Investigating Sleep, Stress, and Mood Dynamics via Temporal Network Analysis

Abstract

Objective/Background:

Prior research has emphasized the bidirectional relationships between sleep, stress, and affective states, such as depression. Given the inherent variability and fluctuations associated with sleep, assessing how sleep and affective variables function within a dynamic system may help further uncover possible causes and consequences of sleep disturbances, as well as find candidate targets for intervention. To this end, we examined dynamic relationships between self-reported stress, depressed mood, and clinically-relevant sleep parameters via temporal network analysis.

Methods:

Participants were 401 nurses (92% female, 78% White, M_age = 39.47 years) who completed 14 days of sleep diaries incorporating self-reported stress and depression, as well as total sleep time, sleep efficiency, sleep onset latency, and wake after sleep onset.

Results and Conclusions:

Overall, total sleep time emerged as a highly influential variable in the context of “outstrength centrality,” meaning total sleep time had numerous outward connections with other variables (e.g., stress and sleep efficiency). The high outstrength centrality of total sleep time suggests this variable is a source of activation within this dynamic system. Conversely, stress showed high “instrength centrality,” suggesting this variable was highly impacted by other variables in the system, such as depressed mood and sleep efficiency. These findings emphasize the importance of assessing unfolding sleep processes within a naturalistic setting, and implicate the role of total sleep time in fueling depressed mood and stress. Discussion emphasizes implications of these results for understanding the connections between sleep, stress, and depression as well as clinical relevance of these findings.

Introduction

The interplay between sleep and affective variables, such as stress and mood, is well-documented.^1–3 Insomnia, mistimed sleep, and sleep deprivation are all associated with an increased risk for medical problems, such as cardiovascular diseases, diabetes, and hypertension, ^4–6 as well as various psychological issues, such as major depression and anxiety ^7,8 and posttraumatic stress.^9,10 In particular, insomnia severity is highly associated with depression severity, which, in turn, can further exacerbate insomnia.¹¹ In addition, longitudinal findings suggest that shorter sleep duration detrimentally affects next day mood, resulting in diminished positive affect.¹² Indeed, the relationship between sleep and mood, similar to the relationship between sleep and stress, is inherently bidirectional.^13–15

Existing theoretical and conceptual models take into account these bidirectional relationships. For example, the diathesis-stress model of insomnia posits that stress serves a precipitating factor that interacts with a predisposing factor (e.g., high sleep reactivity) to predict onset of insomnia, with insomnia persisting as a result of perpetuating factors such as irregular and mistimed sleep.¹⁶ In addition, the cognitive model of insomnia states that insomnia is the result of excessive worry regarding poor sleep and its daytime effects, which in turn leads to increased physiological and psychological arousal that consequently interfere with sleep.¹⁷ Further, Kahn and colleagues¹⁴ document the “vicious cycle” of sleep loss and emotion regulation, emphasizing the associated mechanisms (e.g., cognitive and physiological) of how increased negative emotions lead to disrupted sleep, which in turn leads to further impairments in emotional well-being.

Studies examining mood, stress, insomnia, mistimed sleep, and sleep duration often rely on single-timepoint retrospective questionnaires, or more rarely sleep diaries and actigraphy. Limited data can be extracted from questionnaires, but a host of variables can be extracted from daily mood, sleep diaries and actigraphy, such as mean and intraindivual variability of mood, sleep continuity/efficiency (sleep onset latency, wake after sleep onset, terminal wakefulness), timing (bedtime, waketime), and total sleep time. ¹⁸ Means of these indices have been used to delineate differences between clinically depressed patients and controls using sleep onset latency (e.g., patients with clinical depression exhibit greater sleep onset latency),¹⁹ as well as parse nuanced differences between patients with major depressive disorder or primary insomnia using total sleep time (such as depressed patients showing greater total sleep time compared to patients with insomnia).²⁰ Furthermore, shorter self-reported total sleep time has been associated with higher levels of self-reported daily stress.^3,21,22 Other studies show divergent relationships, such that sleep efficiency and sleep onset latency are unrelated to (or unaffected by) social stressors, ²³ although this study used actigraphy to measure these sleep parameters. Only a few studies have examined relationships between intraindivudal variability of sleep indices and depression. Results from these studies highlight the role of total sleep time in that it tends to be associated with improved mood the next day,²⁴ whereas fluctuations and variability in total sleep time are associated with depressed mood and perceived stress.^25–27

