The Economic Consequences of Increasing Sleep Among the Urban Poor

Pedro Bessone; Gautam Rao; Frank Schilbach; Heather Schofield; Mattie Toma

doi:10.1093/qje/qjab013

. 2021 Apr 8;136(3):1887–1941. doi: 10.1093/qje/qjab013

The Economic Consequences of Increasing Sleep Among the Urban Poor^{^*}

Pedro Bessone ¹, Gautam Rao ², Frank Schilbach ³, Heather Schofield ⁴, Mattie Toma ⁵

PMCID: PMC8242594 PMID: 34220361

Abstract

The urban poor in developing countries face challenging living environments, which may interfere with good sleep. Using actigraphy to measure sleep objectively, we find that low-income adults in Chennai, India, sleep only 5.5 hours a night on average despite spending 8 hours in bed. Their sleep is highly interrupted, with sleep efficiency—sleep per time in bed—comparable to those with disorders such as sleep apnea or insomnia. A randomized three-week treatment providing information, encouragement, and improvements to home sleep environments increased sleep duration by 27 minutes a night by inducing more time in bed. Contrary to expert predictions and a large body of sleep research, increased nighttime sleep had no detectable effects on cognition, productivity, decision making, or well being, and led to small decreases in labor supply. In contrast, short afternoon naps at the workplace improved an overall index of outcomes by 0.12 standard deviations, with significant increases in productivity, psychological well-being, and cognition, but a decrease in work time.

I. Introduction

Understanding the lives of the poor is central to modern development economics. Economists have studied many deprivations associated with poverty, such as lack of access to nutrition, water, education, health care, and clean air. This article considers a previously unexamined challenge faced by the urban poor in developing countries: sleep deprivation. People in these settings face many barriers to a good night’s sleep, such as heat, noise, crowding, physical discomfort, and psychological distress. Sleep could be a crucial input to their productivity, well-being, and cognitive function. Yet we know little about how much and how well people in low-income countries sleep, or the returns to policies that seek to increase sleep.

Using state-of-the-art technology to measure sleep objectively, we uncover widespread sleep deprivation in Chennai, India. Our two samples of low-income adults sleep only 5.5 hours a night on average, far below the minimum level recommended by sleep experts (Hirshkowitz et al. 2015; Watson et al. 2015). This is not due to a lack of trying. People spend about eight hours a night in bed, but their sleep is highly disrupted, with 31 awakenings in a typical night. The implied sleep efficiency—time asleep per time in bed—of 70% is much lower than objective measures from general U.S. populations, and similar to those suffering from disorders such as sleep apnea or insomnia in high-income countries (Hedner et al. 2004; Trauer et al. 2015).

An enormous body of research, mostly conducted in sleep labs in rich countries, documents severe negative effects of sleep deprivation on a range of outcomes, from attention and memory to mood and health (Banks and Dinges 2007; Lim and Dinges 2010). While experimental evidence on the effect of increasing sleep in field settings is scarce,¹ there is a widely held belief among researchers and the public that reducing sleep deprivation would lead to improvements in economic outcomes (Walker 2017). To document these priors, we surveyed 119 experts from sleep science and economics who predicted sizable economic benefits, including a 7% increase in work output, of increasing sleep by half an hour a night from the low levels observed in our setting.

To measure the economic effects of increasing sleep in the field, we conducted a randomized controlled trial with 452 adults in Chennai. We employed participants for a one-month data entry job with flexible hours, allowing us to precisely measure productivity and labor supply, as well as physical and psychological well-being, cognition, and time, risk, and social preferences. Two night sleep treatments gave participants (i) items to improve their home-sleep environments, (ii) information and verbal encouragement to increase their night sleep (the encouragement treatment), and (iii) for a subset of participants, additional financial incentives to increase night sleep (the incentives treatment). These treatments were cross-randomized with a nap treatment that offered participants the opportunity for a daily half-hour afternoon nap at their workplace.²

The night sleep treatments on average increased nighttime sleep by 27 minutes a night (std. err. = 3 minutes), with larger effects for the incentives (33 minutes) than for the encouragement treatment (20 minutes). The increased sleep duration was entirely driven by additional time in bed—on average 38 minutes a night (std. err. = 4 minutes)—rather than higher sleep efficiency. These results demonstrate that people do have substantial ability to adjust their nighttime sleep through changes in time in bed, but may not be able to increase their sleep efficiency. The low sleep efficiency increases the opportunity cost of sleep: raising sleep duration by 1 minute requires 1.4 more minutes in bed.

Similarly, the nap treatment increased daytime sleep by about 14 minutes a day on average (std. err. = 0.3 minutes), while slightly crowding out nighttime sleep. Although both types of treatments increased 24-hour sleep, the 27-minute increase from the night sleep treatments was significantly higher than the 8-minute increase from the nap treatment (p < .01).

We first examine the effects of each combination of treatments on an overall summary index that aggregates all outcomes as in Anderson (2008). Each of the night sleep treatments alone had no significant effect on this overall index: 0.00 standard deviations (std. err. = 0.07) and −0.05 std. dev. (std. err. = 0.07), respectively, for the encouragement only and incentives only groups. In contrast, naps alone had a positive, marginally significant effect of 0.11 std. dev. (std. err. = 0.07, p = .11). Those who received a night sleep treatment in addition to naps had very similar effects to those with naps only.³ This pattern of results suggests that naps have an overall positive effect on outcomes, whereas increases in night sleep do not. To increase statistical power and streamline the discussion of effects on the individual outcomes, we turn to an analysis that pools the two night sleep treatments and does not allow for an interaction effect of night sleep and nap treatments. The effects of each treatment described below should thus be interpreted as conditional on the distribution of the other treatment (Muralidharan, Romero, and Wüthrich 2019).

In the pooled analysis, we find no significant effect of increased night sleep on the overall index (−0.01 std. dev., std. err. = 0.04), or on summary indices corresponding to four families of outcomes: work, well-being, cognition, and economic preferences. In fact, the pooled night sleep treatment had no significant positive effect on any outcome other than sleep itself. It did not significantly increase productivity at the data entry job, a relatively cognitively demanding task intended to be sensitive to sleep deprivation. Instead, increased sleep came at the cost of lowering labor supply by nine minutes a day, leading to a small (but not statistically significant) decrease in earnings. We reject the median expert prediction of a 7% increase in output (p < .001). Similarly, we find no significant effects on detailed measures of physical and psychological well-being or standard measures of cognition and social, risk, and time preferences.

Why does increased night sleep not have benefits in our setting, contrary to expert predictions and a large body of lab studies? One possibility is that the large effects from lab experiments, which typically dramatically reduce sleep for up to a few nights, do not generalize to marginal, policy-relevant increases in sleep in the field. Another possibility is that the low quality of sleep observed in our setting—as proxied by low efficiency and frequent awakenings—explains the lack of benefits of increased nighttime sleep. Returns to increased sleep could be higher in typical rich-country settings. We cannot adjudicate these reasons, but our results highlight the importance of studying sleep in the field, where outcomes have real stakes and sleep is a choice variable with opportunity costs. They also caution against extrapolating sleep science findings across diverse contexts.

In contrast to night sleep, naps significantly improved a range of outcomes. The nap treatment significantly increased the overall summary index by 0.12 std. dev. (std. err. = 0.04, p = .00), as well as the index variables corresponding to well-being (0.08 std. dev., p = .03), and cognition (0.10 std. dev., p = .08). The effect of naps on work outcomes depends on the comparison. Compared with taking a break, naps increased earnings—the summary variable for the work outcomes—by 0.05 std. dev. (p = .05). However, driven by a decrease in labor supply, naps reduced earnings compared to working during the nap time by 0.10 std. dev. (p = .00), highlighting the importance of taking into account the opportunity costs of sleep.

