Light exposure during sleep is bidirectionally associated with irregular sleep timing: the Multi-Ethnic Study of Atherosclerosis (MESA)

Abstract

Exposure to light at night (LAN) may influence sleep timing and regularity. Here, we test whether greater light exposure during sleep (LEDS) is bidirectionally associated with greater irregularity in sleep onset timing in a large cohort of older adults in cross-sectional and short-term longitudinal (days) analyses. Light exposure and activity patterns, measured via wrist-worn actigraphy (ActiWatch Spectrum), were analyzed in 1,933 participants with 6+ valid days of data in the Multi-Ethnic Study of Atherosclerosis (MESA) Exam 5 Sleep Study. Summary measures of LEDS averaged across nights were evaluated in linear and logistic regression analyses to test the association with standard deviation (SD) in sleep onset timing (continuous variable) and irregular sleep onset timing (SD>90 minutes, binary). Night-to-night associations between LEDS and absolute differences in nightly sleep onset timing were also evaluated with distributed lag non-linear models and mixed models. In between-individual linear and logistic models adjusted for demographic, health, and seasonal factors, every 5-lux unit increase in LEDS was associated with a 7.8-minute increase in sleep onset SD (β=0.13 hours, 95%CI:0.09–0.17) and 32% greater odds (OR=1.32, 95%CI:1.17–1.50) of irregular sleep onset. In within-individual night-to-night mixed model analyses, every 5-lux unit increase in LEDS the night prior was associated with a 2.2-minute greater deviation of sleep onset the next night (β=0.036 hours, p<0.05). Conversely, every 1-hour increase in sleep deviation was associated with a 0.35-lux increase in future LEDS (β=0.348 lux, p<0.05). LEDS was associated with greater irregularity in sleep onset in between-individual analyses and subsequent deviation in sleep timing in within-individual analyses, supporting a role for LEDS in irregular sleep onset timing. Greater deviation in sleep onset was also associated with greater future LEDS, suggesting a bidirectional relationship. Maintaining a dark sleeping environment and preventing LEDS may promote sleep regularity and following a regular sleep schedule may limit LEDS.

Graphical Abstract

1. Introduction

Sleep timing is regulated by two primary biological processes: 1) the homeostatic sleep drive and 2) a circadian rhythm in sleep propensity¹. For the circadian regulation of sleep timing, light is the most important zeitgeber, or “time giver”², and in this way, light can contribute to stability or irregularity in sleep-wake timing. Controlled laboratory studies have shown that light exposure shifts circadian rhythms in an intensity-, wavelength-, light history-, and duration-dependent manner to advance or delay the phase of circadian rhythms, dependent on the timing of exposure^2,3. These studies have shown that the human circadian system is most sensitive to the phase-shifting effects of light during the “biological” night^3,4, the time just preceding, during, and following the habitual sleep episode. Light exposure in the earlier part of the biological night delays (moves later), whereas light exposure in the latter half of the night advances (moves earlier), the circadian pacemaker³, a cluster of neurons in the suprachiasmatic nucleus of the hypothalamus that autonomously keeps time and generates timing signals for other body regions⁵. Individuals may differ in sensitivity to the phase-shifting effects of light at night (LAN)⁶. Whether light advances or delays sleep timing, and by how much, depends on the endogenous circadian phase during light exposure². In addition to the phase-shifting effects of light, light also has acute negative effects on sleep propensity, in part through suppression of melatonin⁷.

Our modern light environment is drastically different from the one in which humans evolved, where sunlight provided light during the day and (much dimmer) moonlight or fires were the only nighttime light sources. These natural LAN exposures had a limited impact on the circadian system in comparison with modern electric lighting⁸. In comparison, people now have access to inexpensive, abundant, and bright light at any time of day or night with modern electrical lighting^9,10. Because the phase-shifting effects of light on the human circadian timing system are strongest around the biological night, exposure to LAN in the home environment may influence the circadian system. Therefore, in addition to the acute effects that LAN can have on sleep, LAN may also influence sleep timing. However, epidemiological studies of sleep in naturalistic settings may not have measures of circadian phase that could be used to model the phase-shifting effects of light. In such cases, measuring light exposure during sleep (LEDS) could provide an alternative way to capture the effects of light exposure during a biologically sensitive window.

A particularly important dimension of sleep timing is its nightly regularity or variability. Irregular sleep timing, defined as variability in night-to-night sleep timing, may represent a type of circadian misalignment common in the general population¹¹. Greater irregularity in sleep timing has been linked to mechanisms for chronic disease¹², such as increased inflammatory biomarkers¹³, and metabolic syndrome¹⁴. Experimental sleep restriction studies with and without circadian disruption further support that circadian misalignment itself has adverse cardiometabolic consequences above and beyond any effects of sleep restriction^15–18. In particular, timing and irregularity of sleep onset time have been associated with increased cardiovascular disease risk^19,20 and abnormal metabolic markers¹⁴.

LAN and LEDS are potentially modifiable lifestyle risk factors. However, teasing apart the effects of light exposure on the sleep episode from other effects in the home environment may be difficult due to reverse causation. For example, factors associated with later sleep timing (e.g., evening chronotypes) may lead to more night activity and higher pre-sleep light exposure. Therefore, the evaluation of LEDS rather than pre-sleep light may limit this bias. If LEDS is greater in people with irregular sleep timing, reducing LEDS may present an opportunity to establish more regular sleep patterns and improve downstream health outcomes. Because the interplay of environmental light exposure and sleep timing regularity in the general population has not been well characterized²¹, here, we tested the hypothesis that greater LEDS is bidirectionally associated with less regularity in sleep onset timing in a large, multiethnic cohort studied in their usual home environments.

