The German PCL-5: evaluating structural validity in a large-scale sample of the general German population

ABSTRACT

Background: In attempts to elucidate PTSD, recent factor analytic studies resulted in complex models with a proliferating number of factors that lack psychometrical and clinical utility. Recently, suggestions have been made to optimize factor analytic practices to meet a refined set of statistical and psychometric criteria.

Objective: This study aims to assess the factorial structure of the German version of the PCL-5, implementing recent methodological advancements to address the risk of overfitting models. In doing so we diverge from traditional factor analytical research on PTSD.

Method: On a large-scale sample of the German general population (n = 1625), exploratory factor analyses were run to investigate the dimensionality found within the data. Subsequently, we validated and compared all model suggestions from our preliminary analyses plus all standard and common alternative PTSD factor models (including the ICD-11 model) from previous literature with confirmatory factor analyses. We not only consider model fit indices based on WLSMV estimation but also deploy criteria such as favouring less complex models with a parsimonious number of factors, sufficient items per factor, low inter-factor correlations and number of model misspecifications.

Results: All tested models showed adequate to excellent fit in respect to traditional model fit indices; however, models with two or more factors increasingly failed to meet other statistical and psychometric criteria.

Conclusion: Based on the results we favour a two-factor bifactor model with a strong general PTSD factor and two less dominant specific factors – one factor with trauma-related symptoms (re-experiencing and avoidance) and one factor with global psychological symptoms (describing the trauma’s higher-order impact on mood, cognition, behaviour and arousal).

From the perspective of clinical utility, we recommend the cut-off scoring method for the German version of the PCL-5. Basic psychometric properties and scale characteristics are provided.

HIGHLIGHTS

• We contribute new insights to the debate on the factor structure of the PTSD Checklist (PCL-5) based on a large German general population sample deploying the newest methodological developments in a revised factor-analytical approach.

• Combining theoretical, statistical and practical considerations, we favour a two-factor bifactor model with a strong general PTSD factor and two less dominant specific factors – one factor with trauma-related symptoms and one factor with global psychological symptoms.

• For clinical practitioners, we recommend using the cut-off scoring method.

Antecedentes: En un intento por dilucidar el Trastorno de Estrés Postraumático (TEPT), estudios analíticos factoriales recientes dieron como resultado modelos complejos con un número creciente de factores carentes de utilidad psicométrica y clínica. Recientemente, se han hecho sugerencias para optimizar las prácticas de análisis factorial para cumplir con un conjunto depurado de criterios estadísticos y psicométricos.

Objetivo: Este estudio apunta a evaluar la estructura factorial de la versión alemana del PCL-5, implementando avances metodológicos recientes para abordar el riesgo de sobreajuste de modelos. Al hacerlo, nos apartamos de la investigación analítica factorial tradicional sobre el TEPT.

Método: En una muestra a gran escala de la población general alemana (n = 1.625), se realizaron análisis factoriales exploratorios para investigar la dimensionalidad encontrada en los datos. Posteriormente, validamos y comparamos todas las sugerencias de modelos de nuestros análisis preliminares más todos los modelos factoriales de TEPT alternativos estándares y comunes (incluido el modelo CIE-11) de la literatura previa con análisis factoriales confirmatorios. No sólo consideramos índices de ajuste de modelos basados en la estimación WLSMV (mínimos cuadrados ponderados robustos con media y varianza ajustada), sino que también implementamos criterios como favorecer modelos menos complejos con un número parsimonioso de factores, suficientes elementos por factor, bajas correlaciones entre factores y un número de especificaciones erróneas del modelo.

Resultados: Todos los modelos probados mostraron un ajuste de adecuado a excelente con respecto a los índices de ajuste de modelos tradicionales; sin embargo, los modelos con dos o más factores fallaron progresivamente en cumplir con otros criterios estadísticos y psicométricos.

Conclusión: En base a los resultados, favorecemos un modelo bifactorial de dos factores con un fuerte factor de TEPT general y dos factores específicos menos dominantes: un factor con síntomas relacionados con el trauma (reexperimentación y evitación) y un factor con síntomas psicológicos globales (que describen el mayor impacto del trauma en el estado de ánimo, la cognición, el comportamiento y en el estado de alerta). Desde la perspectiva de la utilidad clínica, recomendamos el método de puntuación de corte para la versión alemana del PCL-5. Se proporcionan propiedades psicométricas básicas y características de la escala.