Considering the individual outcome variables one by one and adjusting for multiple comparisons, naps have significant positive effects on productivity (0.04 std. dev., p = .06), psychological well-being (0.12 std. dev., p = .04), lab measures of cognitive function (0.08 std. dev., p = .06), and attention at work (0.20 std. dev., p = .07).

The nap and night sleep treatments have statistically different effects on the overall index (p = .02). We are less powered to detect differences for the family-level indices. Point estimates are larger for the nap treatment for well-being, cognition, and preferences, but the differences are not statistically significant. At the level of individual outcomes, we find statistically significant differences in effects only for psychological well-being (p = .04). Estimating the per minute effects of the two types of sleep in an instrumental variable analysis, however, we can reject equal per minute effects of naps and nighttime sleep on the summary index (p < .01), on the family indices for cognition (p = .08) and preferences (p = .08), and on some individual outcomes, including labor supply (p = .04) and psychological well-being (p = .05). In every case but labor supply, the effects are more positive for naps.

One possible reason for the different effects of the nap and night sleep treatments is that the timing of sleep may matter. Contrary to hypotheses and some evidence in sleep science (e.g., Nicholson et al. 1985; Mollicone, Van Dongen, and Dinges 2007; Mollicone et al. 2008), naps and night sleep may simply not be close substitutes. An alternative explanation is that sleep quality may play a role, since naps in our study occurred in a more comfortable office environment. We cannot separate these explanations, but hope that future work in similar settings may help answer this question.

Our article makes the following contributions. First, it contributes to a better understanding of the living conditions faced by the poor in developing countries by providing objective measures of sleep. We discover surprisingly low levels of sleep duration and efficiency among the urban poor in Chennai. These findings are consistent with two papers measuring sleep objectively in smaller samples in Sri Lanka and Haiti (Castro et al. 2013; Schokman et al. 2018) and contrast with self-reported measures of sleep, which may fail to capture the low sleep efficiency and its effect on total sleep (Stranges et al. 2012; Gildner et al. 2014; Simonelli et al. 2018).

Second, we build on a recent literature that estimates the causal effect of sleep outside of sleep laboratories. The lack of effects of nighttime sleep we find using a field experiment contrasts with an economics literature that uses natural experiments in rich countries to demonstrate that sleep can have sizable effects on wages (Gibson and Shrader 2018; Giuntella and Mazzonna 2019), hospitalizations (Jin and Ziebarth 2020), accidents (Smith 2016), and civic behaviors (Holbein, Schafer, and Dickinson 2019).⁴ We speculate that the stark difference in sleep efficiency in our setting compared to rich-country populations contributes to this difference. It could also be that increased night sleep simply does not have high benefits at the margin in the field, as in another recent field experiment (Avery, Giuntella, and Jiao 2019). Additional field studies of sleep, including interventions that improve sleep efficiency, may help reconcile these findings.

Third, we show that afternoon naps in a comfortable office environment have positive effects on a range of outcomes, including productivity, well-being, and cognition. Naps are a common feature of life around the world and are particularly prevalent in tropical countries (Dinges 1992). Naps have been studied in sleep labs, but we have little causal evidence on the effects of naps on worker productivity and other real-world outcomes or consideration of the opportunity cost of naps (Lovato and Lack 2010; Ficca et al. 2010). Our work takes a step toward filling this gap, and shows that naps may be an effective way to combat sleep deprivation. The decline of naps as employment in developing countries shifts toward Western schedules could therefore be costly.

Finally, recent research in behavioral and development economics argues that people in developing countries often underinvest in high-return investments such as preventive health, agricultural inputs, or capital investments (Kremer, Rao, and Schilbach 2019). At first glance, the low levels of sleep we discovered appear to tell the same story, and experts predicted substantial effects of increased sleep. Instead, our evidence suggests that—in this context—people do not underinvest in sleep duration given the environmental constraints that they face. The returns to increasing night sleep in their home environments are low and possibly even negative. To paraphrase Schultz (1964), low-income people in Chennai may be poor but efficiently tired.

II. Measuring Sleep in Chennai

II.A. Measuring Sleep Outside the Lab

The gold standard for objectively measuring sleep in labs is polysomnography (PSG), by recording brain waves, blood oxygen levels, eye movements, and body movements to determine sleep/wake cycles and stages of sleep (Marino et al. 2013). Although highly accurate, this bulky technology is impractical for field studies and may interfere with natural sleep patterns at people’s homes, thus making measuring sleep outside of sleep labs challenging. Self-reported measures are unreliable and correlate only moderately with objective measures because people tend to report time in bed rather than hours asleep, leading to overreporting of sleep duration (Lauderdale et al. 2008; Schokman et al. 2018).

Actigraphs, which resemble wristwatches and infer sleep/wake states from body movement, recently emerged as a viable alternative for field studies. These devices allow researchers to objectively measure sleep in participants’ home environments without interfering with sleep, as these devices are portable, comfortable, and unobtrusive. Validation studies show that actigraphs reliably measure sleep duration. Comparisons between actigraphy and PSG measures show high degrees of accuracy in sleep-wake detection, with 90% minute-by-minute agreement between the two (Sadeh et al. 1995; Marino et al. 2013). Actigraphs have been found to provide valid and clinically useful measures of sleep duration even among people with sleep disorders (Kushida et al. 2001; Smith et al. 2018) and reliably capture treatment effects of various interventions on sleep (Sadeh 2011).

Actigraphs also measure sleep efficiency, defined as time asleep divided by time in bed. This measure is available since—in addition to number of hours asleep—actigraphs also detect when an individual is in bed but not asleep. Sleep efficiency is perhaps the most commonly used proxy for sleep quality in sleep science (Ohayon et al. 2017). Disruptions to sleep, such as brief awakenings during the night, drive down sleep efficiency. In addition, sleep efficiency affects the opportunity cost of sleep, since it indicates the time in bed needed to achieve an hour of actual sleep.

II.B. Sleep Deprivation around the World

While sleep scientists recommend seven to nine hours of sleep a night (Hirshkowitz et al. 2015; Watson et al. 2015), numerous studies show that people in high-income countries sleep less than this (Walker 2017).⁵ For instance, Lauderdale et al. (2008) measure sleep via actigraphy among a large, diverse population of healthy young adults in Chicago, and report an average sleep duration of 6.1 hours a night, well below the recommended range.

In contrast, there is scant evidence on sleep patterns in developing countries. Sleep deprivation may be widespread and even more severe in the rapidly growing cities of the developing world, where residential structures are often of low quality and people are exposed to excessive heat, noise, crowding, and pollution—all conditions likely to hinder sleep. Even self-reports—which typically overestimate sleep—suggest a substantial share of people in developing countries sleep less than the recommended seven to nine hours. For example, the 4,500 rural, older Indian adults surveyed in Gildner et al. (2014) self-report 7.1 hours of sleep on average, with about 30% of these individuals reporting 6 or fewer hours a night (Selvamani et al. 2018).

Two recent studies in low-income countries measured sleep using actigraphs and identify significant fractions of the population as sleep deprived. In particular, Schokman et al. (2018) finds an average of only six hours a night among 175 adults from urban Sri Lanka. Knutson (2014) finds that 58 adults in Haiti sleep on average seven hours a night in a rural population without electricity.

II.C. Sleep(less) in Chennai

Our first measure of sleep in Chennai comes from the RCT sample of 452 adults recruited for a full-time data entry job for one month. To capture reliable and objective measures of sleep beyond self-reports, all participants wore actigraphs continuously throughout the study.⁶ Below, we describe sleep during the baseline period (before treatment) in this sample. We then report very similar patterns of sleep in a broader sample in Chennai, which wore actigraphs for three nights.