2. Methods

2.1. Study population

This analysis utilized data from the Multi-Ethnic Study of Atherosclerosis (MESA), designed to study subclinical and clinical cardiovascular disease in a multiracial, multiethnic sample of adults 45–84 years of age at study initiation. The baseline Exam 1 study (2000–2002) enrolled 6,814 participants at six U.S. study sites (Los Angeles, CA; St. Paul, MN; Chicago, IL; Forsyth County, NC; Baltimore, MD; and Manhattan and Southern Bronx, NY)²². To be initially included in the cohort, participants had to be free of known cardiovascular disease or other health conditions that would prevent follow up examination. Subsequent examinations of the cohort were completed over approximately 20 years. Between 2010–2013, actigraphy data that included both sleep-wake and light measurements were collected as part of the MESA Exam 5 Sleep Ancillary study (n=2,261). Of the 2,159 participants with actigraphy and light records, we excluded participants with <6 valid nights. The analysis included demographic and health data as covariates from study questionnaires. The institutional review board at each study site approved the study, and all participants provided written informed consent.

2.2. Actigraphy data processing and derivation of light exposure during sleep (LEDS)

Objective actigraphy and white light data were collected in 30-second epochs with ActiWatch Spectrum actigraphy devices (Philips Respironics, Murrysville, PA) worn on the non-dominant wrist. The ActiWatch Spectrum has a light sensor on the face of the device which captures measures of light illuminance (measured in lux; illuminance intensity measurement range 0.1–35,000+ lux^23,24) and light irradiance (measured in μW/cm²; wavelength spectral range 400 – 700nm). These light data measures are output in the same fashion as activity count data, with values of each provided for every 30-second epoch. The white light illuminance (lux) data were analyzed for all analyses except for an exploratory analysis of blue light LEDS, which used blue light (400–500nm) irradiance (μW/cm²) data. Sleep actigraphy data were scored by trained personnel at Brigham and Women’s Hospital Sleep Reading Center²⁵. Sleep onset for the main sleep episode and nap times were annotated based on hierarchical methods utilizing ActiWatch event marker, participant sleep journal, light levels, and/or activity count data, as previously described²⁶. Measures of light exposure, sleep timing, behavioral activity, and sleep duration were derived for participants with at least 6 valid nights (main sleep episode) of data from the first 7 consecutive days of actigraphy data^27,28. This 6-night threshold was chosen to balance having a sufficient number of days for analysis and to investigate lagged associations while still maintaining adequate sample size. Sleep-wake was determined using the Philips Actiware 5.59 algorithm with sleep onset defined as a 5-minute period of immobility (immobility as <2 activity counts per 30-second epoch) and the intermediate (40 counts per epoch) setting for wake detection. The Phillips algorithm calculated sleep offset as the last epoch of sleep within the sleep interval, which was then compared with sleep journal data by the trained scorer.

Timeseries data for each participant were processed to define a person-specific “day” using the sleep onset timing for the first sleep episode, rather than midnight, as the day start time. Consequently, the first sleep episode began at the beginning of the person-specific day. All daily measures described below refer to the person-specific day. Using this definition of a day, average behavioral activity, average white light (lux) exposure during sleep (LEDS), and the sleep episode duration were calculated. If there were >10 cumulative minutes of missing white light or activity data during the main sleep episode, or if the main sleep episode had an extreme duration of either <2 hours or >14 hours (considered to be outside the standard norm of sleep duration for adults), the night was considered invalid and that night’s LEDS and sleep duration were set to missing. White light lux values within the main sleep episode (from sleep onset to sleep offset) were averaged for a mean nightly LEDS value. For the exploratory analysis comparing light exposure of regular and irregular sleepers, average white lux values in 6-hour increments (midnight-6AM, 6AM-noon, noon-6PM, and 6PM-midnight) were derived. Additionally, participants with an average LEDS value (across all nights) higher than the 99^th percentile were excluded.

2.3. Derivation of sleep regularity outcome variables

The outcome for the analyses of sleep regularity was SD of sleep onset timing across nights. To calculate sleep onset timing SD, the circular standard deviation (SD)²⁹ of the timestamps across each night’s sleep onset was calculated; circular SD was calculated rather than linear SD because time is a circular variable. Sleep onset, rather than other markers of sleep timing such as sleep midpoint or sleep offset, was chosen as the outcome because of its relevance as a marker of the circadian gating of the wake-maintenance zone and propensity for sleep³⁰ and ease of interpretability. Additionally, the temporality of measures ensures that sleep onset occurs prior to the LEDS measure for that sleep episode, while still allowing LEDS to influence sleep onset timing. The outcomes for the night-to-night distributed lag nonlinear model (DLNM) and mixed model analyses were the absolute circular difference or deviation in sleep onset timing from night to night. The absolute difference was used rather than a positive or negative value because the circadian phase of participants was not measured. For example, LEDS exposure could shift subsequent sleep timing to occur later (phase delay) or earlier (phase advance), depending on the timing of light exposure relative to circadian phase.

