Circadian Rhythms, Sleep, and Cognitive Skills: Evidence From an Unsleeping Giant

Abstract

Introduction

Growing evidence on a downward trend in average sleep duration, along with an increased incidence of sleep deprivation, has raised concern about the potential effects on population health and health care costs (Roenneberg 2013). Insufficient sleep is associated with a reduction in daytime alertness; excessive daytime sleepiness, which impairs memory, cognitive ability, and performance (Alhola and Polo-Kantola 2007; Belenky et al. 2003; Katz et al. 2014; Killgore 2010); occupational and automobile injuries (Dinges 1995); and poor health and obesity (see Cappuccio et al. 2010 for a systematic review). Insufficient sleep is increasingly recognized as an emerging global epidemic. According to Roenneberg (2013:427–428), in many countries, people “are sleeping one to two hours less than their ancestors did 50–100 years ago.”

The pressures of a 24-7 life contribute to an increased discrepancy between individuals’ biological needs and the rhythms imposed by social schedules. The detrimental effects of sleep deprivation on health and well-being pose a significant public health challenge given that the proportion of individuals reporting insufficient sleep is rising in many advanced countries and in developing countries undergoing rapid demographic and epidemiological transitions (Hafner et al. 2016). Yet, sleep problems often remain unrecognized in these countries, particularly among older individuals. Indeed, although most sleep research has focused on developed countries, recent studies have suggested that sleep disturbances in the developing world are far higher than previously thought (Stranges et al. 2012). In this study, we analyze the effects of sleep duration on the cognitive skills and depression symptoms of older workers (aged 45 and older) in China. Recent research in neuroscience (e.g., van Praag et al. 2000) has shown that age-related cognitive decline is not a completely exogenous phenomenon but can respond to various types of behavior (see Hertzog et al. 2009 for a review), implying that the health and economic burden of the demographic change may be significantly affected by the health and economic behavior of the population. Sleep may be the most prevalent risky behavior in modern societies (Roenneberg 2013). Therefore, studying the effects of sleep deprivation on the cognitive skills and mental well-being of older workers is a question of interest to anyone studying the challenges of aging. According to the estimates of the China Sleep Research Society,¹ approximately 40 % of Chinese suffer from a sleeping disorder. Luo et al. (2013) also suggested that approximately two of five elderly people living in urban China have sleep problems and that this rate increases rapidly with aging. The rapid growth of the Chinese economy and the “electronification”² of the bedroom have contributed to the observed increase in the number of people reporting insufficient sleep (e.g., Li et al. 2007). Additionally, sleeping is often considered as an unproductive use of time, a cost that should be minimized especially in a period of rapid economic expansion, despite the common wisdom that sleeping a sufficient number of hours is important for health and performance (Pan 2003). In addition, sleep is often regarded as a manifestation of laziness in Chinese traditional culture, in which working hard and diligently is highly praised; thus, people tend to belittle their sleep problems (Pan 2003; Xiang et al. 2008). Paradoxically, the pressures of a rapidly growing economy may have unintended consequences on sleep quality and duration—and, in turn, on individual performance.

Cognitive skills are strongly related to economic outcomes (McArdle et al. 2011); financial decisions, such as portfolio choices (Banks and Oldfield 2007); and economic development (Hanushek and Woessmann 2008). As Lei et al. (2012) argued, the degradation of cognitive skills associated with aging may have important effects in countries such as China with an aging population and lacking intermediary institutions providing support to older people making decisions on income security or health care provision. Medical and psychiatric studies have also provided evidence of an important association between sleep deprivation and depression (Tsuno et al. 2005; Wirz-Justice and Van den Hoofdakker 1999), and recent studies have found high rates of depression among Chinese older adults (Lei et al. 2014).

A voluminous medical literature has studied the associations among sleep deprivation, cognitive abilities, and mental health in laboratory settings. However, the experimental evidence has focused mainly on the effects of total sleep deprivation: for example, being awake continuously for one to three days, or sleeping less than four hours for few days in a row. Only a few studies have evaluated the consequences of chronic partial sleep deprivation—that is, repeated exposure to sleep duration of less than six to seven hours per night—a condition that is far more common in reality but has received less attention because of the limited ability to collect sleep data or observe subjects over time (as discussed in the next section).

Furthermore, as Roenneberg (2013) noted, laboratory studies offer a limited understanding of both the determinants and consequences of sleep deprivation: such studies are usually based on people who have been instructed to follow certain sleep patterns (e.g., bedtime), and laboratory settings are unlikely to reflect real-world conditions (e.g., individuals are often required to sleep with electrodes fastened to their heads). On the other hand, observational studies have mostly relied on descriptive analyses of survey data that do not allow researchers to disentangle the causal effect of sleeping from the effects of other, unobservable confounding factors. Despite a growing interest on the topic, we still know very little about the causal effects of sleep on health, human capital, and productivity. Hence, evidence is lacking on the mechanisms that may underlie the relationship between sleep duration and economic performance.

Our contribution with respect to the extant literature is twofold. First, to the best of our knowledge, no other studies have used a quasi-natural experiment and survey data to identify the effects of sleep duration on cognitive skills. Second, very little evidence exists on the relationship between sleep duration and cognitive skills in developing countries.