1. A Typical Night in Chennai

We first provide an example to highlight key features of participants’ sleep patterns. Figure I, Panel A illustrates a typical night for a study participant, using minute-by-minute actigraph measures of sleep (light gray) and wake (dark red) status. This night closely matches the average time in bed, sleep duration, and sleep efficiency in the RCT sample. The participant spends about 8 hours in bed during this night but only achieves 5.6 hours of highly fragmented and interrupted sleep, involving over 30 awakenings. For comparison, we show a less interrupted night with 90% efficiency in Figure I, Panel B. While this night is unusual in Chennai—only 1% of nights in our sample feature such high sleep—it resembles nights of healthy adults in high-income countries who typically enjoy sleep efficiency of 85%–95% (Cole et al. 1992; Carrier et al. 2001; Walker 2017).

2. Time in Bed versus Time Asleep

The RCT sample spends roughly eight hours a night in bed before treatment begins, with strong congruence between actigraph measures (Figure II, Panel A) and self-reports (Figure II, Panel B). Time in bed in Chennai is quite similar to that found in U.S. samples.⁷ Despite this significant time in bed, study participants only enjoy 5.6 hours asleep per night (Table I and Figure II, Panel C), significantly below time in bed and the recommended seven to nine hours. Ninety-five percent of participants slept less than seven hours a night, and 71% slept less than six hours a night on average. In high-income countries, such low average time asleep is typical in populations with disorders such as sleep apnea (Cole et al. 1992; Kushida et al. 2001; Gershon et al. 2012).

TABLE I.

Sleep Statistics in Two Samples in Chennai

	RCT sample	Broader sample
	(pretreatment)
	(1)	(2)
Panel A: Night sleep
Hours in bed	8.03	7.68
	(0.97)	(1.23)
Hours asleep	5.58	5.46
	(0.87)	(1.15)
Sleep efficiency	0.70	0.71
	(0.08)	(0.10)
Number of awakenings	31.95	27.4
	(7.95)	(10.14)
Fraction sleeping less than 7 hours
Participant-level	0.95	0.93
	(0.22)	(0.26)
Participant-day-level	0.89	0.87
	(0.31)	(0.33)
Fraction sleeping less than 6 hours
Participant-level	0.71	0.69
	(0.46)	(0.46)
Participant-day-level	0.65	0.64
	(0.48)	(0.48)
Self-reported hours asleep	7.20	6.42
	(0.94)	(1.49)
Panel B: Nap sleep
Percent napping on a given day	N/A	0.25
		(0.43)
Duration of naps (conditional on napping)	N/A	0.85
		(0.61)
Panel C: Total sleep
Hours asleep	5.58	5.69
	(0.87)	(1.15)
Participant-nights	3,080	1,367
Participants	452	439

Open in a new tab

Notes. This table presents average sleep characteristics in two samples in Chennai. Standard deviations are in parentheses. Column (1) presents summary statistics from the RCT sample, only using data from the seven nights in the baseline period (i.e., before any treatments were implemented). Column (2) presents summary statistics from the three nights in our complementary sleep survey across a broader population in Chennai (described in Online Appendix F). All outcomes are objectively measured by actigraphs unless indicated otherwise. All means and standard deviations are at the participant level (i.e., we collapse the data at the participant level by averaging across nights) unless indicated otherwise. The variables shown in the table are: (i) hours in bed (per night, regardless of whether awake or asleep); (ii) hours asleep at night (per night); (iii) sleep efficiency (hours asleep/hours in bed); (iv) number of awakenings per night; (v) proportion of participants with less than seven hours of night sleep; (vi) proportion of participants with less than six hours of night sleep; (vii) self-reported hours asleep at night; (viii) proportion of participants napping on any given day; (ix) duration of naps (in hours) conditional on taking a nap; and (x) total hours asleep per 24 hours (the sum of night sleep and nap sleep).

3. Sleep Efficiency

Average sleep efficiency in our sample is 70% (Figure II, Panel E), far below recommended levels by sleep scientists who found that a minimum of 85% is needed to indicate “high-quality” sleep (Ohayon et al. 2017). Like sleep duration, sleep efficiency is much lower than typically found in high-income countries, and instead resembles U.S.-based patients suffering from sleep disorders such as sleep apnea (Roure et al. 2008) or insomnia (Trauer et al. 2015). Sleep efficiency is low throughout the night, remaining around 70% between 1 to 5 a.m. (when almost everyone is in bed), consistent with interrupted sleep throughout the night (Online Appendix Figure A.I). Participants experience about 32 awakenings on an average night (Table I, column (1)), again comparable to insomniacs in the United States (Lichstein et al. 2006).⁸

4. Barriers to Sleep

Why is sleep so inefficient? Survey responses highlight the importance of mental and physical distress (e.g., worries, stress, pain, or hunger) as well as environmental factors (Online Appendix, Figure A.IVa). Over 50% of study participants indicate that their sleep is disrupted by heat, noise, and/or light, which the night sleep treatments described below were intended to address.

5. Napping

Naps are relatively common in this population. Seventy-three percent of participants in our study reported taking at least one nap in the week before enrolling in the study. Conditional on napping, the median time reported for a nap is about one hour. The frequency and length of naps in U.S. populations is similar: Dinges (1992) finds that across a broad population of U.S. adults, 61% reported napping at least once a week with an average nap duration of 73 minutes, while in Pilcher, Michalowski, and Carrigan (2001), 74% of healthy adults report napping during a seven-day period.

6. Self-Reported Sleep

Self-reports significantly overestimate time asleep, relative to the objective actigraph measures, consistent with findings in the United States (Lauderdale et al. 2008; Avery, Giuntella, and Jiao 2019).⁹ Average baseline self-reported sleep duration in our study is 7.2 hours (Figure II, Panel D), quite similar to the average of 7.1 hours found in a representative survey of older adults in rural India described in Gildner et al. (2014). In comparison, average self-reported sleep duration in the United States ranges from 6.8 to 7.9 hours a night (Lauderdale et al. 2008; Watson et al. 2015; Jackson et al. 2018).

7. Broader Population

To investigate the representativeness of our RCT sample, we conducted a supplementary sleep survey with 3,833 individuals across randomly sampled neighborhoods in Chennai.¹⁰ A subset (N = 439) completed three nights of actigraph measurements. Despite not using any of the RCT screening criteria for this survey, the nighttime sleep duration and efficiency in this broader sample are similar to that of RCT participants, with an average of 5.5 hours of sleep a night and 71% sleep efficiency (Table I, column (2)). As in the RCT sample, napping is common, with 25% of participants napping on any given day and an average duration of those naps of roughly 50 minutes.

III. Experimental Design

We designed our experiment with three broad goals in mind. First, we aimed to estimate the effects of increased sleep over a few weeks in relatively natural sleep environments (as opposed to depriving people of sleep in lab settings). Second, we wanted to precisely measure labor supply, productivity, and earnings, and thus we employed study participants full-time in a realistic but closely controlled data entry job. Finally, to provide a broad view of the effects of increased sleep, we collected a range of additional outcomes, including cognition, preferences, and well-being.

Figure III provides an overview of the experimental design and timeline of the study. Four hundred fifty-two participants worked for 28 days in an office in Chennai, spending most of their workdays doing paid data entry work. Enrollment took place on a rolling basis. The office contained computer work stations for data entry, a break room, booths for surveys and experimental tasks, and nap stations on a separate floor.

III.A. Interventions to Increase Sleep

For their first eight days in the study, participants remained in a control condition, allowing us to collect rich baseline data. Then, we cross-randomized participants to two night sleep treatments and a nap treatment, stratified by baseline sleep and earnings.

1. Night Sleep Treatments

Each participant was randomly assigned to one of two night sleep treatment groups (encouragement or incentives) or to a control group in equal proportions.

