Digital Sleep Disruption: Unraveling the Network Structure of Technology Use and Sleep Problems Through Network Analysis

Data were collected from 9443 Chinese adults (54.3% female) aged 18–65 years (M = 28.4, SD = 8.7) through the Credamo platform between September 2023 and January 2024. Participants were recruited using stratified random sampling to ensure demographic representativeness. The stratification variables included age group (18–25, 26–35, 36–45, 46–55, 56–65 years) and geographical region (Eastern, Central, Western, and Northeastern China), with quotas set to approximate the population distribution according to the most recent national census data. Within each stratum, participants were randomly selected from the Credamo panel, which comprises over 2.6 million registered users across China.

The sample size was determined a priori based on recommendations for stable network estimation in psychological research. Following guidelines established by Epskamp et al. [17], we employed Monte Carlo simulations to estimate the minimum sample size required to detect edges with adequate power (≥ 0.80) and precision (correlation stability coefficient ≥ 0.25). Based on these simulations, which assumed a sparse network with approximately 50–60 edges and medium effect sizes, a minimum sample of 500 participants was required. However, to ensure robust estimation of centrality indices and to enable subgroup analyses, we targeted a substantially larger sample. The final sample of 9443 participants provides excellent statistical power and stable network estimates.

Inclusion criteria required regular use of digital devices (> 2 h daily) and proficiency in reading Chinese. Exclusion criteria included: (1) diagnosed sleep disorders (e.g., obstructive sleep apnea, restless leg syndrome, narcolepsy) in the past month, as such conditions represent distinct clinical entities with established pathophysiology that may confound the examination of technology‐related sleep disruption; (2) current shift work or rotating work schedules, which independently disrupt circadian rhythms through mechanisms unrelated to technology use; (3) trans‐meridian travel crossing three or more time zones in the past month, which can induce jet lag and temporary circadian misalignment; and (4) current use of medications known to affect sleep architecture (e.g., sedative‐hypnotics, stimulants, certain antidepressants). These exclusion criteria were implemented to isolate the effects of technology‐related factors on sleep problems while minimizing confounding from established medical conditions and environmental factors with independent effects on sleep.

All measures underwent rigorous translation and back‐translation following established guidelines [18]. For scales originally developed in English, the forward translation was conducted by two bilingual researchers independently, followed by reconciliation and back‐translation by a third researcher blind to the original version. Discrepancies were resolved through discussion with content experts. Responses used 5‐point Likert scales (1 = strongly disagree to 5 = strongly agree) unless otherwise specified.

Participants completed the online survey through the Credamo platform. The survey took approximately 20–25 min to complete. Data quality was ensured through multiple mechanisms: (1) three attention check items distributed throughout the survey (e.g., “Please select ‘Agree’ for this item”); (2) response time monitoring with exclusion of responses completed in less than 8 min (indicating inattentive responding); (3) consistency checks for logically related items; and (4) exclusion of responses with excessive missing data (> 10% of items). These quality control procedures resulted in the exclusion of 847 responses (8.2% of initial completions), yielding the final sample of 9443 participants.

Network analysis was conducted using R software (version 4.2.3) with the qgraph [28], bootnet [17], and mgm [29] packages. A regularized partial correlation network was estimated using the graphical LASSO (least absolute shrinkage and selection operator) method combined with the Extended Bayesian Information Criterion (EBIC) for model selection. This approach produces a sparse network where edges represent unique associations between variables after controlling for all other variables in the network—a critical advantage over simple correlation matrices that conflate direct and indirect relationships.

In this model, known as the Gaussian Graphical Model (GGM), nodes represent observed variables (composite scores from each scale), while edges denote partial correlations between two nodes after accounting for all other nodes in the network. This statistical property is crucial for interpretation: edge weights reflect the unique association between two variables that cannot be explained by any other variable in the network. Positive associations are indicated by green edges, and negative associations by red edges, with edge thickness reflecting the strength of partial correlations.

To assess node importance, we calculated three centrality indices using the centralityPlot function from qgraph [28]: (1) Strength centrality, defined as the sum of absolute edge weights connected to a node, indicating the node's overall connectivity; (2) Betweenness centrality, reflecting the number of times a node lies on the shortest path between two other nodes, indicating its potential as a bridge; and (3) Closeness centrality, calculated as the inverse of the average shortest path length from a node to all other nodes, indicating how quickly changes in one node might spread to others.

Given the relative nature of centrality metrics, we additionally employed the mgm package to estimate node predictability (R²), which quantifies the proportion of variance in each node explained by its neighboring nodes [29]. Unlike centrality indices, predictability provides an absolute measure of interconnectedness that is comparable across studies and indicates how much of each node's variance is accounted for by the network structure.