KEYWORDS:

• Post-traumatic stress disorder
• PCL-5
• LEC-5
• PTSD
• psychometric properties
• factorial structure

PALABRAS CLAVE:

• Trastorno de estrés postraumático
• PCL-5
• LEC-5
• TEPT
• propiedades psicométricas
• estructura factorial

1. Introduction

Post-traumatic stress disorder (PTSD) is one of the most common consequences of trauma exposure, with prevalence rates in the German general population of 1–3% (Glaesmer et al., 2015; Hauffa et al., 2011; Maercker et al., 2008). The long-term social, physical health and economic implications of PTSD are considerable. Therefore, appropriate treatment should be administered to those with PTSD to mitigate this individual and societal cost, beginning with an accurate diagnosis. In this paper, we evaluate the psychometric properties of the German version of the PTSD Checklist (PCL) developed by Weathers et al. (2013) to facilitate the diagnosis of PTSD in German-speaking areas. Specifically, the instrument's factorial structure is evaluated using the factor analytical framework by Schmitt and colleagues (2018). This approach entails a combination of common confirmatory factor analytical practice (CFA) (testing all relevant PTSD models proposed in previous literature) and additional exploratory factor analyses (EFA) paired with refined strict statistical evaluation criteria that move beyond traditional model fit indices.

The PCL (Weathers et al., 2013) is one of the most widely used self-report questionnaires for assessing symptoms of PTSD. It was initially based on the DSM-IV criteria and has since been revised to reflect the DSM-5 criteria (PCL-5), given the existence of traumatic experience (Criterion A). Now, the 20 Items of the PCL-5 correspond to the four DSM-5 symptom clusters (American Psychiatric Association, 2013): re-experiencing (R; Criterion B), avoidance of trauma reminders (AV; Criterion C), negative alterations in cognitions and mood (NACM; Criterion D), and alterations in arousal and reactivity (AAR; Criterion E). The revision to the PCL-5 includes the addition of three items to capture newly added PTSD symptoms, changes to the language of certain items, a change in the response scale from 1–5 to 0–4 and providing only one general version instead of three specific ones (military members, civilians and specific events). The PCL-5 can be used in multiple ways. First, PTSD can be provisionally diagnosed based on an adequate cut-off score (cut-off scoring method). Second, given that at least one B item, one C item, two D items, and two E items at a rating of 2 (moderately) or above are endorsed, provisional PTSD can be diagnosed following the DSM-5 diagnostic rules (symptom scoring method). Third, the PCL-5 can measure symptom severity (sum score) to monitor symptom change across time and in response to interventions (Weathers et al., 2013).

1.1. Psychometric properties and diagnostic utilities of the PCL-5 scale

The revised PCL-5 has undergone comprehensive psychometric evaluation across different languages. A meta-analysis of Forkus et al. (2022) showed satisfactory reliability across all reviewed studies (n = 64). Internal consistencies for the overall scale were high (r = .83–.97) and acceptable to good for the subscales (R: r = .57–.93; AV: r = .69–.91; NACM: r = .74–.94; AAR: r = .71–.90). Likewise retest-reliability ranged from acceptable to high (r = .58–.91). Regarding the convergent validity they found moderate to strong correlations of the PCL-5 total score with other measures of PTSD (r = .44–.89). Correlations of the total score with measures of associated symptoms and comorbid disorders were moderate: traumatic stressors (r = .12–.46), depression (r = .54–.81), anxiety (r = .56–.74), substance use (r = .12–.26), somatic symptoms (r = .50–.61), sleep (r = .51–.62), negative cognitions (r = .47–.61), and dissociation (r = .53–.72). In addition, Forkus et al. found that recommended cut-off scores in most cases ranged between 31 and 33, with a minimum of 22 and a maximum of 49. As for other diagnostic utility estimates, they reported sensitivity (.50 to 1.00), specificity (.35 to .97), positive predictive power (.38–.97), and negative predictive power (.63–1.00), also known as positive/negative predictive value.