2.4. Covariates

Covariates were chosen after reviewing the literature. Race and ethnicity are social constructs; the self-reported MESA-defined groups of race/ethnicity (see categories below) were included as a covariate to attempt to account for the influence of racism and bias on sleep. The term “gender” is used here to reflect the wording used in MESA questionnaires, after which participants were asked to indicate “Male” or “Female”, but we acknowledge that this variable may reflect sex and/or not appropriately capture gender identity. Other covariates included: age (years) during the Sleep Ancillary Exam, self-reported gender from Exam 1 (male/female), self-reported race and ethnicity from Exam 1 using MESA-defined groups (Black, Chinese, Hispanic/Latino, or White), Exam 5 employment status (“no” (ref) or “yes”), partner status (married or living with a partner (ref) vs other), and federal poverty level status, a binary variable derived from self-reported income and household size to reflect 2010–2012 federal poverty level status (above or below federal poverty level). Clinical site location, daylength (duration in hours), and their interaction term were also included; approximate daylength was derived from sunrise and sunset times calculated from the coordinates of the study sites and dates of actigraphy measurement using the “suncalc” package³¹. Health-behavior related covariates included: smoking status at Exam 5 (current=”yes”), chronotype (summary score, modeled as continuous), waist-to-hip ratio (WHR³², continuous), and average behavioral activity (continuous). Self-reported chronotype was measured using the modified Horne-Ostberg Morningness-Eveningness Questionnaire score³³ (higher scores indicate greater morningness) during the ancillary sleep study during Exam 5. Waist and hip circumference were measured during Exam 5. Daily average behavioral activity counts (across both wake and sleep episodes) were calculated from accelerometer activity counts from wrist actigraphs and averaged across person-specific days as an objective measure of behavioral activity (continuous); if >8 cumulative hours of actigraphy data were missing across the day, that day’s average activity was set to missing. Sleep duration was calculated as the nightly duration (hours) of the main sleep episode interval and sleep fragmentation index was calculated as the proportion of immobile to mobile epochs²⁶. For the sensitivity analysis excluding shift workers, work schedules were ascertained during Exam 5 with the question “Which of the following best describes your usual work schedule?” (options: Day/Afternoon/Night/Split/Irregular/Rotating/Don’t work) and participants who responded with a shift other than “Day” or “Don’t work” were considered shift workers. For the sensitivity analysis excluding participants with insomnia, people who responded “yes” during Exam 5 to the question “Have you ever been told by a doctor that you have any of the following?” with “Insomnia?” listed (options: No/Yes) and/or people with a Women’s Health Initiative Insomnia Rating (WHIIR) score of 9 or higher³⁴ were excluded.

2.5. Statistical analyses

This analysis used both cross-sectional and short-term (days) longitudinal study designs. In the primary between-individual linear and logistic analyses, sleep onset SD was analyzed as a continuous and dichotomous outcome with LEDS (averaged across nights) as the exposure of interest while adjusting for covariates in Models 1–5. Model 1 adjusted for age, gender, and race and ethnicity. Model 2 additionally adjusted for employment status, partner status, and federal poverty level status. Model 3 additionally adjusted for smoking, chronotype, WHR, and average behavioral activity. Model 4 additionally adjusted for the site, daylength (hours), and their interaction; results from Model 4 are presented as the main results in the text. Lastly, average sleep duration and sleep fragmentation were identified as possible confounders (which are also possible colliders) and modelled as continuous covariates in Model 5. Sleep duration and sleep fragmentation index were modelled as averaged values in the logistic and lag analyses and as nightly measures in the mixed model analyses.

LEDS was modeled both as a continuous and categorical variable (tertiles), with the lower limit of the third tertile aligning closely with the 1 lux benchmark for sleep environment recommendations³⁵. Sleep onset SD had a right-skewed distribution and was also modeled as a continuous variable after Box-Cox transformation using the “MASS” package³⁶ (Supplemental Material). Sleep onset SD was also modeled as a dichotomized outcome in logistic regression models, with “regular” sleep timing defined as individuals with sleep onset SD less than or equal to 90 minutes and “irregular” sleep timing as individuals with sleep onset SD greater than 90 minutes, in line with prior studies^14,19. This dichotomized category was also used in describing and comparing average white light exposure patterns in “regular” and “irregular” sleepers by 6-hour time blocks; group differences were tested using the Kruskal-Wallis rank sum test.

Because the effect of LEDS on sleep timing may have delayed or non-linear effects, secondary analyses modeled the night-to-night lagged associations of LEDS (e.g., LEDS night 5=lag0, LEDS night 4=lag-1, LEDS night 3=lag-2, LEDS night 2=lag-3, LEDS night 1=lag-4) on absolute deviation in sleep onset timing from sleep onset time on the 5^th night of measurement to sleep onset time on the 6^th night of measurement (circular difference) to identify the optimal window for the exposure-outcome relationship (Figure 1). Lags are denoted with a “-“ to represent going back in time in relation to the modelled outcome. DLNM models were constructed using the “dlnm” package³⁷ (Supplemental Material). Deviation in sleep onset timing was also modeled as the predictor and LEDS (night 6) as the outcome. Associations were estimated with generalized additive models with penalized splines³⁸. The exposure-response function was modeled as a natural spline and the lag-response function as a penalized spline with internal knots selected based on smallest AIC (details in Supplemental Material). While single-lag models were used in the primary analysis, we also conducted an exploratory moving average analysis³⁹ of LEDS (Supplemental Material).

Figure 1.

Depiction of example data timeline and study design. Horizontal yellow bars indicate light exposure and horizontal blue bars indicate the sleep episode; black boxes indicate the window of the sleep episode during which nightly average light exposure during sleep (LEDS) was calculated. Rather than using (A) midnight to demarcate daily measures, this analysis used (B) sleep onset timing to demarcate daily measures. (B) Logistic regression analyses analyzed data averaged across days, whereas (C) the night-to-night analysis distributed lag non-linear model (DLNM) analyzed LEDS on the 5^th night (lag0), the 4^th night (lag-1), the 3^rd night (lag-2), the 2^nd night (lag3-) and the 1^st night (lag-4) as the exposure and the absolute difference in sleep onset timing between the 5^th and 6^th night as the outcome. Likewise, (D) another DLNM analysis modelled the absolute difference in sleep onset timing between the 5^th and 6^th nights (lag0), the 4^th and 5^th nights (lag-1), the 3^rd and 4^th nights (lag-2), and 2^nd and 3^rd nights (lag-3), and the 1^st and 2^nd nights (lag-4) as the exposure and LEDS on the 6^th night as the outcome. Night to night measures were further analyzed in mixed models using the lag identified in the DLNM.

