Data of the nonclinical sample was incidentally derived from online surveys carried out between November 2015 and February 2016, yielding N=1,248 individuals (541 men) aged between 18 and 65 years (M=34.89, SD=13.06). A majority of the subjects indicated to be in a relationship or married (63%) and to have a high school diploma (85%) (Table 1). Data of the clinical sample was incidentally collected between September 2016 and March 2017 from N=575 outpatients (196 men, 19-64 years, M=36.23, SD=12.53) waiting for or currently undergoing psychotherapy. Most subjects were single (57%) and had a high school diploma (28%) or less (48%). Amongst others, the most frequent diagnoses were affective (36%) and anxiety disorders (28%) with most patients having comorbidities as well (Table 1). The samples differed with regard to age (F(1821)=4.24, p=.04) and gender (χ2(1, N=1,823)=248.43, p<.001). As our data collection started before similar PSS versions were introduced (Klein et al., 2016; Reis, Lehr, Heber, & Ebert, 2017), we decided to stay with our version since we aimed at a congruent comparison of our nonclinical and our clinical sample. All subjects indicated German as their mother tongue and gave their written informed consent prior to the study. All studies were conducted in accordance with the Declaration of Helsinki and were approved by the local ethics committee at the University of Mainz.

The PSS-10 was translated and back-translated by two independent native German and English speakers, both with a degree in psychology. Wording incongruities were discussed in an expert committee and conferred with a native speaker. The questionnaire consists of 10 items, each beginning with “In the last month, how often have you …” and has a 5-point Likert scale ranging from 1 (never) to 5 (very often). The PH scale consists of the six negative worded items (1, 2, 3, 6, 9, and 10) and the PSE scale contains the four positive worded items (4, 5, 7, and 8). The total score (PSS-TOTAL) is computed by reversing the PSE items and summing up all items. Higher scores reflect greater levels of perceived stress. Instructions and items can be retrieved from Appendix 1.

To find out about associated constructs in the clinical sample, we used the 53-item Brief Symptom Inventory (BSI; Franke, 2000) with the main subscale Global Severity Index (GSI), as a measure for symptom severity. The BSI has been validated in the German population and internal consistency of the GSI scale is excellent. Furthermore, we asked participants about their overall physical health (1 item) and mental health (1 item) in the last month using visual analogous scales (VAS), both ranging from 1 (very bad) to 9 (very good).

In the nonclinical sample, we also used the VAS named above and a well-being questionnaire to assess a related construct, as perceived stress and well-being are often investigated together (Wersebe et al., 2018). The WHO-5 has been well validated in numerous populations and shows good internal consistency of α=.84 (Bech, Olsen, Kjoller, & Rasmussen, 2003).

CFA was calculated using the Bollen-Stine bootstrapped maximum likelihood (ML) estimation method in SPSS AMOS 23 (Arbuckle & Wothke, 1999). Given the variety of different PSS models claimed in the literature, we tested a (1) one-factor model (overall stress) and (2) a two-factor model (PH and PSE). For model evaluation, we used the comparative fit index (CFI), the Tucker-Lewis index (TLI), the root mean square error of approximation with 90% confidence intervals (RMSEA) and the standardized root mean square residual (SRMR). According to the criteria by Hu & Bentler (1999), values close to or higher than .95 for CFI and TLI; close to or smaller than<.06 for RMSEA and<08 for SRMR indicated a good fit. For competing models, lowest values of the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC) represented measures for parsimony and served as indicators for the favored model.

For group comparisons, we applied hierarchical nested modelling following the MGCFA procedure stated by Vandenberg & Lance (2000). In MGCFA, invariance is tested by restricting different model parameters in a stepwise fashion and comparing the resulting models between several groups (here: clinical vs. nonclinical). If the model fit does not worsen significantly, one can infer that latent group differences are caused by true interindividual differences. Here, we concentrated on equal form, equal factor loadings and equal intercepts and therefore gradually tested configural, metric and scalar invariance (CI/MI/SI) (Putnick & Bornstein, 2016). A more restrictive model was only nested if invariance of the preceding model held. Invariance was considered given if parameter changes for CFI, TLI, RMSEA and SRMR were no bigger than Δ=+/-.01 (Cheung & Rensvold, 2002).

