Network structure of insomnia symptoms in shift workers compared to non-shift workers

ABSTRACT Insomnia is a commonly occurring sleep problem in shift workers. So far, no studies have investigated how insomnia symptoms present differently in shift workers and non-shift workers. The purpose of this study was to compare the network structures and centrality indices of shift and non-shift workers using network analysis and network comparison test. Participants included 1339 hospital employees, where 542 were shift workers and 797 were non-shift workers. Overall, a significant difference between network structures were observed. In particular, daytime dysfunction emerged as a strongly connected symptom in shift workers, as evidenced by strength centrality. Increased use of sleeping medication and decreased habitual sleep efficiency were more strongly associated with increased daytime dysfunction in shift workers. Sleep latency and sleep quality were also more strongly linked in shift workers. These results are in part attributable to differing causes of insomnia in shift and non-shift workers. Furthermore, the results indicate that shift workers are more vulnerable and susceptible to changes in sleep-related indices, such as sleep efficiency and latency. The findings suggest that certain insomnia symptoms are more consequential in shift workers, emphasizing the need for a differentiated approach in treating insomnia according to shift work.


Introduction
Insomnia symptoms are commonly observed in the general population, with a prevalence of 17% in the South Korean general population (Ohayon and Hong 2002). A significant proportion of population suffers from chronic insomnia (Johnson et al. 2006), which severely impairs quality of life. As its prevalence is constantly increasing (Chung et al. 2020), it is important to address risk factors associated with insomnia.
One factor that is often discussed to be strongly associated with insomnia is shift work (Vallieres et al. 2014). Shift work is defined as any type of work with work hours outside the traditional 9 a.m. to 5 p.m. (McMenamin 2007). Shift work is essential in jobs that require a 24-hour service, such as healthcare. However, shift work disrupts circadian rhythm, which poses many risks and vulnerabilities to various health problems, including insomnia (Drake et al. 2004;Vallieres et al. 2014). One study reported prevalence of insomnia symptoms in nurses working shift rotations to be around 28.9% to 40.7% (Flo et al. 2012). Commonly associated insomnia symptoms of shift workers include difficulty in initiating sleep, nocturnal awakening, daytime sleepiness, and poor overall sleep quality (Cheng and Cheng 2017;Ohayon et al. 2002;Yong et al. 2017). In addition, insomnia in shift workers is attributed to be a major cause of work-related accidents (Ohayon and Hong 2002), which can be critical in healthcare settings. Factors that affect insomnia symptoms in shift workers were reported to be off-hours between work shifts (Eldevik et al. 2013), anxiety levels (Leyva-Vela et al. 2018), gender, and circadian type (Flo et al. 2012;Jeon et al. 2017). Other factors that were more prevalent in shift workers include stress and anxiety (Kim et al. 2019). Insomnia symptoms in shift workers also exacerbate existing physical and mental problems, such as chronic pain, anxiety, and depression (Vallieres et al. 2014).
An analysis technique called the network analysis has recently been gaining popularity due to its usefulness in investigating interactions between symptoms. The network perspective conceptualizes symptoms of mental disorder as 'nodes" and relationships between symptoms as "edges," which can be either positive or negative, depending on the type of association (Borsboom and Cramer 2013). The resulting network provides a macroscopic perspective, allowing comparison of symptom presentation and interactions. In addition, network analysis also produces a unique index termed "node centrality." Node centrality can be broadly defined as an index of symptom importance, although such symptoms may not be commonly reported hallmark symptoms of a particular disorder (McNally 2016). To this regard, network analysis is advantageous in investigating symptom interactions, and identify significant symptoms that bridge with other symptoms. Moreover, network comparison test (Van Borkulo et al. 2017) also allows comparison of network structures, centrality indices, and edges between multiple networks. Therefore, network analysis is a suitable tool for investigating the difference between relationships of insomnia symptoms in shift and non-shift workers. Several previous studies have suggested central symptoms in insomnia using network analysis (Bai et al., 2021;Qiu et al. 2021); however, few have investigated insomnia symptoms in shift workers from a network perspective.
Previous studies that compared shift workers and non-shift workers do not provide information on how symptom presentations may differ given a condition where both groups suffer from sleep problems. To address this issue, the current study aimed to investigate and compare networks of insomnia symptoms in shift workers and non-shift workers, both with high levels of insomnia symptoms. This would allow identification of unique insomnia symptoms in shift workers that are otherwise not present in non-shift workers.