Using the lag period(s) identified with DLNM modeling, night-to-night within-individual associations between LEDS and irregularity in sleep onset timing were further modeled using mixed linear regression modeling. The absolute deviation in sleep timing from night 1 to night 2, night 2 to night 3, night 3 to night 4, night 4 to night 5, and night 5 to night 6 were modeled as the repeated-measure dependent variable, while LEDS during night 1, night 2, night 3, night 4, and night 5 were modeled as the repeated-measure exposure. In Model 5, nightly sleep duration and sleep fragmentation index (nightly measures from nights 1–5 instead of averaged values) were included as fixed effects. To explore bidirectionality, deviation in sleep onset (nights 1–2, 2–3, 3–4, 4–5, 5–6) was also modeled as the exposure and LEDS (nights 2, 3, 4, 5, 6) modeled as the dependent variable. The covariance structure was specified as first-order autoregressive structure with heterogenous variances (correlation=corAR1(form=~1|ID)) and ID modeled as a random effect to allow for variability between participants using the “nlme” package⁴⁰ in R. A sensitivity analysis excluding shift workers and a sensitivity analysis excluding people with insomnia were also conducted for the logistic and mixed model regressions. An exploratory mixed-model analysis stratified by gender and an exploratory mixed-model analysis of blue light LEDS were also conducted. All statistical analyses were performed in R version 4.1.1. and results considered statistically significant if p-value <0.05.

3. Results

3.1. Demographic characteristics

After excluding 206 participants with <6 valid nights of data, 1,933 MESA participants were included in the analysis (Supplemental Table 1). A representation of the study design is provided in Figure 1. On average, participants were 69 years of age, 55% female, and the majority were retired or unemployed. The average sleep onset timing among the sample was approximately 11:35PM (median=10:19PM, sd=1.37 hours). The range in LEDS at the 1^st, 2^nd, and 3^rd tertiles were 0–0.33, 0.33–0.95, and 0.95–39.4 lux, respectively. Approximately 43% of the sample had a sleep onset variability (SD) of 1 hour or more and 31.5% had an average LEDS value of 1 lux or brighter (Supplemental Table 2). Overall, participants in the highest tertile of LEDS exposure (≥0.95 lux) were older, had a lower MEQ score, and differed by daylength and site (Table 1). Greater LEDS exposure was also more common among Black participants and less common among Hispanic/Latino participants (Table 1). Visual inspection of averaged light levels stratified by sleep regularity group suggested greater LEDS among irregular sleepers (>90 minutes SD) compared to regular sleepers (Figure 2). When light levels averaged by clock time were compared between these groups, irregular sleepers had greater light exposure from midnight to 6AM compared to regular sleepers (average white light 6.67 lux vs. 2.55 lux, p<0.001; Supplemental Table 3), but otherwise there were no other statistically significant differences by group.

Table 1.

Descriptive statistics of the study sample, stratified by LEDS tertiles.

	LEDS Tertiles
Variables	0–0.33 lux (n=645)	0.33–0.95 lux (n=644)	>0.95 lux (n=644)
Age, years (mean (SD))	68.77 (9.09)	68.95 (8.81)	70.76 (9.39)
WHR (mean (SD))	0.93 (0.08)	0.94 (0.08)	0.94 (0.09)
Gender (N (% Male))	282 (43.7)	299 (46.4)	297 (46.1)
Race/ethnicity (N (%))
Chinese	68 (10.5)	72 (11.2)	79 (12.3)
Hispanic/Latino	182 (28.2)	160 (24.8)	118 (18.3)
Black	131 (20.3)	178 (27.6)	211 (32.8)
White	264 (40.9)	234 (36.3)	236 (36.6)
Currently employed (N (% Yes))	154 (24.0)	153 (23.9)	155 (24.2)
Income below federal poverty (N (% Yes))	45 (7.2)	46 (7.4)	41 (6.6)
Partner status, married or currently living with partner (N (% Yes))	389 (61.3)	402 (63.3)	368 (58.0)
Current smoking (N (% Yes))	43 (6.7)	46 (7.2)	36 (5.6)
Chronotype MEQ score (mean (SD))	17.60 (3.48)	17.58 (3.49)	16.61 (3.77)
Daylength (hours)	11.92 (1.93)	12.32 (1.93)	12.67 (1.89)
Clinical Site (N (%))
Wake Forest University	108 (16.7)	105 (16.3)	95 (14.8)
Columbia University	126 (19.5)	92 (14.3)	125 (19.4)
Johns Hopkins University	72 (11.2)	105 (16.3)	106 (16.5)
University of Minnesota	160 (24.8)	125 (19.4)	79 (12.3)
Northwestern University	90 (14.0)	113 (17.5)	133 (20.7)
University of California, Los Angeles	89 (13.8)	104 (16.1)	106 (16.5)
Shiftwork (N (% Yes))	75 (11.6)	77 (12.0)	83 (12.9)
Insomnia (N (% Yes))	236 (36.6)	237 (36.8)	244 (37.9)
Average LEDS, mean (SD)	0.14 (0.10)	0.58 (0.17)	4.65 (6.23)

Figure 2.

Average daily log10-lux illuminance values in MESA by time of day, stratified by the regular sleep onset timing group (n=1,540 participants) and the irregular sleep onset timing group (n=393 participants).