Our study also contributes to a small but growing strand of economics literature analyzing the determinants and consequences of sleeping. Although people spend a large part of their time sleeping, and despite great variation in sleeping patterns in the population, the determinants of sleeping choices and their consequences have been largely understudied in the economics literature. Most economic models, indeed, consider sleeping as a predetermined constraint on individual’s time use or more generally as standard leisure time that can be traded off when relative returns to market activities or to other leisure activities increase. A few notable exceptions include studies analyzing how sleeping duration responds to market incentives (Biddle and Hamermesh 1990); the effect of daylight saving time on car crashes, work accidents, health, and financial markets (Barnes and Wagner 2009; Jin and Ziebarth 2015; Kamstra et al. 2000; Monk 2012; Smith 2016; Sood and Ghosh 2007); and differences in productivity between different chronotypes (morning vs. evening types; Bonke 2012). Our study is closely related to two recent studies analyzing the effects of sleep deprivation on productivity (Gibson and Shrader 2014) and health (Giuntella and Mazzonna 2017). Our study speaks to the previous literature analyzing the determinants of cognitive skills and mental health (Banks and Dinges 2011; Mazzonna and Peracchi 2012; Schneeweis et al. 2014), and in particular to those studies focusing on older workers and on aging and cognitive skills in China (Huang and Zhou 2013; Lei et al. 2012, 2014).

To identify the effects of sleep duration, we adopt an instrumental variable (IV) strategy following Gibson and Shrader (2014), who examined the effects of sleep on productivity in the United States by exploiting variation in sunset time and its effects on bedtime within U.S. time zones. Our IV strategy exploits the same relationship between circadian rhythms and bedtime in the particular context of China, a country roughly as broad as the continental United States but using a single time zone. Due to circadian rhythms, the human body reacts to environmental light, producing more melatonin when an environment becomes darker. Thus, when sunset occurs at a later hour, individuals tend to go to bed later (Roenneberg and Merrow 2007). Although, in principle, individuals could compensate for a later bedtime by waking up later in the morning, social schedules (e.g., work schedules, school start times) are less responsive to solar cues (Hamermesh et al. 2008). Thus, a later sunset can have important effects on sleep duration.

Because of the single Chinese standard time zone, large sunrise and sunset time disparities exist between cities, far larger than those exploited by Gibson and Shrader (2014) in their U.S. study. For instance, on June 21, in the city of Urumqui (northwest China), the sun rises at 6:27 a.m. and sets at 9:56 p.m., with the solar noon occurring at 2:11 p.m. Comparatively, in Harbin (northeast China), the sun rises at 3:43 a.m. and sets at 7:27 p.m., with the solar noon occurring at 11:35 a.m. (see Fig. 1).

Our identification strategy exploits this variation in sunset time to identify the effects of sleep duration on cognitive outcomes. Using data from the Chinese Health and Retirement Longitudinal Study (CHARLS), we focus on the effects of sleeping on cognitive skills of a relatively old population (aged 45 and older). We show that individuals living in cities where the sun sets at a later time, ceteris paribus, sleep less than individuals living in early sunset cities. Using sunset time as an instrument for sleep duration, we show that increasing sleep duration significantly increases the scores reported in cognitive tests measuring mental and numerical skills. A 30-minute change in sleep duration produces effects on cognitive outcomes ranging from 0.2 to 0.3 standard deviations. Results, using dummy variables for particular sleep ranges as dependent variables, are qualitatively similar. These effects are consistent with the growing evidence from neurobiological studies that even relatively moderate sleep restriction can severely affect waking neurobehavioral functions and that repeated exposure to partial sleep deprivation (“sleep debt”) may increase the severity of age-related chronic disorders (Spiegel et al. 1999; Van Dongen et al. 2003).

To repeat, the effects are driven by the employed population living in urban areas. These results are consistent with the premise that in rural areas, schedules tend to adjust to the daily solar cycle. In urban contexts, however, employed individuals are constrained in the morning by work schedules and cannot fully compensate for a later bedtime. These estimates represent local average treatment effects in which compliers are those workers who face binding morning schedule constraints and engage in less evening adaptation to reduce the effects of environmental light.

A natural concern is that the geographical distribution of economic activity may confound our effect of interest. To address this issue, we use only within-region variation and control for local geographic and economic conditions. Furthermore, we provide several indirect tests of the exclusion restriction assumption, and it is reassuring that the relationship between sunset time and sleeping time is driven by individuals whose time use is constrained by work schedules. In particular, we find that the results are statistically significant and economically important among the employed living in urban areas; effects are smaller and nonsignificant among the unemployed, the self-employed, and farmers, whose schedules are more flexible or more likely to respond to the solar cycle. Finally, we find nonsignificant effects on health outcomes that are usually correlated with economic activity but should not be affected by sleeping (or sunset) time, such as vision, hearing, and speech problems; lung diseases; individual’s knee height; and individual’s health as a child (Huang et al. 2013).

Overall, our results suggest that sleeping has important effects on cognitive performance and depression symptoms. The heterogeneity of our results by occupational and demographic characteristics is consistent with recent findings from other countries (Giuntella and Mazzonna 2017) and suggests that sleeping duration is importantly affected by social constructs, such as work schedules, that create a misalignment between social and biological time (Wittmann et al. 2006). Our findings also suggest that although coordination has clear economic advantages (Hamermesh et al. 2008), its costs should not be neglected.

Sleep Deprivation, Cognitive Skills, and Depression Symptoms

The effects of sleep deprivation have been studied extensively in laboratory settings. Most experiments have focused on acute sleep deprivation, keeping participants awake continuously for generally one to three days and testing their cognitive performance before, during, and after deprivation. Alhola and Polo-Kantola (2007) and Killgore (2010) reviewed the effects of acute sleep deprivation on a wide range of cognitive processes, including basic cognitive functions such as attention, multiple aspects of sensory perception, emotional processing, learning, memory, and decision-making. Despite the different methodologies, most studies have suggested that sleep deprivation induces cognitive impairment and that individual characteristics affect the tolerance of sleep deprivation.