Study design and participants
This study was cross-sectional, where data were obtained from the Sleep and Related Health Problem Survey at Seoul National University Bundang Hospital Sleep Center, Korea. The survey was conducted as per employee health promotion program, where employees of the hospital voluntarily participated. Participants included nurses, technicians, other paramedics, and office workers. Surveys were conducted in May 2015 over the course of 4 weeks. A total of 1807 participants completed the survey. In accordance with the aim of this study, only the data of poor sleepers were investigated. Therefore, participants that scored PSQI global score 5 or above were selected (Buysse et al. 1989), resulting in 1339 participants. Of the 1339 participants, 542 were shift workers (121 males and 421 females) and 797 were non-shift workers (29 males and 768 females). Written informed consent was waived, and the procedure was approved by the Institutional Review Board of Seoul National University Bundang Hospital (IRB number: B-1607/356-301). This study complied with Declaration of Helsinki.

Pittsburgh Sleep Quality Index (PSQI)
The Pittsburgh Sleep Quality Index (PSQI) is a 19-item self-report scale that measures sleep quality and disturbances over the past month (Buysse et al. 1989). Each item is scored using a 4-point Likert scale ranging from 0 to 3, where 0 indicates "very good" and 3 indicates "very bad." The items are grouped into seven components, where each component represents different aspects of sleep quality and disturbances. The components are subjective sleep quality, sleep latency, sleep duration, habitual sleep efficiency, sleep disturbances, use of sleeping medication, and daytime dysfunction. The sum total of component scores produces the global PSQI score. Global score of 5 or above has demonstrated high sensitivity in distinguishing poor sleepers (Buysse et al. 1989). The Korean version of PSQI used for this study was translated and validated, with internal consistency of 0.84 (Sohn et al. 2012).

Descriptive statistics
Basic information regarding participants such as gender, age, and marital status were investigated. To compare any significant differences between groups, t-tests and Χ 2 tests were conducted accordingly. All analyses for descriptive statistics were conducted using IBM SPSS Statistics for Windows, Version 27.0 (Armonk, NY: IBM Corp.).

Network estimation
Two separate networks for each group (shift workers vs non-shift workers) were estimated using the qgraph package in R (Epskamp et al. 2012). Networks were nondirected as this study was cross-sectional. The resulting networks contained seven nodes each, corresponding to the seven PSQI components. For all nodes included in the networks, item informativeness, redundancy, and predictability were calculated using R package psych (Revelle and Revelle 2015), networktools (Jones and Jones 2017), and mgm (Haslbeck and Waldorp 2020), respectively. As each component score was ordinal, polychoric correlation matrix was used to estimate the network structures. The polychoric correlation matrix was regularized using graphical least absolute shrinkage and selection operator (glasso). A sparse and parsimonious network model was produced with the extended Bayesian information criterion (EBIC) .
For each node, centrality indices of strength, betweenness, and closeness were also calculated for each network. Correlation stability coefficient (CS-coefficient), an index of stability for each centrality index were calculated using the bootnet package . As a result, only the CS-coefficient for strength exceeded 0.25, indicating appropriate stability for statistical interpretation. Strength of a node is defined as the absolute sum of edge weights connecting the node (Bringmann et al. 2019). Thus, high strength centrality of a node indicates that its connections with adjacent nodes are stronger among others.
For each edge, the bootnet package ) was implemented, using non-parametric bootstrap with 1000 samples. This allowed to determine edge-weight accuracies and whether certain edges are significantly stronger than others.

