Introduction

Necessity of the research

Artificial light at night (ALAN) refers to light generated by human activities during nighttime, typically emitted from sources such as streetlights, building illumination, vehicle headlights, and electronic devices (Wang et al., 2023b). Since ALAN is powered by electricity and often serves residential lighting and socioeconomic functions, its variability is closely linked to economic growth and urbanization (Mellander et al., 2015; Proville et al., 2017). As a hallmark of modernization, ALAN has expanded rapidly (Gaston and Sánchez de Miguel, 2022). Over the past few decades, satellite-detected nighttime light emissions have increased globally at a rate that surpasses population growth (Gaston and Sánchez de Miguel, 2022). While this growth reflects social and economic development, it has also disrupted the natural circadian rhythms of many organisms, including humans (Helbich et al., 2024). For these reasons, some scholars argue that ALAN should be considered a focus of global change in the 21st century (Davies and Smyth, 2018).

In this context, growing concerns have emerged regarding the potential negative effects of nighttime lighting. Studies have linked ALAN to sleep disturbances (Helbich et al., 2024), increased risks of cancer (Palomar-Cros et al., 2024), cardiovascular issues (Boakye et al., 2025; Hu et al., 2024), and mental health problems (Liu et al., 2025). However, as a human invention, ALAN can also offer several benefits that contrast with its adverse effects (Wang et al., 2023b). Specifically, ALAN can help overcome time constraints (Wang et al., 2023b), such as facilitating walking or other forms of outdoor activity during evening hours (Boyce, 2019; Painter and Farrington, 1997; Svechkina et al., 2020).

Among various nighttime behaviors, physical activity at night is especially relevant to public health and closely tied to ALAN, particularly in Asian countries such as China. In line with global trends (Guthold et al., 2018, 2020), China is currently experiencing a public health challenge marked by insufficient levels of physical activity (Zhang et al., 2023). One contributing factor is the increasing amount of time people spend on sedentary tasks, such as office work (Parry and Straker, 2013). Many Chinese workers follow the so-called “996” schedule—working from 9 a.m. to 9 p.m., 6 days a week (Liu and Chen, 2023; Wang, 2020). This demanding schedule leaves little room for engaging in leisure-time physical activity outside of nighttime hours (Jiang et al., 2025; Su et al., 2024). Therefore, it becomes particularly important to study environmental factors—such as ALAN—that may encourage or support physical activity at night.

Research gap

Research has shown that street lighting, a primary source of ALAN, increases the opportunities for residents to walk or cycle at night (Fotios et al., 2019; Lee et al., 2021). Subjective perceptions of ALAN may also influence residents’ engagement in outdoor activities, including nighttime exercise (Wang et al., 2023a). Similarly, positive perceptions of street lighting have been associated with more moderate to high-intensity physical activity (Evenson et al., 2007).

Although these studies suggest a potential role of ALAN in supporting nighttime physical activity, the existing body of research remains limited and preliminary. For example, the aforementioned studies merely recorded whether a street was lit and the number of people engaging in a single type of activity, such as cycling. This gap highlights the need for more nuanced approaches to quantifying the relationship between ALAN and physical activity, more specifically, approaches that account for different measurement methods of ALAN and capture various aspects of physical activity. In addition, little is known about the underlying mechanisms that might explain how ALAN may associate with nighttime physical activity.

In the present study, we aimed to address these gaps by thoroughly measuring both ALAN and physical activity and investigating their hypothesized association pathways using mediation models. We propose and test environmental restorativeness and perceived safety as mediators. The rationale for this model is discussed in the following theoretical framework section.

Theoretical framework

Mediating role of environmental restorativeness

Modern lifestyles require high levels of directed attention, creating a strong need for psychological restoration in the evening. Publicly lit outdoor spaces may help meet this need (Nikunen, 2013; Nikunen and Korpela, 2009). According to attention restoration theory (ART), restorative environments provide four key elements: “escape” (experiencing a new or reinterpreted space apart from daily life), “extent” (a broad setting to immerse in), “compatibility” (alignment with personal goals), and “soft fascination” (gentle attraction without overstimulation) (Kaplan and Kaplan, 1989; Kaplan, 1995).

Outdoor lighting at night may support at least two ART dimensions. First, ALAN transforms the cityscape, helping residents mentally escape daily routines. Second, as part of urban nighttime landscapes, ALAN may offer mild visual appeal, fostering “soft fascination.” Some scholars suggest nighttime lighting enhances environmental attractiveness (Boyce, 2019). Studies indicate that ART-based restorative qualities, such as “being away,” may relate to perceived brightness (Nikunen et al., 2013). A Chinese study found that participants rated nighttime urban scenes with artificial lighting as more restorative than daytime urban or green spaces (Chen et al., 2011), supporting the idea that ALAN can enhance the perception of restorativeness.

Additionally, restorative environments may promote physical activity. For example, older adults with access to comfortable outdoor spaces are more likely to engage in walking (Roe and Roe, 2018), and neighborhood restorative quality positively predicts physical activity (Dzhambov et al., 2018a; Dzhambov et al., 2017; Dzhambov et al., 2018c). Based on this, we hypothesized that ALAN is associated with nighttime physical activity through environmental restorativeness.