Devices + Encouragement: This treatment involved a bundled intervention to increase night sleep. Individuals were offered (a) information about the benefits of sleep (in particular, generic health benefits) and tips to improve their sleep (such as going to bed at the same time every day, avoiding caffeine after 4 p.m., and avoiding screens before bed), (b) encouragement to increase their sleep as well as daily feedback on their night sleep duration as measured by the actigraph, and (c) loaned devices to improve their sleep environment. The offered devices included eye shades, earplugs, a cot, a mattress, sheets, pillows, and a fan (see Online Appendix Figure A.IIb).¹¹
Devices + Incentives: This group received the same bundled intervention as the Devices + Encouragement group plus financial incentives to increase their actigraph-measured sleep during the treatment period. Each day, participants were paid Rs. 1 per minute of increased sleep for up to two hours of increased sleep (Rs. 120, about $1.70), relative to their baseline period sleep. There was no penalty for sleeping less than in the baseline period.¹² To control for any income effects due to the sleep incentive payments, participants in the control and encouragement groups were randomly and anonymously matched to participants in the incentives group and received the same stream of payments, independent of their own sleep.
Control: This group did not receive any of the above treatments. To deal with the concern that loaning items might generate reciprocity effects or affect reported well-being directly, we offered placebo household goods, unrelated to sleep, to a subset of control participants. The total value of these goods was roughly the same as that of the sleep devices and included items such as small kitchen devices, a chair, decorative figurines, and a flashlight. These goods were also returned at the end of the study.¹³

Given the difficulty of increasing sleep in the field, we took a bundled approach in designing our treatments, working to increase sleep through multiple channels. Participants could respond to the encouragement and incentives by spending more time in bed or by taking steps to increase their sleep efficiency. The tips to improve sleep, such as avoiding caffeine in the evening, turning off the television, or putting away one’s cellphone at night, could plausibly increase sleep efficiency. Finally, the loaned devices could increase sleep efficiency and time in bed, if the devices made it easier to fall asleep or reduced awakenings or if they made time spent in bed more enjoyable.

2. Nap Treatment

Motivated by existing lab evidence that naps can be effective in boosting cognitive function (Lovato and Lack 2010) and can make up for limited nighttime sleep (Mollicone et al. 2008), we cross-randomized the night sleep treatments with a nap intervention. Starting on day 9 of the study, a random subset of individuals were given the opportunity to take a short afternoon nap every day between 1:30 and 2 p.m. Located in a quiet and gender-separated part of the study office, the 25 private nap spaces included a bed, blanket, pillow, table fan, ear plugs, and eye shades (see Online Appendix Figure A.IIc). The actigraphs show that roughly 90% of study participants did sleep during their allotted nap time. Those who did not want to nap were asked to sit quietly or rest in their nap area; they did not have the option to work during this time.

The remaining (no-nap) participants were randomized each day with equal probability to either a work day, on which we allowed them to work through the nap period, or a break day, on which we enforced a half-hour break from data entry during the same period. Break day participants were allowed to engage in any leisure activity they chose, including sitting in a comfortable office break room. By comparing nap and break participants, we isolate the effect of a nap relative to a break of the same length. By instead comparing the nap and work participants, we can estimate the net effect of naps on work output, including the lost work time.

III.B. Outcome Measures

Sleep and work are the two key sets of outcomes of this study. We measured each of them daily using actigraphs and the data entry platform, respectively. Study participants also completed a series of short surveys and experimental tasks throughout the study (see Online Appendix Table A.III). Described in greater detail in Online Appendix C, these measures fall into three broad categories: (i) physical and psychological well-being, (ii) cognition, and (iii) economic preferences.

1. Measures of Sleep

We measure night and nap sleep—sleep duration, time in bed, efficiency, and interruptions—using actigraphs, as described in Section II.A. Ninety-four percent of participants wore their actigraph on any given day, balanced across treatments. We complement these measures with daily self-reports of time in bed, time asleep, and number of awakenings during the night.

2. Work-Related Outcomes

Participants were engaged each day in data entry work. We designed a software interface that presented participants with images containing alphanumeric text and asked them to transcribe the data by typing into text boxes (see Online Appendix Figure A.III). The task was designed to mimic a real-world data entry job.¹⁴ Participants were paid for time spent typing as well as the amount of data entered, as described below. This design allows us to precisely measure labor supply, productivity, and earnings.

Labor Supply. Our preregistered measure of labor supply is the active typing time as automatically measured by the data entry software. As in many workplaces, participants were not forced to arrive or leave the office at precise times. On most days, participants could arrive or depart from the office as they chose between 9:30 a.m. and 8 p.m. When in the office, participants could take breaks from work. We can precisely measure even short breaks: if a participant spent two consecutive minutes without typing, the software automatically paused and the break period did not count toward the labor-supply measure. Thus, participants had a fair amount of control over their labor supply, except for time slots set aside for surveys, experimental tasks, and the lunch break.

Earnings. Earnings in the data entry work is our preregistered measure of performance at work and was used as our summary measure of work because it combines labor supply and productivity. It has two components: an “attendance pay” per hour of active typing (one-third of work earnings) and a “performance pay,” a piece rate for each correct character and a penalty per mistake (two-thirds of data entry earnings). Each half hour, piece rates were randomly varied between a low value (Rs. 5 per 1,000 correct characters) and high value (four times as large) with equal probability. The penalty rate remained constant throughout at Rs. 1 per 10 mistakes. The variation in piece rates allows us to benchmark any productivity effects of the sleep treatments against monetary incentives. The participants were paid daily, just before leaving the office for the day.¹⁵

Productivity. Our preregistered measure of worker productivity is output divided by active typing hours. Output is the number of correct entries minus (a weighted) number of mistakes. The weight on mistakes was defined as the ratio of the average piece rate and the penalty rate.

3. Well-Being

We collected a wide range of measures of psychological and physical well-being. As preregistered, we examine these variables both as indices and individually. The preregistered measures of mental well-being are happiness, sense of life possibilities (Cantril Scale), life satisfaction, stress, and depression. The measures of physical well-being are performance in a stationary biking task, reported days of illness, self-reported pain, activities of daily living, and blood pressure.¹⁶

4. Cognition

Sleep scientists have documented a strong relationship between sleep and cognition in numerous laboratory studies in rich countries (Lim and Dinges 2010; Killgore 2010). We collected (i) laboratory measures of cognitive function borrowed from cognitive psychology and sleep medicine; and (ii) a measure of attention to incentives at work embedded in the data entry task.

Lab Measures of Cognition. Each afternoon, participants completed the Psychomotor Vigilance Task (PVT), a standard measure of alertness and attention used in sleep medicine (Basner and Dinges 2011). Every other day, they completed cognitive tasks measuring memory (Corsi blocks task) and inhibitory control (Hearts and Flowers task), described in detail in Dean, Schilbach, and Schofield (2019) and briefly in Online Appendix C.5. All cognitive tasks were incentivized for performance (e.g., speed, accuracy).

Attention to Work Incentives. To test whether sleep affects the ability to attend to important aspects of one’s work environment, for example, the incentives faced, we randomized the visual salience of piece rates across days starting on day 6 of the baseline period. In the salient condition, the current piece rate was highlighted in different colors for each rate and displayed on the screen at all times. We consider this condition the “full-attention” benchmark, as in Chetty, Looney, and Kroft (2009). In the nonsalient condition, noticing and remembering the piece rate was more challenging. A single muted color was used for both piece rates, and (in the second half of the study) the rate was only visible for the first 15 seconds of a half-hour slot, fading out slowly. Online Appendix Figure A.III provides screenshots of each condition described below. The participant-level attention measure is the difference in average response to piece rate incentives in the full-attention benchmark and in the nonsalient condition.¹⁷

5. Preferences

Sleep may affect preferences through its impact on cognition or directly. For instance, sleep has been hypothesized to play a critical role in replenishing self-control (Vohs and Baumeister, 2016) and sleep deprivation has been correlated with cyberloafing at work (Wagner et al. 2012) and cheating (Barnes et al. 2011). Similarly, sleep could alter the weight placed on sure things versus gambles or on others versus the self (McKenna et al. 2007; Anderson and Dickinson 2010; Holbein, Schafer, and Dickinson 2019). To examine such effects we study time preference via financial savings outcomes and choices on a real-effort task, and risk and social preferences via standard experimental economics measures described below.

