Overall, 754 individuals (62% female, mean age = 48.69 years; SD = 14.00; range 18–92) from the German nationwide online panel PsyWeb (N = 12.317 in 2017; https://www.uni-muenster.de/PsyWeb) participated in the study. PsyWeb is a panel that provides information about psychological topics of common interest and offers possibilities to take part in anonymous psychological tests and studies for registered members from the general population (i.e., panel members). Panel members were contacted via e-mail by the panel organization and were invited to take part in an online survey study investigating influencing factors on sleep and sleep problems. We specifically invited all panel members, independent of existing sleep problems or previous traumatic experiences. Study participants did not receive any monetary compensation but were offered automated feedback regarding their sleep quality and depression scores after completion of the survey. Participants were included if they were 18 years or older and proficient in the German language, but were excluded from all analysis if they did not give written informed consent.

It is worth noting that the data collected in this study was also used to validate the German version of the Fear of Sleep Inventory-Short Form (FOSI-SF; Drexl, Kunze, & Werner, 2019). Both projects were preregistered specifying their different research foci and analytic approaches (Kunze, Drexl, Metodiev, & Werner, 2017; Werner, Metodiev, Drexl, & Kunze, 2017).

The survey included several measures assessing FoS, insomnia symptoms, nightmares, traumatic experiences, PTSD symptoms, and other aspects of trauma-related sleep disturbances, like arousal and a proxy for fragmented REM sleep, with higher scores indicating increased symptomatology. Traumatic experiences and PTSD symptoms were measured by the German version of the Life Events Checklist (LEC-5, including the extended criterion A assessment), followed by the PTSD Checklist for DSM-5 (PCL; range 0-80; Krüger-Gottschalk et al., 2017) if any traumatic experience was indicated by the participant. In the present sample, internal consistency for the PCL was excellent (Cronbach’s α = .95). Insomnia severity was measured via the German version of the Insomnia Severity Index (ISI; range 0-28; Gerber et al., 2016). It assesses difficulties with initiating or maintaining sleep as well as early morning awakenings and related worries, and there is good internal consistency in our sample (α = .84) and a clinical cut-off at 15 for moderate insomnia. Furthermore, nightmares were assessed using the German version of the Nightmare Distress Questionnaire (NDQ; range 13-65; Böckermann, Gieselmann, & Pietrowsky, 2014) with excellent internal consistency in the present sample (α = .91). Additionally, FoS was measured via the German version of the FOSI-SF (Drexl et al., 2019). The FOSI-SF contains 13 items (range 0–52) on the fear of being particularly vulnerable during sleep, fear of experiencing nightmares, fear of darkness, and related behaviors, such as sleeping with lights on. The FOSI-SF showed good internal consistency in this sample (α = .86). Further measures linked to trauma-related sleep disturbances included the German versions of the Pre-Sleep Arousal Scale (PSAS; range 15-75; Gieselmann, de Jong-Meyer, & Pietrowsky, 2012; somatic arousal [8 items]: α = .80; cognitive arousal [7 items]: α = .92) as well as Nocturnal Mentations as a proxy for REM sleep fragmentation (NMS; range 3-27; Wassing et al., 2016); however internal consistency was questionable for this 3-item scale (α = .63). Depression and anxiety were assessed for exploratory purposes using the German versions of the depression module of the Patient Health Questionnaire (PHQ-9; range 0-27; Löwe, Spitzer, Zipfel, & Herzog, 2002) and the General Anxiety Disorder Screener (GAD-7; range 0–21; Löwe et al., 2008).

Only participants with at least one potentially traumatic experience (according to DSM-5) – and therefore valid values for PTSD symptom severity (PCL) – were further included in the analyses. Potentially traumatic experiences were defined on the basis of the LEC-5 and the extended criterion A assessment if one of the traumatic events was personally experienced or witnessed (Weathers et al., 2013). However, if the indicated index traumatic event for the PCL did not include any of the following criterion A characteristics, the participant was assigned to the no-trauma group and excluded from further analyses. The criterion A characteristics were: Danger of life, serious injury, sexual violence, or – in the case of the death of a close family member – accident or violence. After exclusion, the remaining sample consisted of 471 trauma-exposed individuals (see Table 1 for demographic variables of the sample).

