The Effects of Active and Passive Leisure on Cognition in Children: Evidence from Exogenous Variation in Weather

Abstract

Leisure time activity is often positioned as a key factor in child development, yet we know relatively little about the causal significance of various specific activities or the magnitude of their effects. Here, we couple individual fixed effects and instrumental variable approaches in trying to determine whether specific forms of leisure contribute to gains in test performance over time. We merge a restricted access version of the Panel Study of Income Dynamics (PSID) Child Development Supplement (CDS), longitudinally collected from 1997 to 2007, with a database of over three million county-day observations of sunlight. We use this proxy for weather to instrument for the variation in physical, outdoor, sedentary, and screen time behaviors based on CDS time diaries. We find evidence that physical and outdoor activity positively influence math performance, while sedentary behavior and screen time exhibit the opposite effect. Moreover, the effect sizes range from a quarter to three quarters of a standard deviation per additional daily hour of activity, rendering them meaningful in a real world sense. Our stratified results indicate that children from less educated mothers and girls seem to be most sensitive to the effects of active and passive forms of leisure. We conclude with a descriptive examination of the trend lines between our data and the new 2014 CDS cohort, providing relevant contemporary context for our findings.

Time use patterns among children have long been a popular empirical concern in light of their intuitively influential contributions to developmental trajectory. In the social sciences, work tends to focus on time use as outcome, tacitly positioning specific behaviors as meaningfully contributing to the mechanics of social reproduction and individual life chances (Gershenson 2013; Kalil et al. 2016; Weininger, Lareau, and Conley 2015). Other work in the realm of public health customarily considers time use as a predictor, focusing on the associations between specific behaviors and phenomena like obesity or attention deficit disorder (Falbe et al. 2013; Nikkelen et al. 2014). Unfortunately, because of the practical difficulties of inferring causal relationships using observational data, much of what we suspect to be involved in altering developmental trajectory is ultimately based on intuition and theory rather than robust empirical evidence. While some work does offer experimental evidence of the effects of select activities using randomized control trials (RCT)—e.g. Loprinzi and Kane 2015—it is difficult to infer how routinized behaviors affect outcomes of interest solely based on interventions among carefully targeted respondent populations. That is, RCT studies are by definition unable to reproduce the real-world conditions in which behaviors emerge and unfold, while their limited scale prevents inferences on how effects vary among subgroups. Moreover, the paucity of plausibly causal estimates based on observational data hamstrings our ability to approximate the magnitude of effects and thus the real-world significance of certain behaviors, whether intuitively or associatively positive or negative.

Here, we exploit the robust association between weather patterns and a select subset of leisure activities in attempting to draw a clearer connection between behavior and cognition, a crucially important factor in child development and achievement. We link a database of more than three million county-day measures of sunlight compiled by the U.S. government to corresponding time use dairy and cognitive assessment records longitudinally collected in the first Child Development Supplement (CDS) module of the Panel Study of Income Dynamics (PSID). By exploiting the plausibly exogenous variation in weather on the day of measured activity, we can deliver estimates of how more ‘active’ and ‘passive’ forms of leisure—which are particularly sensitive to weather and sunlight—contribute to shaping cognitive assessment over time.

We find that physically active and outdoor leisure activity (for American children, largely synonymous) positively contribute to growth in math skills, while sedentary activity and screen time in its various manifestations exhibit the opposite effect. These effects are meaningfully large in a real world sense, ranging from a rise or fall of a fifth to more than half a standard deviation in math scores per additional daily hour spent on the specific activity. We then present stratified results, with the clearest differences based on caregiver’s education (in our data, almost always mothers)—with strong effects of active and passive leisure on cognition evident among the children of high school graduates, and smaller, statistically insignificant relationships among those with college-educated parents. We also highlight relevant differences in time use among families in our data which may contribute to some the patterns found in our stratified results, and relate to achievement gaps among children more generally. We conclude by descriptively examining time use among a comparably aged subset of the new CDS cohort in 2014 to adolescents in our main data, offering suggestive evidence of recent trend lines in behavior among US children, and providing relevant contemporary context for our findings.

Childhood Inequality, Time Use, and Cognitive Growth

Much of the literature on child development is animated by longstanding gaps in academic achievement and cognitive assessment. Baseline differences in test performance (i.e. before entering kindergarten) between white and black children, to take one example, are often found to be as large as a full standard deviation (Bond and Lang 2013). Differences patterned on race are significantly narrower than they were in the mid-20th century by any reasonable estimate, yet there is evidence this convergence has stagnated since the 1990s (Neal 2006), and may be attributable to cohort-specific gains which manifested in the 70s and 80s (Chay, Guryan, and Mazumder 2009). These differences in measured cognition reverberate through the life course, and have been implicated in reduced intergenerational mobility and flatter wage trajectories among African Americans, controlling for other significant factors in achievement like grades and attitudinal traits (Bhattacharya and Mazumder 2011; Hall and Farkas 2011). The differences between high- and low-socioeconomic status (SES) children are even starker. Despite some evidence of a modest convergence in baseline readiness between 1998 and 2010, differences in aptitude between the richest and poorest children have increased by over a third since 1970s born cohorts and are currently about twice as large as the black-white achievement gap (Reardon and Portilla 2015). Class inequalities have primarily been driven by gains among high-SES daughters compared to poorer girls and boys of all backgrounds, highlighting an emerging and interrelated inequality based on sex (Bailey and Dynarski 2011). These initial advantages extend into college enrollment and degree attainment, with sex-based achievement gaps continuing to widen both within and between birth cohorts (Buchmann and DiPrete 2006).

While the etiology of comparative disadvantage is difficult to disentangle, significant differences in cognition before kindergarten and the tendency for gaps to widen during the summer months highlight the crucial importance of the home environment (Alexander, Entwistle, and Olson 2007; Condron 2009). Significant baseline differences among high- and low-SES toddlers have been found as early as eighteen months (Fernald, Marchman, and Wiesleder 2013), and indeed may be large enough in magnitude to largely account for the lag in US educational assessment compared to similar industrialized countries (Merry 2013). While institutional and geographic factors are doubtlessly crucial to the formation of intellectual capital, these findings suggest that variations in home activity (i.e. outside of formal school or childcare settings) may have cumulative effects on development that rival or even exceed those of school or neighborhood quality (Potter and Roksa 2013).