Specifically for the German PCL-5, Krüger-Gottschalk et al. (2017) conducted a first psychometric evaluation in a German clinical sample (n = 352). They found high internal consistency (α = .95), high re-test reliability (r = .91) and a high correlation with the total severity score of the Clinical-Administered PTSD Scale for DSM-5 (CAPS-5) of r = .77 – satisfying both evaluations of reliability and validity. They recommend a cut-off score between 31–33 for clinical populations in Germany. Both evaluation methods resulted in acceptable diagnostic utility estimates, with the cut-off scoring method outperforming the symptom score method (sensitivity = .84–.88, specificity = .66–.68, positive predictive power = .81, negative predictive power = .73–.77, overall efficiency = .78–.79).

1.2. Structural modelling of PTSD and the PCL-5

The factor structure of PTSD has substantial implications on the diagnostic algorithms (e.g. the symptom scoring method). Ultimately, diagnoses derived from these algorithms directly inform the treatment pathways of patients in both clinical and research settings. Subsequently, it is essential to have a good understanding of the structural validity of the PCL-5 (Forkus et al., 2022).

Over the years, several studies on military and civilian samples could not support the initially intended four-factor DSM-5 model (Armour et al., 2015; Blevins et al., 2015; Wortmann et al., 2016). Instead, various alternate models have evolved (for an overview, see Table 1). The 4-factor Dysphoria Model (Simms et al., 2002) kept the re-experiencing and avoidance factors and restructured the remaining AAR and NACM cluster items to favour a more prominent dysphoria factor. This conceptualization might explain the comorbidity with distress-based disorders (Contractor et al., 2018). The 5-factor Dysphoric Arousal Model (Elhai et al., 2011) kept the first three factors but differentiated between dysphoric and anxious arousal symptoms in the AAR cluster. The 6-factor Anhedonia Model (Liu et al., 2014) additionally splits the NACM cluster into anhedonia and negative affect. In contrast, the Externalizing Behavior Model (Tsai et al., 2015) focuses on emotion dysregulation difficulties by suggesting an additional externalizing behaviours factor based on the AAR cluster. The Hybrid Model (Armour et al., 2015) integrates previously postulated factors within the DSM-5 conceptualization in a 7-factor model. In the aforementioned recent review (Forkus et al., 2022) of existing factor analytical literature, the Hybrid Model found solid empirical support. It has been shown that of 34 studies that tested the Hybrid Model, 82.3% found it to be optimal compared to alternate models. The Anhedonia and Hybrid Models performed equally well in six studies.

Table 1. Overview of Standard and Common Alternative DSM-5 Plus ICD-11 Derived PTSD Models With Symptom Mapping.

		DSM-5 Numbing Model^1	Dysphoria Model^2	Dysphoric Arousal Model^3	Anhedonia Model^4	Externalizing Behavior Model^5	Hybrid Model^6	ICD-11 Model^7
	PTSD symptom	4-factor	4-factor	5-factor	6-factor	6-factor	7-factor	3-factor
B1	Unwanted upsetting memories	R	R	R	R	R	R
B2	Nightmares	R	R	R	R	R	R	R
B3	Flashbacks	R	R	R	R	R	R	R
B4	Psychological cue reactivity	R	R	R	R	R	R
B5	Physiological cue reactivity	R	R	R	R	R	R
C1	Avoidance of thoughts of trauma	AV	AV	AV	AV	AV	AV	AV
C2	Avoidance of reminders of trauma	AV	AV	AV	AV	AV	AV	AV
D1	Memory impairment	NACM	D	NACM	NA	N	NA
D2	Negative beliefs	NACM	D	NACM	NA	N	NA
D3	Blaming self or others	NACM	D	NACM	NA	N	NA
D4	Negative affect	NACM	D	NACM	NA	N	NA
D5	Decreased interest in activities	NACM	D	NACM	AN	N	AN
D6	Feeling isolated	NACM	D	NACM	AN	N	AN
D7	Difficulty experiencing positive affect	NACM	D	NACM	AN	N	AN
E1	Irritability or aggression	AAR	D	DA	DA	EB	EB
E2	Risky or destructive behaviour	AAR	AAR	DA	DA	EB	EB
E3	Hypervigilance	AAR	AAR	AA	AA	AA	AA	SOT
E4	Heightened startle reaction	AAR	AAR	AA	AA	AA	AA	SOT
E5	Difficulty concentrating	AAR	D	DA	DA	DA	DA
E6	Difficulty sleeping	AAR	D	DA	DA	DA	DA

Note. AAR = alterations in arousal and reactivity. AA = anxious arousal. An = anhedonia. AV = avoidance. D = dysphoria. DA = dysphoric arousal. EB = externalizing behaviours. N = numbing. NA = negative affect. NACM = negative alterations in cognitions and mood. R = re-experiencing, SOT = sense of threat.