Mediating role of perceived safety

Perceived safety may mediate the relationship between ALAN and nighttime physical activity. This link is intuitive; one main role of outdoor lighting is to enhance the safety of pedestrians and cyclists after dark (Fotios et al., 2014). Lighting helps people see others and obstacles, reducing the risk of collisions, falls, and crime (Boyce, 2019; Cheng et al., 2016; Uttley et al., 2015). The prospect-refuge theory explains how environmental features like lighting shape our sense of safety through three dimensions: prospect (clear visibility), refuge (potential hiding places), and escape (barriers to fleeing danger) (Fisher and Nasar, 1992; Nasar et al., 1993; Nasar and Jones, 1997). Lighting improves visibility, enhancing prospect and reducing refuge by making hiding places more visible (Loewen et al., 1993; Nasar and Jones, 1997). Moreover, there may be an acquired association between lighting and safety regardless of the presence of others or clarity of view (van Rijswijk and Haans, 2017).

Empirical studies support this connection. Well-lit streets often feel safer (Peña-García et al., 2015), and increasing nighttime lighting has been shown to reduce both crime and fear of crime (Painter et al., 1988), especially for women and the elderly (Painter, 1991). Adequate illuminance can even evoke a daytime-like sense of safety (Boyce et al., 2000). These findings suggest higher ALAN leading to increased perceived safety.

Perceived safety also positively influences physical activity. A meta-analysis of 16 studies found that people who feel safe are more likely to be physically active, while those in high-crime areas are less likely (Rees-Punia et al., 2018). Other studies suggest that perceived safety can mediate the effects of neighborhood and individual factors on physical activity (Pratt, 2015; Timperio et al., 2015). Thus, we propose that ALAN may be associated with nighttime physical activity via perceived safety.

Since perceived safety may also affect how people experience restorative environments, we included a covariance link between these two variables to account for their potential interaction (Collado et al., 2017; Stragà et al., 2023). Considering that socioeconomic factors (such as urban-rural location, income, and education level) may influence lighting and physical activity resources near peoples' residences and lifestyles, these potential confounding effects were also considered in the mediation model.

Research orientation and hypotheses

Based on the literature discussed above, we proposed a theoretical model suggesting that ALAN is positively associated with nighttime physical activity through perceived safety and environmental restorativeness (Fig. 1).

Fig. 1
figure 1

Conceptual framework.

Full size image

To test this model, we conducted a quantitative study using a cross-sectional online questionnaire to collect data.

Our study was guided by the following hypotheses:

  • H1: A higher proportion of Chinese residents choose to engage in physical activity at night compared to during the day.

  • H2: Higher levels of ALAN are associated with increased nighttime physical activity among Chinese residents.

  • H3: Environmental restorativeness and perceived safety serve as mediating variables in the relationship between ALAN and nighttime physical activity.

Materials and methods

Participants

We conducted an online survey in early December 2023, distributing recruitment information through various online chat groups. Participants were required to read the study instructions and provide informed consent before proceeding. Eligibility criteria included: (1) being 18 years or older; (2) having no serious medical conditions that could interfere with physical activity; (3) having continuously lived in their primary residence over the past month—to ensure that the assessment of neighborhood ALAN and nighttime physical activity occurred within the same timeframe; (4) being physically located in that residence while completing the survey, so their location could be recorded via smartphone for objective ALAN measurement; and (5) being capable of using a smartphone for location data collection. Participants were excluded if they submitted incomplete surveys or failed to meet the quality control criteria (see section “Mediating role of perceived safety”). The study was conducted with the approval and oversight of the Ethics Committee of Southwest University.

Sample size calculation

To enhance the representativeness of the nighttime physical activity population, the required sample size was calculated based on established recommendations. Our formula was as follows, based on past work (Charan and Biswas, 2013; Pourhoseingholi et al., 2013; Tumuhamye et al., 2013):

$$N=\frac{p\left(1-p\right){Z}^{2}}{{d}^{2}}$$
(1)

In the formula, N denotes the required sample size, while p represents the prevalence of the target variable. Z corresponds to the z-score associated with the desired confidence level, and d indicates the level of precision (i.e., the acceptable margin of error or effect size). Most researchers report their results using a 95% confidence interval (CI), in which case Z is typically set to 1.96. When the prevalence of the condition ranges between 10% and 90%, some scholars recommend setting the precision (d) at 5% (Pourhoseingholi et al., 2013).

According to preliminary investigations (Jiang et al., 2025; Su et al., 2024) and our estimation, we calculated the required sample size based on a conservative prevalence rate of 50%. As a result, the minimum sample size needed was determined to be 385 individuals.

Quality control

We followed recommendations for online surveys (Buchanan and Scofield, 2018) and implemented several quality control measures. To prevent duplicate submissions, we restricted participation by both IP address and account. After submitting the survey, the same device could not access the participation link again. Certain questions, such as age and gender, were asked twice throughout the survey. Inconsistent answers (e.g., reporting an age between 15 and 25 at the beginning of the survey but a different age range at the end) led to exclusion. Additionally, we included simple verification questions (e.g., “What is the common color of plants?”) to assess attentiveness. Participants selecting obviously incorrect answers (e.g., “Plants are usually black”) were excluded. Participants were also required to pass a CAPTCHA test and complete real-name authentication, which was automated by the WeChat platform and not visible to the survey administrators. In total, 1158 questionnaires were completed, and 646 met the eligibility criteria for inclusion in the study.