Descriptive analyses (means, standard deviations) as well as Student's t-tests, Spearman's rank correlations and internal consistencies (Cronbach's alpha) and were performed using SPSS 23.0. p-values<05 indicated statistical significance. Correlations of r=.20, r=.50 and r=.80 were interpreted as small, moderate and strong (Cohen, 1992).

Since most studies on the PSS have been conducted in nonclinical individuals, we first tested competing models in the nonclinical sample. The one-factor model fit the data poorly. The two-factor solution fit the data well with the factors correlating at r=-.85 (see Table 2 for the statistical indices of every model). Since the two-factor solution was the most parsimonious, it was chosen as most viable PSS-10 model.

In a next step, we checked modification indices for significant residual correlations to uncover potential method bias. Accordingly, the residuals of items 1 and 9 and items 1 and 3 were freed to correlate (all belonging to the PH subscale; see Fig. 1). Following these adjustments, the overall model fit improved and appeared to be excellent, resulting as final PSS-10 model (χ232=167.316; p<.001; CFI=.975; TLI=.965; RMSEA=.058[.050-.067]; SRMR=.031).

Fig. 1.

Final PSS model retrieved from CFA (standardized solution) in nonclinical and clinical subjects; statistical indices from nonclinical subjects are shown in brackets. Factor loadings are shown above the arrows, communalities are shown next to the Items. Items 4, 5, 7 and 8 are reverse-scored (R). Items 2 and 8R served as marker variables. Residuals of items 1/3 and items 1/9 were freed to correlate.

(0.09MB).

For the clinical sample, we calculated the same models (Table 2). Again, we found the 2-factor solution with correlating residuals to have the best overall fit (χ232=151.644; p<.001; CFI=.954; TLI=.935; RMSEA=.081[.068-.094]; SRMR=.046) with both factors correlating at r=-.70 (for the final model, see Fig. 1). Item loadings ranged from .48 to .83 (nonclinical) and .54 to .84 (clinical).

First of all, the baseline CI model fit the data equally well in clinical and nonclinical individuals. Nesting the metric model into the configural model did not produce a worse fit meaning that the relationship between items and factors were equal across both populations. Finally, differences in goodness-of-fit indices in the scalar model were excellent (Table 2) except for CFI and TLI which yielded a minor change of ΔCFI=-.024 and ΔTLI=-.022. As this was due to differential item functioning (DIF), we did stepwise hierarchical nesting for every single item in order to find out which item(s) might be affected. Out of all 10 items, only item 7 revealed a violation of the criteria offered by Cheung amp; Rensvold (2002) with a difference of ΔCFI=-.017 and ΔTLI=-.020. This was regarded as minor issue since other studies have reported statistical deviations of item 7 as well (Cole, 1999; Reis et al., 2017) and the fit indices themselves were still adequate. Thus, SI and strong measurement invariance was given. It is important to note that multigroup analyses are sensitive to sample size. For this reason, we calculated an additional MGCFA with a random subsample of n=575 nonclinical subjects, which was retrieved using the SPSS random sample command. However, results did not differ substantially from the original analyses (SI: ΔCFI=-.033; ΔTLI=-.030; ΔRMSEA=.013; ΔSRMR=.003).

Internal consistencies in both samples were good to very good, yielding αnonclin=.88 and αclin=.89 for PSS-TOTAL, αnonclin=.85 and αclin=.85 for PH and αnonclin=.79 and αclin=.82 for PSE. For nonclinical individuals, PSS-TOTAL correlated negatively with well-being (r=-.71). In the clinical sample, PSS-TOTAL was positively associated with GSI (r=.49). In both samples, PSS-TOTAL correlated moderately with reported physical health (rclin=-.40; rnonclin=-.41) and strongly with mental health (rclin=-.72; rnonclin=-.73; all p's<.001) (Table 3). Mean scores of all PSS scales were significantly higher in the clinical population than in the nonclinical population (PSS-TOTAL: Mnonclin=28.33, SDnonclin=6.97,Mclin=31.61, SDclin=7.17, t(1821)=-9.24, p<.001, d=0.47; PH: Mnonclin=17.88, SDnonclin=4.77,Mclin=19.43, SDclin=4.86; t(1821)=-6.39, p<.001, d=0.32; PSE: Mnonclin=13.55, SDnonclin=2.80,Mclin=11.82, SDclin=3.26; t(1821)=10.99, p<.001, d=0.59).