Savings. We measured savings behavior by providing participants an opportunity to save money in a lockbox at the study office, as in Schilbach (2019). At the end of each work day, after receiving their earnings, individuals had the opportunity to deposit or withdraw money. Participants were randomly assigned to receive daily interest rates between 0% and 2% for any money saved in the box.¹⁸ For participants receiving the positive interest rate, at least, the savings account we offered was quite lucrative. The deposits made in this account constitute our main savings outcome.

Effort Discounting. We measured present bias using real-effort choices, following Augenblick and Rabin (2019). Participants made decisions about how many pages to type at the end of the day on a particular date under different piece rates. Using choices elicited both in advance and on the day of the work itself, we structurally estimate an individual-level present bias parameter β_i, once each in the baseline and treatment periods. A complete description of the task is in Online Appendix C.6.3.

Social and Risk Preferences. We measured risk and social preferences via standard tasks in the behavioral economics literature. Risk aversion and loss aversion are captured via a multiple price list elicitation similar to those in Holt and Laury (2002), and Charness, Gneezy, and Imas (2013). Social preferences are measured via dictator, ultimatum, and trust games (Camerer 2003).

6. Realism and External Validity

Conducting the study in the context of a month-long data entry job in a controlled environment follows Kaur, Kremer, and Mullainathan (2015) and allows for the provision of afternoon naps and precise measurement of the outcomes described above. However, it also comes with some costs. First, the work environment has some unusual and artificial features, such as regular surveys and laboratory measures of cognition and preferences. Second, any labor supply responses we find might be muted in environments where employers more strictly control schedules. In practice, participants tend to spend about 8 hours at the office each day, with an average arrival time of 10:32 a.m. (std. dev. = 43 minutes) and average departure time of 6:20 p.m. (std. dev. = 57 minutes). This is quite similar to other jobs in our context, with long commutes and unreliable transportation, such that arriving strictly at a given time is difficult.

III.C. Expert Predictions

To quantify how our results compare with existing scientific understanding, we conducted surveys of experts in sleep science and economics to elicit their prior beliefs about the treatment effects of the night sleep interventions included in this study (DellaVigna, Pope, and Vivalt 2019). Participation in the survey was solicited via emails to experts in each field. The survey provided information on the design of the study, the magnitude of the increase in night sleep reported in Section IV.B, and the outcome measures described above.¹⁹ Three versions of the survey were tailored to different respondents: development and labor economists; behavioral economists; and sleep medicine experts. A total of 28 labor and development economists, 19 behavioral economists, and 77 sleep medicine experts responded to the survey. In an effort to keep the survey short, we did not elicit predictions about the effects of the nap treatment. All experts were provided with relevant benchmarks (e.g., the elasticity of effort with respect to the piece rate) and made predictions on labor supply and work output effects. Both types of economists responded with their beliefs about savings. Only behavioral economists were asked to predict changes in present bias, and only sleep experts were asked to predict changes in sustained attention and physical health. The expert predictions are shown in Figure IV and in Online Appendix Table A.IV and are discussed when presenting results. Further details are provided in Online Appendix C.1.

Figure IV — Expert Predictions versus RCT Results

This figure summarizes the predictions made by experts in economics and sleep science about the expected treatment effect of our (pooled) night sleep interventions. We normalize each prediction, dividing them by the control group’s standard deviation. For comparison, we also present the actual estimated treatment effect. The bars show the interquartile range (25th to 75th percentiles) of the predictions for a given outcome variable. We also show the median prediction (X) and the actual point estimates (diamond) of the treatment effects. N refers to the number of responses for each outcome variable. This number varies by outcome because different types of experts (e.g., sleep researchers, behavioral economists) were asked about some different outcomes. *Correct Entries* refers to the number of daily correct characters in the data entry task, that is, a measure of overall output per day. *Labor Supply* refers to the daily number of hours working in the typing task, that is, time at the office excluding voluntary and scheduled pauses. *Physical Health* refers to a variable that averages (normalized) systolic and diastolic blood pressure. We flip the sign of this variable so a positive value means an improvement in health (i.e., a reduction in blood pressure). *Attention* refers to an index pooling inverse response times and minor lapses in the Psychomotor Vigilance Task, a standard lab measure of attention in the sleep literature. Signs are flipped such that higher values refer to increased attention. *Savings* refers to the daily amount deposited minus the amount withdrawn in the savings box during the experiment. *Present Bias* refers to the present-bias parameter β. Unlike the other variables, the predictions and point estimate refer to the reduction in present bias (increase in β) rather than a normalized outcome, for ease of interpretation.

III.D. Study Population and Balance Checks

We followed two strategies to recruit our study sample. First, recruiters went to low-income neighborhoods in Chennai and spread information about the study, advertising a one-month data entry job. Second, past participants could refer potential new participants to the study. In both cases, recruiters interviewed individuals to determine their eligibility to participate in the study.

1. Eligibility Criteria and Selection

Interested individuals participated in a two-stage screening process, involving a brief unpaid survey and a home visit to check whether the individual met the study’s eligibility criteria: (i) being 25 to 55 years old; (ii) fluency in the local language (Tamil) and the ability to read and write numbers; (iii) having worked fewer than five days per week and earning an average of Rs. 700 ($10) or less per day worked in the previous month; (iv) living in a dwelling able to accommodate the sleep devices used in night sleep treatments and ownership of three or fewer of the sleep devices being offered in the study; (v) the intention to be in Chennai for the following five weeks; and (vi) no children in the household younger than three years.

Importantly, this recruitment and screening procedure does not seem to select participants on average levels of sleep quantity and efficiency. In Table I, we find very similar patterns of sleep among individuals in Chennai in the broader sleep survey, as described in Online Appendix F.

2. Informed Consent

All participants went through a detailed informed-consent procedure including information about the work task, other experimental measures and surveys, the actigraphs and the randomized treatments. The specific research hypotheses were not shared with participants to avoid demand effects. Instead, the goal of the research was described as work to understand the “difficulties underprivileged people in India face, and how these problems affect their lives.”

3. Sample Characteristics

Online Appendix Table A.I shows sample characteristics. A typical study participant was about 35 years old with one or two children and 10 years of education. Two-thirds of study participants were women. Although only 30% of participants had prior computer experience, participants were eager to learn and improved rapidly in their data entry speed during the baseline period.

4. Balance Checks

We test for baseline imbalances in demographics and baseline measures of outcome variables across the experimental conditions in Online Appendix Tables A.I and A.II. Whether we separately consider each treatment cell (Online Appendix Table A.I) or compare the pooled night sleep treatment groups with the control and the nap and no-nap groups (Online Appendix Table A.II), the treatment groups were well-balanced across key characteristics. For each treatment arm, a joint F-test comparing it to the control group indicated no systematic differences on observable characteristics across groups.

As is expected given the large number of comparisons, a few statistically significant differences across treatment groups did emerge. Most notably among those, participants in the night sleep treatment groups were about a year younger than those in the control group, and baseline productivity and earnings were about 3%–4% lower in the nap group than in the no-nap groups (Online Appendix Table A.II). We control for age and for the participant-level baseline average of the outcome variables, so these imbalances should not affect our results.²⁰

IV. Experimental Results

IV.A. Empirical Framework

Most of our empirical analyses, including all work-related outcomes, estimate treatment effects on outcomes measured at the participant-day level using variants of the following equation:

(1)

where y_itd is the relevant outcome for participant i on their tth day in the study on calendar date d. T_i is a vector of indicator variables capturing the treatment(s) that participant i was assigned to. β is the vector of coefficients, capturing the effect of each treatment on the outcome of interest.

Following McKenzie (2012), we control for the average baseline value of the outcome variable Inline graphic in all specifications, and drop the baseline days from the regression. We also drop days on which participants were absent, since attendance was balanced across groups. X_it includes controls for participants’ age (quartiles) and gender and, where specified, a dummy variable for whether a given no-nap participant i was assigned to work through the nap period or instead to take an enforced break on day t. This allows us to estimate the effect of naps separately compared with working and taking a break.²¹ Finally, we include day-in-study and calendar date fixed effects, captured by δ_t and λ_d, respectively. All standard errors are clustered at the participant level.