Measurements

Time characteristics of physical activity

We used a binary question to inquire about participants’ primary time for physical activity over the past month. The question was: “In the past month, during which time period do you mainly engage in physical activity (e.g., walking)? Daytime or nighttime? (Nighttime refers to the period after sunset when city nighttime lighting is on).” Following this, we included an open-ended question to allow participants to explain the main reasons for choosing this particular time period for their activities. No word count limit was set for the responses.

ALAN

We measured ALAN from both objective and subjective perspectives.

Objective ALAN: We utilized the data from Chen et al. (2021), which contains an extended time series of ALAN measurements similar to the National Polar-orbiting Partnership Visible Infrared Imaging Radiometer Suite (NPP-VIIRS), using a novel cross-sensor approach. This dataset is calibrated against DMSP-OLSNTL and monthly NPP-VIIRSNTL results, demonstrating excellent spatial patterns and temporal consistency (Chen et al., 2021). The spatial resolution in the WGS84 coordinate system is 15 arc-s (~500 m). We used the most recent annual data, updated by the authors through 2022.

To assign ALAN measures to participants’ locations, we obtained their latitude and longitude coordinates from their addresses (retrieved via Gaode Map, see: https://mobile.amap.com/). A buffer zone with a radius of 1 km was established around each set of device coordinates, following previous research (Campbell and Janssen, 2023; Hu et al., 2022). This range also aligns closely with prior studies on neighborhood physical activity (Kaczynski et al., 2009; Rundle et al., 2019; Rundle et al., 2016). We then computed the average ALAN levels within this buffer zone to represent neighborhood ALAN.

Subjective ALAN

We adapted the approach from Wang et al. (2023a) to assess participants’ perceptions of ALAN in their neighborhood environment. Participants’ responses were measured using an 11-point scale, where a score of 0 indicated complete darkness and a score of 10 indicated extremely bright conditions. To ensure consistency with objective measurements, we prefaced each question with a contextual statement such as, “Within a 1-kilometer radius around your neighborhood (approximately a 15-min walk), how would you rate the nighttime brightness?” The choice of a “15-min walk” as a measure was based on urban research norms, as it typically corresponds to a distance of ~1 km (Dong et al., 2023; Ma et al., 2023).

Nighttime physical activity

We considered both the intensity and domains of physical activity, acknowledging that nighttime may be particularly associated with certain types or parts of physical activities.

Intensities of nighttime physical activity

We used three questions to assess the intensity of physical activity (Terracciano et al., 2021). The specific question asked was: “In the past month, how often did you engage in walking, moderate physical activity (such as carrying light loads or cycling at a normal pace), or vigorous physical activity (activities that make you sweat or breathe hard, such as fast cycling, aerobics, or heavy lifting) lasting at least 10 min during the night?” Each survey item was measured on a 7-point ordered categorical scale: 1 indicated 1 day a week or less on average, 2–6 indicated 2–6 days a week on average, and 7 indicated almost every day.

Domains of nighttime physical activity

Given that nighttime typically corresponds to non-working hours, we concentrated on physical activities related to leisure and transportation. We adapted a questionnaire designed for community physical activity surveys by referencing Troped et al. (2003) and emphasizeding nighttime scenarios in the survey items.

For leisure physical activities, we asked, “In the past month, on average, how many days per week did you engage in leisure physical activities during the night?” The results were measured using an 8-point response scale: 0 indicated never, 1–6 indicated 1–6 days per week, and 7 indicated nearly every day. Additionally, we asked, “On average, how many minutes did the above activities last each day?” to measure the duration of daily activities. Responses were measured using a 10-point scale: 1 indicated 0–10 min, 2–9 indicated 11–20 to 81–90 min, and 10 indicated more than 90 min. The multiplied score of these two items was considered the total score for leisure physical activities (Troped et al., 2003).

For transportation-related physical activities, we considered both walking and cycling, following the method used by Troped et al. (2003). We asked, “In the past month, on average, how long did you walk at night for transportation purposes each day?” For the cycling question, we substituted “walk” with “ride a bike.” The results for both questions were measured using a 10-point ordered scale: 1 indicated 0–10 min, 2–9 indicated 11–20 to 81–90 min, and 10 indicated more than 90 min.

Environmental restorativeness

Following methods used by other scholars and based on our foundational hypotheses, we employed the “Away” and “Fascination” dimensions from the Perceived Restorativeness Scale (PRS) to measure environmental restorativeness (Dzhambov et al., 2017; Dzhambov et al., 2018b; Dzhambov et al., 2018c). Adapting Dzhambov et al. (2018b)’s survey, we used the following statements: “Near my home/residence (within 1 km), there are places where I can relax and escape from the daily chores that demand my attention” and “Near my home (within 1 km), there are places that look attractive or appealing, capturing my attention with many interesting things.” Responses were measured on a 1–10 agree/disagree scale. The average score of these two items was considered the summative score for environmental restoration quality. These two questions demonstrated high internal consistency (Cronbach’s α = 0.9).