Chronic partial sleep deprivation is more common in reality because of several factors, including sleep disorders, work demands, and social and domestic duties; however, it has received much less attention given the difficulty of analyzing individual sleep duration over a long enough period. Earlier studies yielded mixed findings as a result of defectively designed experiments; but recent studies, with improved experimental control groups, have consistently found that chronic sleep deprivation adversely affected cognitive performance—in particular, behavioral alertness (e.g., Banks and Dinges 2011). For instance, the most extensive, controlled dose-response experiment on chronic sleep restriction was based on 14 days of sleep limitation to no more than four, six, or eight hours (Van Dongen et al. 2003). Those with 4 or 6 hours of sleep reported negative effects comparable with those individuals forced to stay awake for 24 or 48 hours. Moreover, the experiment showed evidence of cumulative dose-response effects on neurobehavioral functions. In other words, additional days of sleep restriction resulted in additional neurobehavioral impairments. Unfortunately, the external validity of these experimental studies is rather limited. In the real world, people are often exposed to partial sleep deprivation for a long time (months or years)—a chronic condition that cannot be reproduced and examined in a laboratory setting.

On the other hand, several observational studies have analyzed the relationship between sleep duration and cognitive skills. However, survey-based studies cannot establish any causal relationship (see Yaffe et al. 2014 for a comprehensive review). Overall, these survey-based studies have yielded mixed results (Benito-León et al. 2009; Hahn et al. 2014; McEwen 2006).

Several studies have shown that sleep deprivation is also associated with negative mood states, anxiety, and depression (Sagaspe et al. 2006; Tsuno et al. 2005). However, as Goel et al. (2009) noted, experimental evidence on the negative effects of sleep deprivation on depression and anxiety is limited.

In the economics literature, a few studies have attempted to determine the causal effect of school start times on academic performance. Carrell et al. (2011) identified the causal effect of school start times on academic achievement by using two policy changes in the daily schedule at the U.S. Air Force Academy along with the randomized placement of freshman students to courses and instructors. They found that starting the school day 30 minutes later had a significant and economically important positive effect on student achievement. However, they could not observe sleep. Furthermore, this study did not provide any evidence regarding the effect of chronic sleep deprivation and did not consider the cumulative and long-run effects of partial sleep deprivation.

An important contribution of our study is the focus on a developing economy. Most studies analyzing the effect of sleep deprivation have focused on individuals living in advanced economies. In particular, because of the ongoing aging of the population, there is increased attention on the sleep deprivation and sleep quality of older adults in low- and middle-income countries. Indeed, evidence from the medical literature suggests a greater occurrence of sleep disorders among older adults because of deterioration of the suprachiasmatic nucleus region of the brain (Van Someren 2000). The deterioration of sleep quality among the elderly may contribute to cognitive decline. Furthermore, given that family members (spouses and adult children) are the predominant caregivers in China, the degradation of cognitive skills of the elderly may impose an additional burden on prime-age workers (United Nations Economic and Social Commission for Asia and the Pacific 2015).

Yet, very little evidence exists on the causal effects of sleep deprivation on cognitive skills, and sleep deprivation has been largely understudied in developing countries. Gildner et al. (2014) and Stranges et al. (2012) are two notable exceptions, analyzing the association between sleep deprivation and cognitive outcomes in low- and middle-income countries. In addition, Luo et al. (2013) analyzed 1,086 community residents aged 60 and older who completed the Chinese version of the Pittsburgh Sleep Quality Index (CPSQI), finding that poor sleep quality is highly prevalent among elderly Chinese residents in urban Shanghai. However, like most other studies on the topic, they did not attempt to recover the causal effect of sleep deprivation.

Identification Strategy and Empirical Specification

Background: Time Zones in China

China is the second largest country in the world by land area and is the most populous. Its territory spans more than 60 degrees of longitude. As described in Fig. 2, under a standard time zone scheme, Chinese territory would cover an area corresponding to five time zones ranging from UTC+5 to UTC+9. In 1912, the year after the collapse of the Qing Dynasty, the newly empowered Republic of China established five time zones in the country, ranging from 5.5 to 8.5 hours past Greenwich Mean Time (GMT). In 1949, however, Mao Zedong decreed that all of China would henceforth be on “Beijing time.” As the Communist Party consolidated control of the country, the one time zone was meant to foster national unity (Gilley 2004). Ever since, all of China has shared a single official time zone: the official national standard time is eight hours ahead of Greenwich Mean Time (UTC+8), referred to as Beijing time domestically and as China standard time internationally. This common national time produces some geographic distortions. For example, when a daily flag raising ceremony takes place in Tiananmen Square at sunrise in Beijing, easternmost China has already experienced an hour of daylight, and the sun will not rise for another three hours in the westernmost part of China.

Mainland China is administered by 31 provincial-level administrative divisions (22 provinces, 5 autonomous regions, and 4 provincial-level municipalities). Figure 2 shows that most provinces are located in UTC+7 and UTC+8, where 95 % of the Chinese population lives. Thus, for most people in China, the single time zone simply requires slight or moderate adjustment. The provincial-level administrative divisions located within UTC+8 longitudes include Fujian, Jiangxi, Zhejiang, Anhui, Shanghai, Jiangsu, Shandong, Hebei, Beijing, Tianjin, and Liaoning; at least half of the areas of Hubei, Henan, Shanxi, Guangdong, Jilin, Inner Mongolia, and Heilongjiang; and about one-third of Hunan. As reported in Table 1, with an estimated population of 0.77 billion people (56 %), the UTC+8 zone produced 65 % of China’s GDP in 2014.³ Almost all the remaining population and economic activity is located in UTC+7.