For some outcomes, such as preferences, we only have one observation in the baseline and one in the treatment period per participant. In those cases, we run participant-level regressions:

(2)

where again the outcome variable only uses the observations from the treatment period and we control for the baseline observation of the outcome, Inline graphic . The vector X_i includes the same gender and age controls. Because these outcome measurements span multiple days (e.g., present bias) or are on a fixed day-in-study (e.g., the endline survey), this specification does not include day-in-study or calendar-date fixed effects or control for whether no-nap participants worked or took a break.

1. Combining Outcomes into Indices

Given the large number of outcomes, we divide them into four major families: work, well-being, cognition, and preferences. We construct a single summary outcome for each family. The work outcomes are naturally summarized by (standardized) earnings in the data entry task, which combines productivity and labor supply into a single quantity. For the other families, we create standardized index variables by residualizing each constituent outcome with respect to day in study and calendar date, standardizing it by the control group’s mean and standard deviation, and then taking a weighted average to form the index. Following Anderson (2008), the weights are the inverse of the covariance matrix of the residualized, standardized outcomes. This ensures that outcomes that are highly correlated receive less weight than outcomes that capture independent information. Signs of outcome variables are flipped when necessary, so positive treatment effects imply an improvement in the outcome.²² We also report treatment effects on an overall index, which combines the four family-level summary outcomes into a single variable. We use the same procedure to create the overall index.

2. Multiple-Hypothesis Testing

We report three approaches to dealing with multiple-hypothesis testing issues caused by observing many outcomes. Our simplest approach is to examine a single overall index variable which combines all outcomes, as described above. Our intermediate approach is to consider outcomes at the level of the four families, using one summary variable for each family, while correcting for the existence of multiple families. Finally, at the level of the individual outcomes, we correct for the existence of multiple outcomes within each family. In each case, we report adjusted p-values that control the family-wise error rate—the probability of at least one false rejection—using a step-down permutation procedure based on Westfall and Young (1993). The adjusted p-values are displayed for the main results in Table IV and Online Appendix Tables A.VII, A.VIII, and A.IX.²³

TABLE IV.

Pooled Treatment Effects of Night Sleep and Nap Treatments

	Overall	Work				Well-being
	Index	Earnings	Productivity	Labor supply	Output	Index	Physical	Psychological
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
Night sleep treatments	−0.01	−0.04	0.02	−0.08	−0.04	0.01	0.04	−0.05
	(0.04)	(0.02)	(0.02)	(0.02)	(0.02)	(0.03)	(0.04)	(0.06)
	{.79}	{.08}	{.30}	{.00}	{.05}	{.69}	{.35}	{.36}
		[.30]	[.32]	[.00]	[.09]	[.97]	[.36]	[.36]
Nap treatment	0.12	−0.02	0.04	−0.09	−0.01	0.08	0.05	0.12
	(0.04)	(0.02)	(0.02)	(0.02)	(0.02)	(0.03)	(0.04)	(0.05)
	{.00}	{.23}	{.03}	{.00}	{.77}	{.01}	{.18}	{.02}
		[.25]	[.06]	[.00]	[.77]	[.03]	[.19]	[.04]
Participants	451	451	451	451	451	452	452	452
Unadjusted p-value NS versus nap	{.02}	{.62}	{.49}	{.65}	{.23}	{.16}	{.83}	{.03}
FWER-corrected p-value NS versus nap		[.62]	[.66]	[.66]	[.45]	[.41]	[.84]	[.04]
Night sleep treatments	−0.00	0.03	−0.01	−0.00	0.04	−0.04	−0.04
	(0.05)	(0.04)	(0.11)	(0.04)	(0.07)	(0.06)	(0.08)
	{.93}	{.51}	{.95}	{.95}	{.58}	{.45}	{.65}
	[.98]	[.76]	[.95]	[.98]	[.64]	[.64]	[.64]
Nap treatment	0.10	0.08	0.20	0.07	0.13	0.03	0.07
	(0.05)	(0.04)	(0.11)	(0.04)	(0.06)	(0.05)	(0.08)
	{.03}	{.03}	{.07}	{.05}	{.04}	{.52}	{.33}
	[.08]	[.07]	[.07]	[.15]	[.20]	[.51]	[.51]
Participants	452	452	429	452	452	415	415
Unadjusted p-value NS versus nap	{.12}	{.31}	{.21}	{.22}	{.36}	{.30}	{.32}
FWER-corrected p-value NS versus nap	[.41]	[.31]	[.31]	[.41]	[.36]	[.36]	[.36]

Open in a new tab

Notes. This table shows treatment effects of the pooled night sleep and the nap interventions on the overall index as well as the four families of outcomes: work, well-being, cognition, and preferences. Each row shows coefficients of treatment cells compared to the control group that receives no sleep-related treatments. All work-related regressions are conducted at the participant-day level using equation (1). All other regressions are at the participant level using equation (2). All outcome variables are the same as in Table III. In contrast to Table III, in this table (i) we pool the Devices + Encouragement and the Devices + Incentives treatments to estimate the pooled effect of the two night sleep treatments, and (ii) we do not include a separate indicator for the group that receives both the night sleep and nap treatments. Hence, these estimates should be interpreted as weighted averages of treatment effects in the relevant cells. For instance, the coefficient on the night sleep treatment is the average effect of being assigned to one of the two night sleep treatments (with equal probability), in a population which either receives naps or does not (with equal probability). Below each coefficient, we report (i) the corresponding standard errors in parentheses (·), robust to heteroskedasticity and clustered at the participant-level when applicable, (ii) the unadjusted p-value in curly brackets {·}, and (iii) the Westfall-Young FWER-adjusted p-value in square brackets [·], as described in more detail in Online Appendix E.

3. Pre-Analysis Plan

This study was preregistered on the AEA Registry and ClinicalTrials.gov, including a pre analysis plan (PAP). We deviate from the PAP in some instances. The main deviations (in our view) are the following. First, we prespecified a regression model that included all interactions of treatments. We soon came to realize we were not well-powered for this analysis and that it would lead to a large number of coefficients and comparisons that would be difficult to present and interpret. We still present the prespecified estimates in Tables II and III but complement them with a simplified but higher-powered specification that pools the two night sleep treatments and omits the interactions between nap and night sleep treatments in Table IV. Second, we had not fully specified our approach to multiple-hypothesis testing and made some changes after receiving comments and discussing with experts. We added the overall index variable to parsimoniously aggregate all outcomes. We also redefined the four families of outcomes (work, well-being, cognition, and preferences rather than work and decision making) and created a summary variable for each family. Other deviations are detailed in Online Appendix Section D.

TABLE II.

Treatment Effects on Sleep

	Night sleep	Time in bed	Sleep efficiency	Nap sleep	24-Hr sleep
	(1)	(2)	(3)	(4)	(5)
Devices+Encouragement only	0.32***	0.53***	−0.54	0.00	0.33***
	(0.08)	(0.09)	(0.66)	(0.00)	(0.08)
Devices+Incentives only	0.49***	0.84***	−1.14	0.00	0.49***
	(0.08)	(0.09)	(0.70)	(0.00)	(0.08)
Nap only	−0.13*	−0.11	−0.72	0.25***	0.09
	(0.08)	(0.09)	(0.72)	(0.01)	(0.08)
Devices+Encouragement and nap	0.21***	0.40***	−1.06	0.24***	0.42***
	(0.07)	(0.08)	(0.66)	(0.01)	(0.08)
Devices+Incentives and nap	0.48***	0.58***	0.90	0.24***	0.70***
	(0.08)	(0.09)	(0.68)	(0.01)	(0.08)
Control mean	5.61	8.07	69.86	0.00	5.61
Control std. dev.	1.20	1.37	11.28	0.00	1.20
Participant-nights	8,454	8,454	8,454	7,191	8,035
Participants	451	451	451	450	451

Open in a new tab

Notes. This table reports the treatment effect of the night sleep and nap interventions on sleep patterns. Each column shows the OLS estimates of equation (1) including dummies for each treatment arm and controlling for the average baseline measure of the dependent variable (ANCOVA), age, sex, and day-in-study and date fixed effects. Each row shows coefficients of treatment cells compared to the control group that receives no sleep-related treatments. Night sleep, time in bed, nap sleep, and 24-hour sleep (columns (1), (2), (4), and (5)) are measured in hours. Sleep efficiency (column (3)) is the ratio of night sleep and time in bed (multiplied by 100 for clarity). 24-hour sleep is the sum of nap sleep in the office and night sleep. Standard errors are clustered at the participant level. Stars next to coefficients reflect unadjusted p-values (* significant at 10%; ** at 5%; *** at 1%).