Perceived safety

We followed the approach of previous studies on neighborhood safety (Dallago et al., 2009; Juul and Nordbø, 2023; Shenassa et al., 2006) and measured perceived safety with the following question: “Do you feel safe when walking near your residence (within 1 km)?” Responses were measured using a 7-point Likert scale (1 = very unsafe, 7 = very safe).

Covariates

We included several covariates in our study. Individual-level variables included gender, age, education level, monthly household income, marital status, and student status. Gender may influence light sensitivity and participation in physical activity (Chellappa et al., 2017; Rosenfeld, 2017), while age may affect exposure to ALAN and motivations for physical activity (Chen et al., 2022; Renner et al., 2007). Low-income individuals tend to live in areas with lower socioeconomic status, which are often associated with poorer physical activity environments (Wilson et al., 2004). Those with higher education levels may be more likely to engage in physical activity (Droomers et al., 2001), while students, due to being in a unique period of emotional and environmental change, may have different physical activity patterns compared to non-students (Clemente et al., 2016).

At the regional physical level, we included population density within a 1 km radius (as a proxy for urbanization (Li et al., 2023a)), daytime land surface temperature, and the Normalized Difference Vegetation Index (NDVI, as a proxy for vegetation). Urbanization is closely linked to ALAN (Zhang and Seto, 2013) and affects people’s lifestyles, including physical activity (Ojiambo et al., 2012). Similarly, outdoor temperature also influences physical activity (Ho et al., 2022), with extreme temperatures in winter or summer potentially reducing people’s willingness to go outside. Finally, neighborhood greenery may influence environmental restorativeness, perceived safety, and physical activity (Dzhambov et al., 2018a; Markevych et al., 2017; Rahm et al., 2021). Population density data were obtained from the WorldPop website (details: https://www.worldpop.org/), expressed as the adjusted number of people per square kilometer based on national totals (data DOI: 10.5258/SOTON/WP00675). Land surface temperature and NDVI data were sourced from the Google Earth Engine (GEE) platform. We used the MODIS/006/MOD11A2 dataset to derive the annual mean daytime temperature for China from January 1, 2022, to December 31, 2022, with a spatial resolution of 1 km. We calculated the maximum NDVI values using the MODIS/006/MOD13A2 dataset for the same period. To exclude the effects of water bodies, we removed negative values from the images (Liu et al., 2019; Markevych et al., 2017). Ultimately, we computed the average temperature and NDVI values within the 1 km buffer zone around the geolocation of responses.

At the regional socioeconomic level, we obtained provincial-level data on per capita disposable income, the proportion of the population with undergraduate education, and the employment rate from the National Bureau of Statistics of China (details: https://www.stats.gov.cn/sj/ndsj/). These factors are closely related to regional socioeconomic status and may influence the physical activity resources and opportunities available to individuals.

Statistical analysis

Detection of common method bias

To reduce the common method bias inherent in questionnaire studies, we conducted Harman’s single-factor test (Fuller et al., 2016). The results revealed that the first factor accounted for 23.4% of the variance, which was below the recommended threshold of 50%, indicating a low risk of common method bias.

Keywords extraction

We used the Quanteda package for R to process open-ended responses and identify keywords related to the reasons for choosing daytime or nighttime for physical activity (Benoit et al., 2018). First, we created a corpus from participants’ responses using the “corpus” function. Then, we used the “token” function to filter out non-alphanumeric characters. Finally, we created a document-feature matrix with the “dfm” function to calculate word frequencies and identify keywords.

Correlations

Spearman’s rank-order correlation tests were conducted to explore general correlations between two continuous variables of interest. In addition, we constructed a correlation network using Spearman correlation and the EBICglasso method (Zeng et al., 2023). This approach is data-driven and allows for the exploration of complex patterns of association among variables in cross-sectional data. To further illustrate the potential positions of variables in the network, we reported three centrality indices (Zeng et al., 2023): (1) Strength (sum of the weighted number and strength of all connections a node has relative to others); (2) Closeness (how closely a node is connected to all others, considering indirect connections); and (3) Betweenness (indicating the importance of a node in the average pathway between other node pairs) (Hevey, 2018). We used the following intervals to interpret correlation coefficients: 0–0.10 (no significant correlation), 0.11–0.39 (weak correlation), 0.40–0.69 (moderate correlation), 0.70–0.89 (strong correlation), and 0.90–1.00 (very strong correlation) (Labata-Lezaun et al., 2022; Schober et al., 2018).

Collinearity diagnosis

Before performing regression analyses, we examined collinearity issues among predictors. Variance inflation factor (VIF) values less than 5 and Pearson’s r values less than 0.8 were considered evidence of the absence of harmful multicollinearity (Rogerson, 2019; Shrestha, 2020). After screening, we excluded the covariate “proportion of the population with undergraduate education” due to its strong correlation with per capita disposable income (Pearson’s r > 0.8). After this exclusion, the VIFs for the remaining variables were kept below 3.0.