3.2. LEDS averaged across days is associated with irregular sleep onset timing

In the primary linear and logistic regression between-individual analyses, greater average LEDS was associated with greater irregularity in sleep timing (Tables 2–3). When sleep timing variability was modeled as a continuous outcome, every 1-lux and 5-lux unit increase was associated with a 1.8-minute (β=0.03 hours, 95% CI: 0.02, 0.03) and 7.8-minute (β=0.13 hours, 95% CI: 0.09, 0.17) increase in sleep onset SD, respectively (Table 2, Model 4 results). When analyzed as tertiles, the highest LEDS tertile (≥0.95 lux) was associated with a 13.8-minute increase in sleep onset SD (β=0.23 hours, 95% CI: 0.15, 0.31). Results from linear models with Box-Cox transformed sleep onset SD were similar (Supplemental Figure 1, Supplemental Table 4). Likewise, when sleep timing variability was dichotomized and analyzed in logistic regression, every 1-lux and 5-lux unit increase was associated with approximately 6% (OR=1.06, 95% CI: 1.03, 1.08) and 32% (OR=1.32, 95% CI: 1.17, 1.50) higher odds of irregular (>90 minutes SD) sleep timing, respectively (Supplemental Table 5, Model 4 results). This association decreased to approximately 5% and 29% higher odds of irregular sleep timing when average sleep duration and the sleep fragmentation index were included. When analyzing LEDS as tertiles, those in the highest exposure group had 64% (OR=1.64, 95% CI: 1.23, 2.21) higher odds of irregular sleep timing compared to the lowest tertile (Supplemental Table 5).

Table 2.

Linear regression results for light exposure during sleep (LEDS, continuous or tertiles) with sleep onset timing variability (SD, hours) as the outcome.

	Crude [95% CI]	Model 1 [95% CI]	Model 2 [95% CI]	Model 3 [95% CI]	Model 4 [95% CI]	Model 5 [95% CI]
LEDS as continuous*:
LEDS (per 1-unit lux)	0.03 (0.02–0.04)	0.03 (0.02–0.04)	0.03 (0.02–0.03)	0.03 (0.02–0.03)	0.03 (0.02–0.03)	0.02 (0.02–0.03)
LEDS as tertiles*:
LEDS T1 (0–0.33 lux)	(ref)	(ref)	(ref)	(ref)	(ref)	(ref)
LEDS T2 (0.33–0.95 lux)	0.01 (−0.07–0.08)	−0.01 (−0.09–0.06)	−0.02 (−0.09–0.06)	−0.01 (−0.08–0.07)	0.00 (−0.08–0.07)	0.02 (−0.05–0.09)
LEDS T3 (0.95–39.4 lux)	0.25 (0.17–0.32)	0.23 (0.16–0.3)	0.23 (0.16–0.31)	0.22 (0.15–0.3)	0.23 (0.15–0.31)	0.21 (0.14–0.29)

^*Sleep onset SD (hours) as outcome

Crude and adjusted linear regression model estimates with 95% confidence intervals are presented, with p<0.05 in bold. Model 1 adjusted for age, gender, and race/ethnicity; Model 2 adjusted for all the covariates included in Model 1 in addition to poverty, employment status, and partner status; Model 3 adjusted for all the covariates included in Model 2 in addition to smoking, chronotype, waist to hip ratio, and average behavioral activity; Model 4 adjusted for all the covariates included in Model 3 in addition to daylength, Exam 5 site, and their interaction; Model 5 adjusted for all the covariates in Model 4 in addition to average sleep episode duration, and average sleep fragmentation index.

Results from Model 4 are presented as the main results in the text.

Table 3.

Results of mixed model regression for night-to-night associations between LEDS (exposure) and absolute deviation in sleep onset (outcome).

		Deviation in Sleep Onset (hours)
	LEDS	β	SE	t	p-value
Crude	1-lux units	0.008	0.002	4.347	1.40E-05
Model 1	1-lux units	0.008	0.002	4.327	1.53E-05
Model 2	1-lux units	0.008	0.002	4.344	1.42E-05
Model 3	1-lux units	0.007	0.002	4.157	3.26E-05
Model 4	1-lux units	0.007	0.002	4.113	3.95E-05
Model 5	1-lux units	0.008	0.002	4.395	1.13E-05

Results from mixed model regression. Model 1 adjusted for age, gender, and race/ethnicity; Model 2 adjusted for all the covariates included in Model 1 in addition to poverty, employment status, and partner status; Model 3 adjusted for all the covariates included in Model 2 in addition to smoking, chronotype, waist to hip ratio, and average behavioral activity; Model 4 adjusted for all the covariates included in Model 3 in addition to daylength, Exam 5 site, and their interaction; Model 5 adjusted for all the covariates in Model 4 in addition to nightly sleep episode duration and nightly sleep fragmentation index.

Results from Model 4 are presented as the main results in the text.

3.3. LEDS up to two nights prior is associated with future sleep onset timing deviation in lagged analyses

In the secondary analyses, to identify the appropriate lag for the exposure-response relationship, we investigated night-to-night lagged associations between LEDS (predictor) and deviation in sleep onset timing (night 5 to night 6 as the outcome, Figure 1C). The single-night lag-response curves for predicted LEDS on sleep timing are shown in Supplemental Figure 2, with a positive association between greater lux and greater deviation in sleep timing. Increasing LEDS from 0 to 0.5 lux was predicted to increase sleep deviation by 0.6 minutes at lag0 (β=0.01 hours; 95% CI:0.002,0.018; Supplemental Figure 2A) and 0.3 minutes at lag-1 (β=0.005 hours; 95% CI:0.0004, 0.01), but not at other lags. Likewise, increasing LEDS from 0 to 1 lux was predicted to increase deviation in sleep timing by 1.1 minutes at lag0 (β=0.019 hours; 95% CI:0.003,0.036; Supplemental Figure 2B) and 0.6 minutes at lag-1 (β=0.01 hours; 95% CI:0.0007, 0.02). Greater increases from 0 to 3 lux or 0 to 10 lux were associated with deviations of 3.4 and 10.7 minutes at lag0, respectively (β=0.057 hours; 95% CI:0.009, 0.105; β=0.178 hours; 95% CI:0.031, 0.325; Supplemental Figure 2C–D), and 1.8 and 5.5 minutes at lag-1 (β=0.03 hours; 95% CI:0.002, 0.06; β=0.09 hours; 95% CI:0.007, 0.177). However, other preceding timepoints (lag-2 to lag-4) were not associated. The exposure-outcome response surface had a saddle shape, with the greatest slope at lag0 (Supplemental Figure 2E). Supplemental Figure 2F shows the cumulative influence of LEDS on sleep deviation (across lags), with confidence bands becoming progressively wider as lux increases. The results of the exploratory moving average lag analysis are provided in Supplemental Table 6.