TABLE III.

Fully Disaggregated Treatment Effects of Night Sleep and Nap Treatments

	Overall	Work				Well-being
	Index	Earnings	Productivity	Labor supply	Output	Index	Physical	Psychological
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
Devices+Encouragement only	0.00	−0.07**	−0.02	−0.06*	−0.07**	0.13***	0.16**	0.02
	(0.07)	(0.03)	(0.03)	(0.04)	(0.03)	(0.05)	(0.07)	(0.09)
Devices+Incentives only	−0.05	−0.07*	−0.02	−0.08**	−0.08**	0.05	0.08	−0.02
	(0.07)	(0.04)	(0.03)	(0.04)	(0.03)	(0.05)	(0.06)	(0.09)
Nap only	0.11	−0.07	−0.01	−0.07*	−0.05	0.18***	0.16**	0.19**
	(0.07)	(0.04)	(0.03)	(0.04)	(0.04)	(0.06)	(0.07)	(0.10)
Devices+Encouragement and nap	0.13*	−0.07*	0.04	−0.15***	−0.04	0.13***	0.11*	0.11
	(0.07)	(0.04)	(0.03)	(0.04)	(0.03)	(0.05)	(0.06)	(0.09)
Devices+Incentives and nap	0.08	−0.07*	0.04	−0.17***	−0.07**	0.10**	0.11*	0.07
	(0.06)	(0.04)	(0.03)	(0.04)	(0.03)	(0.05)	(0.07)	(0.09)
Participants	451	451	451	451	451	452	452	452
Devices+Encouragement only	−0.00	−0.00	0.04	−0.09	−0.05	−0.10	−0.15
	(0.07)	(0.07)	(0.17)	(0.06)	(0.12)	(0.09)	(0.13)
Devices+Incentives only	−0.03	0.04	−0.12	−0.04	0.11	−0.16*	−0.06
	(0.08)	(0.07)	(0.19)	(0.07)	(0.12)	(0.08)	(0.13)
Nap only	0.09	0.07	0.15	−0.01	0.12	−0.09	−0.01
	(0.07)	(0.07)	(0.17)	(0.07)	(0.12)	(0.09)	(0.14)
Devices+Encouragement and nap	0.05	0.13*	0.03	0.09	0.23*	0.03	0.03
	(0.08)	(0.07)	(0.18)	(0.07)	(0.12)	(0.10)	(0.13)
Devices+Incentives and nap	0.14**	0.09	0.32**	0.01	0.11	−0.10	0.01
	(0.07)	(0.07)	(0.16)	(0.07)	(0.12)	(0.09)	(0.14)
Participants	452	452	429	452	452	415	415

Open in a new tab

Notes. This table shows the treatment effects of the five fully disaggregated treatment arms on the overall index as well as the four families of outcomes: work, well-being, cognition, and preferences. Each row shows coefficients of treatment cells compared to the control group that receives no sleep-related treatments. All work-related regressions are conducted at the participant-day level using equation (1). All other regressions are at the participant level using equation (2). The overall index (column (1)) aggregates across the four family-level outcomes. The work outcomes include data entry earnings (the summary variable for the work family, column (2)); productivity (output per hour typing, column (3)); active typing time (column (4)); and output (column (5)). Well-being outcomes include an overall index (column (6)) of the two broad measures of well-being: a physical well-being index (column (7)) and a mental well-being index (column (8)), as described in Section III.B. Cognition measures include an overall index (column (9)) of two measures: lab measurements of attentiveness, memory, and inhibitory control (column (10)); and attention to piece rates in the data entry task (column (11)). Preference measures include an index (column (12)) of three different categories: time preferences (savings, which additionally control for the surveyor on site, and present bias, column (13)), social preferences (column (14)), and risk preferences (column (15)). All indices are a weighted average of their components, in which the weights take into account the covariance structure of the components (Anderson 2008). All dependent variables are normalized with respect to the control group’s mean and standard deviation. When required, outcomes are flipped so that a positive value aligns with what would be considered a “better” outcome. Standard errors in parentheses are robust to heteroscedasticity and clustered at the participant level when applicable. Stars next to coefficients reflect unadjusted p-values (* significant at 10%; ** at 5%; *** at 1%). Online Appendix Table A.VII shows p-values that take into account multiple-hypothesis corrections.

IV.B. Effects on Sleep

1. Overview

Figure V and Table II show the impacts of our treatments on the different measures of sleep. We find that the two night sleep treatments quickly and substantially increase night sleep by 27 minutes a night on average. Offering short afternoon naps increases daytime sleep by 14 minutes a day. Thus, it is possible to substantially increase sleep in the sleep-deprived population we study through encouragement, incentives, and nap opportunities.

2. Night Sleep Treatments Only

Both night sleep treatments sharply increased sleep, as measured by the actigraphs. On average, individuals who received the Devices + Encouragement and Devices + Incentives treatments only—that is, without also receiving naps—increased their time asleep at night by 19 and 29 minutes a night compared with the control group, respectively (Figure V, Panel A, and Table II, column (1)). Including those who might have additionally received naps, the night sleep treatments increased night sleep by 27 minutes a night on average (Online Appendix Table A.VI). This is a larger gain than typically achieved by sleeping pills and cognitive behavioral therapy for insomnia (Riemann and Perlis, 2009; Trauer et al. 2015).

The increase in sleep was driven entirely by additional time in bed rather than improved sleep efficiency. Both night sleep treatment groups increased their time in bed significantly throughout the treatment period—32 minutes for the encouragement only group and 50 minutes for the incentives only group (Figure V, Panel B, and Table II, column (2)).²⁴ We find no significant changes in sleep efficiency compared to the control group (Figure V, Panel C, and Table II, column (3)), even in the middle of the night when all participants are likely to be in bed (Online Appendix Figure A.Ia).

Increasing time spent in bed is a simple and practical way for our study participants to increase their sleep duration. However, improving sleep efficiency appears to be difficult for participants, even with the aid of the devices and tips surrounding sleep hygiene. Participants faced substantial (implicit) incentives for improvement. For example, a participant in the incentives group who improved their sleep efficiency from 70% to 80% would on average earn an additional Rs. 48—about 20% of average typing earnings—in sleep incentives each night (holding fixed eight hours a night in bed). Yet we saw no changes in sleep efficiency. Improving sleep efficiency may require more substantial interventions, including ways to overcome barriers to sleep—such as mosquitoes, crowding, or psychological distress—that remained unaddressed by our treatments (Online Appendix Figure A.IVa).

3. Nap Treatment Only

The nap intervention was effective at increasing participants’ daytime sleep (Figure V, Panel D, and Table II column (4)). Nearly all participants in the nap treatment (92%) reported falling asleep during their allotted nap time, consistent with actigraph data showing that participants fell asleep in 93% of all nap sessions. The mean actigraph-recorded (unconditional) time asleep during the nap period was 14 minutes, and the median duration was 16 minutes (Online Appendix Figure A.IId).