Research in the theoretical tradition of Bourdieu has long stressed the importance of activities geared toward curating a disposition suitable for succeeding in school and navigating institutions later on in life, like engagement with high art, music, and literature (DiMaggio 1982). More recently, the theory of ‘concerted cultivation’ has been developed as an extension of this work to describe how the childrearing practices and time uses of children systematically differ by race and social class. Using extensive qualitative evidence, Lareau (2000; 2002) describes how the leisure time of children in middle class families is highly structured and organized compared to the more informal, family-oriented working class home environment. Quantitative work generally supports the notion that material and cultural resources are predictive of enrollment in more formalized activity (Weininger, Lareau, and Conley 2015), and that these ‘cultivated’ time uses positively predict school achievement and test scores (Gaddis 2013; Jæger 2011).

Meanwhile, public health research in this vein is largely concerned with the division between active and sedentary forms of leisure time expenditure, rather than the formal-informal distinction. Interestingly, many of the same systematic differences between classes that characterize enrollment in music lessons or museum attendance are mirrored in sedentary activity and media consumption. Recent reviews have illustrated that low-SES children are significantly more likely to engage in sedentary behavior than their peers in the U.S. and other high-income countries (Leech, McNaughton, and Timperio 2014; Mielke et al. 2017). Related research also highlights a growing SES disparity in adolescent obesity prevalence that is masked by a recent trend of overall stabilization, consistent with gaps in relevant time uses (Frederick, Snellman, and Putnam 2014).

Systematic reviews generally find substantial positive associations between physical activity and cognitive wellbeing across the life course, but particularly among children and seniors (Esteban-Cornejo et al. 2015; Prakash et al. 2015). Aside from cross-sectional correlates, which are suggestive of better overall health profiles and the long-run lifestyle differences they reflect, other work finds that even relatively modest short-term interventions in physical activity can manifest in changes in measured cognitive function. Usually, these consist of interventions which try to capture the ‘acute’ effects of exercise (e.g. studies estimating the immediate effects on cognitive performance of 30 minutes of moderate-intensity exercise) or more medium-run behavioral modification (e.g. engaging in physical activity for 20 minutes a day over two weeks, and comparing pre- and post-treatment cognitive assessment). Meta-analyses of research on the effects of acute exercise intervention suggest a positive effect on executive functioning (i.e. skills that allow for planning, multitasking, self-control, etc.), particularly in children (Ludyga et al. 2016). Because executive function seems particularly sensitive to interventions in activity, this suggests that test performance could be affected by, for instance, impulse inhibition more so than improvements in working memory or reaction time.

Other reviews of medium-term interventions find more general positive cognitive effects—i.e. across domains that include academic skills—but effect sizes differ widely among the studies (Lees and Hopkins 2013). These experimental results illustrate that the benefits of exercise may partly reflect a direct dose-response relationship that can manifest in the short- or medium-term, rather than simply reflecting long-term differences in wellbeing. In addition to differences in cognitive performance, some cross-sectional and RCT studies also find physical, structural differences in brain integrity between more- and less-fit children (e.g. white or gray matter volume) (Erickson, Hillman, and Kramer 2015). The extant evidence illustrates that activity (and inactivity) may be related both to physical changes in the brain itself, along with cognitive improvements that can positively affect test performance through numerous channels—including those that relate to behavior (e.g. ability to focus) rather than intellectual skill per se. These observations are nevertheless preliminary and provisional, as researchers have tended to focus on the effects of physical behavior on cognition later in the life course (i.e. older adults) rather than in childhood (Prakash et al. 2015).

On the other end of the time use spectrum, research has often focused on the effects of screen time, as it constitutes an outsized proportion of American children’s leisure activity. Still, because screen time and sedentary behavior are so intertwined in the American context, much of this work focuses on physical outcomes like body mass index, and extant findings on cognition are rather limited and equivocal. Several longitudinal studies have found early exposure to media and television in both infants and toddlers to predict worse cognitive outcomes later (Tomopoulos et al. 2010; Zimmerman and Christakis 2005). Yet other research which specifically examines the impact of educational programming (e.g. Sesame Street) finds positive and significant effects, signaling the importance of content (Mares and Pan 2013). Quasi-experimental studies on the cognitive effects of computer use have found that effects are largely null (Fairlie and Robinson 2013), or negative but modest in magnitude (Vigdor et al. 2014). With respect to video games, Suziedelyte (2015) uses a fixed effects research design and finds that gaming (both console- and computer-based) is positively associated with math ability. Small-scale experimental studies have also shown that gaming may positively affect spatial reasoning and executive functioning (Kühn et al. 2014). Yet other experimental research has found that even very short exposures (i.e. minutes, not hours) to high-intensity media can depress the executive functioning in children (Lillard and Peterson 2011), negatively mirroring related RCT study findings in physical activity.

Cognition: Practical Significance and Empirical Measurement

Establishing the real world significance and construct validity of cognition as it is commonly measured in aptitude tests is less straightforward than it may seem. For instance, some have provocatively argued that the primary driver of labor market success and attendant social reproduction is not cognition or skills per se but the attitudes and personality traits which are inculcated in the higher strata (Bowles and Gintis 1976; 2002). Other researchers have similarly argued for the primacy of noncognitive skills in fostering long-run outcomes in educational attainment (Heckman and Kautz 2013). While recent work that attempts to econometrically isolate the causes of achievement supports the importance of intangibles like attitude and disposition, cognition still appears to be the primary determinant of economic success on both individual and national scales (Hanushek 2013; Lundborg, Nystedt, and Rooth 2014).