Most of the previous studies assessing links between mood, stress, and sleep have primarily assessed these parameters in participants with normally timed sleep; there is limited research examining the relationships between sleep, stress, and affective variables in individuals with mistimed sleep. For example, night shift workers (i.e., those whose work schedules are outside of a typical 9 a.m. to 5 p.m. workday) tend to report more sleep disturbances, ^28,29 as well as greater levels of stress compared to day shift workers.^30,31 As such, this line of research suggests there may be additional moderating factors that may have been ignored in previous studies. Indeed, these findings suggest a need to move to an approach that best embraces the complexity of these parameters and their dynamic relationships with affective variables that may exacerbate sleep difficulties.

Within the past decade, a promising new theoretical and statistical approach to psychopathology, termed the network (or complex systems) approach, has emerged as an alternative view on the structure of psychological disorders and other related psychological phenomena.^32,33 This view stands in opposition to the historical prevailing latent variable approach, which postulates that symptoms of a specific disorder, for example, stem from an underlying unobservable cause (e.g., anhedonia, sleep difficulties, and concentration difficulties are related only as they stem from a “depression” entity). According to the network approach, depression may best serve as a descriptive label for the complex interactions between its individual symptoms. In addition, the network approach also considers external conditions that may impact or trigger these individual symptoms (e.g., stress). For example, consider an instance in which someone has recently lost a loved one. This person may be ruminating over this recent loss, causing them to be unable to fall and stay sleep, growing more weary and tired as time goes on. As a result, this person may fall into a state of hopelessness, helplessness and anhedonia, causing them to withdraw from their loved ones, ultimately developing more virulent depressive symptoms. At night, continued worries may cause more stress and loss of sleep, reinforcing this vicious cycle that keeps this person in a disordered state. As such, the focus of conceptualization and treatment may then shift to an understanding of the factors that maintain this system – for example, the relationship between sleep difficulties and stress or rumination.

Network analysis aims to approximate the tenets of the network approach by statistically examining relationships (or connections, termed edges) between symptoms (or other entities, termed nodes), assessing the extent to which a specific node is highly central (or influential within the network). Centrality metrics are one unique feature of network analysis in that a highly central node is one that is likely to spread activation throughout the network via the edges connecting it to other nodes.³⁴ Dynamic networks have typically been constructed using ecological momentary assessment (EMA) data and have shown promising clinical utility.³⁵ Extending to sleep, dynamic networks may be an ideal method to assess the inherent variability and fluctuations within these processes. For example, prior research has shown that variability in sleep disturbances (e.g., sleep onset latency variability) predict acute suicidal ideation above and beyond other depressive symptoms in a high-risk sample.³⁶ Further, greater sleep variability and nightly sleep fragmentation have been found in patients with heightened HPA-axis activity, ³⁷ as well as in patients and community adults with moderate depressive symptomology.^18,38

The Current Study

The goal of the present study was to understand the associations between nightly self-reported sleep (i.e., sleep timing, sleep onset latency, wake after sleep onset, sleep efficiency, and total sleep time) and daily self-reported stress and depressed mood in a large sample of nurses who completed 14 days of repeated assessments. We examined whether dynamic network analysis could provide additional insights into these processes by examining the complex interrelationships between these factors. Examining these processes within a dynamic network may further uncover the possible causes and consequences of daily sleep disturbances, or which process effectively “holds” the network together. Further, such findings may elucidate which dimensions of sleep are important to prioritize in future clinical interventions. Of note, network analysis is best seen as a hypothesis-generating approach. As such, given the use of network analysis as an exploratory procedure, the data analytic plan was not pre-registered in an independent, institutional registry.

Novelty of Analytic Approach

Traditional dynamic (i.e., longitudinal) network analysis focuses on extending multilevel regression models to multivariate data. In this sense, each variable (i.e., node) in the network predicts one another (and itself) at each time point. When visualized, these networks provide a sense of the uni- or bidirectional relationships between certain nodes. Nodes that have multiple outward projections are often implicated as central within a temporal network, allowing one to infer that this node may be a primary source of activation influencing the state of other nodes. These highly central nodes are often candidate intervention targets in psychopathology networks, with the assumption that intervening on a specific node with reduce the connectivity of the network as a whole. For example, a model may suggest less sleep at one night will predict greater depressed mood the next day. This relationship can be graphically represented by a red arrowhead (indicating a negative association) stemming from “sleep” and leading to “depressed.” Indeed, the utility of representing these associations graphically increases when multiple nodes are conceptualized within a network, allowing the research to more clearly pinpoint inward and outward connections for key variables.