Network stability and accuracy were evaluated using a case‐dropping bootstrap procedure with 1000 iterations [17]. This procedure estimates the correlation stability (CS) coefficient, which indicates the maximum proportion of cases that can be dropped while maintaining a correlation of 0.7 between the original centrality indices and those from the bootstrap samples. A CS coefficient above 0.25 is considered acceptable, and above 0.5 is considered good. Edge weight accuracy was assessed through 95% confidence intervals derived from nonparametric bootstrapping.

Figure 1 presents the EBICglasso network model estimation for Sleep Problems (SP), comprising 11 nodes. Across the entire network, 53 out of 55 possible edges (96.4%) were nonzero, indicating a highly connected network structure. The strongest positive edge weight in the network was observed between Electronic Device Dependence (EDD) and Online Gaming Addiction (OGA) (r = 0.41, p < 0.001), while the strongest negative edge weight was between Social Media Anxiety (SMA) and Online Gaming Addiction (OGA) (r = −0.19, p < 0.001).

The network model comprises 11 nodes representing technology use factors and sleep problems. Each node is labeled with its abbreviation: SP = Sleep Problems; ST = Screen Time; BED = Before‐bed Electronic Device Use; EDD = Electronic Device Dependency; SMA = Social Media Anxiety; DIO = Digital Information Overload; VSP = Virtual Social Pressure; BLE = Blue Light Exposure; CRD = Circadian Rhythm Disruption; OGA = Online Gaming Addiction; WLDI = Work‐Life Digital Integration. Green edges indicate positive partial correlations, while red edges indicate negative partial correlations. Edge thickness reflects the magnitude of partial correlations (edge weights). The colored ring surrounding each node represents node predictability (R²), indicating the proportion of variance explained by neighboring nodes.

Among all nonzero edges connected to the SP node, the strongest positive edge weights were with Blue Light Exposure (BLE) (r = 0.31, p < 0.001) and Circadian Rhythm Disturbance (CRD) (r = 0.26, p < 0.001), followed by Social Media Anxiety (SMA) (r = 0.20, p < 0.001), Electronic Device Dependence (EDD) (r = 0.12, p < 0.001), and Online Gaming Addiction (OGA) (r = 0.12, p < 0.001). Weaker positive edge weights were found between SP and Virtual Social Pressure (VSP) (r = 0.09, p < 0.001), Work‐Life Digital Integration (WLDI) (r = 0.09, p < 0.001), Bedtime Electronic Device Use (BED) (r = 0.05, p = 0.002), and Screen Time (ST) (r = 0.03, p = 0.041). Digital Information Overload (DIO) showed a small negative edge weight with SP (r = −0.07, p < 0.001).

A systematic comparison between the zero‐order correlation matrix and the partial correlation (edge weight) matrix revealed substantial attenuation of associations when controlling for other network variables (see Table 1). For example, the zero‐order correlation between Screen Time (ST) and Sleep Problems (SP) was r = 0.42, whereas the corresponding edge weight in the network was r = 0.03—a reduction of 93%. Similarly, Bedtime Electronic Device Use (BED) showed a zero‐order correlation of r = 0.38 with SP but an edge weight of only r = 0.05 (87% reduction). In contrast, Blue Light Exposure (BLE) showed a zero‐order correlation of r = 0.52 with SP and maintained a substantial edge weight of r = 0.31 (40% reduction). This pattern indicates that much of the variance shared between distal technology factors (ST, BED) and sleep problems is explained by other variables in the network, particularly intermediate factors like CRD and BLE. The relatively preserved edge weight for BLE suggests that its association with SP reflects a more direct relationship that is not fully accounted for by other network variables.

The edge weight between BLE and CRD was moderate (r = 0.18, p < 0.001), which may appear counterintuitive given the established physiological pathway through which blue light affects circadian rhythms. These findings warrant explanation. First, the network model estimates partial correlations controlling for all other variables, and the shared variance between BLE and CRD is partly explained by their mutual associations with other network nodes (particularly BED and EDD). Second, the BLE scale primarily captured behavioral patterns of light exposure (timing, brightness settings), while the CRD scale assessed subjective experiences of circadian misalignment; these represent different levels of the causal pathway. Third, individual differences in circadian photosensitivity may introduce heterogeneity in the BLE‐CRD relationship. Nevertheless, both BLE and CRD showed strong direct associations with SP, consistent with their complementary roles in technology‐related sleep disruption.

As Table 1 shows, the predictability of node SP was R² = 0.735, indicating that 73.5% of the variance in Sleep Problems could be explained by its neighboring nodes in the network. This high predictability demonstrates that the network structure effectively captures the factors contributing to sleep problems. The mean node predictability across all nodes was R² = 0.65 (range: 0.57–0.74), suggesting that, on average, 65% of the variance in each node could be explained by its network neighbors—indicating a well‐connected and interdependent network structure.

