light of the substantial overall improvement in cognition with normative development that is unhindered by month-to-month variation. Second, the effects of SES on children’s cognition are substantially larger than time-of-year variation in cognitive performance. Countless studies have documented positive associations between SES and cognition (25–28), including in the ABCD Study dataset used here (29–35). The large magnitude of the effect sizes reported in these studies points to the critical importance of understanding SES-related factors that may impact child cognition such as the availability of educational resources (36), families’ financial resources and distress (37), access to healthcare (38, 39) and nutrition (40), and neighborhood safety (41). Again, though end-of-summer minima in cognitive performance are observed consistently even across SES brackets, the magnitude of this seasonal variation is only a small fraction of the effect size of SES on cognition. Accordingly, children stand to benefit much more from policies and practices aimed at increasing socioeconomic resources than those aimed at decreasing the summer slide. Although we found that time-of-year was significantly associated with cognition across two countries (the United States and Singapore) with different climates and average temperatures across the calendar year, it is still worth noting the potential effects of climate and daily temperature on cognitive performance in children. It has been well documented in a plethora of prior studies that cognitive functioning tends to worsen with hot temperatures (42) in both adults and children, often with substantial effects of summertime heat on scholastic test performance and academic outcomes (43). It may be difficult to think clearly in times of extreme heat, when a higher portion of the brain’s metabolic energy is diverted toward thermoregulation (44). Such extreme heat days often occur during the summer school vacation months in many parts of the United States (45) and occur regularly throughout the year in Singapore (46). Importantly, due to the rapid pace of climate change and global warming (47), high temperatures are predicted to become increasingly prevalent with substantial psychological effects on children (48). Together, these observations suggest that our consistent finding of decreasing cognition during school vacation in three datasets from the United States may potentially reflect the influence of higher temperatures during the summer months, warranting future research into the effects of rising temperatures on children’s cognitive development. While the large cross-sectional datasets used in the present study reveal important population-level patterns in cognitive behavior across the calendar year, future studies including more detailed longitudinal assessments of individual children’s environments and day-to-day experiences during school vacation are needed to further characterize the holistic effects of school vacation for development. Critically, there may be substantial benefits of school vacation time for children’s development that were not assessed in these datasets. For example, studies have shown that taking breaks for physical exercise can facilitate memory recall (49) and that time spent with peers outside of school can be essential for developing healthy affective and social skills (50). Moreover, prior evidence supports the idea that unstructured play is an important aspect of children’s cognitive, physical, social, and emotional well-being (51). Exploration in particular may be an important aspect of children’s protracted cognitive development (52, 53), particularly for skills like creativity (54) that are not well measured by standardized cognitive assessment batteries. We therefore caution against changes to current school vacation schedules without careful investigation of their potential beneficial effects. The results of the present study should be viewed in light of several key limitations that motivate future work in this area. First, as the present study is cross-sectional and correlational, we are not able to establish whether there is a causal relationship between school vacation and summertime cognitive performance minima, nor are we able to quantify longitudinal changes in cognition within individuals. Indeed, our results demonstrated that the time-of-year of participation in these large-scale studies was associated with key sociodemographic factors that are also associated with cognitive performance. Notably, we found that seasonal variation in youth cognition was present even when statistically covarying for sociodemographic effects, suggesting these are additive effects. Future longitudinal studies are warranted to investigate the many potential underlying causes of seasonal variation in youth cognition. Numerous factors could be at play, and the relative influence of these factors is likely to vary across children. For example, future work may investigate differences in stressors, enrichment programs, sunlight exposure, social interactions, exercise, etc. between the school year and school vacation months. Second, dense longitudinal sampling of individual children across the calendar year is warranted for future studies to quantify the magnitude of the drop in cognitive performance during summer vacation as well as the magnitude of the potential rebound in an individual child’s cognitive performance at the start of the next academic school year. Future research may seek to investigate potential mechanisms by which cognitive performance “recovers” and continues to improve throughout development in order to better understand what may drive these smaller-scale increases and decreases in youth cognition. Furthermore, previous work has posited that the extent of gains in cognition during the school year may actually be related to the amount of change in performance during the summertime (55); that is, some