3.4. Night-to-night within-individual analysis suggests LEDS affects future sleep onset timing deviation

Repeated nightly measures of LEDS and absolute sleep onset deviation were next analyzed within individuals with mixed linear models using a single-night lag window. LEDS was significantly associated with deviation in sleep onset the following night, with or without adjusting for that night’s sleep fragmentation and sleep duration. Every 1-lux and 5-lux increase in nightly LEDS was associated with approximately 0.42 minutes (β=0.007 hours, p<0.05; Model 4) and 2.2 minutes, respectively, increase in sleep deviation the following night (Table 3). Blue LEDS was also associated with future deviation in sleep timing in an exploratory analysis of blue light (Supplemental Table 7). An exploratory analysis stratified by gender was also performed, but the interaction between gender and LEDS on future sleep deviation was not statistically significant (Supplemental Table 8).

3.5. Large deviations in sleep timing the night before is associated with LEDS the following night in lagged analyses

A secondary lagged analysis modelled the lagged effects of sleep onset timing deviation (predictor, night 5 to 6=lag0, night 4 to 5=lag-1, etc.) on LEDS (outcome, night 6; Figure 1D). The single-night lag-response curves did not show associations between sleep deviation and LEDS at lag0 until sleep timing deviation became 4 hours or greater (Supplemental Figures 3A–C, Supplemental Table 2). A large deviation of 4 or 5 hours was predicted to increase LEDS (β=1.059 lux; 95% CI:0.248, 1.870; β=1.598 lux; 95% CI:0.704, 2.492; Supplemental Figure 3D). However, deviations of 1–3 hours showed some associations at longer lags (lag-2 to lag-4); for example, a 2 hour deviation a few nights prior (lag-2 to lag-4), but not the night or two before (lag0 to lag-1), was predicted to increase night 6 LEDS by 0.46 lux (lag-2), 0.55 lux (lag-3), and 0.64 lux (lag-4), respectively (β=0.464 lux; 95% CI:0.186, 0.743; β=0.552 lux; 95% CI:0.163, 0.941; β=0.640 lux; 95% CI:0.032, 1.247). These inconsistencies may be due to underlying variation in the data, as shown in the curve of the exposure-outcome response surface (Supplemental Figure 3E). Supplemental Figure 3F shows the cumulative influence of sleep deviation on LEDS, with increasing effects and confidence bands widening after approximately 6 hours deviation.

3.6. Night-to-night within-individual analysis also suggests sleep onset timing deviation affects future LEDS

To evaluate whether there may be bidirectional effects whereby deviation in sleep onset timing may also affect LEDS, repeated nightly measures of absolute sleep onset deviation were next analyzed within individuals with mixed linear models using the same lag window (lag0) as the prior mixed model analysis. Deviation in sleep onset was significantly associated with LEDS the following night, with or without adjusting for sleep fragmentation and sleep duration. Every 1-hour increase in sleep timing deviation was associated with approximately 0.35 greater LEDS (p<0.05; Model 4) the following night (Table 4). Results stratified by gender are also provided, but the interaction between gender and LEDS in the primary (not stratified by gender) mixed model analysis was not statistically significant (Supplemental Table 9).

Table 4.

Results of mixed model regression for night-to-night associations between absolute deviation in sleep onset (exposure) and LEDS (outcome).

		LEDS (lux)
	Deviation	β	SE	t	p-value
Crude	1-hour units	0.369	0.063	5.888	4.09E-09
Model 1	1-hour units	0.364	0.063	5.788	7.44E-09
Model 2	1-hour units	0.364	0.063	5.776	7.97E-09
Model 3	1-hour units	0.351	0.063	5.560	2.79E-08
Model 4	1-hour units	0.348	0.063	5.526	3.39E-08
Model 5	1-hour units	0.349	0.063	5.539	3.15E-08

Results from mixed model regression. Model 1 adjusted for age, gender, and race/ethnicity; Model 2 adjusted for all the covariates included in Model 1 in addition to poverty, employment status, and partner status; Model 3 adjusted for all the covariates included in Model 2 in addition to smoking, chronotype, waist to hip ratio, and average behavioral activity; Model 4 adjusted for all the covariates included in Model 3 in addition to daylength, Exam 5 site, and their interaction; Model 5 adjusted for all the covariates in Model 4 in addition to nightly sleep episode duration and nightly sleep fragmentation index.

Results from Model 4 are presented as the main results in the text.

3.7. Excluding shift workers or participants with insomnia does not abrogate the association of LEDS or sleep timing deviation

To evaluate the influence of shift work on the measured outcomes, we conducted sensitivity analyses excluding the 235 participants who reported working an afternoon, night, split, irregular/on-call, or rotating shift. The timeseries plot was similar, although the gap between regular and irregular sleepers became smaller (Supplemental Figure 4). The logistic regression results remained largely the same, if slightly attenuated in a few of the models (Supplemental Table 10). Regression estimates were also largely unchanged in the night-to-night mixed model regression results (Supplemental Table 11).

Because sleep onset timing may be less accurately determined and the frequency of wake episodes during sleep may be greater in participants with insomnia⁴¹, we also conducted a sensitivity analysis excluding the 717 participants with a history of insomnia and/or with a WHIIR score ≥ 9. Timeseries plots of the log10-transformed averaged lux values did not substantially change (Supplemental Figure 5). Overall, logistic regression results had slightly greater estimates for the effects of LEDS on sleep irregularity among the 1,116 participants without insomnia (Supplemental Table 12). However, regression estimates were somewhat attenuated in the night-to-night mixed model regression results, decreasing from an estimate of 0.007 hours to 0.006 hours (β=0.006 hours, p<0.05; Model 4; Supplemental Table 13).