Figure 2 and Table 1 display the centrality indices for each node. The node with the highest strength centrality was Online Gaming Addiction (OGA) (z = 1.52), followed by Blue Light Exposure (BLE) (z = 0.91), indicating that these nodes have the strongest overall connections within the network. The node with the highest betweenness centrality was Blue Light Exposure (BLE) (z = 1.88), followed by Sleep Problems (SP) (z = 1.60), suggesting that these nodes serve as critical bridges through which other variables in the network are interconnected. The node with the highest closeness centrality was Sleep Problems (SP) (z = 1.86), indicating that changes in this node would propagate most rapidly to other variables in the network.

Centrality indices are presented as standardized z‐scores (x‐axis), with higher values indicating greater centrality. The y‐axis lists all 11 network nodes. Strength centrality reflects the sum of absolute edge weights connected to each node. Betweenness centrality indicates how often a node lies on the shortest path between other nodes. Closeness centrality represents the inverse of average shortest path length to all other nodes. Node abbreviations are defined in Figure 1 legend.

The bootstrap analysis indicated good stability of the network estimates. The correlation stability coefficient (CS‐coefficient) for strength centrality was 0.67, for betweenness centrality was 0.44, and for closeness centrality was 0.52, all exceeding the recommended threshold of 0.25 and indicating that the centrality estimates are sufficiently stable for interpretation [17].

The results provided support for all three hypotheses. Consistent with H1, Blue Light Exposure demonstrated the strongest direct edge weight with Sleep Problems (r = 0.31), supporting its role as the primary proximal factor. Consistent with H2, Screen Time, Bedtime Electronic Device Use, Electronic Device Dependence, Virtual Social Pressure, and Work‐Life Digital Integration showed substantially weaker direct edge weights with SP (r range: 0.03–0.12) compared to their zero‐order correlations, indicating that their associations with sleep problems are largely mediated through other network variables. Consistent with H3, Online Gaming Addiction, Digital Information Overload, Social Media Anxiety, and Circadian Rhythm Disturbance showed moderate centrality indices and served as intermediate nodes connecting distal technology factors to sleep problems.

Consistent with our first hypothesis, Blue Light Exposure exhibited the strongest direct edge weight with Sleep Problems and demonstrated high centrality across all three indices. This finding aligns with established chronobiological research demonstrating that evening blue light exposure suppresses melatonin secretion through activation of intrinsically photosensitive retinal ganglion cells. Because melatonin serves as the primary physiological signal for sleep initiation, its suppression delays the onset of sleepiness and prolongs sleep latency, ultimately leading to delayed sleep onset and reduced sleep efficiency [2, 3, 30]. The high betweenness centrality of BLE indicates that it serves as a critical bridge in the network, through which effects of other technology use factors may be channeled.

The strong direct association between Blue Light Exposure and Sleep Problems, even after controlling for all other network variables, suggests that light‐based interventions may be particularly effective for technology‐related sleep disruption. Practical strategies include the use of blue light filtering software, hardware‐based blue light reduction features, and behavioral recommendations to limit screen exposure during the evening hours [31]. The network structure also suggests that reducing blue light exposure may have downstream effects on other network nodes through its bridging role.

The network analysis revealed that Online Gaming Addiction, Digital Information Overload, Social Media Anxiety, and Circadian Rhythm Disturbance function as intermediate factors with direct pathways to Sleep Problems and connections to more distal technology use variables. This finding supports our third hypothesis and provides insight into the mechanisms through which broader patterns of digital behavior translate into sleep disruption.

The significant negative partial correlation between Online Gaming Addiction and Social Media Anxiety (r = −0.19) suggests a potential compensatory relationship—individuals with high gaming engagement may use gaming as a coping mechanism to manage or escape from social anxiety [32]. This interpretation is consistent with research showing that immersive gaming activities can provide temporary relief from real‐world stressors and social pressures. However, gaming addiction independently contributes to sleep problems through mechanisms including pre‐sleep cognitive arousal, time displacement, and disrupted sleep schedules [33].

The positive association between Digital Information Overload and Circadian Rhythm Disturbance (r = 0.21, p < 0.001) highlights the cognitive pathway through which excessive digital information consumption affects sleep. Information overload may maintain cognitive activation during evening hours, interfering with the natural wind‐down process necessary for sleep initiation [34]. This pre‐sleep hyperarousal state characterized by racing thoughts, difficulty disengaging from information streams, and heightened alertness represents a key mechanism through which psychological factors mediate the technology‐sleep relationship.