Not all Chinese citizens observe official Beijing time: for example, a number of minority Muslim Uighur citizens of Xinjiang Uyghur Autonomous Region province. Indeed, Xinjiang, located in the westernmost part of the country, operates on a two-time zone system: the official Beijing time and the unofficial Xinjiang time (UTC+6). Although schools, government offices, public service offices, airports, and railway stations all adopt Beijing time, some bus lines and local shops use local Xinjiang time. For those following Beijing time, they implement a modified time schedule. For instance, schools start at 10 a.m. Beijing time, which equals to 8 a.m. Xinjiang time. For this reason, we excluded Xinjiang and all individuals living in UTC+5 and UTC+6 from our sample, instead focusing on individuals living in UTC+7, UTC+8, and UTC+9. In these three areas, work schedules and economic activity are coordinated. Offices, shops, public transportation, and schools follow Beijing time.

Identification Strategy

The goal of this study is to recover the causal effects of sleep duration on cognition and depression. A simple comparison of people with different sleeping behavior would not allow us to identify a causal relationship. A natural concern is that there may be omitted variables related to both sleep duration and our outcomes of interest. Moreover, depression and cognitive impairment can affect sleep duration, generating a reverse causality problem. For these reasons, we rely on an IV strategy to identify the effects of sleep on cognitive skills and depression symptoms using information on sunset time. More specifically, our identification strategy exploits the geographical variation in sunset time across Chinese cities as sources of exogenous variation in sleep duration.

Our strategy is the same as the one used by Gibson and Shrader (2014) to analyze the effects of sleep duration on productivity in the United States, and it is closely related to the approach used by Giuntella and Mazzonna (2017), who used discontinuities in sunset at a time zone border to recover the causal effects of sleep deprivation on health. The main idea underlying these strategies is that circadian rhythms are important determinants of human sleeping patterns. Our internal pacemaker, the brain’s suprachiasmatic nucleus (SCN)—also thought of as the body’s master clock—regulates the body’s biological rhythms by changing the concentrations of the molecular components of the clock to levels consistent with the appropriate stage in the 24-hour cycle. This process is known as entrainment. In practice, when the sun sets and it becomes darker, the SCN produces more melatonin, facilitating sleep (Aschoff et al. 1971; Duffy and Wright 2005; Roenneberg and Merrow 2007; Roenneberg et al. 2007). Within a time zone, people organize their lives according to common social time, yet differences in sunrise and sunset time can be very large as dawn and dusk progress from east to west (Gibson and Shrader 2014; Roenneberg and Merrow 2007). Previous studies have shown that wake-up time is less affected by solar cues than bedtime. Instead, wake-up times are significantly affected by work schedules and other social constraints (such as children’s school start times), which in turn respond to social conventions, economic incentives, and regional coordination (Giuntella and Mazzonna 2017; Hamermesh et al. 2008; Roenneberg 2013). Thus, sunset time can have important effects on sleep duration.

Figure 1 illustrates the variation in sunset time across Chinese cities on summer solstice, June 21. Because China follows a unique time zone, the differences between the easternmost and the westernmost regions of China are marked in the figure.⁴ If people could/would compensate by waking up later, the differences would have no effect on sleep duration. However, because schedules tend to be less responsive to solar cues, many individuals are not able to fully compensate in the morning by waking up at a later time. In particular, we expect sunset time to have larger effects on sleep duration among employed people in urban areas because they are more likely to be constrained by standard office hours. Unfortunately, CHARLS data do not allow us to observe individuals’ work schedules. However, consistent with our hypothesis, Giuntella and Mazzonna (2017) used a similar strategy with U.S. time use data to show that later sunset times have significantly larger effects on individuals who start working early in the morning or have school-aged children.

On the other hand, we expect that the sleeping behavior of people who are not employed or live in rural areas would be less affected by these differences in sunset time because their daily activities are more likely to respond to the daily solar cycle. In the upcoming Main Results section, we provide evidence consistent with this hypothesis.

The economic activity in China is clustered on the eastern side of the country. Thus, one may confound the positive effect of an early sunset on the eastern side of the country with the economic development of these areas. For this reason, we control for regional fixed effects in all our specifications and control for city-level GDP and population. We consider four regions in accordance with the definition used in the Twelfth Five-Year Plan for National Economic and Social Development of China (see Fig. 3).⁵ The eastern region, the most developed area, consists of 10 provincial-level administrative divisions, 8 of which are entirely within UTC+8. The northeastern region covers three provinces located either in UTC+8 or UTC+9. The central region includes four provinces stretching over UTC+7 and UTC+8 as well as two provinces located in UTC+8 only. The remaining 12 provincial-level administrative divisions constitute the western region, a vast territory characterized by lower socioeconomic status and higher poverty rates. The western region ranges from UTC+5 to UTC+7, but only Xinjiang, Tibet, and part of Gansu and Qinghai provinces are located in UTC+5 and UTC+6, all of which have a low population density.⁶ Therefore, focusing only on individuals living in UTC+7, UTC+8, and UTC+9 excludes only a tiny fraction of the population. In addition, we include controls for other factors that may affect exposure to the sunlight as indicators for the level of air pollution and the type of landscape (e.g., mountain, plateau).