In the night-to-night mixed model regression analyses modeling the effect of sleep onset deviation on LEDS the following night, excluding shift workers led to attenuated effect estimates, decreasing from 0.35 to 0.20 (p<0.05; Model 4; Supplemental Table 14). However, excluding participants with insomnia led to greater effect estimates, increasing from 0.35 to 0.54 (p<0.05; Model 4; Supplemental Table 15).

4. Discussion

In this large sample of adults from 6 geographically diverse communities across the U.S., studied with multi-day assessments of sleep-wake patterns and ambient light exposure, we found a bidirectional positive association between greater light exposure during the main sleep episode (LEDS) and greater irregularity in sleep onset timing. While recent lighting guidelines for chronobiological health recommend a sleeping environment that is darker than 1 lux illuminance³⁵, 31.5% of our sample of older adults had an average illuminance during sleep above this threshold, suggesting the pervasiveness of suboptimal light exposure during sleep. Averaged across nights, greater LEDS was associated with greater sleep onset irregularity. In repeated measures night-to-night analyses, greater LEDS was associated with greater deviation in sleep onset timing the next night. Likewise, greater deviation in sleep onset timing was associated with greater future LEDS. Sleep onset timing appeared to have a greater magnitude of effect on future LEDS than the reverse, but this may be due to limitations in exposure measurement and/or study sample characteristics. Overall, our results suggest a bidirectional association between LEDS, a form of LAN during the sleep episode, and sleep irregularity, suggesting the importance of both maintaining a dark environment during sleep and maintaining regular sleep timing.

This study investigated individual-level light exposure during sleep (LEDS) in a naturalistic setting, which compliments the growing body of research on outdoor LAN exposure and sleep. Prior studies of satellite-measured outdoor LAN have reported associations between greater LAN and shorter sleep duration⁴², as well as greater daytime sleepiness and delayed sleep timing⁴³. Similarly, studies of objective personal measures of indoor LAN have reported associations between greater LAN and worse sleep quality (LAN measured with headband device)^44,45 and delayed sleep timing (LAN measured with light sensor placed near head)⁴⁶. However, only limited studies examine sleep irregularity, which increasingly is linked to cardiometabolic disturbance. Phillips et al. investigated the association between LAN and the sleep regularity index in 12 regular and 12 irregular college students and reported lower daily rhythm amplitude of light exposure in irregular sleepers. While there were no differences in the daily rhythm amplitude of light exposure between the two groups, the irregular sleepers did have greater light exposure during the biological night (defined as a 10-hour window from beginning of dim light melatonin onset) when light exposure was normalized to their average daily illuminance¹¹. Mead et al. analyzed repeated measures of light exposure 1-hour before sleep onset as well as light exposure during the sleep episode with sleep offset, sleep duration, sleep percentage, and sleep fragmentation index in a sample of 124 adults (light measured with wrist-worn ActiWatch) with mixed model regressions⁴⁷. While their results did not support an association between pre-sleep light exposure with any of the sleep outcomes, a 1-unit increase in average lux during sleep (LEDS) was associated with a 0.12-hour later sleep offset timing, a 5% reduction of sleep percentage, and an increase by 0.03 of sleep fragmentation index⁴⁷. While more research is necessary, these findings suggest LAN and LEDS may be an upstream factor contributing to sleep irregularity.

Because the circadian system may respond differently to light spectra, we also conducted an exploratory analysis investigating the effect of blue wavelength (400–500nm) LEDS on sleep regularity. Melanopsin, the opsin utilized by the intrinsically photosensitive retinal ganglion cells in the retina responsible for relaying light signals to the circadian pacemaker, is most sensitive to light exposure in the blue light range^48,49. Our analysis supports an association between greater blue light exposure and more sleep irregularity. While studies have reported associations between blue light, such as that emitted by personal devices, and suppression of melatonin^50–52, further research on the spectral effects of light at different times of day, angle of incident light exposure⁵³, and how exposure to different wavelengths may affect sleep are needed⁵⁴.

Our analysis results did not support a gender-specific effect of LEDS on future sleep timing or sleep timing on future LEDS. Prior studies conducted in young men and women report greater effects of light on melatonin among men (with women having larger melatonin amplitude)^6,55–57. However, these studies were conducted in small samples, and men in general have been shown to have greater bright light exposure during the day compared to women⁵⁸, which could serve as a protective factor against the detrimental effects of LEDS on sleep. The effects of gender and menopause itself on sensitivity to LAN is not well established⁵⁹ and could be further investigated in future research.

If LEDS does contribute to sleep timing regularity and disease outcomes related to circadian disruption⁶⁰, understanding which populations have greatest exposure and why will help inform targeted sleep health interventions. In our sample, LEDS increased with age; further research is needed to investigate whether LAN or LEDS may play a role in the increased risk of developing sleep disorders and chronic health conditions among older adults, or whether chronic health conditions could lead to greater LEDS exposure. LEDS did not appear to be linked to shiftwork or insomnia in our sample, although this may differ for cohorts with higher prevalence of current employment and younger age groups. LEDS was also more prevalent among Black participants and less prevalent among Hispanic participants; this finding partly aligns with past research reporting greater outdoor neighborhood LAN exposure for Asian, Hispanic, and Black Americans compared to White Americans⁶¹. LEDS may also have a seasonal and geographical component, as LEDS was more prevalent as daylength increased (e.g., summer months) and less prevalent for the Minnesota study site; this may be due to greater sunlight intrusion at different times of the year and/or occlusion of the light sensor by clothing or bedsheets during colder months, as well as proximity to urban areas and nighttime light pollution. Future studies should investigate causal factors for LAN and LEDS and investigate impacts in populations which may be more sensitive to the effects, such as children, adolescents, and young adults^62–64. The high inter-individual variability in light-induced melatonin suppression⁶ also suggests that the same LAN or LEDS may have different effects within populations.