Exploration of relationship forms

Following the recommendations of Mansournia and Nazemipour (2024), we employed non-parametric methods to explore the form of relationships before conducting linear regression. Specifically, given our focus on ALAN, we used Generalized Additive Models (GAM) and drew on prior research to construct a smooth spline function based on ALAN indicators, while incorporating other covariates in their untransformed forms to obtain their estimated parameters (Smiley et al., 2019).

The GAM equation used was:

$${\rm{E}}({\rm{Y}})=\alpha +{\rm{s}}({\rm{ALAN}},{\rm{k}})+{\rm{Covariate}}\,1+{\rm{Covariate}}\,2+\ldots +{\rm{Covariate}}\,{\rm{n}}$$
(2)

Here, E(Y) represents the expected value of the physical activity indicator, α is the intercept, and s() denotes the smoother based on the penalized spline. For subjective ALAN, The number of allowed knots “k” was set to 4 based on previous recommendations (Baayen and Linke, 2020; Xue et al., 2018) and also that it generally led to lower AIC values. For objective ALAN, the k was set as default (k = 10). The reported effective degrees of freedom (EDF) values indicate the curvature of the smoothing term. An EDF value equal to 1 indicates a linear relationship, values between 1 and 2 suggest some non-linear trend, and values greater than 2 indicate a strongly non-linear relationship (Hunsicker et al., 2016; Köthe et al., 2023). This analysis was conducted using the mgcv package in R (v.4.2.1).

Regression analysis

We also conducted linear regression models to analyze associations between ALAN indicators and physical activity. This was done both to test Hypothesis 2 and select appropriate independent variables for the mediation analysis, considering that we had multiple ALAN measures. According to strict mediation theory, if the independent variable is not related to the dependent variable, further testing for mediation effects is unnecessary (Wen and Ye, 2014).

Given our relatively large sample size (over ten observations per variable), we did not transform the variables (Schmidt and Finan, 2018). To mitigate the impact of heteroscedasticity, we employed robust standard error linear regression based on the HC3 method (Hayes and Cai, 2007). All analyses were adjusted for the included covariates.

Mediation analysis

Structural equation modeling (SEM) was employed to investigate the pathways and mediation effects between the variables of interest. To avoid excessive parameter estimation and ensure statistical efficiency, we established six separate mediation models based on different physical activity variables.

The results of the measuring scales were processed as continuous summary scores to increase statistical power (Dzhambov et al., 2018b). The analysis was conducted using the Maximum Likelihood (ML) estimator. We used the bias-corrected bootstrap method (Hayes and Scharkow, 2013) with 10,000 replications to generate corresponding standard errors and CIs for all paths (Brown, 2015; Haukoos and Lewis, 2005; Kelley, 2005), which can cope with nonnormality.

After specifying the covariance chains, our core model essentially became a saturated model similar to a chain-mediation model (Agler and De Boeck, 2017); therefore, it was unnecessary to report the model fit indices. An indirect effect (i.e., a product of coefficients for the constituent links) that is significantly different from zero was regarded as evidence of mediation (Hayes, 2017; Zhao et al., 2010).

Correlations, linear regressions, and SEM were conducted in SPSS 26.0 and AMOS 26.0 software. A p-value smaller than 0.05 was regarded as evidence of statistical significance.

Sensitivity analysis

The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines recommend additional analyses along with the primary outcomes (Cuschieri, 2019; Vandenbroucke et al., 2007; Von Elm et al., 2007). Given that females may be more sensitive to nighttime safety concerns, we stratified the SEMs by gender.

Results

Participant characteristics

A total of 646 participants were included in this study, with 61.8% male and 38.2% female. The age distribution was primarily between 15 and 35 years, accounting for 87.6% of the total sample. Among them, 42.7% were aged 15–25 years, and 44.9% were aged 26–35 years. The majority of participants were unmarried (66.3%), and more than half were non-students (56.7%). Regarding educational background, 62.8% of participants had a bachelor’s degree, and 13.3% held a degree higher than a bachelor’s. The majority of participants had a monthly household income between 5001 and 15,000 RMB (63.7%) (1 RMB is equal to 0.14 USD as of 28 July 2024), as shown in Table 1.

Table 1 Participants’ characteristics (N = 646).
Full size table

Time characteristics of physical activity

More than 70% of participants indicated that they primarily engage in nighttime physical activities (Fig. 2). For this group, the primary reason for choosing nighttime physical activity was timing. Additionally, keywords such as “exercise” and “maintaining physical fitness” highlighted their primary motivations for engaging in nighttime physical activity. These motivations were similar to those of participants who performed physical activity during the daytime.

Fig. 2
figure 2

Preferred time to exercise and reasons for exercising outdoor at night or during the day (N = 646).

Full size image

Correlations

No correlations were observed between objective ALAN and the other study variables (Fig. 3). However, subjective ALAN was correlated with nighttime walking (ρ = 0.153, p < 0.001), moderate physical activity (ρ = 0.220, p < 0.001), vigorous physical activity (ρ = 0.212, p < 0.001), leisure physical activity (ρ = 0.239, p < 0.001), transportation walking (ρ = 0.258, p < 0.001), environmental restorativeness (ρ = 0.451, p < 0.001), and perceived safety (ρ = 0.381, p < 0.001). These correlation strengths were weak to moderate.