In practice, we exploit cross-sectional variation in average sunset time within regions that are considered to be homogeneous with respect to socioeconomic characteristics. Tables S1 and S2 (Online Resource 1) illustrate the within-region variation in sunset time for our baseline sample. The average within-region standard deviation in sunset time is approximately 30 minutes.

Our identification relies on the assumption that conditional on our set of regional controls, sunset time is orthogonal to other characteristics that can affect our outcomes of interest: conditional exogeneity. As with any identification assumption, ours is directly untestable. We acknowledge that other unobserved confounding factors may be correlated with average sunset time and our main outcomes of interest. Yet, we think that this concern is substantially mitigated by the inclusion of regional fixed effects and city-level economic controls. Furthermore, we provide several indirect tests of the exclusion restriction. In particular, the heterogeneity of our findings by employment status and urban area appears to be consistent with our identification assumption. As previously noted, we expect and show that the largest effect of sunset time on sleeping behavior is observed among employed individuals living in urban areas. We find no evidence of significant effects of sunset on populations that have more flexible individual schedules (the self-employed) or are more likely to adapt to the daily solar cycle (rural residents). For these groups, we find no evidence of significant reduced-form effects on cognitive abilities and depression symptoms. Furthermore, we show that individuals living in early and late sunset cities are well-balanced on most covariates, and we implement some robustness checks on predetermined characteristics and health outcomes not associated with sleep that should not be affected by our instrument but that are likely to be correlated with the level of economic development of the city.

Note that the cognitive differences that arise from contrasting cities with different sunset times are likely to be the result of a long-term exposure to sleep deprivation and are not the result of the effect of differences in sleeping behavior in the last month as measured in the data used in this study (see the upcoming Data section). In other words, what we measure with our two-stage least squares (2SLS) strategy is not the effect of a one-hour difference in average sleeping in the previous month, but rather the average effect of a long-term exposure to a one-hour difference in sleeping.

Another potential concern is that our identification strategy could reflect differences in sunlight exposure, thus, violating the exclusion restriction assumption. In other words, if the differences in sunset time are correlated with sunlight exposure, we might confound the effect of sleep duration with that of the sunlight. For instance, sunlight exposure increases the production of vitamin D, which is usually associated with mood and depression (e.g., Kjærgaard et al. 2012). Moreover, previous research on the effect of daylight saving time has shown that additional light in the evening increases the level of physical activity (Wolff and Makino 2012).⁷ If we consider that the average sleep duration in our sample is 6.3 hours, it is very likely that most Chinese people are usually awake (and then potentially exposed to sunlight) during all sunshine hours, regardless of their location. Nevertheless, any difference in sunlight exposure found across Chinese cities should advantage individuals living in cities with a late sunset. Yet, we show that these individuals tend to sleep less. Thus, the difference in daily exposure to sunset light may introduce a downward bias in the estimated effect of sunset time on sleep duration and cognitive abilities.

Residential sorting and commuting may be other important confounding factors. The household registration system (Hukou), implemented in China since 1949, substantially mitigates the concern of residential sorting. Under this system, every citizen must register a Hukou location and type (i.e., a single permanent place of residence and its classification as urban or rural), which are passed on by one’s parent(s) at birth regardless where one physically resides (Chan and Zhang 1999). Access to most public services and welfare benefits—such as children’s education, housing, social security programs—are attached to one’s Hukou. In our data, less than 1% of respondents live in a city different from their Hukou; more generally, less than 3 % live in a province different from their birthplace. Finally, a potential concern is that if commuting tends to go in a certain direction (e.g., west to east), commuting times may be correlated with sunsets as well as with sleep duration. Unfortunately, the CHARLS data do not contain information on individual commuting time. However, in 2015, Baidu (the country’s largest Internet search company) used smartphone data to estimate average commuting time and distance in China. According to Baidu’s analysis, the national average commuting distance in 2015 was 9.18km, with an average traveling time of 28 minutes. The individuals commuting the longest were traveling from Tongzhou to Beijing, approximately 50km (Chen 2015). Although commuting may substantially affect sleeping time, these distances are unlikely to be reflected in significant differences in sunset time and thus are unlikely to be correlated with our instrument.

Empirical Specification

Another relevant concern is data quality, given that we use self-reported measures of sleep (see also the Data section). Because self-reports are subject to biases and measurement error, self-reported sleep duration might poorly represent the real sleeping habits (Lauderdale et al. 2008) and may introduce large measurement error in the analysis. However, if measurement error is uncorrelated with our instrument (sunset time), our IV strategy allows us to obtain consistent estimates of the parameter of interest. Furthermore, in the data used in our analysis, the question on sleep duration explicitly asked the respondents to report only the time spent sleeping, excluding time spent in bed awake (e.g., watching a movie).

Data

The Chinese Health and Retirement Longitudinal Study (CHARLS) is a nationally representative longitudinal survey of Chinese residents aged 45 and older and their spouses. CHARLES was harmonized with the Health and Retirement Study (HRS). This was meant to guarantee high-quality standards and the comparability of results also with the other related ageing surveys such as the English Longitudinal Study of Ageing (ELSA) and the Survey of Health, Ageing and Retirement in Europe (SHARE).

The first national wave of CHARLS was fielded in 2011 and includes approximately 10,000 households and 17,500 individuals in 126 cities, 150 counties/districts, and 450 villages/resident committees. The data includes 28 provinces. The second wave was fielded in 2013.⁸ The survey includes information on demographics, family structure/transfer, health status and functioning, biomarkers, health care and insurance, work, retirement and pension, income and consumption, assets (individual and household), and community-level information.⁹ The interview is conducted using a face-to-face, computer-assisted personal interview (CAPI).