children who experience greater dips in performance may also exhibit greater rebound. Our cross-sectional results appear to align with this idea, in that our models predict a full recovery of cognitive performance after the drop aligned with school vacation. Future longitudinal studies may investigate whether lower cognitive performance during the summer months reflects a trait-like decrease in children’s abilities or a more temporary state-like change by further investigating the timing of this recovery and its dependence on more transient factors like ambient temperature. Finally, while the present study made use of several independent datasets of youth, it will be important in future work to examine these effects across an even wider range of large-scale international datasets to more comprehensively characterize how different school vacation patterns might be associated with differences in cognition across the calendar year. Our results highlight a systematic, generalizable, small-magnitude variation in performance on standardized cognitive assessments across the calendar year in four independent datasets of youth. Although neurocognitive development is often conceptualized as a linear process with steadily increasing skills from 1 y to the next as the brain matures, our observation of cyclical variation adds complexity to this picture. Our characterization of seasonal variation in youth cognition extends previous work by demonstrating that small decreases in cognition following school vacation are consistent across cognitive domains, geographic locations, and participant demographics, but are specific to school-aged children and not young adults. Furthermore, our work situates seasonal variation in cognition within the broader context of cognitive neurodevelopment, highlighting that the consistent, small-magnitude decreases in cognitive performance following school vacation are dwarfed by environmental factors such as SES that have much larger effects on youth cognition. Together, this detailed characterization of seasonal variation in youth cognition provides a solid foundation to quantify cyclical variability in human behavior and inform future investigations of cognitive development. Materials and Methods Participants. We first investigated n = 11,040, 9 to 10 y-olds from the ABCD study (15) baseline cohort collected across 21 sites in the United States. These data were split into demographically matched subsamples (set1: n = 5,366; set2 n = 5,322) as in prior work using the ABCD Reproducible Matched Samples (ARMS) (17, 18). We then examined n = 9,416, 8 to 22 y-olds from the PNC (19) and n = 843 participants, 2,453 observations from youth ages 7 to 21 y from the Oregon ADHD-1000 Study (20), all collected in the United States. To test for generalizability of our results, we also investigated n = 342, 7-y-olds from the Growing Up in Singapore Towards healthy Outcomes (GUSTO) study (21) collected in Singapore. Participant demographics across all four datasets can be found in Table 3. All four datasets used in the present study had study visit days distributed across the full calendar year (SI Appendix, Fig. S7). Table 3.Sample demographicsCharacteristicABCD, N = 11,040PNC, N = 9,416ADHD-1000, N = 2,453GUSTO, N = 342Age (y), Mean (SD)9.9 (0.6)14.2 (3.7)11.8 (2.9)7.4 (0.1)Sex, n (%)Female5,288 (48)4,871 (52)931 (38)158 (46)Male5,752 (52)4,545 (48)1,522 (62)184 (54)Abbreviations: ABCD = Adolescent Brain Cognitive Development Study; PNC = Philadelphia Neurodevelopmental Cohort; ADHD-1000 = Attention-Deficit Hyperactivity Disorder-1000 dataset; GUSTO = Growing Up in Singapore Toward healthy Outcomes study; SD = standard deviation. Sample demographics Abbreviations: ABCD = Adolescent Brain Cognitive Development Study; PNC = Philadelphia Neurodevelopmental Cohort; ADHD-1000 = Attention-Deficit Hyperactivity Disorder-1000 dataset; GUSTO = Growing Up in Singapore Toward healthy Outcomes study; SD = standard deviation. Cognitive Domains. Cognitive performance in the ABCD dataset was quantified by three principal components derived in a previous study (16), which capture broad domains of general cognition, executive function, and learning/memory. These scores were derived from performance across nine cognitive tasks, including seven tasks from the NIH Toolbox (Picture Vocabulary, Flanker Test, List Sort Working Memory Task, Dimensional Change Card Sort Task, Pattern Comparison Processing Speed Task, Picture Sequence Memory Task, and the Oral Reading Test) (56) as well as the Little Man Task and the Rey Auditory Verbal Learning Task (57). To investigate potential differences in seasonal variation across cognitive domains, we also investigated cyclical patterns in performance on each of these nine tasks independently (SI Appendix, Fig. S2). Cognitive performance in the PNC dataset was quantified as factor scores on five cognitive factors (overall cognition, executive functioning, complex cognition, social cognition, and memory) (58). An overall cognitive score for the Oregon ADHD-1000 was calculated as the first principal component across tasks that assessed inhibitory control (D-KEFS color-word interference task, stop-signal reaction time), set-shifting (D-KEFS trailmaking number-letter switching task), working memory (the 1-back and 2-back conditions of the N-back task, the backward spatial span task), and signal detection (continuous performance task d-prime). For the D-KEFS scores and the stop-signal reaction time, the sign of each measure was flipped so that higher scores represent better performance. Study visits that did not include data for all tasks were excluded from the PCA analysis (total N = 1,096). Analyses of individual tasks used all available observations. Cognitive performance in the GUSTO dataset was quantified as performance on the Wechsler Abbreviated Scale of Intelligence (WASI) matrix reasoning and block design scales (59). GAMs with Cyclic Cubic Splines. GAMs