There is also the related question of what cognition as it is customarily rendered in the social sciences signifies. Tests which measure aptitude are only imperfect proxies of cognitive ability, and necessarily signal other noncognitive factors like motivation or goal orientation (Heckman and Kautz 2013). This is why, as Heckman and colleagues point out, achievement tests are better predictors for future success than fluid IQ, as they are capturing these intangibles which help mediate the effects of ability on outcomes. While we recognize that our measures of cognition are almost certainly capturing other dimensions of overall ability like motivation, these tests are purer reflections of aptitude than grades, to take one example, and are the closest we can reasonably hope to get in capturing intellect using secondary data (Borghans et al. 2011). Moreover, even if behaviors are affecting cognitive performance through a more nebulous channel than raw intellectual ability (i.e. through more robust executive function and impulse control), ultimately they still produce measurable effects on assessment and can be considered net positive or negative.

Methods and Research Design

Dataset and Variables

Our core data come from the original CDS module of the PSID (Panel Study of Income Dynamics 2017). The PSID is a nationally representative, longitudinal study, with respondents surveyed annually from 1968–1997, and biennially thereafter up to present. In 1997, the CDS component collected additional data on children (newborns to 12-year-olds) and their parents or guardians. The second and third waves followed up on eligible minors in 2002 and 2007, respectively, with any child 18 or over transferring to the Transition to Adulthood (TA) supplement, which consists of a more limited array of data that do not include time diaries or cognitive assessments. A major strength of the CDS is the detailed measurement of daily behavior offered by the time diary module. Parents and children listed the child’s activities over 24 hours in an open ended diary with no predetermined minimum time increment on a randomly selected day during the school week and weekend in each CDS wave. These activities were then coded by PSID researchers, and in their raw form may be aggregated to obtain a detailed snapshot of how and where children spent their time.

While we aggregated many individual activities into cohesive categories in the earlier stages of our analyses, we ultimately include only those that exhibit a robust first-stage association with our weather measures in the full specifications. These consist of the total time on the diary day doing: 1) Any leisurely physical activity that is not formal or organized (e.g. walking or hiking, but not baseball practice); 2) Any outside activity (including most forms of physical activity); 3) Any passive sedentary activity (‘hanging out,’ relaxing, or media consumption, but not reading or homework); 4) Television shows or movies across any platform; 5) Video games across any platform; and 6) Total screen time (media, gaming, and any additional non-educational computer or handheld use). A full table of our time use variables and their stratified descriptive statistics across the three waves is located in the online appendix (Appendix Table S1). We use weekday measures of time use, and necessarily exclude summer vacation as the CDS was administered only during the school year. The CDS also collected assessments of math and verbal ability using the Woodcock Johnson Revised (WJ-R) tests, which are well-established, age-standardized metrics of reasoning ability in children as young as three years old (see: Woodcock and Johnson [1989] for more detail). We use both applied problems (open-ended math word problems) and broad reading (reading, writing, and oral exercises) age-standardized scores as outcomes, but present the results for the latter in the online appendix (Appendix Table S2) because our time use variables of interest fail to significantly predict verbal assessment.

Analytic Strategy

We begin by leveraging the longitudinal aspect of the data and employ an individual fixed effects (FE) approach, which zeroes out any time invariant factors. In the FE methodological framework, each respondent effectively acts as their own control by only considering within-child variation in behaviors, covariates, and test scores over time, and as a matter of course accounts for stable characteristics like race, birthweight, or sex. Still, even if we assume that time varying unobservables were not biasing results, the potential for reverse causation could complicate our interpretation of them. That is, parents may deliberately modify their child’s behavior because they are not performing as well in school in a way that is reflected in cognitive tests. Specific activities would then incorrectly be identified as aiding or impeding development when the causal directionality is actually reversed. To better address issues of endogeneity, we use a combined fixed effects instrumental variable (FE-IV) approach, where we exploit the variation in weather to estimate the effect of time uses on cognition across waves.

Consider we have a variable Z that is associated with an endogenous variable X, but has no direct causal relationship with the outcome Y except through its effect on X. The variable Z—the instrument—could then be used to estimate the causal effect of X on Y because Z would only affect Y through the mediator. In the first stage, the endogenous variable of interest—here, a given time use measure—is regressed on the instrument (sunlight) and any relevant covariates:

$$X_i={\beta }_0+{\mathit{\beta }}_1{\mathit{Z}}_{\mathit{i}}+{\beta }_2{K}_i+{\epsilon }_1$$ (1)

Where Z is the instrumental variable, K is a suite of covariates, and ε is the error term. In the second stage, the parameter X_i* —which represents the first stage effect of Z on X_i, adjusting for covariates—is then incorporated into the second stage equation which captures the effect of the instrumented endogenous variable on the outcome:

$$Y_i={\beta }_0+{\mathit{\beta }}_1{\mathit{X}}_{\mathit{i}}^{\mathbf{*}}+{\beta }_2{K}_i+{\epsilon }_2$$ (2)

These second stage estimates provide the effect sizes of time use on cognition, but only through the influence of the exogenous instrument sunlight. For our estimates to be unbiased, however, it is crucial that our instrument meets the exclusion restriction criteria of not having a direct effect on cognition. That is, that there is no relationship between sunlight or shade and cognition except through how the intervening behavior is influenced.

Challenges to this assumption include the possibility that season itself is predictive of cognitive performance. Recent research suggests that while there does appear to be a seasonal pattern to cognitive processing presumably based in circadian rhythms, these do not manifest in actual differences in measurable ability (Meyer et al. 2017). That is, the brain seems to work harder in certain seasons than it does in others to exhibit stable performance, with the only difference being in energy expenditure. Another possibility is that restrictions in sunlight depress cognitive performance through the channel of negative mood or depression, popularized in the psychological literature as Seasonal Affective Disorder (SAD). Recent work finds that rainfall seems to depress subjective life satisfaction in cross-sectional but not longitudinal models, with effect sizes in the former that are trivial compared to other determinants of mood like economic status (Barrington-Leigh and Behzadnejad 2017). Other work finds that temperatures above 70°F tend to induce negative emotions, while even particularly cold weather (<20°F) is associated with higher subjective wellbeing (Noelke et al. 2016). In the only large-scale study that directly examines the effect of weather on cognition that we are aware of, Kent et al. (2009) find that among adults over 45, sunlight is positively associated with cognitive assessment, but only among individuals suffering from depression. In their small scale experimental study, Keller et al. (2005) found that time outdoors moderates the relationship between weather and affective and cognitive outcomes, which suggests that behavior (i.e. going outside, rather than general abstract satisfaction with positive weather conditions) is the linchpin in the causal process. Taken together, these findings suggest that the impact of the weather on mood is rather ambiguous, and that sunlight seems to affect outcomes chiefly through the act of being in or outdoors—i.e. through activity and behavior.