Social Media Anxiety showed a moderate direct edge weight with Sleep Problems (r = 0.20), indicating that anxiety specifically related to social media use contributes to sleep disruption beyond its associations with other network variables. This effect likely operates through psychological mechanisms including rumination about social interactions, anticipatory anxiety about missing updates, and social comparison processes that elevate emotional arousal before sleep [6, 35]. Importantly, this psychological pathway is distinct from the physiological pathway of light exposure—while light affects sleep through melatonin suppression and circadian signaling, social media anxiety affects sleep through emotional and cognitive arousal mechanisms that impair the psychological wind‐down necessary for sleep initiation.

Supporting our second hypothesis, Screen Time, Bedtime Electronic Device Use, Electronic Device Dependence, Virtual Social Pressure, and Work‐Life Digital Integration demonstrated substantially weaker direct edge weights with Sleep Problems compared to their zero‐order correlations. This pattern indicates that these factors influence sleep primarily through indirect pathways mediated by intermediate variables.

The dramatic attenuation of the Screen Time‐Sleep Problems association (from r = 0.42 to r = 0.03) illustrates a key insight from network analysis: much of what appears to be a direct effect of total screen time on sleep is actually mediated through specific aspects of technology use, particularly Blue Light Exposure and Circadian Rhythm Disturbance. Notably, our Screen Time measure assessed general daily device use across all contexts rather than evening‐specific use, which may explain its weak direct association with sleep problems—the sleep‐disruptive effects of screen time appear to operate primarily through evening‐specific behaviors captured by other variables in the network (BED, BLE). This finding has important implications for intervention design, suggesting that general recommendations to “reduce screen time” may be less effective than targeted interventions addressing specific mediating mechanisms, particularly those involving evening technology use.

Electronic Device Dependence showed moderate direct and indirect effects on sleep problems, consistent with research linking technology addiction to sleep disruption through multiple pathways including time displacement, pre‐sleep arousal, and disrupted routines [36]. Virtual Social Pressure and Work‐Life Digital Integration contribute to sleep problems through their effects on evening technology use patterns, psychological arousal, and boundary management between work and rest periods [37].

The network approach employed in this study advances theoretical understanding of technology‐induced sleep disruption by revealing the interconnected structure of contributing factors. Rather than conceptualizing technology effects on sleep as a set of independent risk factors, the network model suggests a complex system in which physiological mechanisms (blue light, circadian disruption), psychological factors (anxiety, information overload), and behavioral patterns (gaming, device dependence) mutually influence each other and collectively determine sleep outcomes.

For clinical practice and public health intervention, these findings suggest several priorities. First, interventions targeting Blue Light Exposure through technological solutions (filtering software, device settings) or behavioral recommendations (screen‐free periods before bed) may be particularly effective given its strong direct association with sleep problems and its bridging role in the network. Second, addressing intermediate factors such as Social Media Anxiety and Digital Information Overload through cognitive‐behavioral techniques may help interrupt the psychological pathways through which technology use affects sleep. Third, the finding that general screen time shows minimal direct effects on sleep after controlling for specific mechanisms suggests that interventions should focus on modifying specific aspects of technology use rather than advocating for blanket reductions in technology engagement.

Several limitations warrant consideration. First, the cross‐sectional design precludes causal inference about the direction of relationships. While network analysis reveals the conditional independence structure among variables, it cannot distinguish between causal effects, reverse causality, or reciprocal relationships. Longitudinal network analyses employing time‐series data would enable examination of temporal dynamics and more confident causal inference [38].

Second, all measures were self‐reported, which may introduce recall bias and social desirability effects. While we employed validated scales with demonstrated psychometric properties, future research should incorporate objective measures of technology use (e.g., smartphone tracking data, blue light sensors) and sleep (e.g., actigraphy, polysomnography) to strengthen the validity of findings [39].

Third, our sample comprised general population adults recruited through an online platform, which may limit generalizability to clinical populations or individuals without internet access. The exclusion of individuals with diagnosed sleep disorders, while methodologically appropriate for examining technology‐specific effects, means our findings may not generalize to clinical samples. Future research should examine whether the network structure differs in populations with established sleep pathology.

Fourth, the network analysis approach, while providing valuable insights into the structure of relationships among variables, does not capture potential nonlinear effects or threshold phenomena that may characterize technology‐sleep relationships. Future research employing machine learning approaches may complement network analysis by identifying nonlinear patterns and interaction effects.

Fifth, while our analysis focused on technology use as a predictor of sleep problems, the relationship may be bidirectional. Individuals with pre‐existing sleep difficulties may engage in increased digital behavior as a coping mechanism or due to an inability to fall asleep. This reverse causality cannot be ruled out in our cross‐sectional design and warrants investigation through longitudinal studies that can disentangle the temporal sequence of technology use and sleep disturbances.