CHARLS contains information on average sleep duration in the last month. Respondents are asked to report how many hours they slept per night, on average, over the month preceding the interview. Using this question, we constructed both a linear measure of sleep duration in hours and indicators for whether individuals slept at least seven hours, at most six hours, at least eight hours, or between seven and nine hours.¹⁰ We computed yearly average sunset time for Chinese cities using the National Ocean and Atmospheric Administration (NOAA) solar calculator¹¹ and the information on the city of residence available in the survey.

Because the sleeping variable in CHARLS measures the self-reported average sleep duration in the last month, in principle, we could also use the average sunset time in the last month preceding the interview as an instrument to exploit the seasonal variation in sunset time. Unfortunately, 87 % of the interviews took place in July and August, so we do not have sufficient seasonal variation to exploit. Still, in the robustness checks, we also exploit this small variation across interview dates to show that short-term effects are somewhat negligible in our setting. Moreover, because our aim is to evaluate the long-term effects of chronic sleep deprivation on cognitive abilities, we believe that the average sunset time should better capture these long-term effects than short-term sleep restriction caused by seasonal variation in sunset time.

The CHARLS data contain several questions measuring cognition and depression. In particular, we measure the individual-level cognitive abilities using the results from all cognitive tests available in the survey: Mental, Numerical, Memorial, and Draw. Interviewers administer these tests to all respondents (excluding only proxy respondents) and follow a protocol aimed at minimizing the potential influences of the interviewer and the interview process. The questions were adapted from the HRS and are comparable with similar cognitive tests implemented in ELSA and SHARE. The HSR research team developed the tests based on the psychology literature on intelligence and cognition as well as the geriatric and neurological literature on cognitive impairment and dementia (Ofstedal et al. 2005). The initial evaluation suggested that the response rates, psychometric properties, and construct validity were reasonable (Crimmins et al. 2011; Wallace and Herzog 1995). These cognitive tests strongly predict important economic outcomes (McArdle et al. 2011) and financial decisions such as portfolio choices (Banks and Oldfield 2007). These tests have been widely used in the literature, especially to analyze the effects of retirement on cognitive abilities at older ages (Bonsang et al. 2012; Mazzonna and Peracchi 2012).

The Mental test measures awareness of the date, the day of the week, and the season of the interview to capture the respondent’s mental intactness (McArdle et al. 2011). The score is based on the number of correct answers, ranging from 0 to 5.

The Numerical test measures the individual’s ability to compute simple mathematical subtractions (e.g., “successively subtract 7 from 100, five times”). The score is based on the number of correct answers, ranging from 0 to 5.

A well-known indicator used in the literature analyzing cognitive skills is the Telephone Interview of Cognitive Status (TCIS), which is the sum of the scores from the Mental and Numerical tests. The TICS test can be administered in person or by telephone—in CHARLS, the interview is conducted using a face-to-face CAPI—and is highly correlated with the Mini-Mental State Exam (MMSE) (Folstein et al. 1975), a screening tool frequently used by health care providers to assess overall brain function. Because TICS is simply obtained by the sum of the scores from the Mental and Numerical tests, we do not include it in the analysis. However, results for TICS are available upon request.

The Memorial or Word Recall test measures episodic memory (McArdle et al. 2011). Interviewers read a list of 10 nouns to respondents and asked them to recall the words immediately and then 10 minutes later. The score is based on the average number of correct answers during two recall sessions, ranging from 0 to 10.

Following Huang et al. (2013), we also measure the individual’s depression status by using a Chinese version of the CES-D 10 questionnaire and constructing an indicator ranging from 0 to 30 based on the respondent’s answer. The higher the CES-D score, the more severe the depression.

As mentioned earlier, we restrict our analysis to individuals living in UTC+7, UTC+8, and UTC+9. Although Xinjiang, Tibet, and part of Gansu and Qinghai provinces are located in UTC+5 and UTC+6, CHARLS was fielded in only one city (Aksu from Xinjiang) within this area, where a two-time zone system—the official Beijing time and the unofficial Xinjiang time—operates.¹² We exclude individuals older than 70 mainly to avoid selection issues, given that the average life expectancy in China is 75 years for men and 787 years for women (WHO n.d. 2013).¹³ Furthermore, because we focus on older workers in our main analysis, the share of individuals older than 70 reporting to work is less than 2 % and thus represents a very selected sample of the population.¹⁴

Table 2 reports summary statistics for all the variables used in the analysis. Column 1 presents the summary statistics for the entire sample. In column 2, we focus on the employed population living in the urban areas, who (as explained earlier) are more likely to be affected by social constraints (e.g., work schedules) and thus less able to compensate a later bedtime with a later wake-up time (see also Giuntella and Mazzonna 2017).

Note that the average sleep duration reported in our sample is less than 6.5 hours. Thus, the average sleep duration in China is far below the sufficient number of hours of sleep usually recommended in the literature (between seven and nine hours of sleep; see Cappuccio et al. 2010). Therefore, according to the results of the experimental literature on partial sleep deprivation discussed earlier, we might expect negative health effects from such sleeping habits.

In Table S4 (Online Resource 1), we report the summary statistics for the employed and urban sample. We perform a balancing test comparing late sunset areas with early sunset areas. Controlling for regional fixed effects, GDP, and population, individuals living in late sunset areas tend to sleep significantly less and have poorer cognitive outcomes than those in early sunset areas. We also find evidence of a marginal difference in age: individuals living in early sunset areas tend to be slightly older than those living in late sunset areas. The balance failure on age would actually bias this research design against finding positive effects of sleep on cognition. Indeed, late sunset, low-sleep workers are slightly younger and presumably have slightly stronger cognitive ability. However, most of the other individual covariates are balanced across the two samples, and we control for age in all our estimates.