There is also the possibility that our results are driven by differences in the bioavailability of Vitamin D, which aside from specific foods like fortified milk is largely endogenously produced via sunlight. Though Vitamin D is integral to skeletal health, it does play a role in neurological development and its deficiency tends to be associated with neuropsychiatric disorders (Eyles et al. 2009). Conceptually, we do not necessarily see this as a violation of the exclusion restriction because in order to get the benefits of sunlight on cognitive performance via the vitamin D pathway, children do still need to go outside and expose their skin to the sun as even conventional window glass blocks most ultraviolet rays. This would merely alter the causal pathway, such that physiologically endogenous vitamin synthesis acts as the causal mechanism between behavior and performance instead of through some other physical or psychological channel.

There are also other more clinically significant sun-related factors which work in the opposite direction, and would bias our estimates downward. Cutaneous exposure to ultraviolet light has long been recognized as degrading folic acid—or B9, an essential vitamin that acts as a precursor to DNA methylation—in both clinical trials and in vivo (Off et al. 2005). Indeed, the degradation of folate induced by ultraviolet light has been positioned as a prominent factor in the evolution of melanin content in humans because of its role in fetal neural tube development and related reproductive success (Jablonski and Chaplin 2000). Recent reviews have argued that the consumption of folate in particular and other essential B vitamins in general are important to cognitive performance and brain health at every stage of the life cycle, in addition to during pregnancy (McGarel et al. 2015). Sunlight also plays a key role in the formation of ozone (O₃) from the environmental precursors of volatile organic compounds (VOC), carbon dioxide (CO), and nitrogen dioxide (NO₂), and these chemical reactions are accelerated in higher temperatures. The links between pollution and cognitive development are rather clear and well-established empirically, with prenatal exposures being particularly damaging (Peterson et al. 2015). While more immediate effects are difficult to estimate, extant work tends to show that childhood exposure to these compounds is associated with reduced academic achievement and cognitive functioning, as well as with accelerated decline later in the life course (Clifford et al. 2016). Researchers have also illustrated related links between seasonal warm-weather allergens and decreased performance on high-stakes exams (Bensnes 2016).

Because time diaries were assigned based on randomly selected dates, we assume no systematic relationship between when behavior was measured and any relevant demographic correlates. Additionally, we include a suite of household shocks which are linked in the literature to either childhood development or other fundamentals like family SES status, including indicators of an additional birth (Sandberg and Rafail 2014), the departure of a parent from the household (Tach 2015), and whether the family changed residence from a prior wave (Jelleyman and Spencer 2008). We also include a five year rolling average of real family income preceding the measurement year, and the poverty rate of the home census tract as a proxy for neighborhood conditions. We found parental education to be so stable across waves as to function as time invariant, and thus dropped it from the specifications. Finally, we control for ‘educational time,’ consisting of the total time in school and doing homework summed together. Children experience systematic differences in their exposure to formal academic instruction due to the geographic variation in the minimum number of hours which states and other localities mandate students receive. We found no noticeable difference in results by excising homework time and confining the measure to the length of the schoolday alone.

To capture the effect of sunlight on behavior, we use daily historical records of insolation (sunlight measured in kilojoules per square meter at the county level), collected as part of the North America Land Data Assimilation System (NLDAS) and disseminated by the Centers for Disease Control (CDC) Wonder database (Centers for Disease Control 2017). Raw insolation values were calculated by the CDC on grid cells, then aggregated up to the county level and undergoes quality control before being publicly released. We then coupled these daily county-level values of sunlight to the PSID CDS based on the individual days when the time diaries were completed and the places children lived using a restricted access version of the core data. Because the NLDAS measures sunlight at the ground level, in practice it also acts as a proxy for broader weather patterns, picking up precipitation, overcast skies, etc. In Figure 1, we compare the sunniest (Arizona) and cloudiest (Vermont) places to the national average in daily insolation over the same time periods, and over the course of the year. We also show the variation in average annual insolation at the county level across the U.S. from 1997–2008 graphically in Figure 2.

Figure 1. Extremes in Sunlight Compared to National Average.

Figure 2. Geographic Variation in Average Sunlight, 1997–2007.

There are three possible sources of variation in sunlight from wave to wave in our data: 1) The effect of children moving to a qualitatively different climate across waves (e.g. testing in May 1997 and 2002, but having moved from Vermont to Arizona in the interim); 2) The broader changes in weather due to taking the test in different seasonal conditions (e.g. testing in December 1997 and June 2002); and 3) The random weather perturbations that the jagged edges of the curves in Figure 1 represent (e.g. engaging in activity and testing during a seasonably sunny few days in May 1997, and then again during a tropical depression in May 2002). The first two represent the most obvious potential sources of bias. For instance, there could in theory be some character trait that prompts a simultaneous migration to the Sunbelt and an outsized growth in cognition, which would bias our estimates upward. Yet because we found that only a small minority of families move across waves (about 10% of the sample over the ten year span), and even then mostly to geographically proximate locales, we feel this is an unrealistic candidate confounder. Further, the ‘whether moved’ control is insignificant in every one of our specifications, while restricting the sample to ‘never-movers’ produces substantively similar results, with slightly larger coefficients.