Notwithstanding the ease of interpretation that network models allow, temporal networks are often considered to serve as a helpful proxy for network theory.³⁹ As mentioned previously, the main tenets of the network approach assume that psychological phenomena or constructs (such as depression) are emergent as a result of the causal interactions between the facets that make up the phenomenon of interest. Using cross-sectional networks as proxies for network theory is hampered by these models’ inability to confidently identify possible causal connections,⁴⁰ whereas by modeling temporal associations, the researcher or clinician gains further insight into the dynamic processes which maintain a disordered state. Thus, temporal networks models show great promise at both a nomothetic and idiographic level.

Novelty of Sample

Keeping in mind the conclusions that can be drawn from the findings presented below, another novelty of the present research is that the sample (described below) consists entirely of nurses. Nurses are prone to sleep disturbances due to demanding work environments and frequent changes in work schedules.³¹ Furthermore, nurses report short and poor sleep quality, as well as high levels of stress, which likely exacerbates these sleep-related issues.⁴¹ As such, the inherent variability in this sample is particularly useful, as our analyses will allow for an examination of the differential relationships among various sleep parameters and their associations with other relevant affective variables (i.e., stress and depressed mood).

Methods

Procedure

This study was part of a larger study examining the effects of sleep on antibody response to the influenza vaccine (R01AI128359) that took place between September 2018 and November 2018. Participants for this study were recruited from two Dallas, TX regional hospitals via flyers, employee email systems, and nursing staff presentations directing them to an initial online consent form. Nurses (N = 461) provided online consent and were requested to complete initial online Qualtrics surveys collecting demographic information, as well as retrospective self-report estimates of their recent health status. Next, participants were invited to enroll in the main part of the overarching study in early fall (i.e., at the beginning of the influenza season), which included completion of in-person informed consent approximately one month later. During this process, participants were given instructions on completing stress surveys and sleep diaries. Participants provided daily measures for the next 14 days (N = 401). The study protocol and procedures were approved by the Medical City Plano and University of North Texas Institutional Review Boards. The present study used only diary data (i.e., stress surveys and sleep diaries) from the larger study.

Participants

Inclusion criteria for the larger study were 1) not yet having received the current season’s influenza vaccine; 2) being between the ages of 18 and 65; and 3) that registered nurses were actively working at least part-time at one of two regional hospitals. Exclusion criteria were 1) either being currently pregnant or currently nursing, or planning to become pregnant; and 2) having an egg allergy. See Table 1 for an overview of demographics measures from this sample.

Table 1.

Participant Demographics (N = 401)

Measure
Age (M (SD))	39.47 (11.14)
Gender (%)
Male	32 (8.0)
Female	369 (92.0)
Marital Status (%)
Married	253 (63.1)
Single	105 (26.2)
Divorced	33 (8.2)
Separated	7 (1.7)
Widowed	3 (0.7)
Race (%)
White	312 (77.8)
African-American/Black	27 (6.7)
American Indian/Alaskan Native	6 (1.5)
Asian	42 (10.5)
Multiracial	7 (1.7)
Other	7 (1.7)
Ethnicity = Hispanic/Latinx (%)	43 (10.8)
Night Shift Worker (%)	106 (26.4)
Part-time Employment (%)	26 (6.5)
ISI Total (M (SD))	5.77 (4.50)
PHQ-9 Total (M (SD))	3.64 (3.97)
GAD-7 Total (M (SD))	2.80 (3.48)

Note. ISI = Insomnia Severity Index ⁷⁰; PHQ-9 = nine-item Patient Health Questionnaire ⁷¹; GAD-7 = seven-item Generalized Anxiety Disorder questionnaire.⁷²

Measures

Daily Sleep Diary-Determined Sleep

An electronic version of the Consensus Sleep Diary – Core ⁴² was completed by participants each morning upon awakening using REDCap.⁴³ Diaries were used to compute the following variables used in analyses: sleep onset latency (time to fall asleep), wake after sleep onset, and terminal wakefulness (amount of time spent in bed after the final awakening), total sleep time (time in bed [with the intention of sleeping] minus the sum of sleep onset latency, wake after sleep onset, and terminal wakefulness), and sleep efficiency (total sleep time divided by time in bed, multiplied by 100). Lastly, to account for sleep timing, we constructed a “midpoint” node representing the halfway point between bedtime and waketime (expressed numerically). For example, if a participant went to bed at midnight and woke up at 6 a.m., then the midpoint value would be 3.0.