Fig. 3: Spearman’s rank correlations between the variables of interests (N = 646).
figure 3

Note: Numbers in cells indicate spearman’s rho; MPA moderate physical activity, VPA vigorous physical activity, Leisure PA leisure physical activity.

Full size image

The centrality indicators from the network analysis can help identify which variables are more crucial and influential within the entire variable system, thereby revealing the core of the variable group. Our network analysis indicated strong centrality for environmental restorativeness, suggesting its potential role as a mediator between the variables (Fig. 4). Subjective ALAN exhibited lower centrality, indicating its associations with other variables (as indicated by the pairwise correlation analysis) might be realized by some mediators. Consistent with pairwise correlation analyses, objective ALAN showed no correlations with other variables.

Fig. 4: Correlations network between variables of interest (N = 646).
figure 4

Note: Blue lines indicate positive correlation while red lines indicate negative correlations. The thickness of lines indicates the strength of correlations. MPA moderate physical activity, VPA, vigorous physical activity, Leisure PA leisure physical activity.

Full size image

Examining the forms of associations with GAMs

The associations between objective ALAN and other variables were not significant (Table 2). Subjective ALAN also showed significant nonlinear relationships with all study variables, showing U-shaped patterns, with more consistency when subjective ALAN values were above 4 (Fig. 5).

Table 2 Results of GAM regressions (N = 646).
Full size table
Fig. 5: Relationships between study variables as visualized by GAMS (N = 646).
figure 5

Note: (a)–(h) represent the associations of subjective ALAN with walking, MPA, VPA, leisure PA, transportational walking, transportational cycling, environmental restorativeness, and subjective safety, respectively; the associations were adjusted for gender, family income, education status, student status, marital status, land surface temperature, NDVI, population density, per capita disposable income, and employment rate; MPA moderate physical activity, VPA vigorous physical activity, Leisure PA leisure physical activity.

Full size image

Linear regression

Based on the distinct association detected by GAMs, we performed linear regression (Table 3). Subjective ALAN showed significant relationships with both physical activity and the mediators that could be generally divided into two stages (scores of 0–3 and 4–10); therefore, segmented linear regressions were performed. The results revealed that subjective ALAN exhibited significant positive associations with all mediators and physical activity variables in the 4–10 score range. By contrast, no significant relationships were found in the 0–3 score range. This negated the necessity of further mediation analysis in the 0–3 score stage.

Table 3 Linear associations between ALAN and study outcomes/mediators.
Full size table

Mediation analysis

Given the significant linear relationship observed between subjective ALAN in the 4–10 score range and nighttime physical activity, we performed mediation analyses for this score range, encompassing 613 participants. The effects and paths in the final models are summarized in Fig. 6 and Table 4.

Fig. 6: Mediation models loaded with different physical activity outcome (N = 613).
figure 6

Note: (a)–(f) represent the associations of subjective ALAN with walking, MPA, VPA, leisure PA, transportational walking, transportational cycling, environmental restorativeness, and subjective safety, respectively; dotted lines indicate non-significant pathways (p > 0.05). All pathways were controlled for gender, age, education level, monthly household income, marital status, student status, population density, daytime land surface temperature, NDVI, per capita disposable income, and employment rate.

Full size image
Table 4 Pathways between subjective ALAN and physical activity outcomes from SEMs (N = 613).
Full size table

In the 4–10 score range, subjective ALAN was significantly associated with all nighttime physical activities, consistent with the earlier linear regression results. Furthermore, the indirect associations between subjective ALAN and all nighttime physical activities were significant. It is noteworthy that after accounting for the indirect effects, the direct associations between subjective ALAN and nighttime walking, moderate physical activity (MPA), vigorous physical activity (VPA), and leisure physical activity (leisure PA) were not significant. This suggests that we observed some full mediation effects.

Regarding specific indirect pathways, those mediated by environmental restorativeness were all significant, whereas those mediated by perceived safety were partially non-significant (for the outcome variables MPA, VPA, leisure PA, and transportation walking).

Sensitivity analysis

After stratifying the mediation models by gender, a few subtle differences emerged (Supplementary Table 1). Among females, the “ALAN→Safety→Walking” pathway became non-significant (p = 0.805). Among males, the “ALAN→Restorativeness→Transportation cycling” pathway became marginally significant (p = 0.059). Meanwhile, the other indirect pathways were not substantially changed.

Discussion

We investigated Chinese citizens’ nighttime physical activity participation and examined the associations between ALAN and nighttime physical activities. We also explored two potential pathways for these associations: environmental restorativeness and perceived safety. We assessed nighttime lighting using both subjective (self-reported) and objective (satellite imagery-based) methods, and distinguished physical activities by intensity and domain. Our results indicate that the vast majority of our participants (Chinese adults) engaged primarily in physical activities at night. We also found that within certain subjective ALAN score ranges (≥4), ALAN was positively associated with nighttime physical activities. In mediation models, environmental restorativeness emerged as a significant mediator in the relationship between subjective ALAN (≥4) and nighttime physical activity. Generally, these findings support our Hypothesis 1 (H1) and partially support our Hypothesis 2 (H2) and Hypothesis 3 (H3).