The prospect of a seasonally-based confounder is more problematic considering the realities of the data. Estimates based on a convergence in active and sedentary time uses as the school year progresses and the weather gets sunnier would be biased if the real driver of improving test scores is simply having had a longer time in school since the prior summer within each wave. For instance, we may find that children who have their time uses and cognitive performance measured in December in the first wave and June in the second exhibit an increase in active behavior and a co-occurring growth in cognitive performance. Yet these cognitive improvements may be attributable merely to the timing of the test further along in the school year in the second wave, and the substantial accumulation of knowledge children exhibit by virtue of the additive gains they experience as the school year progresses (Fitzpatrick, Grissmer, and Hastedt 2011).We control for the cumulative gains realized by the progression of the school year by including a ‘time in school’ variable backdated to the day after Labor Day (largely but not universally coincident with the start of the school year in the US) from the date of the WJ-R test in each CDS wave. As we will discuss in the results, time in school was a significant predictor in the verbal but not math specifications. We also perform a robustness check whereby we restrict our sample to sunlight outliers in an attempt to capture the influence of more random as opposed to seasonally-based weather, finding substantively similar results—with the exception of gaming, which is no longer significantly predictive (see: Table S3 in the online Appendix).

Conceptually, we intuit that short- and medium-run changes in sunlight result in consequent reductions or increases in specific forms of leisure activity over the same time period, which then may have short- and medium-run cognitive effects. For instance, we might expect less sunlight during a storm spell to result in more sedentary time at home, and less time engaging in physical activity outdoors. Prior work on leisure time expenditure expectedly illustrates a positive association between sunlight and physical and outdoor activity, with the opposite for sedentary behavior and screen time (Lee, Gino, and Staats 2014; Zivin and Neidell 2014). Others have also leveraged the random occurrence of weather shocks to explore electoral outcomes (Madestam et al. 2013), criminal behavior (Jacob, Lefgren, and Moretti 2007), and civil conflict (Miguel, Satyanath, and Sergenti 2004), to name a few.

Though we match sunlight data to the single time diary day, we recognize that these discrete daily averages will usually reflect the time of year they take place in. Thus while we instrument with the average sunlight in KJ/m² over the 24 hour period consistent with a time diary recorded on, say, Wednesday June 5th, we recognize by virtue of seasonality we will in many cases be approximating sunlight on Tuesday the 4th, Thursday the 6th, etc. Because we view this in theory as more or less a dose-response relationship in the short and medium terms—like a milder cousin to other exposures which induce changes in cognitive assessment, like neighborhood violence (e.g. Sharkey 2010)—we exclude observations where the WJ-R test was administered before the time diary day, and cap the lag time at nine days so as to avoid an unrealistic coupling of exposure and outcome. About 16% of our observations were cases where children either took the test on the same or following day as the time diary was recorded, while 80% were lagged by less than 10 days. Specifications which were not bounded by lag time restrictions produced substantively similar results, but with smaller coefficients and greater precision, consistent with the growth in sample size and reduced signal after including cases where testing substantially lags or precedes treatment.

Missing Data and Survey Weights

Another crucially important potential source of bias arises from missing data. Even when attrition is properly accounted for, estimates will be biased if missingness is not randomly determined, and is conditioned on a factor that also affects cognitive performance. Data missingness is not a trivial concern using PSID-CDS data; in the inaugural 1997 wave, about 20% of families otherwise eligible and contributing to the study did not complete time diaries. Here, we use multiple imputation (MI) to handle missing cases, which involves an iterative process where observed data are used to predict missing values multiple times, resulting in a pre-specified number of imputed datasets which are then combined to obtain optimal parameter estimates (McKnight et al. 2007). We follow the best practices suggested by Graham, Olchowski, and Gilreath (2007) and conservatively use 150 imputations. We also include the dependent variables in our imputation procedure, as suggested by Young and Johnson (2010). We present imputed results alongside our main estimates, as well as conventional OLS and FE specifications. We find that weighting our data do not affect our estimates.

Results and Discussion

In Table 1, we present our core results on the effects of physical, outdoor, and sedentary activity on math scores. In our FE-IV models, each additional daily hour of physical and outdoor activity produced a gain of about 40 percent of a standard deviation in math scores. The negative effect of sedentary behavior is less substantial, but still accounts for a reduction of about a fifth of a standard deviation with each additional hour. We also find substantial effect sizes for television and gaming, with each additional hour resulting in a loss of 38 and 63 percent of a standard deviation, respectively. For total screen time, the effect size is a more modest reduction of about 1/5th of a standard deviation with each additional daily hour. The effects of sedentary behavior and screen time are particularly meaningful given that by the 2007 wave, the adolescents in our data spent over 2.5 hours each day performing these passive leisure activities. This suggests that even halving the time spent engaging in sedentary behavior or consuming electronic media could manifest in significant math gains.

Table 1.

Behaviors and Math Outcomes

	Outcome: Applied Problems Standardized Scores (Math Word Problems); IV: Sunlight (KJ/m2) on Activity Day
	OLS	FE	FE-IV	Imputed	OLS	FE	FE-IV	Imputed	OLS	FE	FE-IV	Imputed
Educational Time	.457*** (.093)	.501*** (.104)	.819*** (.158)	.905*** (.166)	.459*** (.093)	.506*** (.104)	.870*** (.174)	.930*** (.174)	.296** (.100)	.445***(.117)	−.457 (.345)	−.200 (.257)
Days in School	.008 (.004)	.010* (.005)	−.004 (.007)	−.003 (.006)	.008 (.004)	.010* (.005)	−.004 (.007)	−.003 (.006)	.007 (.004)	.010* (.005)	−.007 (.008)	<.001 (.005)
Physical Activity	.171 (.317)	.376 (.321)	6.587** (2.116)	6.927** (2.280)
Outdoor Activity					.186 (.311)	.434 (.308)	6.854** (2.240)	6.766** (2.231)
Sedentary Activity									−.550*** (.133)	−.127 (.142)	−3.269** (1.090)	−2.876** (.927)
Controls/Shocks	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
N	3646	3646	2210	6693	3646	3646	2210	6693	3646	3646	2210	6693
F	52.46***	4.41***	4.41***	4.91***	52.63***	4.48***	4.17***	4.81***	54.69***	4.48***	4.40***	5.26***
First Stage F statistic	—	—	36.26	—	—	—	31.88	—	—	—	30.00	—
Educational Time	.326** (.096)	.500*** (.114)	−.741 (.534)	−.735 (.533)	.457*** (.095)	.487*** (.107)	−.101 (.298)	−.040 (.250)	.362*** (.099)	.489***(.116)	−.410 (.340)	−.181 (.263)
Days in School	.008* (.004)	.011* (.005)	.007 (.006)	.006 (.006)	.008* (.004)	.011* (.005)	−.002 (.007)	.002 (.006)	.008 (.004)	.011* (.005)	−.004 (.007)	.001 (.005)
Television	−.643*** (.175)	.097 (.188)	−6.419* (2.621)	−6.951* (2.880)
Video Games					.192 (.282)	.097 (.282)	−10.735* (4.508)	−11.340** (4.235)
Total Screen Time									−.344* (.144)	.030 (.157)	−3.651** (1.250)	−3.208** (1.049)
Controls/Shocks	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
N	3646	3646	2210	6693	3646	3646	2210	6693	3646	3646	2210	6693
F	54.14***	4.11***	2.80**	2.81**	51.89***	4.15***	2.69**	3.28**	52.63***	4.15***	4.16***	5.09***
First Stage F statistic	—	—	11.54	—	—	—	10.69	—	—	—	26.49	—