Daily Stress and Mood

Upon awakening, participants reported their stress severity levels during the previous day using the item “I felt stressed,” rated on a scale ranging from 0 (“not at all”) to 4 (“extremely”). Similarly, participants also reported their feelings of depression using the item “I felt down, depressed, or hopeless” on a similar Likert-type scale. Previous work investigating stress and sleep using diary have used similar single-item measures of stress.⁴⁴ Similarly, previous network investigations have also relied on single-item indicators of depressed mood.⁴⁵ The present analyses included stress and depressed mood as nodes.

Statistical Analysis

Network Estimation and Visualization

We analyzed our 14-day longitudinal data via the multilevel vector autoregressive (mlVAR) model, a commonly-used method to analyze multivariate time series data from multiple subjects.⁴⁶ Importantly, when examining repeated measures from multiple subjects, it is likely that consecutive responses are dependent on one another (e.g., sleep difficulties at one measurement occasion predict sleep difficulties at the next measurement occasion), thus violating typical statistical assumptions of routinely-used statistical analyses (e.g., repeated measures ANOVA). mlVAR models often employ “Lag-1” factorization to estimate temporal effects within an autoregressive model. The autoregressive model explicitly models the non-independence of repeated measures data, as this model regresses a variable at time t on a lagged version of that variable (measured at the previous time point, t-1), which is a feature of Lag-1 factorization. Importantly, Lag-1 factorization indicates that measurements are independent given only the previous case.⁴⁶ “Lagging” in time series refers to specifying a fixed amount of passing time (e.g., an interval). Lag-1 factorization estimates the effect of one variable on another one time interval forward. Put another way, a variable at one time point is only predicted by the variable precedes it in the time series, which itself is predicted by previous time point’s variable. For example, variable Y at time 1 predicts variable Y at time 2, which predicts variable Y at time 3, and so forth. This specific effect is known as an autoregression, wherein one variable predicts itself at subsequent time points. Variable X at time 1 predicting variable Y at time 2 (and so on) would be represented by cross-lagged effects, which are commonly represented by the edges between nodes in the temporal network figures.⁴⁶ An alternative, Lag-0, assumes that repeated measures are independent of one another, which is usually implausible in time series designs (e.g., a participant’s mood assessed at 12 p.m. is likely highly correlated with their mood assessed at 8 a.m. and/or at 4 p.m). A vector autoregressive (VAR) model is simply a multivariate extension of the univariate autoregressive model. That is, in the vector autoregressive model, all variables are regressed on a lagged version of the same variable and all other variables. A multilevel extension to the vector autoregressive model (mlVAR) allows these parameters of this model to vary randomly across individuals. Averaged parameters from the multilevel model represent “fixed effects,” and subject-level deviations from these fixed effects are noted as “random effects.” As such, both fixed and random effects inform separate levels of analysis, with the former showing average intraindividual effects, and the latter showing individual differences.