Main Results

First-Stage Estimates: Sunset and Sleep

Panel A of Table 3 illustrates the relationship between sunset time and sleep duration. We report the coefficients of all the covariates in Online Resource 1, Table S5. Consistent with our conjecture, these effects are driven by employed people (column 1), nonsignificant among the non-employed (column 2), and larger among the employed living in urban areas than among individuals in nonfarming occupations in rural areas.¹⁵ Thus, our estimates represent local average treatment effects. Compliers are those workers who face binding morning schedule constraints and adapt less in the evening (e.g., adjusting exposure to white and blue light) to reduce entrainment.

Among employed individuals living in urban areas, a half-hour increase in average sunset time (a 1.16 standard deviation increase) reduces sleep duration by 16.5 minutes (0.19 of a standard deviation). Hence, sunset time is a strong and significant predictor of sleep duration, particularly if we consider that the average sleep duration is roughly 6.5 hours. This finding is also relevant from a medical perspective because the experimental evidence on sleep deprivation finds the presence of sizable negative health effects when people are constrained to sleep less than seven hours per night.

Indeed, in our upcoming section on robustness checks, we show no significant effect among the farmers in the CHARLS sample. Our findings confirm our conjecture that farmers are less affected by time zone settings and more likely to follow the natural light in their daily activities, consistent with research showing that farmers in China have longer sleep duration and better sleep quality than Chinese employed as blue-collar workers or civil servants (Sun et al. 2015). We also find no evidence of significant effects among the self-employed (see the section on robustness checks), who are likely to have more flexible work schedules than employed individuals.

As an alternative measure of sleep deprivation, we use an indicator for whether respondents reported to have slept at least seven hours, a commonly used metric in sleep research (Cappuccio et al. 2010). The effects, shown in panel B of Table 3, are qualitatively similar to those reported in panel A, and the point estimates are even more precisely estimated. A half-hour increase in sunset time decreases the likelihood of sleeping at least seven hours by 6.5 percentage points among the employed and by 11.5 percentage points among employed individuals living in rural areas. These results suggest that given the relatively low average sleep duration, the 15-minute difference implied by a half-hour increase in sunset time has large effects on the likelihood of sleeping less than the recommended amount of time.

Reduced-Form Analysis

Table 4 analyzes the reduced-form relationship between sunset time and a battery of cognitive outcomes and a depression test (CES-D). Each set of rows reports the estimated reduced-form effect on a specific cognitive outcome (along with the number of observations and the R²), starting from the Mental skills test. The last row reports the estimated effect on the depression test (CES-D). It is worth noting that given the small sample size, not all the coefficients are precisely estimated.

With the exception of the effect on the Memorial test for a few subgroups, our point estimates show that a later sunset time is always negatively associated with the average test score. Consistent with our first-stage results, point estimates are larger (in both relative and absolute terms) for employed respondents in urban areas. For this subgroup, a half-hour increase in the average sunset time is significantly associated with a 0.135-point reduction in mental abilities (3.5 % reduction with respect to the mean), a 0.26-point reduction in numerical skills (7.5 % reduction with respect to the mean), and a 0.5-point reduction in the average TICS score (7.5 % reduction with respect to the mean). The estimated effects on the memorial score are rather small and not significantly different from 0. Consistent with our first-stage estimates, effects are largest among employed respondents living in urban areas.

Finally, the last set of rows of Table 4 show that a later sunset is associated with a higher depression score among the employed population in urban areas. A 30-minute increase in average sunset time increases the depression score by 11.5 % with respect to the mean observed in the sample. The heterogeneity of findings across sociodemographic groups provides an indirect test for the exclusion restriction supporting the causal interpretation of our results. Results are robust to the inclusion of four-digit occupation fixed effects (see Table S6, Online Resource 1).

2SLS Estimates

Having shown the effect of the average sunset time on sleep duration and its reduced-form effect on cognition and depression, we show the 2SLS estimates to provide an estimate of the average effect of one hour of sleep on cognitive abilities and depression symptoms. However, as discussed throughout this article, these estimates must be interpreted with caution because they represent the effect of both short- and long-term effect of sleep deprivation. In particular, although the reduced-form effects capture the effect of long-run exposure to sleep deprivation, the first-stage analysis is based on the short-run effect of average sunset time on sleep duration in the month before the interview.

Panel A of Table 5 reports 2SLS estimates for the effects of sleep duration on cognitive skills and depression symptoms in our main sample. Focusing on employed individuals in urban areas mitigates the concerns regarding weak instrument problems. Because of some item nonresponse on cognitive questions, the value of the F test on the excluded instruments slightly decreases with respect to the value shown in Table 3 and is below the value of 10, the rule-of-thumb indicator for weak instrument problems. However, this should not represent a concern for two main reasons. First, all values between 8.09 and 13.77 still correspond to a maximum 2SLS size distortion of no more than 10 % (see the critical value for weak instrument test in Stock and Yogo 2005). Second, our estimates pass the Anderson-Rubin test, indicating that our measure is robust to a weak instrument (Moreira 2003).

Our 2SLS estimates show that an increase in average sleep duration increases cognitive abilities (column 2) and lowers the depression symptoms score (column 4). In particular, a 15-minute increase in average sleep duration—the sleep variation generated by a 30-minute increase in sunset time—would increase numerical skills by 7.5 % and mental scores by 3 % (although not precisely estimated) with respect to the mean of each dependent variable.¹⁶ To gauge a sense of the magnitude, Huang and Zhou (2013) estimated that concluding primary education increases cognition scores by 20 %.