Note: Cluster-robust standard errors are in parentheses.

^*p < .05

^**p < .01

^***p < .001 (two-tailed tests).

Sample sizes decline slightly in our FE-IV specifications because of a small number of observations with missing geocodes.

Standard OLS estimates (pooled across waves) do evince significant, negative relationships between sedentary behavior, television, total screen time, and math scores, but with noticeably smaller coefficients. Our conventional FE estimates, however, render null effects across the board (though removing the ‘education’ control [length of school day plus homework] produces significant results for sedentary behavior and screen time in the expected direction [estimates not shown]). This raises the possibility of a selection bias mechanism whereby higher-ability children spend more time sedentary or consuming media. Another possibility is that the effects we find in FE-IV specifications reflect relationships among ‘compliers’ in our natural experiment that are not common across the CDS sample (more on this later in the results). Our imputed specifications in Table 1 produce substantively similar estimates, which suggest that data missingness is not systematically biasing our core results.

Among the stratified results, the clearest differences are among the college- and non-college-educated caregiver subsamples (Table 2). For each of our time use variables, coefficients are about 5–6 times as large for the children of non-college-educated mothers compared to their peers, and the null results for the latter are likely at least partly due to more modest first-stage associations, which are near or under commonly proposed weak-instrument thresholds. (While there is no clear consensus value, specialists generally recommend first stage F statistics of the excluded instrument above 10 [Stock, Wright, and Yogo 2002].) That is, sunlight is a noticeably poorer predictor of behavior among children with college-educated mothers compared to their peers, and this results in higher standard errors (but cannot account for the lower point estimates).

Table 2.

FE-IV Results Stratified by Primary Caregiver’s Education

	Outcome: Applied Problems Standardized Scores (Math Word Problems); IV: Sunlight (KJ/m2) on Activity Day
	College Educated Primary Caregiver						Non-College Educated Primary Caregiver
Educational Time	1.130*** (.288)	1.171 (.384)	.783 (.571)	.396 (1.500)	.913 (.385)	.818 (.525)	.818 (.200)	.862 (.213)	−.782 (.499)	−1.251 (.884)	−.213 (.344)	−.761 (.506)
Days in School	.017 (.019)	.017 (.019)	.018 (.017)	.023 (.014)	.019 (.016)	.018 (.017)	−.010 (.009)	−.011 (.009)	−.014 (.010)	.002 (.009)	−.002 (.008)	−.009 (.009)
Physical Activity	2.413 (4.760)						8.286** (3.043)
Outdoor Activity		2.782 (4.956)						8.370** (3.093)
Sedentary Activity			−1.047 (1.803)						−4.059* (1.572)
Television				−3.922 (8.283)						−8.268* (4.187)
Video Games					−2.224 (3.748)						−12.130* (5.591)
Total Screen Time						−1.107 (1.896)						−4.634* (1.858)
Controls/Shocks	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
N	634	634	634	634	634	634	2779	2779	2779	2779	2779	2779
F	2.76**	2.62*	3.29**	2.54*	3.20**	3.32**	2.78**	2.72**	2.54**	1.42	1.81	2.32*
First Stage F Statistic	7.80	4.46	7.77	.77	9.39	6.88	19.60	18.86	18.05	6.24	9.08	15.50

Note: Cluster-robust standard errors are in parentheses.

^*p < .05

^**p < .01

^***p < .001 (two-tailed tests).

Further, as a matter of course, we cannot produce estimates for non-compliers who, say, mostly stay inside regardless of weather conditions. We suspect that behavior may be less sensitive to weather conditions among the children of college-educated caregivers because they are engaged or enrolled in more formalized activities (Lareau 2000; 2002; Weininger, Lareau, and Conley 2015), which would be more likely to take place ‘rain or shine.’ Intuitively, one would suspect that children who have unstructured free time would be more sensitive to weather conditions compared to those taking piano or violin lessons, for example. This leaves open the possibility that a different estimation strategy may uncover significant relationships among time use and cognition among the children of college-educated parents that we do not observe here.

We also see variation in effect sizes stratified by sex and race. (Stratifying on sex or race does not reveal differences in the responsiveness of behavior to weather conditions, unlike the caregiver college/no college estimates.) Effect sizes are at least twice as large for girls compared to boys, with all time uses exhibiting significant relationships with math assessment except for video games (Table 3). For whites, the effects of physical and outdoor activity was greater in magnitude than for African Americans, with the opposite true of sedentary behavior and screen time—though the latter two are only significantly predictive at the 10% level for the black subsample (Table 4). Our stratified results based on imputed data, however, deliver larger coefficients and smaller p values for the black subsample, and effect sizes are about 30–60% larger compared whites in those specifications depending on the behavior (online Appendix Table S6). Because other imputed stratified estimates (see Tables S4-S5 in the online appendix) deliver similar or more conservative estimates compared to those using listwise deletion, it is possible that our main race-stratified results are downwardly biased due to nonresponse. One possibility in explaining the differences in effect sizes (or whether behaviors are significantly predictive) stratified by race or sex is the presence of a threshold effect, whereby the marginal consumption of a specific behavior becomes either more or less impactful past a certain point. For instance, girls are over three times more sensitive to the effects of screen time than boys, yet the latter consume 25–40% more depending on the wave. Boys may thus reach a saturation point where after each additional unit of electronic media consumption no longer appreciably affects performance.