Choosing to engage in physical activity at night

Over 70% of our respondents indicated that they primarily engage in physical activities at night, with “time” constraints being the main reason for this choice. Physical activity is widely recognized as a key factor in promoting health, and the World Health Organization recommends it for all populations (Bull et al., 2020; Carty et al., 2021). However, common physical activity guidelines typically focus on the type, intensity, quantity, and duration of activities, rather than addressing the distribution of physical activity between day and night (Yi et al., 2022).

While the human body’s physiological processes follow a circadian rhythm, which may influence responses to and motivations for physical activity, nighttime physical activity may still theoretically help reduce obesity (Saidi et al., 2021), regulate mood (Widyowati et al., 2023), and in some cases, even improve sleep (Bulckaert et al., 2011). Therefore, physical activity performed at night due to time constraints may still contribute positively to public health rather than detracting from it.

The extent to which Chinese individuals engage in physical activity at night has not been thoroughly studied. One early study reported that at certain universities in Anhui Province, China, the proportion of students choosing to exercise during the day and at night was quite similar (Yu et al., 2013). A more recent study on Chinese college students also found that over 70% of participants chose to exercise at night, with “time” being cited as the primary reason (Su et al., 2024). These results, along with our own findings, underscore the importance of optimizing the environment for nighttime physical activity in China.

ALAN and nighttime physical activity

Our GAMs indicate an approximately U-shaped relationship between subjective ALAN and nighttime physical activity. Based on this finding, our further linear regression and SEMs indicated significant associations between subjective ALAN in the 4–10 score range and all nighttime physical activities. It is noteworthy that no significant linear associations between subjective ALAN and nighttime physical activity were observed in the lower score ranges (<4). This finding suggests that there might be a threshold effect, where subjective ALAN begins to positively associate with nighttime physical activity only beyond a certain level. However, due to the small sample size in the lower score range (n = 33), which led to low statistical power, the interpretation of this result should be approached with caution.

Unlike subjective ALAN, we did not observe significant results for objective ALAN. Moreover, we did not observe any substantial correlation between subjective and objective ALAN. While there is limited research directly addressing this topic, other studies have reported inconsistencies between subjective and objective measures of environmental factors (Leslie et al., 2010). We speculate that this discrepancy may arise from differences in how perceived and measured neighborhood areas are defined. As some researchers have suggested, areas that are not frequently visited or used by residents might be invisible to them (Leslie et al., 2010). The neighborhood areas that individuals truly interact with on a daily basis may be highly personalized (Christensen et al., 2022), and their size and shape might differ greatly from the standard buffer zones we used in this study. Therefore, individuals’ evaluations of neighborhood ALAN could be based solely on the areas they regularly frequent, such as a specific square or street, and such partial areas may not be representative of the overall lighting level of the entire neighborhood. Additionally, satellite imagery may only capture the brightness of ALAN, whereas subjective ALAN may encompass additional dimensions such as color temperature, which may influence participants’ overall perception of brightness and have complex effects on both behavior and psychological responses. These factors may collectively contribute to the different results observed between objective and subjective ALAN.

Mediating roles of environmental restorativeness and perceived safety

Our hypothesized mediators fully mediated the association between subjective ALAN and walking, moderate physical activity, vigorous physical activity, and leisure physical activity at nighttime, as evidenced by the close-to-zero direct association after introducing the mediator (Gunzler et al., 2013). While this finding does not rule out the possibility of other mediating factors (Morera and Castro, 2013), it does suggest we may have identified critical pathways to explain the relationship between ALAN and nighttime physical activity.

Among the two hypothesized mediators, environmental restorativeness emerged as the stronger one, as its mediating effect accounted for a substantial proportion of the overall association. Many studies have examined green space as a supporter of both environmental restorativeness and physical activity (Markevych et al., 2017). We should emphasize that restorative opportunities exist beyond green natural settings. For instance, restorative environments can include non-green natural settings (Li et al., 2023b) as well as manmade environments (Bornioli and Subiza-Pérez, 2023), such as libraries (Stragà et al., 2023) and urban areas with night lighting (Chen et al., 2011).

Regarding perceived safety, we only observed its mediating role between subjective ALAN and nighttime walking, although this relationship was weak (β = 0.040). Interestingly, we also found that perceived safety suppressed the relationship between subjective ALAN and nighttime cycling for transportation, as reflected in the indirect effect opposite to the total effect (MacKinnon et al., 2000), though the effect also remained weak (β = −0.033). Further inspection of the model paths revealed that the relationship between perceived safety and nighttime physical activity was generally weak and non-significant, which explained the near-zero indirect effects. One possible explanation is that, although ALAN helps to enhance feelings of safety, participants generally perceived high levels of safety, which are unlikely to restrict people’s outdoor activities, thereby diminishing the role of increased safety. This is likely common in China, where people highly rate the perceived safety of their communities (Hill et al., 2016). Some scholars have also suggested that outdoor activities in China may not be heavily influenced by perceived safety (Zhang and Yao, 2022). Overall, our findings emphasize that environmental restorativeness might serve as a more important mediator compared to perceived safety. This suggests that attention should be paid to how lighting improves the “psychological appeal” of the environment, rather than focusing solely on enhancing safety.