All analyses were carried out using the Statistical Package for the Social Sciences (IBM SPSS Statistics, Version 24). Cluster analysis is a data-driven approach seeking to identify specific subgroups of individuals within a larger sample on the basis of shared characteristics. Specifically, cluster analyses aim to group individuals that are similar to each other on specified variables into distinct groups. In the present study, cluster analyses were used to explore different symptomatic profiles of trauma-exposed individuals. In order to identify subgroups within our sample, we first performed a non-hierarchical two-step cluster analysis. This type of cluster analysis is advantageous, as it automatically uses standardized variables and chooses the optimal number of clusters based on Schwarz’s Bayesian Criterion (BIC) and the ratio of distance measures (Schendera, 2010); in this case the Euclidian distance measure was used. The resulting cluster quality was automatically rated based on the silhouette measure for cohesion and separation. For PTSD symptomatology we used the PCL score without items referring to sleep (“PCL-”, i.e., PCL without items 2 and 20) to decrease overlap with other measures assessing sleep-relevant variables.

As non-hierarchical cluster analyses only detect main clusters, we also conducted a hierarchical cluster analysis, using Ward’s method (Ward, 1963). This method, which has been broadly used in the social sciences (Clatworthy, Buick, Hankins, Weinman, & Horne, 2005), seeks to minimize the total within-cluster variance (leading to more homogeneous subgroups) and tends to create approximately equally sized, non-overlapping clusters (Schendera, 2010). Using this agglomerative approach, each individual initially represents its own cluster and clusters then progressively merge with others (as a function of their relative distance, i.e., the squared Euclidian distance) until one cluster including all cases is formed. The ideal number of clusters was determined by inspection of the resulting dendrogram and agglomeration coefficients, where a large increase between two consecutive cluster solutions indicates an unfavorable combination of two heterogeneous clusters and should therefore be abandoned. If the dendrogram and agglomeration coefficients supported more than one cluster solution, they were all treated as final solutions and further examined. Differences in clustering variables between the resulting clusters of the final cluster solution were then explored via subsequent analyses of variance (ANOVAs) followed by post-hoc analyses. Differences in secondary and demographic variables were investigated for exploratory purposes.

The psychometric characteristics of the sample with regard to variables used in the cluster analyses are given in Table 2. PTSD symptomatology is reported both overall (PCL) and without items referring to sleep disturbances (PCL-), as the latter was used for cluster analyses.

The lowest BIC (2067.80) and the largest ratio of distance (1.39) both supported a 2- cluster solution, which was automatically chosen. Cluster quality was rated as good, based on the silhouette measure for cohesion and separation. Cluster 1 was characterized by low values of all variables (“healthy cluster”; n = 418), and cluster 2 was characterized by high values of all variables (“clinical cluster”; n = 53) (see Table 3). Subsequent t-tests revealed significant differences between the two clusters on all grouping variables, ps ≤ .001.

The dendrogram of the hierarchical cluster analysis using Ward’s method and squared Euclidian distances showed possible solutions of two, three, and four clusters (see Figure S1, Supplementary Material). There was a smaller increase in agglomeration coefficients between the 4- and 3-cluster solutions (155.73) and a larger increase between the 3- and 2-cluster solutions (525.09), indicating a stronger increase in the heterogeneity within clusters between the latter solutions (see Table S1, Supplementary Material). Therefore, the 2-cluster solution was abandoned and the 3- and 4-cluster solutions were further described.

The 3-cluster solution revealed two bigger clusters and one smaller cluster (Cluster 1: n = 199; Cluster 2: n = 223; Cluster 3: n = 49). One-way ANOVAs and Scheffé’s post-hoc comparisons indicated that all clusters differed significantly from each other regarding all variables (see Table 4 for descriptive values and inferential statistics). Specifically, Cluster 1 was characterized by low levels, Cluster 2 by medium levels, and Cluster 3 by high levels of all variables. Based on these results, and taking the clinical cut-offs for insomnia (ISI ≥ 15) and PTSD symptoms (PCL ≥ 33) into consideration, these clusters were named “healthy cluster”, “subclinical cluster”, and “clinical cluster” (see Table 4). In line with the two-step cluster analysis, the clinical clusters of both analytic approaches are comparable with respect to cluster size (n = 53 vs. n = 49) and mean scores of all variables.