Table 3.

FE-IV Results Stratified by Sex

	Outcome: Applied Problems Standardized Scores (Math Word Problems); IV: Sunlight (KJ/m2) on Activity Day
	Girls						Boys
Educational Time	.787** (.240)	.821** (.250)	−1.115 (.586)	−1.422 (.875)	−.387 (.471)	−1.070 (.618)	.813*** (.212)	.853*** (.237)	.163 (.418)	−.012 (.606)	.267 (.371)	.185 (.402)
Days in School	−.011 (.008)	−.011 (.010)	−.018 (.013)	.004 (.010)	−.002 (.013)	−.011 (.012)	.003 (.009)	.003 (.009)	.002 (.009)	.011 (.008)	.002 (.010)	.003 (.009)
Physical Activity	10.403** (3.418)						3.149 (2.555)
Outdoor Activity		10.494** (3.491)						3.348 (2.733)
Sedentary Activity			−4.876** (1.731)						−1.727 (1.421)
Television				−8.264* (3.692)						−4.052 (3.695)
Video Games					−30.782 (18.233)						−3.931 (3.455)
Total Screen Time					−5.795** (2.227)							−1.846 (1.523)
Controls/Shocks	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
N	1807	1807	1807	1807	1807	1807	1815	1815	1815	1815	1815	1815
F	2.30*	2.24*	2.68**	1.68	.93	2.22*	3.37***	3.31**	3.15**	2.33*	3.07**	3.14**
First Stage F statistic	21.58	20.53	17.45	8.04	3.79	13.63	19.96	16.77	13.70	4.08	7.00	13.07

Note: Cluster-robust standard errors are in parentheses.

^*p < .05

^**p < .01

^***p < .001 (two-tailed tests).

Table 4.

FE-IV Results Stratified by Race

	Outcome: Applied Problems Standardized Scores (Math Word Problems); IV: Sunlight (KJ/m2) on Activity Day
	White Subsample						Black Subsample
Educational Time	.967*** (.256)	1.055** (.306)	−.262 (.437)	−.662 (.754)	.132 (.276)	−.213 (.429)	.636** (.242)	.667** (.252)	−.580 (.524)	−.539 (.542)	−1.087 (1.948)	−.674 (.603)
Days in School	−.001 (.011)	−.002 (.011)	−.004 (.012)	.008 (.010)	.001 (.011)	−.003 (.012)	−.001 (.009)	−.001 (.009)	−.002 (.009)	.008 (.009)	.003 (.016)	.003 (.009)
Physical Activity	7.115* (3.389)						5.849* (2.797)
Outdoor Activity		7.754* (3.837)						5.916* (2.836)
Sedentary Activity			−2.786* (1.342)						−3.031† (1.562)
Television				−5.717 (3.325)						−4.869 (2.766)
Video Games					−9.281 (4.814)						−20.456 (25.129)
Total Screen Time						−3.028* (1.495)						−3.951† (2.164)
Controls/Shocks	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
N	1755	1755	1755	1755	1755	1755	1460	1460	1460	1460	1460	1460
F	2.71**	2.45*	3.38***	2.35*	2.14*	3.25**	1.49	1.47	1.46	1.03	.45	1.25
First Stage F statistic	15.69	12.11	19.42	7.18	11.14	16.69	15.17	14.63	10.85	6.61	.75	7.83

Note: Cluster-robust standard errors are in parentheses.

^*p < .05

^**p < .01

^***p < .001 (two-tailed tests).

The differences in the effects of physical and outdoor activity are more difficult to speculate on because all children exhibit relatively low levels of either behavior. If children are customarily indoors and relatively inactive whatever their background, why might we see such a difference in effect sizes? One reason could be a systematic difference in opportunity cost structures, such that some children are substituting active behavior for less productive time uses. While we seek to determine the effects of qualitatively orthogonal behaviors—i.e. active v. sedentary—we do not code these in such broad ways as to render them in zero-sum opposition. That is, because we deliberately exclude many kinds of behavior that could fall under the outside/active (e.g. walking to or from school) or sedentary (e.g. doing homework on the couch) categories, reductions in one do not necessarily lead to increases in the other. Thus some children may be substituting physically active behavior for, say, watching television, while for others that same active leisure may be impinging on studying or practicing violin.

With respect to screen time behaviors, we also cannot account for the content of what children were watching or what games they were playing, which could also influence our results. (The CDS does include general information on the genre of television programs children consumed, but only for the 1997 and 2002 waves.) If higher-SES children are watching National Geographic and their peers are watching daytime talk shows or cartoons, one might expect a qualitatively different effect. Another consideration is why we did not see similar effects using verbal outcomes (see online Appendix Table S2 for broad reading results). First, fewer children were eligible to take the reading tests because of different age restrictions. Still, we suspect the reason may lie not in the structure of the data but in how math and verbal ability change over the course of the school year. In our verbal outcome specifications, time in school is highly significant in each. Verbal ability may thus be more sensitive to the additive cognitive gains children experience as the school year progresses, while math may be more ‘sticky’ in comparison.