How LEDS and variability of sleep onset timing were defined in this study should be considered when interpreting results. LEDS was defined as light during each individual’s main sleep episode, separately for each night, as a marker of light exposure during the individual’s biological night. This definition was chosen because the focus of this study was on light exposure during the biological night, rather than the solar night, and because people’s activities are shaped by their rest-wake patterns, rather than by clock time. Modeling pre-sleep light exposure may have shown larger effect estimates; however, we chose to model LEDS rather than pre-sleep light to limit possible biases of reverse causation. By focusing on the sleep episode, we may be targeting each person’s sensitive window to LAN. This contrasts with other studies of LAN that have relied on local clock time irrespective of interindividual differences or LAN during periods of low activity, such as during the 5 hours of lowest average daily activity (L5), which may miss windows where circadian phase is most sensitive to light. Thus, how and when LAN is measured depends on how “night” is defined. LAN during the night defined by the clock on the wall and in the same way for all individuals may not appropriately reflect individual circadian phase sensitivity to light. However, these measures may be more feasible for large studies where scored sleep onset and offset information is not available. Variability of sleep onset timing was characterized in this study using person-specific measurements; each person-specific night-to-night variability was defined as the difference between two sequential starts of the main sleep episodes.

Measuring personal light exposure for large numbers of individuals and in the home environment can be challenging, and we are limited in our understanding of the light sources and reasons for LEDS exposure patterns in this study. For example, LEDS exposure (calculated as average illuminance across the sleep episode) could be due to a constant dim light source, such as a television or dim room light, or due to a brief exposure to brighter room light, such as turning on a light to use the bathroom. For example, a 5-minute exposure to 100-lux illumination during an 8-hour sleep episode (with 0 background lux) would translate to ~1.04 average lux; the same 5-minute 100-lux exposure during an 8-hour sleep episode where the background illuminance is equal to 1 lux would translate to ~2.03 average lux. The duration of the sleep episode would also influence the LEDS measurement. For example, a 5-minute exposure to 100-lux during a 6-hour sleep episode (with 0 background lux) would translate to ~1.39 average lux. Both constant dim background light and/or a short pulse of bright light may plausibly influence sleep timing⁶⁵, and future studies should consider incorporating mixed methods research to better understand the causes and sources behind LAN exposure⁶⁶, as well as measurement of LAN/LEDS with static light sensors in the sleeping area.

There are a number of strengths and limitations for this study. Analyses were adjusted for possible confounders, including daylength and study site to control for seasonal-related (and latitude / climate-related) effects of light and other environment exposures on sleep, as well as physical activity, sleep fragmentation, and sleep duration. The within-individual night-to-night analysis is also a strength of the study, further supporting the logistic regression findings and testing the temporal association to provide evidence for a cause-and-effect relationship. An additional strength is the measurement of objective light and sleep patterns in a naturalistic setting of participants in their home environments and a sensor that detects whether the device is being worn or not to determine missingness. While the magnitude of effect for LEDS in this study was relatively small, this may be due to measurement limitations and/or study sample characteristics. Participants were not given special instructions to avoid covering the device, and therefore there may be measurement error due to occlusion by clothing or bedding. As a wrist-worn device, the ActiWatch does not measure light exposure at the eye level, and there is measurement error when comparing it to head-worn devices and/or laboratory-grade photometers^67,68; it may also not adequately capture light or blue light from personal device usage. However, the measurement error for LEDS (such as due to occlusion by bedsheets) would be expected to bias results towards the null⁶⁹, although further work could simulate the direction and amount of bias given different study parameters⁷⁰. While we evaluated light during the sleep episode as a measure of light exposure during the biological night, a marker of circadian phase (such as dim light melatonin onset) was not available; as such, we were also unable to evaluate the direction of shift in response to LEDS. This analysis was also conducted in an older population, which may be less sensitive to the effects of inopportune light exposure due to physiological age-related changes. For example, yellowing of the eye lens with age may reduce blue-wavelength light and the amount of light intensity that reaches the retina^71,72, suggesting that older adults may require more daytime bright light to entrain and may be less sensitive to inopportune light exposure than younger adults. We also did not have measures of other exposures that could contribute to sleep irregularity, such as noise, bedroom temperature, or humidity, or contextual information about sleep fragmentation. Another limitation is the exclusion of participants with missing data and the potential for selection bias. Actigraphy has been shown to have some misclassification of wake and sleep episodes⁷³; however, a sensitivity analysis excluding participants with insomnia did not substantially alter results.

5. Conclusion

Light exposure during the sleep episode (LEDS), a form of LAN, is associated with irregular sleep timing in a large sample of U.S. older adults in both averaged and night-to-night analyses, with greater light exposure predictive for larger subsequent irregularity in sleep onset timing. These results support the importance of maintaining a dark sleeping environment and maintaining regular sleep timing. Future research should aim to utilize personal measures of LAN and/or static light sensors in the sleeping area, conduct night-to-night analyses of LAN and LEDS with sleep and health outcomes, and evaluate the causes and behaviors behind LAN exposure.

HIGHLIGHTS.

• Light pollution and light at night may disrupt sleep and circadian rhythms

• Light exposure during sleep is associated with irregular sleep timing

• Irregular sleep timing is also associated with greater light exposure during sleep

• Both dark sleeping spaces and regular sleep schedules are important to maintain

Data availability:

Actigraphy data and data from the MESA sleep ancillary study are available from the NHLBI-funded National Sleep Research Resource (NSRR): https://sleepdata.org/. Study data can also be obtained through a data use agreement with the MESA Data Coordinating Center: https://www.mesa-nhlbi.org.