¹ American Psychiatric Association (2013), ²Simms et al. (2002), ³Elhai et al. (2011), ⁴Liu et al. (2014), ⁵Tsai et al. (2015), ⁶Armour et al. (2015), ⁷Brewin et al. (2009).

However, empirical support for the Hybrid Model remains debated due to substantial statistical and practical limitations of contemporary factor analytical approaches to PTSD. The theory-testing and post-hoc adjustment approach tends to favour increasingly complex models, raising concerns about statistical model robustness because of (1) diminished suitability of model fit indices for performance evaluation, (2) a greater prevalence of under-identified factors with fewer than three items per factor, (3) increased inter-factor correlation leading to decreased construct validity (Schmitt et al., 2018). Practically, the heightened complexity of the model’s structure and diagnostic algorithm diminishes its clinical utility (Silverstein et al., 2018). Furthermore, previous work emphasized the relatively nonspecific PTSD factors (‘negative alterations in cognitions and mood’ and ‘hyperarousal’), neglecting unique factors such as ‘re-experiencing’ and ‘avoidance’ (Krüger-Gottschalk et al., 2022; Rasmussen et al., 2019). As a solution, Schmitt et al. (2018) suggest a combination of EFAs and CFAs on the same data set as a viable solution to create a model that is both theory-informed and reflective of the underlying factorial structure. They also provide detailed statistical methods and a framework to evaluate factor structures and address any potential limitations (see Method section).

Efforts to simplify the PTSD construct include reducing the number of factors, eliminating items and adopting the streamlined ICD-10 conceptualization. For instance, Ferrie et al. (2022) identified a three-factor model (AN, R, NACM) as the best-fitting model through exploratory factor analysis on a general British population sample, excluding items with poor factor loadings. Brewin et al. (2009) and Schellong et al. (2019) suggest a three factor solution (re-experiencing, avoidance and perceived sense of threat) based on the ICD-11 through post-hoc item allocation (see Table 1). This approach, applied to various instruments, including the Harvard Trauma Questionnaire (Hansen et al., 2015), the Impact of Event Scale-Revised (Hyland et al., 2017) and the Clinician-Administered PTSD Scale (CAPS-5) (Krüger-Gottschalk et al., 2022) has shown some success. Additionally, evidence of the specifically developed International Trauma Questionnaire (Ben-Ezra et al., 2018; Cloitre et al., 2013; Ho et al., 2020; Redican et al., 2021) supports the factorial validity of the ICD-11 model.

American veteran samples exhibit evidence for both a unidimensional structure (Jenkins-Guarnieri et al., 2023) and two- and four-bifactor models with one decisive general and several weaker specific factors (Schmitt et al., 2018).

In a German clinical sample (Krüger-Gottschalk et al., 2017) evidence for the latent factor structure of the PCL-5 remained inconclusive. Furthermore, the methodological suggestions for DSM models and the application of the ICD-11 model, as put forth by Schmitt and colleagues, have yet to be tested.

In summary, critique on contemporary factor analytical procedure questions the clarity of traditionally identified PCL-5 factorial structures. Newer attempts, exhibiting heterogeneity, alter the questionnaire or adopt different theoretical frameworks without consistent use of updated statistical methods. Previous studies often involved small, homogeneous samples—primarily white, female college students, veterans, or treatment-seeking individuals—limiting generalizability.

1.3. Research objective

To establish robust and generalizable psychometric evidence, we propose evaluating the optimal PCL-5 factor model by (1) deploying statistically rigorous methods, (2) using a large-scale, general-population sample, and (3) testing and comparing theory and empirically-derived models within a clinical utility framework. Our study aims to execute this by implementing Schmitt et al. (2018) statistical algorithm in a substantial, non-clinical, population-based sample in Germany.