Our results also have implications for achievement gaps documented in the social science literature. In our data, black-white and caregiver-college/non-college math score gaps are about 4/5ths and 2/3rds of a standard deviation, respectively, which is broadly consistent with recent estimates of school readiness differences by race and income among similarly aged cohorts (Reardon and Portilla 2015). Extrapolating from effect sizes and average time uses/math scores across waves, back-of-the-envelope calculations suggest that differences in behavior may account for nontrivial proportions of the black/white and caregiver college/no college gaps in assessment. For instance, differences in sedentary behavior alone account for about 15 and 10 percent of the black/white and caregiver college/no college assessment gaps in our data. (We use sedentary behavior as an example because it is an ‘umbrella’ category that includes screen time, television, etc.) Thus for certain subpopulations, narrowing time use differences among children may aid to helping shrink gaps in assessment and achievement more generally.

In Figure 3, we show how time uses in the original CDS cohort we use in this analysis (1997–2007) to the newer 2014 module. We restricted the 2014 CDS descriptive estimates to children above 10 so they are comparable in age to the 2007 wave (when the youngest children in our sample were about ten years old). Somewhat encouragingly, we did not see any dramatic overall changes from 2007 to 2014, particularly in screen time or sedentary behavior. We actually see a small reduction in total screen time, most likely because we code these behaviors as primary rather than secondary activities. With the emergence of smartphones and tablets, many children may be consuming media while they are primarily waiting for their parents at school, being driven home, etc. In terms of the trends in behavioral and assessment gaps among adolescents from 2007–2014, some results are modestly encouraging (See online Appendix Table S1). For instance, black-white and college-no college gaps in math performance and sedentary behavior shrank from 2007 to 2014. On the other hand, gaps in screen time largely persisted or grew. Sex-based gaps also increased noticeably, as girls were less sedentary and consumed less media in 2014 compared to 2007, while boys’ behavior was stable or only trivially declined. Whether the convergence in time uses among subgroups or positive trends in behavior compared to earlier cohorts manifest in any measurable effect on test performance gaps either in the PSID or more broadly is an open question. One conclusion is rather clear: In 2014, American children are still engaging in far more passive than active leisure, whatever their background.

Figure 3. Time Use 1997–2014.

Our research suffers from various limitations which render our results provisional. For instance, we do not capture the qualitative details of what kinds of physical, sedentary, etc. behavior children are engaging in. We may see differences among our subsamples because, for example, higher-SES kids are engaging in relatively productive or cognitively active sedentary behaviors like conversing with their siblings as opposed to sitting alone watching television. Using the weather as an instrument, as we have described, also has its pitfalls and potential biases. Just as sunnier weather may allow us more time outside and the opportunity to be physically active, it varies systematically with the school year, produces a critically important secosteroid in humans, co-occurs with ground-level pollution, and so on. While we make every attempt to account for these factors conceptually and empirically, because one cannot ‘prove’ that the exclusion criteria are met, the possibility remains that there is some latent bias that drives our results. It is also important to remember that our findings consist of local average treatment effects, where only children that actually accepted the treatment—’compliers,’ who are induced to go outside in sunny weather but who otherwise would not, for example—contribute to our estimates.

It may be the case that population-level estimates (i.e. average treatment effects) are lower than what we find here, if, for example, there are declining marginal returns to going outside for ‘outdoor oriented’ children or adaptations to staying indoors for their ‘indoor oriented’ counterparts. The signal between behavior and cognitive achievement would be weakened in these subgroups, and thus perhaps not broadly relevant in terms of policy because changing behaviors would only affect children whose behavior is sensitive to environmental conditions. Finally, due to the realities of the data we use, we cannot detail the mechanisms of the causal process we are trying to capture. That is, we cannot distinguish whether, for example, the connection between physical activity and cognitive performance is more straightforwardly physiological (e.g., through concrete channels like brain structure or enhanced executive functioning) or psychological (by being beneficial in themselves or substituting for other behaviors which may offer negative psychic and cognitive value).

Conclusion

In this paper we use individual-level fixed effects coupled with a quasi-experimental research design, instrumenting active and sedentary behaviors with sunlight in trying to determine the cognitive effects of time use. We find robustly positive effects of physical activity and outdoor activity on math scores, with the opposite true for sedentary behavior and screen time. We also find that sedentary and screen time uses exhibit more substantial effects among children with less educated caregivers and girls, and speculate that time use gaps, opportunity cost structures (i.e. the range of alternative activities), and the differential sensitivity of behavior to weather conditions among subgroups may play a role in these stratified results.

In the realm of public health, small-scale RCT and other experimental studies could help uncover what mechanisms are at play in deconstructing causal pathways, while other research using observational data and quasi-experimental methodology can lend support to or undermine our findings. While it is only in its earliest stages, the new PSID CDS will be a powerful tool in the near future for researchers who seek to understand how behavior affects assessment and achievement. Other longitudinal observational studies like the National Longitudinal Study of Adolescent to Adult Health (Add Health) can also be leveraged to examine these relationships, though each dataset has its relative weaknesses (e.g. the inability for researchers to obtain geocoded data in Add Health). For practical purposes, we believe our findings lend further support to the general notion that, ceteris paribus, physical activity seems to generally be a positive factor in child development, while sedentary and screen behavior seem to evince the opposite effect. Whether this is more purely physiological in origin or a reflection of the different psychological effects certain behaviors have on development is an open question, and fertile ground for future research.

Somewhat encouragingly, we did not see an overall uptick in total screen time, television consumption, or sedentary activity when comparing adolescents in our module to the new 2014 CDS cohort. Still, children today seem to be exhibiting similarly low levels of both physical and outdoor activities as their peers did in 2007. Gaps in more passive leisure behaviors patterned on race, sex, and particularly SES as captured by mother’s education persist, but seem to have attenuated. Still, even with modest convergence, adolescents with educated caregivers in 2014 spend more than a half hour less engaging in sedentary behavior per day than their peers. Future work in the social sciences could integrate these differences in active and passive leisure into theoretical models which seek to explain why life chances so noticeably diverge among children. In this sense, playing outside or watching television can be seen as extensions of the formal-informal dichotomy many sociologists sketch out in examining patterns in achievement and childhood trajectory. Class, race, and sex structure not only the level of formalization but often whether those activities are active or passive, and these findings are provisional evidence that these behaviors are significant not only in terms of weight gain or physical health but cognitive performance as well, implicating them in broader trends in childhood inequality in achievement.