The data were analyzed using the R package mlVAR, ⁴⁷ which estimates three networks: (1) a temporal network, which is a directed network of regression coefficients depicting lagged associations between symptoms from one measurement occasion to the next; (2) a contemporaneous network, which is a Gaussian graphical model (GGM) depicting associations between variables that remain after controlling for temporal associations; and (3) a between-subjects network, a GGM depicting the variance-covariance structure of the participants’ means. ^46,48 As such, each of these three networks allows for different insights in the covariation among symptoms and potential dynamics between these symptoms. For example, the temporal network provides insight into Granger causality, ⁴⁹ where past values of variable X contain information that helps predict current values of variable Y, above and beyond the predictive information contained in past values of Y itself.³⁹ By default, mlVAR employs within-person (aka person-mean) centering to estimate temporal models. In multilevel models, person-mean centering rescales values such that they are representated as deviations from the individual’s mean, which helps to disaggregate within- and between-subjects effects.³⁹ More specifically, mean values are rescaled to “0” such that any other value now represents a deviation from the subject’s mean (in either direction, positive or negative). As an example, consider the utility of within-person centering for the variables in this study. If total sleep time is rescaled to have a mean of 0 for each participant, then one can readily assess fluctuations in total sleep time across the course of the study. Assessing reduced sleep time predicting greater sleep time repeatedly throughout the assessment period would be noted by a negative autoregressive effect in the temporal network (e.g., a “bad” night of sleep is usually followed by a “better” night in terms of total sleep time). Cross-lagged effects (i.e., associations between variables) can also be similarly estimated throughout the assessment period (e.g., a less-than-average total sleep time the previous night is associated with greater-than-average stress levels the following day). Although this procedure may bias autoregressive parameters (which is explored as a function of this procedure in our supplemental materials), within-person centering in the mlVAR package is used to obtain estimates for the temporal network which are represented as aggregated effects of the lagged associations among all participants.

Contemporaneous networks, on the other hand, provide insight into relationships between nodes that possibly occur during a shorter interval of time. The edges in these networks are based on residuals from their corresponding temporal networks, and as a result, these relationships are often interpreted as being reflective of momentary changes that are otherwise not specified given the parameters of the time series.^39,50 For example, in this study, temporal effects would assess day-to-day associations. Contemporaneous networks, however, may provide insight into processes between these variables that occur over the course of several hours. Importantly, contemporaneous networks are not analagous to cross-sectional networks. Graphical depictions of contemporaneous networks are non-directional but are simply based on information from their corresponding temporal networks. The edges in these networks also depict partial correlations between nodes, which depict associations between nodes within the same window of measurement after controlling for all other variables in the same window of measurement and all variables of the previous window of measurement.⁵¹ As an example, total sleep time may share a connection with sleep effeciency in a contemporaneous network. A positive association in this network suggests that, when controlling for all other variables (e.g., sleep parameters), greater total sleep time would be associated with greater sleep efficiency throughout the same night. Lastly, between-subjects networks depict correlations between the mean levels of variables while controlling for all other variables, providing insight into how, on average, different variables are related to one another.⁴⁶ In a sense, between-subjects networks depict the fixed effects from the mlVAR model.

These networks were visualized using the R package qgraph.⁵² In the present visualizations, blue lines (representing associations or coefficients, termed “edges”) between variables (termed “nodes”) represent positive associations, whereas red edges between nodes represent negative associations. The thicker the edge, the stronger the connection between two nodes. Lastly, node influence was examined via node centrality. For temporal networks, instrength centrality is the sum of all incoming connections to a node from other nodes, suggesting that a node with high instrength centrality is highly impacted by other nodes.³⁴ Outstrength centrality is the sum of all direct connections from a node to other nodes, indicating that a node with high outstrength centrality may serve as a node that is a source of activation in the network.³⁴ Contemporaneous and between-subjects networks rely on node strength centrality, which is defined as the sum of all absolute edge weights connected to a given node.⁵³ The R code for analyzing and visualizing these networks are listed as supplemental materials. In addition, a detailed write-up of imputation procedures to handle missing data and examining assumptions of these statistical models are also included in the supplemental materials. See Figure 1 for a graphical overview of diary assessment in this study and how these data can be used to construct networks via mlVAR.

Figure 1.

Graphical representation of how diary data are used to construct temporal and contemporaneous networks via mlVAR. In this study, (a) repeated measures (14 days) from multiple participants (N = 401) were assessed using REDCap. (b) Data are structured into “long” format, where each row represents a participant’s value for a certain measure at each individual assessment period (denoted by T). (c) These data can then be analyzed via mlVAR to construct a temporal network depicting directed associations between variables used in the network. In this hypothetical example, “TST” assessed at one timepoint predicts “SE” and “Stress” at the following timepoint. A contemporaneous network is constructed based off residuals from the temporal network. In this hypothetical example, the association between TST and SE suggests these variables are associated with one another within the same repeated measures window. See the main text for a more detailed discussion of the models estimated via mlVAR.