2. Method

2.1. Subjects

The present cross-sectional study is part of a representative survey of the general German population conducted in 2016 with the help of an independent institute for market and social research (USUMA, Berlin). Germany was separated into 258 sample areas representing the country's different regions. Within these regions, 4902 households were selected using the ‘random-route-procedure.’ A participating member of each household fulfilling the inclusion criteria (age at or above 14, able to read and understand the German language) was selected randomly via a Kish selection grid. A maximum of three attempts were made to contact the selected person in their private homes. Subjects were visited by a trained study assistant of USUMA and informed about the investigation. When they provided written informed consent the study assistant then administered self-rating paper-pencil questionnaires which were filled out privately and offered help where needed. A total of 2510 people between the ages of 14 and 91 agreed to participate, completing the self-rating questionnaires between September and November 2016 (participation rate: 52.0%, refusal to be interviewed: 30.0%, target person repeatedly unavailable: 18.0%). The final sample consisted of 1625 participants (excluding 86 underage participants, 20 subjects with missing data on all items of the PCL-5 and 779 participants without trauma exposure). Characteristics of the study sample are presented in Table 2. In our pursuit to derive scientific inferences concerning the factorial structure of the PCL-5, we used the unweighted sample, acknowledging the trade-off in sample representativeness and the consequent limitations in making statistical inferences, such as the determination of psychometric norms and prevalences. Yet the analysis sample was approximately representative regarding age (M = 48.36; SD = 18.2; Range: 18–94) and sex for the Federal Republic of Germany when compared to the microcensus of the Federal Office of Statistics. The Ethics Review Committee of the Medical Faculty of the University of Leipzig determined the proposed study to be outside their scope but expressed no ethical concerns for the survey.

Table 2. Distribution of Sociodemographic Characteristics in the Sample.

Characteristic		n	%
Sex	Male	752	53.7
	Female	873	46.3
Age	18–30	283	17.5
	31–50	523	32.2
	51–70	588	36.1
	>71	231	14.2
Education	No higher education entrance qualification	1253	77.1
	A-levels/higher education entrance qualification	189	11.6
	University degree	183	11.3
Household income(per month)	< 1250 €	286	17.6
	1250–2500 €	636	39.1
	> 2500 €	631	38.8
	Missing data	72	4.5
Residence	Rural area	238	14.6
	Urban area	1387	85.4
Partner	Living with partner	899	55.3
	Living without partner	713	43.9
	Missing data	13	0.8
Exposure to traumatic experiences	Motor vehicle accident	1099	45.7
	Severe human suffering	953	39.6
	Life-threatening injury/illness	743	30.9
	Other serious accident	674	28
	Sudden, unexpected death of a loved one	584	24.3
	Fire/explosion	568	23.6
	Natural disaster	553	23
	Physical assault	512	21.3
	Witness violent death	452	18.8
	Another very stressful event	414	17.2
	Sexual assault	317	13.2
	Other unwanted sexual experience	317	13.2
	Assault with weapon	301	12.5
	Combat	234	9.7
	Exposure to a toxic substance	232	9.7
	Captivity	156	6.5
	Caused severe injury/death of another	144	6

2.2. Instruments

Life-Events-Checklist 5 (LEC-5). The LEC-5 (Krüger-Gottschalk et al., 2017; Weathers et al., 2013) is a self-report measure assessing the lifetime prevalence of trauma. The German version of the LEC-5 and PCL-5 (Krüger-Gottschalk et al., 2017) has been computed by a professional translation and back-translation process from and to the English version. Sixteen predefined traumatic events and the option to describe any other very stressful event or experience not listed before are presented. The A-criterion of the PTSD diagnosis is endorsed when at least one of the events is answered by ‘Happened to me’, ‘Witnessed it’, ‘Learned about it’, and ‘Part of my job’. The answer options ‘Not sure’ and ‘Doesn't apply’ are not considered affirmative. Based on this metric, prevalence rates of traumatic events in the sample (including individuals without traumatic exposure, n = 2,404) are reported in the lower part of Table 2.

Post-Traumatic Disorder Checklist-5 (PCL-5). The Posttraumatic-Disorder-Checklist 5 (PCL-5; Weathers et al., 2013) consists of 20 items covering the DSM-5 PTSD symptoms. Symptom severity in the past month can be rated on a 5-point Likert scale from 0 (‘Not at all’), 1 (‘A little bit’), 2 (‘Moderately’), 3 (‘Quite a bit’) to 4 (‘Extremely’). Following the DSM-5 diagnostic criteria, a preliminary diagnosis can be given.