Network comparison test
Next, an exploratory test comparing the networks was conducted using the NetworkComparisonTest package (Van Borkulo et al. 2017). Network comparison test (NCT) conducts hypothesis tests on network structure invariance, edge invariance, and global strength invariance. Network structure invariance is based on the premise that unequal edges between the networks would indicate different network structures. Therefore, it is an omnibus test that investigates overall difference between edges for each network. Significant results for network structure invariance would mean that network structures of symptoms are significantly different. Edge invariance evaluates the difference between specific edges between nodes. Global strength invariance tests the difference between the absolute sum of all edges in networks (Van Borkulo et al. 2017). Centrality invariance tests significant difference of node centrality between the groups. It should be noted that all analyses regarding network comparison were exploratory, and thus lacked any specific hypotheses about certain edges or node centralities. Therefore, multiple correction was not mandatory (Van Borkulo et al. 2022). However, the results are to be interpreted with caution, taking into consideration that additional research are needed to corroborate any significant results.

Network estimation
Item informativeness and predictability for each node in both networks are included in Supplementary Table S1. Item redundancy showed that all nodes were less than 25% statistically correlated with all other nodes within the network. Estimated networks for shift workers and non- shift workers are illustrated in Figure 1. As only the CScoefficient for strength centrality exceeded 0.25 for both groups (0.52 and 0.67 for shift and non-shift workers, respectively), only node strengths were interpreted. Figure 2 illustrates the standardized values of strength centrality for each node in shift and non-shift workers. For shift workers, C1 (sleep latency) and C7 (daytime dysfunction) had the largest strength centrality values. For non-shift workers, C1 (sleep latency), C3 (sleep duration), and C5 (sleep disturbances) had the largest strength centrality values. Further bootstrapping indicated that in shift workers, edges showing significant differences with most other edges were C6 -C7, C4 -C7, C2 -C3, C1 -C2, and C3 -C4. Of those, the listed first three edges had negative associations (C6 -C7, C4 -C7, and C2 -C3) and positive associations for the others. In non-shift workers, edges that showed significant differences with most other edges were C3 -C4, C1 -C5, and C3 -C5. Of those, only edge C3 -C5 had a negative association. Edge differences are depicted in Figure 3.

Network comparison test
NCT for network invariance revealed that the network structures were significantly different (M = 0.37, p < .046). However, no significant difference between global strengths was observed (S = 0.26, p = .49). Edge invariance indicated that edges C1 -C2, C4 -C7, and C6 -C7 were significantly different between the networks (p = .023, p = .003, and p = .003, respectively). C1 -C2 is significantly more positive, and C4 -C7 and C6 -C7 are significantly more negative in shift workers compared to non-shift workers. Centrality invariance test showed that difference between strength was only significant for C7 (daytime dysfunction; p = .029), where strength for shift workers was significantly larger than  non-shift workers. In sum, centrality and edges associated with node C7 (daytime dysfunction) appears to have contributed the most to difference between the networks.