Regarding the 4-cluster solution, the clinical cluster was further divided into two clusters (Cluster 3: n = 38; Cluster 4: n = 11). One-way ANOVAs and Scheffé’s post-hoc comparisons were again used to explore differences between the identified clusters regarding all clustering variables as well as some additional exploratory and demographic variables (see Table 5 for descriptive values and inferential statistics). In line with the 3-cluster solution, Scheffé’s post-hoc comparisons demonstrated that both clinical clusters were characterized by significantly higher levels of all variables compared to the subclinical cluster and the healthy cluster; only nocturnal mentations did not differ significantly between one of the clinical clusters (Cluster 4) and the healthy cluster (Cluster 1), M_Diff = 0.92, p = .948 (see Table 5). However, comparing both clinical clusters, Cluster 4 additionally showed significantly increased levels of somatic pre-sleep arousal, M_Diff = 3.41, p = .033, as well as much higher levels of FoS than Cluster 3, M_Diff = 18.16, p ≤ .001 (see also Figure 1 for cluster group mean z-scores). Moreover, the amount of variance in FoS that can be explained by clustergroup membership (η²) was much higher in the 4-cluster solution (η² = .64) than in the 3-Cluster solution (η² = .33), while the amount of explained variance of all other variables did not differ between solutions. Therefore, Cluster 4 was named “clinical cluster with FoS”, while the third cluster remained a more general “clinical cluster”.

Additionally, one-way ANOVAs and Scheffé’s post-hoc comparisons were used to explore differences between the final four clusters regarding secondary and demographic variables that were not used as clustering variables (see Table 5 for descriptive values and inferential statistics). Both clinical clusters were characterized by significantly higher levels of depression and anxiety compared to the subclinical cluster and the healthy cluster. However, comparing both clinical clusters, no differences in the levels of depression and anxiety were found. Considering age, only a difference between the healthy and the clinical cluster was found indicating higher age in the healthy cluster. Furthermore, there was a significant association between cluster group membership and gender, χ²(3) = 23.67, p < .001. Overall, clusters with higher symptom severity were associated with female gender (healthy cluster: 53.8% women; subclinical cluster: 70.7% women; clinical cluster: 84.2% women; clinical cluster with FoS: 90.9% women).

Some limitations must be considered when interpreting the current findings. First, although the LEC-5 and specific items regarding the index traumatic experience were used to identify trauma-exposed individuals according to PTSD criterion A as defined in the DSM-5 (Weathers et al., 2013), traumatic experiences and PTSD symptomatology were based solely on online self-report measures. Second, this is the first study to classify individuals on the basis of FoS, insomnia symptoms, nightmares, arousal, a self-report proxy for REM sleep fragmentation, and PTSD symptoms. Although meaningful cluster solutions were found, we were not able to validate cluster stability and meaningfulness of cluster membership with an external criterion. Therefore, it is essential to replicate and extend the present findings in diagnosed PTSD samples. For this purpose, we are currently collecting data in PTSD patients before trauma-focused treatment with a twofold aim: 1) to investigate whether the results of the current study also hold for clinically diagnosed PTSD samples and 2) to validate the meaningfulness of the identified subgroups by using treatment outcome as an external criterion. Third, somatic pre-sleep arousal and sleep difficulties were only measured via self-report. Especially in sleep research the use of self-reported versus objectively measured sleep is an often discussed topic. However, the subjective “sleep quality” experience seems to cover aspects that cannot be exhaustively captured via objective indices yet (see Krystal & Edinger, 2008, for a discussion on this topic). The diagnosis of insomnia disorder is currently only based on subjective complaints (e.g., Harvey & Spielman, 2011), therefore focusing on subjective indices in clinical studies is a common approach. Fourth, the 3-item measure of nocturnal mentations, which was used as a self-report proxy for REM sleep fragmentation, showed low internal consistency and was the only variable that did not consistently differentiate between the healthy, subclinical, and clinical clusters. To increase the validity of this self-report proxy, future research should include physiological measures of arousal and REM sleep fragmentation. Finally, medication and substance use as well as other sleep disturbances that might occur in PTSD (e.g., sleep apnea, parasomnias, and disruptive nocturnal behaviors) were not assessed.

In sum, the data-driven, cluster-analytic approach used in this study clearly supports FoS as an important characteristic and possible additional treatment target of trauma-related sleep disturbances in individuals with PTSD. Current standard trauma-focused and/ or sleep-focused treatments seem to only moderately reduce trauma-related sleep disturbances, and residual sleep symptoms often remain. The present data support the proposition that FoS might offer an important construct involved in the development and maintenance of sleep disturbances after exposure to a traumatic event, at least in a subgroup of individuals suffering from PTSD. However, research about FoS is still in its infancy and additional studies are needed to investigate whether directly targeting FoS during treatment – particularly in PTSD subgroups with high FoS scores – might enhance treatment efficacy.