Appropriateness of Temporal Network Estimation for Diary Data

As we emphasize further in our limitations at the end of this paper, mlVAR, as well as other researchers^39,54 suggest using this package with at least 20 time points to construct temporal networks. As detailed above, we assessed our parameters over the course of 14 days, suggesting that the number of time points (T) may not be adequate for mlVAR estimation (although this is a testable hypothesis, as we detail below). However, longitudinal network estimation is a burgeoning methodology, with Blanchard and colleagues’ recent review⁵⁴ documenting a wide range of completed time points used to construct longitudinal networks (i.e., from 7 to 140 time points). Further, recent advances to longitudinal network estimation now allows for analyzing panel data (i.e., repeated measures taking place over longer intervals, such as months), although the developers of this procedure recommend no more than 10 time points as convergence of the model will not be possible.⁵¹

Thus, we sought to further assess the appropriateness of a 14-day assessment period for mlVAR estimation via its built-in simulation function (“mlVARsim”). A more detailed write-up of these stimulations is provided in the supplemental materials, along with an explanation of how temporal networks may be biased with fewer time points. Nonetheless, results from the simulations suggest that estimation of a temporal network with 401 participants, 14 time points, and 7 nodes is not unduly biased by a lower-than-recommended number of time points in the data.

Results

Temporal Network

See Figure 2 for the temporal network depicting dynamic relationships between stress, depression, and sleep. See Figure 3 for the centrality plot depicting outstrength and instrength centrality for each node in this network. Total sleep time had direct, negative connections with stress and sleep efficiency, suggesting that shorter sleep was associated with greater self-reported stress, whereas longer sleep was associated with lower sleep efficiency (i.e., more fragmented sleep). Further, total sleep time shared a positive connection with sleep onset latency, suggesting that longer sleep was associated with increased sleep onset latency (i.e., time to fall asleep). Lastly, total sleep time had a positive connection with the midpoint node, whereas the midpoint node shared a negative connection with total sleep time. Overall, total sleep time was highly central in the context of outstrength, suggesting that this node was a prominent source of activation in the network. Conversely, stress was highly central in the context of instrength, suggesting this node was highly impacted by other nodes within the network. Specifically, both depressed mood and sleep efficiency positively predicted stress, whereas total sleep time and the midpoint node negatively predicted stress.

Figure 2.

Temporal network of stress, depression, and sleep dynamics. Blue arrows represent positive associations, whereas red arrows represent negative associations. Thicker arrows represent stronger coefficients. Note. This graph was constructed using the “circle” layout in qgraph for interpretability and thus the spatial distance between specific nodes is insignificant.

Figure 3.

Outstrength and instrength centrality plots (standardized) from the temporal network presented in Figure 1.

Contemporaneous Network

Figure 4 details the direct associations between nodes after controlling for all temporal relationships within the network. As such, contemporaneous connections may indicate relationships that occur within the same window of measurement that cannot otherwise be explained by temporal effects. In this network, stress and depression shared strong positive connections, as did total sleep time and sleep efficiency. Sleep efficiency, however, shared strong negative associations with both sleep onset latency and wake after sleep onset. Notably, wake after sleep onset was not associated with any node in the temporal network. Decomposing the variance in these networks and examining a contemporaneous network allows for insight into the more momentary changes between sleep variables.

Figure 4.

Contemporaneous network of stress, depression, and sleep variables.

Between-Subjects Network

See Figure 5 for an analogue of a between-subjects network depicting partial correlations between nodes while controlling for all other nodes in the network. Due to a possible statistical artifact, a between-subjects network was unable to be estimated from the mlVAR package. A between-subjects network aims to model associations between the means of the subjects’ scores within the time span of measurement. As such, the number of subjects in the time-series is similar to the number of “observations” (here, 401) for a seven-node partial correlation matrix (i.e., the number of nodes used in our temporal and contemporaneous networks). Indeed, this sample size is considered to be relatively modest for a cross-sectional network analysis.⁵⁵ However, to approximate a between-subjects network, we used the bootnet R package to estimate a partial correlation network based on nodes that reflect the average value of each parameter across the 14-day period. Overall, this network closely resembled the contemporaneous network, although this network was more sparsely connected. As in the contemporaneous network, sleep efficiency shared strong, negative connections with wake after sleep onset and sleep onset latency.

Figure 5.