are an extension of generalized linear models that can be used to fit smooth curves to a set of data. GAMs are fit using a set of smooth functions (e.g., splines), with penalties imposed on the flexibility of the curve to avoid overfitting. Cyclic cubic splines are one particular set of smooth functions that can be used to model cyclical data such as time-of-year (placing December 31st and January 1st side-by-side along a circle). We assessed the association between time-of-year (date of cognitive assessment, as day-of-year) and each cognitive measure using GAMs that modeled time of year with a cyclic cubic smoothing spline. We began with an upper limit of 52 evenly placed knots along the calendar year (k = 52) to reflect an upper bound of (at most) weekly variation, and penalized this flexibility to avoid overfitting, leading to models that retained less than 9 effective degrees of freedom of the smooth function across all subsamples. Each model included covariates for age (modeled as a penalized thin plate smooth), biological sex, and day of the week (modeled as a penalized cyclic cubic spline with seven evenly placed knots). For the ABCD dataset, we accounted for random intercepts and random smooths for ABCD sites using a factor-smooth interaction term that modeled deviations from the global time-of-year smooth for each ABCD site. Given that the ABCD Study dataset also contains siblings, we also trained GAMs with siblings excluded and found no differences in results (SI Appendix, Table S3). For the PNC and Oregon ADHD-1000 datasets, we included an interaction term for age and time of year to account for changes in nature of the time-of-year from late-childhood to early-adulthood. Statistical significance was assessed using the P-values associated with the F-statistic for each term in the GAM. The F-statistic tests the null hypothesis that the term does not contribute to the model, and it is inherently one-sided, as the F-distribution is bounded below by zero. To account for multiple comparisons, P-values were adjusted using the false discovery rate (FDR) method. Effect size in partial R-squared was calculated by comparing the full model to a nested model that omitted the terms of interest. To visualize the smooth terms, we plotted the partial effects of time-of-year and, where relevant, other covariates or moderators. Thus, plots display the effect of time-of-year (and other plotted covariates) after accounting for all other model terms. CI for the smooth terms were calculated using the SE provided by the model’s predictions. Specifically, 95% CI was derived from the predictions and SE from the predict function from the mgcv package, with type set to “terms.” The day of minimum performance was estimated from the model, and a credible interval was derived from 1,000 draws from the posterior distribution of the fitted model using a Gaussian approximation to the posterior. Analysis code is publicly available at https://github.com/bart-larsen/Cognition-time-of-year.git. Exposome Bifactor Analysis. To characterize multidimensional environmental factors associated with SES, we leveraged a measure of each child’s exposome (12–14) that has been previously validated in prior research (22). For the ABCD Study dataset, we used exposome factor scores derived in our prior work (23), where we applied longitudinal bifactor analysis to derive one general exposome factor (referred to here as exposome) and six specific exposome subfactors (“School,” “Family Values,” “Family Turmoil,” “Dense Urban Poverty,” “Extracurriculars,” and “Screen Time”) capturing a multitude of aspects of each child’s environment. Full details describing the derivation of these exposome factors and evaluation of model fit and reliability can be found in our previous work and are summarized briefly here (23). We began with 354 variables that were consistently assessed at the baseline, 1-y follow-up, and 2-y follow-up timepoints of the ABCD Study. These variables capture multiple dimensions of each child’s unique, complex environment, including child-reports, caregiver-reports, and geocoded data. These variables were then reduced in a data-driven manner to a set of 32 summary scores using the psych package in R (60). We then applied exploratory structural equation modeling (ESEM) (61) with bifactor rotation (BI-CF-QUARTIMAX) using Mplus version 8.4 (62) while accounting for clustering by families and stratification by data collection site. Factor loadings were constrained to be equal across timepoints to ensure longitudinal measurement invariance, such that the model would have the same configuration, loadings, and intercepts across all three timepoints (not relevant for the present cross-sectional study). We determined the appropriate number of factors to extract by a combination of model fit indices and assessments of interpretability. Additionally, we evaluated reliability indices such as factor determinacy, which indicates how representative each factor score is of the factors from which it was derived. The factor determinancy score for the general exposome factor exceeded conventional minima, suggesting that this score is sufficiently determined. For the PNC dataset, we also used exposome factor scores derived from an analogous bifactor analysis validated and described in detail in our previous work (22). To align the directionality of exposome scores across studies, we inverted the exposome scores derived in the PNC dataset such that higher exposome scores are associated with higher SES as in the ABCD dataset. As noted in our prior work, the features loading highest on the general exposome factor across studies are those typically associated with SES (e.g., household income, neighborhood-level area deprivation index, and parental education), such that higher exposome scores reflect multidimensional environmental characteristics associated with higher SES (23). Supplementary Material Appendix 01 (PDF)