Finally, we observed slight differences between males and females in the mediation analysis. Specifically, among females, perceived safety did not significantly mediate the relationship between subjective ALAN and nighttime walking. Further examination of the model revealed that this was primarily due to the lack of a significant association between perceived safety and walking behavior. In addition to the generally high levels of perceived safety mentioned above, we speculate that this may be related to differences in physical activity preferences. Compared to males, females may be less likely to choose nighttime walking as their primary mode of transportation or recreation. Their nighttime physical activity is more likely to occur indoors or in organized settings. Therefore, although improved nighttime lighting may enhance their sense of safety, it might not be an essential influencer of their walking behavior. This phenomenon suggests that future lighting infrastructure projects should consider the behavioral habits and actual needs of different groups.

Contributions

This study has several contributions:

  1. (1)

    We add to the evidence regarding nighttime physical activity among Chinese residents, offering insights for urban planning and construction to support China’s “National Fitness” program.

  2. (2)

    We quantify the relationship between ALAN and different intensities and domains of physical activity using both subjective and objective methods, supporting the positive aspects of ALAN and enriching the understanding of ALAN.

  3. (3)

    We identify the mediating role of environmental restorativeness in the relationship between subjective ALAN and certain physical activities. This provides a new mechanism for understanding the phenomenon and highlights the potential for creating restorative urban environments using artificial lighting.

Limitations and future directions

This study also has several limitations. First, the sample was relatively small and may not represent the general population of China. In previous representative Chinese samples, there were slightly more men than women, and only about 5% of participants had a university degree or higher (Zhang et al., 2022), which differs from our sample and limits the external validity of our findings. Second, there are several types of tools available to measure environmental restorativeness, but they are often used in experimental designs with specific scenarios (Han, 2018). Due to feasibility constraints, we used simple items provided by previous studies to assess the general and broad neighborhood environment, which may limit the effectiveness of the measurement. Third, there is currently no established questionnaire specifically for assessing perceived ALAN. We considered the dimension of brightness, but other aspects such as color, glare, and distribution were not measured (Wang et al., 2023a; Wu et al., 2023), potentially limiting our understanding of subjective ALAN. Fourth, our mediation models were based on cross-sectional data, meaning the effects in the model should only be interpreted as associations rather than any causal influences (Dzhambov et al. 2023; Dzhambov et al., 2019). The assumed regression order (e.g., subjective ALAN predicting nighttime physical activity) is based on theoretical reasoning, but the real association could be reversed or even bidirectional (e.g., people who enjoy nighttime physical activity are more likely to notice well-lit neighborhood areas, which in turn further increases their likelihood of going out at night). Nevertheless, cross-sectional mediation is still a valuable starting point (Stewart et al., 2021), providing a direction for future exploration of causal pathways. Future research should use more causally indicative mediation techniques, such as longitudinal mediation, within-subject mediation, and instrumental variable mediation to explore the process through which ALAN affects nighttime physical activity. Fifth, our sample recruitment was primarily online. While this is an efficient and reliable survey method (Fricker and Schonlau, 2002; Ritter et al., 2004), it can introduce bias (Zhang et al., 2017), and the accuracy of the physical location data may have been compromised. Moreover, individuals who did not use the internet or smartphones could not participate in the survey, reducing the representativeness of the sample. Finally, to align all variables within the same time frame, we limited the time period for the subjective ALAN and nighttime physical activity surveys to the past month, and participants were only required to have lived in their reported residences for the past month. We did not collect data on the length of time participants had lived in the area. This may overlook cumulative environmental effects. Future research should consider the length of residence and the long-term impact of ALAN on physical activity (e.g., over a period greater than one year). Given the novelty and interdisciplinary nature of this study, future research should draw on insights from related fields such as transportation, economics, energy, and consumer behavior to enrich the theoretical framework surrounding the nighttime environment (Bhuniya et al., 2023; Datta et al., 2025; Guchhait et al., 2024; Sarkar et al., 2024).

Conclusion

This study explored the relationship between ALAN and nighttime physical activity in China, as well as the mediating roles of environmental restorativeness and perceived safety. We found that a large proportion of participants reported primarily engaging in physical activity at night and time constraints were the main reason for doing so. We found no significant association between objective ALAN and nighttime physical activity. However, within higher ranges, subjective ALAN was significantly positively associated with various intensities and domains of nighttime physical activity, with environmental restorativeness acting as the main mediator of this relationship. These findings highlight the importance of nighttime lighting in providing psychological restoration and opportunities for physical activity. Policymakers should focus on optimizing lighting design and parameters to enhance urban restorativeness, ultimately promoting public health. However, it is important to note that this study is based on cross-sectional data and purely theoretical assumptions. Future research should employ more causally indicative methods to validate these findings, such as conducting experimental studies to explore the impact of ALAN and using longitudinal models to investigate the strength of these mediation effects over time. We also recommend that future scholars use well-established tools with better measurement characteristics and recruit more representative samples to improve both the internal and external validity of related findings.