Between-subjects network analogue based on mean levels of nodes throughout the 14-day period of the study. See the main text for further details regarding construction of this network.

Discussion

Summary of Findings

In an experience sampling study of 401 nurses who completed daily measurements across 14 days, using dynamic network analyses, we found that total sleep time and sleep timing were strong drivers of stress, mood, and other sleep variables in the network, and that stress was strongly impacted by other variables in the network. Specifically, later sleep timing was associated with subsequent shorter sleep, whereas longer sleep was associated with later sleep timing. Total sleep time also shared strong, negative connections with stress and sleep efficiency, and a strong, positive connection with sleep onset latency. Total sleep time also evinced a strong, negative autoregressive effect, such that shorter sleep one night was associated with longer sleep the next night, and vice versa. This relationship is consistent with previous findings documenting that sleep fluctuates across short periods of time, particularly in samples with clinically-significant insomnia symptoms.^27,37,56,57 Similarly, our finding that shorter sleep was associated with greater self-reported stress is also consistent with research emphasizing the detrimental effects of sleep deprivation or restriction on stress and mood.^12,58,59 Importantly, however, our study details the directionality of these relationships, revealing that shorter sleep leads to greater self-reported stress, implicating total sleep time as an important source of activation within this network.

The relationship between total sleep time and depressed mood appeared evident via an indirect effect of sleep efficiency, such that longer sleep was associated with more fragmented sleep. This is consistent with the Spielman 3-factor model of insomnia, which posits that one of the main contributors to the development of insomnia is sleep extension/excessive time in bed.¹⁶ In addition, total sleep time had a direct connection with stress, and also appeared to indirectly impact stress through other sleep parameters, such as sleep efficiency. High sleep efficiency may encapsulate recent sleep deprivation, such that the individual falls asleep quickly due to exhaustion. In turn, higher levels of sleep efficiency (reflecting well-consolidated sleep due to this exhaustion) may then exacerbate one’s ability to cope with stress, possibly as a result from recovering from this exhaustion. Overall, these results are consistent with previous findings documenting the indirect effects or mechanisms by which sleep disturbances impact depressed mood; ^60,61 however, the temporal relationships discussed in our study highlight the importance of simultaneously examining dynamic associations between stress, depressed mood, and multiple meaningful clinically-relevant sleep parameters (i.e., not just one index of sleep disturbance).

As emphasized previously, the corresponding contemporaneous and between-subjects network allow for a different level of analysis of the temporal relationships discussed above. Contemporaneous relationships are interpreted as faster-moving processes within the network literature, ⁴⁶ albeit the lack of temporal precedence does not allow for causal insights into these relationships. However, sleep efficiency was a highly central node in the contemporaneous network, suggesting the impact that sleep efficiency had on these nodes (or the impact these nodes have on sleep efficiency) may be reflective of intranight variability. Further, wake after sleep onset had numerous connections in the contemporaneous network, whereas this node was unrelated to any other node in the temporal network. As such, the influence of wake after sleep onset in this network implicates its importance as a sleep parameter affecting other nightly sleep processes. For example, consider wake after sleep onset’s strong, negative connection with sleep efficiency. Throughout the night, multiple periods of wakefulness that occur after sleep onset may continuously diminish the overall efficiency of sleep within the night. Such an association is important to document, as sleep efficiency provides an overall sense of how well the patient slept and does not distinguish frequent episodes of wakefulness.⁶² However, it is important to note that sleep efficiency is calculated using wake after sleep onset, so this may explain their strong association.

Lastly, the between-subjects network was relatively sparse, albeit these networks are helpful in that they depict partial correlations between average levels of nodes (i.e., the relationship between two nodes while controlling for all other nodes in the network). One notable difference in this network was the lack of connections between stress and any other node, whereas stress was highly impacted by other nodes in the temporal network. This difference suggests that, at a mean level (specifically when controlling for all other variables), self-reported stress levels were unrelated to depressed mood or any sleep parameter. Overall, the covariation between means of participants in this model is highly similar to the connections depicted in the contemporaneous network, with this network providing a useful analogue to a cross-sectional analysis of these relationships.⁴⁶

Importance and Relevance of Findings

To our knowledge, this is the first study assessing temporal network dynamics among sleep parameters and self-reported stress and depressed mood. Prior investigations of putative intranight sleep dynamics have relied on electroencephalographic signatures captured in a laboratory environment, ^63,64 but this study is the first to examine unfolding dynamic sleep processes over time in a naturalistic setting. Assessing these causes and consequences of sleep in everyday life enhances the ecological validity of our findings. Further, incorporating multiple sleep parameters within a temporal network allowed us to examine the relative importance of specific sleep parameters and their relationships with self-reported stress and depressed mood. As such, we were able to highlight the importance of total sleep time as a driving force behind the interplay of other sleep variables, stress, and mood.