The “quality” of nap sleep in the office appears to be higher than that of night sleep and naps at home. For instance, sleep efficiency during naps in the office (85%) is higher than efficiency in night sleep (66%) and in naps at home (72%, similar to night sleep), if one excludes in all cases the time taken to first fall asleep.²⁵ The average number of awakenings per minute of sleep achieved is also lower for the office naps. Better sleep quality during naps in the office—compared to both naps and night sleep at home—is consistent with a more comfortable sleep environment in the office.

4. Interactions, Crowd-Out, and Heterogeneity

We find only modest interactions between the sleep treatments in terms of their effects on the various sleep measures. The effect on 24-hour sleep of receiving the encouragement treatment and the nap treatment together is very similar to the sum of the effects of receiving each treatment alone (25 minutes versus 25 minutes, p = .98). The same is largely true for providing incentives and naps together (42 minutes versus 35 minutes, p = .32).

We do find evidence that napping modestly crowds out nighttime sleep: those treated with naps spent seven minutes less in bed at night and slept, on average, eight minutes less per night (Table II, columns (1)–(2)). In contrast, the night sleep treatments do not interfere with naps. Participants randomized to the night sleep treatments did not nap any less when offered a nap (Table II, column (4)). Both treatments increased total time asleep in 24 hours, although naps alone had a substantially smaller and insignificant effect (Table II, column (5)).²⁶

Finally, the effect of the night sleep treatments on sleep quantity and efficiency did not differ significantly by baseline sleep patterns, nor by characteristics such as participants’ gender, age, or baseline earnings (Online Appendix Table A.V). Nor did these factors predict meaningful differences in nap duration for the nap treatment group. The treatments thus seem to have been equally effective at increasing sleep (and leaving efficiency unchanged) for different categories of participants.

IV.C. Overall Effect of Each Treatment on Outcomes

Table III presents the treatment effects for the five combinations of night sleep and nap treatments. Given the large number of outcomes and treatments, we focus on the effects on the overall summary index (column (1)) which parsimoniously and efficiently averages our outcomes.²⁷ Each of the night sleep treatments alone had no effect or a slightly negative (but insignificant) effect on participants: 0.00 std. dev. (std. err. = 0.07) and −0.05 std. dev. (std. err. = 0.07), respectively, for the encouragement only and incentives only groups. In contrast, participants in the nap only treatment experienced positive and marginally significant effects of 0.11 std. dev. (std. err. = 0.07, p = .11). The effect of naps only differs significantly from that of incentives only (p = .02) and suggestively from that of encouragement only (p = .11). Those who received a night sleep treatment in addition to naps had very similar overall effects to those who received naps only: 0.13 and 0.08 std. dev. for the encouragement and nap and incentives and nap groups respectively, compared to 0.11 std. dev. for nap only.

These results provide evidence that naps have an overall positive effect on outcomes, while increases in night sleep do not. However, since each treatment cell has only around 75 participants, this analysis has limited statistical power. The combination of five treatment cells with numerous outcomes also makes discussion of detailed results unwieldy. We therefore turn to a simplified but higher-powered version of this analysis, which pools the two night sleep treatments—which typically have similar and statistically indistinguishable effects —and does not include a separate indicator for the group that receives both the night sleep and nap treatments. The resulting estimates in Table IV should be interpreted as weighted averages of treatment effects in the relevant cells. For instance, in this fully pooled specification, the coefficient on the night sleep treatment is the average effect of being assigned to one of the two night sleep treatments (with equal probability), in a population which either receives naps or does not (with equal probability). In Online Appendix Table A.VIII, we instead pool the two night sleep treatments but include a separate indicator for individuals who received a combination of either night sleep treatment along with the nap treatment.

IV.D. Effect of Night Sleep Treatments

1. Overview

Experts from sleep science and economics predicted that increased night sleep would result in higher work output and labor supply, improved health and attention, increased financial savings, and reduced present bias (Figure IV). In contrast to these predictions and an influential literature in sleep science, we find no effect of the pooled night sleep treatments on any of these outcomes. More generally, we find no positive effects of the night sleep treatments on the four family-level summary variables, or on any of the individual outcomes in our pooled specification (Figure VI and Table IV). Instead, increases in night sleep come at the cost of significantly reduced labor supply and therefore a marginally significant reduction in work output.

Figure VI — Summary of Treatment Effects

This figure summarizes the treatment effects in our study. We plot the point estimates and 90% confidence intervals for the pooled night sleep interventions in Panel A and the nap intervention in Panel B. All outcomes are standardized, that is, we subtract the mean and divide by the standard deviation of the individuals receiving neither the night sleep nor the nap interventions. The coefficients and confidence intervals are based on the estimates and standard errors in Table IV. The comparison group for the nap treatment is the pooled nap control group, that is, participants not assigned to the nap intervention. The outcome variables, described in more detail in Section III.B, are as follows. Overall index: aggregates across all the outcomes in the table. Work: (i) earnings; (ii) productivity; (iii) active typing time; and (iv) output from the data entry task. Well-being: (i) “Well-being Index,” a composite index of the physical and mental well-being indices; (ii) “Physical,” a physical well-being index, a composite of performance in an endline biking task, self-reported illnesses, self-reported pain, self-reported health, and blood pressure; and (iii) “Psychological,” a mental well-being index, a composite of self-reported depression, happiness, life possibility, life satisfaction, and stress. Cognition: (i) “Cognition Index,” composite index of a lab-based and a work-based measure of cognitive function; (ii) “Lab Tasks,” index of lab measures of attention, memory, and inhibitory control; and (iii) “Work Task,” measure of attention to piece rates in the data entry task. Preferences: (i) “Preferences Index,” composite index of time, social, and risk preference indices; (ii) “Time,” index capturing time preferences, including savings and present bias; (iii) “Social,” index representing social preferences; and (iv) “Risk,” index representing risk preferences.

2. Work Outcomes

The night sleep treatments did not cause significant improvements in productivity, labor supply, output, or earnings (Table IV, columns (2)–(5)). Although the night sleep treatment groups were 1.3% more productive than the control group (column (3)), this difference of 0.02 std. dev. (std. err. = 0.02) is not statistically significant even without multiple-hypothesis testing corrections.

The night sleep treatments reduced labor supply by approximately 10 minutes a day (0.08 std. dev. with std. err. = 0.02, Table IV, column (4)) with no change in days worked (Online Appendix Figure A.VII). This effect on labor supply remains significant with p < .01 when correcting for multiple outcomes in the work family. Given the additional time in bed induced by the night sleep treatments, participants have less time available for work and leisure, which comes at the cost of reduced labor supply. Specifically, participants arrive at work six minutes later in the morning, take two minutes more of breaks at work, and leave for home three minutes earlier on average (Online Appendix Table A.X). While perhaps obvious ex post, the opportunity costs of increasing sleep are typically neglected in the sleep literature. Indeed, the mean expert prediction was an increase in labor supply of 7%, which is strongly rejected by the data (p < .001).

The small increase in productivity was not enough to outweigh the reduction in labor supply, leading to a small and marginally significant decrease in earnings and output, respectively (each 0.04 std. dev., std. err. = 0.02).²⁸ This finding is again in contrast to the mean (median) expert prediction of a 12% (7%) increase in output. The discrepancy can be partly explained by experts overestimating the productivity effects of increased sleep, and partly by their mispredicting that more sleep would increase the time allotted to work. Eighty-three percent of experts made point predictions outside of the 95% confidence interval of our estimate of the effect on output.

Our results also contrast with those from natural experiments studying the economic consequences of sleep. Gibson and Shrader (2018) exploit variation in sunset times in the United States and estimate that 8.5 minutes of additional sleep per night increases earnings by 1.1% in the short run. Giuntella and Mazzonna (2019) use time zone border discontinuities in the United States and find that 19 fewer minutes of sleep are associated with 3% lower earnings. Extrapolating these estimates linearly to our experiment would predict 3.5% and 4.3% increases in earnings, respectively, which we firmly reject (p < .01).