Discussion
This study employed network analysis to produce two networks of insomnia symptoms in shift and non-shift workers, as measured by PSQI. Only those with high levels of insomnia symptoms were evaluated in order to compare differential presentations of insomnia symptoms between the groups. As evidenced by significant results of the network invariance test, shift workers revealed different patterns of insomnia symptoms compared to non-shift workers. Notably, shift workers had a considerably stronger strength centrality of daytime dysfunction (C7). In addition, edges relating to C7, such as those connected to use of sleeping medication (C6) and habitual sleep efficiency (C4), had stronger links in shift workers' network. Sleep latency (C2) and subjective sleep quality (C1) were also more strongly connected.
The main finding from this study was the significant structural difference between networks of shift workers and non-shift workers. In particular, high strength centrality of daytime dysfunction symptom was observed in shift workers compared to non-shift workers. A node with high strength centrality has stronger connections with adjacent nodes, implying that large changes in that node will likely result in large changes in the neighboring nodes. As a result, daytime dysfunction appears to be more strongly linked to other insomnia symptoms in shift workers. This result has two implications. First, it supports the hypothesis that presentation and interactions between insomnia symptoms differ across shift and non-shift workers. Second, it indicates that daytime dysfunction is more frequently and strongly associated with other types of insomnia symptoms in shift workers. In other words, insomnia symptoms in shift workers likely manifest during waking hours, such as feeling sleepy during work, rather than during sleeping hours, such as tossing and turning in bed. This is also supported by significantly higher scores of sleep latency and sleep disturbances components in non-shift workers in our study.
Another important aspect to consider when assessing centrality index of a node is its connected edges. In particular, daytime dysfunction in shift workers was significantly more related to the use of sleeping medication, where increased usage of sleeping medication was related to decreased daytime dysfunction. However, no statistical difference was observed for sleeping medication usage between the shift and non-shift workers. This is indicative of higher effectiveness of sleeping medication in shift workers in reducing daytime dysfunction. One possible explanation for these results is that the causes of insomnia vary. In non-shift workers, insomnia is attributed to hyperarousal (Riemann et al. 2010), dysfunctional beliefs about sleep (e.g., fearing consequences of not sleeping at least 8 hours each night; Morin et al. 2002), and counterproductive safety behaviors as coping strategies (Harvey 2002). On the other hand, insomnia in shift workers is predicted to be a result of circadian rhythm disruption (Wright et al. 2013), which can be alleviated by taking sleep medication to induce sleep at its assigned time. It is worth noting that strength centrality of sleep disturbances (C5) was high in non-shift workers. Sleep disturbance component measures factors that interfere with sleep, such as nocturnal urination, sleep apnea, nightmares, and pain. These symptoms require other types of medication or treatment (Aurora et al. 2010;Gottlieb and Punjabi 2020) and are not easily alleviated, or even exacerbated by using sleeping medications. These findings emphasize the need to investigate work schedule during assessment of sleep problems. Furthermore, it also implies that different treatments are required for insomnia in shift workers and non-shift workers.
Daytime dysfunction was also significantly more negatively linked to habitual sleep efficiency in shift workers. In addition, shift workers showed greater positive relationship between sleep latency and subjective sleep quality. Taken together, these results suggest that daytime dysfunction and subjective sleep quality in shift workers are more heavily influenced by sleep efficiency and sleep latency, compared to those in non-shift workers. This implies that shift workers are more vulnerable, and thus react more sensitively to sleep-related indices such as sleep efficiency and sleep latency. Furthermore, daytime dysfunction could also be affected by other factors, such as decreased access to supporting system due to shift work and stressful working environment. To this regard, future studies may investigate mechanisms of how some factors affect daytime dysfunction in shift workers.
The limitations of this study are as follows. First, to investigate those with insomnia, only participants with PSQI total score 5 or above were selected. While this allows to exclude healthy population without insomnia symptoms, it also results in a sample that potentially suffers from a variety of sleep disorders, warranting caution when interpreting the results. To this regard, future studies may investigate how symptoms would present differently in those diagnosed with the same sleep disorder. Second, because the participants were hospital employees, the majority were females in their twenties. This presents a challenge for generalization of the results on males. Third, current results could have been enriched by also considering the shift work schedules. Depending on the work schedule, some shift workers may face greater stress compared to others, which can contribute to difference in results. Finally, it should be noted that correction for multiple comparisons was not applied for NCT. While multiple comparison correction is not mandatory in the case of exploratory analysis (Van Borkulo et al. 2022), our results should be interpreted with care, and further studies are necessary to corroborate our results.
To summarize, this study's results showed that the networks of insomnia symptoms in shift workers and non-shift workers differ. In particular, the main symptom of shift workers was daytime dysfunction, suggesting this symptom to be a core target symptom to reduce other symptoms within the network. Additionally, its stronger connection with sleeping medication was explained by the fact that shift and non-shift workers experience insomnia for distinct reasons. This necessitates different treatments depending on whether one has a shift working job. Moreover, stronger effects of sleep efficiency and sleep latency on daytime dysfunction and sleep quality in shift workers suggest increased vulnerability and sensitivity to sleep-related symptoms in shift workers. Taken together, the findings from this study holds important clinical implications, highlighting the effect of shift work on how insomnia symptoms may present, indicating the need for different types of treatments according to shift work occupation.

Disclosure statement
No potential conflict of interest was reported by the authors.

Funding
The author(s) reported there is no funding associated with the work featured in this article.