These findings are also clinically relevant. That is, these findings implicate the role of total sleep time as a potential candidate target of intervention, particularly for patients reporting high stress and depressed mood. Prioritizing an optimal level of total sleep time may have important ripple effects on subsequent psychological functioning. Although empirically-supported treatments for insomnia and related sleep issues already exist (e.g., cognitive behavioral therapy for insomnia), these treatments primarily target sleep efficiency and sleep quality, with less of a focus on total sleep time as an intervention target. Thus, more research is needed that focuses on developing interventions to help better achieve sleep extension in patients with suboptimal total sleep time. Naturally, these findings rely on subjective total sleep time, and it is possible that subjects may over- or underestimate their total sleep time compared to objective metrics.⁶⁵ As such, future research may benefit from examining these temporal relationships with actigraphy or EEG measures of sleep. An additional caveat relates to the effectiveness or ecological validity of intervening on highly central nodes highlighted within temporal networks. A burgeoning line of research has sought to address this issue empirically. For example, Elliot and colleagues⁶⁶ found that baseline central symptoms of anorexia nervosa, such as fear of weight gain, were prognostically indicative of recovery and impairment post treatment. However, the extent to which intervening on highly central nodes and assessing whether the network becomes less dense (i.e., connected) over time still remains an empirical question.

Limitations

This study is not without its limitations. These analyses were conducted using a relatively homogeneous sample of nurses; thus, these findings may not generalize to the population or patients with clinical levels of depressive and insomnia symptomology. Although nurses report frequent insomnia symptoms and work-related stress,^41,67 future research should continue to examine these temporal relationships in other samples. Importantly, and in line with our caveats regarding the use of temporal network analysis for treatment and conceptualization, these findings should be not taken as reflective of idiographic or intraindividual processes. However, network analysis is best seen as a hypothesis-generating approach, and these findings may still be especially helpful when examining relationships between sleep, stress, and mood at a more idiographic level.

In addition, we were somewhat limited by the nature and timing of the daily assessments. That is, nodes were constructed the following morning or day based on information from the previous day or night (e.g., the sleep parameters). Ideally, future research would assess relevant parameters at two different time points (such as mood and stress in the evening and the next morning) to better discern between temporal and contemporeous effects, as well as to reduce common methodological bias of diary assessments. Indeed, there is great heterogeneity in sampling rates across several studies that have employed temporal network analysis for ESM data and at present, there are no formal guidelines as to the ideal number of measurement occasions (or number of subjects) for constructing temporal networks.³⁹ As we detail in our supplemental materials, however, we were able to run native simulations within the mlVAR package as a quasi-post-hoc power analysis to determine whether the subject to measurement occasion ratio was appropriate for this analysis, with these simulations suggesting that these study parameters approximated the “true” temporal network structure. In addition, recent work has suggested that network estimation procedures for panel data (i.e., repeated measures data with fewer measurements across a longer time scale) are too computationally complex for data with more than 10 time points.⁵¹ Thus, given the parameters of our data, we sought to strike a balance between an appropriate level of analysis and the nuances of cutting-edge methodology. Nonetheless, future research may benefit from examining these relationships with a greater number of repeated measures and over a longer period of time. Lastly, replication of these findings over longer periods of time is also highly encouraged (e.g., over the course of three to four weeks). Replication in this manner may help provide a consensus as to the appropriate number of time points for temporal network estimation that includes daily sleep diary data.

Highlights.

• Sleep’s heterogeneous nature may be best deconstructed with network analysis

• Temporal network analysis aims to examine directional relationships between variables over time

• Using diary data, total sleep time (TST) emerged as a strongly impactful variable in a sample of nurses

• Self-reported stress levels were highly impacted by TST and other sleep parameters

Data Availability

The data underlying this article will be shared on reasonable request to the second author of this paper.