Finally, we find that a 15-minute increase in average sleep duration would reduce the CES-D score by 10.5 % with respect to the mean. This is an economically important effect. Again, to gauge a sense of the magnitude, Lei et al. (2014) showed that having completed junior school is associated with a 20 % lower depressive symptom score.

Panel B reports 2SLS estimates using the indicator for having slept at least seven hours. Effects go in the same direction of those presented in panel A. As mentioned earlier, the first-stage results are slightly stronger when using this metric, and thus point estimates are more precisely estimated. Having slept at least seven hours increases the mental skills score by 30 % and numerical skills score by 65 % with respect to the mean (columns 1 and 2). We also find large significant effects on the depression symptoms score.

Robustness Checks

We report the results of a set of tests that we implement to verify the robustness of our estimates and the validity of our identification strategy. We indirectly test our identification assumption by verifying whether the average sunset is significantly associated with other predetermined characteristics that should not be affected by a different sunset time. In Table 6, we test the association of the average sunset time with predetermined characteristics (knee height and child health) and health problems that, to the best of our knowledge, have not been associated with sleep duration (vision, hearing and speech impediments, and respiratory diseases). For instance, lower leg length is not correlated with height shrinkage, which may be affected by adult lifestyle behaviors and is more likely to reflect predetermined anthropometric characteristics (Maurer 2010; Roubenoff and Wilson 1993). We adopt the same specification that we used to analyze the effect of the average sunset time on sleep duration. The results clearly show no evidence of any significant effect of the average sunset time on these variables. This is reassuring because a significant correlation with predetermined characteristics or health metrics that are unlikely to be affected by sleep duration would cast doubts on our identification strategy.

As mentioned earlier, we also test whether sleeping behavior, cognitive skills, and depression symptoms of farmers and the self-employed are affected by variation in sunset time. Panel A of Table 7 shows that the effects are nonsignificant and smaller in magnitude among farmers (despite the large sample size). Similarly, we do not find evidence of significant effects when we focus on the self-employed in urban areas (panel B). Given the small sample size and the rather large standard errors, we cannot conclude that this group is not affected at all. However, point estimates are far smaller than those found among employed people, suggesting that the self-employed have more flexible schedules that allow them to, at least partially, compensate for the effect of the large differences in sunset time across Chinese cities.

Our results are not explained by within-region variation in temperatures (see Table S7, Online Resource 1), nor are they driven by one particular province. When we replicate all our estimates by excluding one province at a time, the results hold and remain consistent with those reported in the main text (available upon request). Finally, using the retrospective questions of the survey, we are able to recover respondents’ migration histories. Our findings are substantially unchanged if we exclude individuals who were born in a different province or individuals who reported working in a different province. This is not surprising given that we focus on individuals older than 45. Internal migration was long restricted by the Hukou system, which began to loosen only in the mid-1980s in response to the demands of both the market and rural residents wishing to seek greater economic opportunity in cities (Au and Henderson 2006).

We also analyze alternative nonlinear measures of sleep, such as sleeping no more than six hours and sleeping at least eight hours. These metrics have often been used in the medical literature analyzing sleep deprivation (Cappuccio et al. 2010). The results are fully consistent with those reported in Table 3. Regardless of the metric considered, a later sunset time has a negative effect on sleeping, particularly for employed people living in urban areas (see Table S8, Online Resource 1).

Finally, we exploit the longitudinal nature of the data to show that our results are driven by long-term effects of sleep restrictions. We restrict the sample to the respondents interviewed in both waves and use a fixed-effects model; the sample size is clearly reduced but it is sufficient for a robustness check. Because by construction, none of the longitudinal respondents changed their residence across the two waves, we use only the variation in sunset time due to changes in the interview date across waves. The results are reported in Table S9, Online Resource 1. Despite the small variation in the month of the interview, the seasonal (short-term) variation in sunset time significantly affects sleeping time. However, such short-term variation in sleeping induced by changes in the interview month has no significant effects on cognitive skills and depression symptoms.

Conclusions

Economists have largely ignored the effects of sleep on health and human capital. The medical literature provides extensive evidence for the association between sleep deprivation and health. However, most of these studies have not attempted to analyze the causal relationship between sleep duration and health outcomes. Furthermore, relatively little is known about the effects of sleep duration on cognitive skills in developing countries. However, in many low- and middle-income countries, sleep deprivation is increasingly recognized as a public health challenge, particularly because of its effects on the elderly population.

In this article, we analyze the effects of sleep duration on the cognitive skills of older workers in China. We show that sleep duration has nontrivial effects on cognitive and mental skills. To identify the effects of sleep on cognitive skills, we use an instrumental variable that exploits the relationship among sunset time, circadian rhythms, and average sleep duration. Our findings indicate that an increase in average sleep duration significantly increases cognitive skills and reduces depression symptoms scores at older ages. We find no evidence of significant effects on memorial skills.¹⁷ The effects are larger among employed individuals in urban areas and are robust to the use of nonlinear metrics of sleep duration. Importantly, a later sunset does not have significant effects on outcomes that are not affected by sleep duration.

Acknowledgments

We are thankful to James Fenske, Daniele Paserman, and Guglielmo Weber for their comments and suggestions. We also thank seminar attendees at the University of Oxford and at the Conference on Health, Demography and Ageing in China, Stanford Center for International Development, October 2015.

Notes

References