2.3. Statistical procedure and considerations

For our analyses we applied the detailed recommendations of Schmitt et al. (2018) regarding the selection of appropriate estimators, the factor analysis framework, and the use of a comprehensive set of indicators to assess model fit (including model fit indices, misspecification indices, interfactor correlations, factor identification, and the number of factors). In their work, they advocate selecting the best model based on a combination of theoretical and empirical considerations and have defined an improved standard for factor structure validation.

1.	Factor Analysis Framework

We chose a combined approach of EFA, CFA and Bifactor Modelling for our factor analysis framework. We ran EFAs to gain a comprehensive understanding of the empirical data in order to determine the appropriate number of factors to extract. Supplementing CFAs with EFA helps to detect and correct misspecifications in CFAs, thereby avoiding overly complex and overfit models that may result from relying solely on confirmatory models and assuming a priori theory. We ran CFAs to test and compare the most common theoretically postulated models, the ICD-11 model and the models derived from EFAs. We then ran a bifactor model analysis based on the outcomes to test whether a unidimensional construct with facets (interpretation of the sum score with merely descriptive facets / cut-off scoring method) or a multidimensional construct (reliable and interpretable subscales / symptom scoring method) is more appropriate.

2.	Model Estimation

We used the Weighted Least Square Mean and Variance Adjusted estimation (WLSMV) and a polychoric correlation matrix if not stated otherwise in the result section. The WLSMV estimator has been proven to be the most efficient and robust estimation method for coarse-ordered categorical scaled and skewed data. Since all of the PCL-5′s items are based on an ordinal Likert scale and were highly skewed (skew statistics range between 1.57 (Item 1) and 3.02 (Item 14)), we deemed the WLSMV the most appropriate. In CFAs, we fixed all variances of latent variables to one. In EFAs, the mean and the variance were fixed at 0 and 1, respectively. For the bifactor model, we set the variances of the latent variables to 1 and the covariances between latent variables to 0 (i.e. standard bifactor model practice, Reise et al., 2007). Preliminary analyses to determine the appropriate number of factors to extract for the EFAs contained an evaluation of eigenvalue magnitudes, parallel analysis and model fit statistics. We used an oblique Geomin rotation to obtain a simple approximate structure with the EFA. The programme used for all statistical analyses was R (R Core Team, 2021). The CFAs, EFAs and the bifactor model were conducted with the lavaan package (Rosseel, 2012). Parallel analyses were run with the psych package (Revelle, 2022). Missing data have been imputed by the mice package (Buuren & Groothuis-Oudshoorn, 2011).

3.	Model Fit and Factor Model Selection

We conducted a comprehensive, balanced evaluation of the models that incorporated multiple model fit indices and other criteria to assess the amount of variance explained (good model fit from a statistical perspective) and the validity of the factorial structure (good model fit from a practical standpoint).

Since the χ² statistic is known to produce statistically significant values for good fitting models when the factor structure is complex and the sample size is large, we additionally evaluated the following Approximate Fit Indices (AFI). According to Hu and Bentler (1999), the Comparative Fit Index (CFI) and Tucker-Lewis Index (TLI) statistics greater than 0.90 are deemed adequate, and values greater than 0.95 are good. Root Mean Square Error of Approximation (RMSEA) values less than 0.10 and 0.06 are considered mediocre and good, respectively. Both fit indices are commonly used for maximum likelihood estimated models and must be interpreted carefully with WLSMV estimation (Sass et al., 2014). Robust to the estimation method is the Standardized Root Mean square Residual (SRMR), whose values are deemed adequate if less than 0.08 (Savalei, 2021). We employed maximum likelihood estimation to calculate the Akaike Information Criterion (AIC; Akaike, 1974) and Bayesian Information Criterion (BIC; Schwarz, 1978) to compare the relative fit of nested and non-nested models to the same data. Several indices assess the reliability, stability and dimensionality of bifactor models. Although there is no universal agreement on the standards for these bifactor indices, we provide an overview of the most common indices (Reise et al., 2013; Rodriguez et al., 2016; Schmitt et al., 2018). Omega (ω), as a measurement for reliability, provides coefficients for the general/hierarchical factor (ω), specific factors (ω_S) and after accounting and isolating the specific factors, only the general factor (ω_H) and vice versa only the specific factors (ω_HS). When the ω_H is large (e.g. > .80) and the ω_HS are small (e.g. < .50), this indicates that the general factor is more reliable than specific factors. Explained Common Variance (ECV) values greater than .85 on the general factor (and smaller specific factor ECVs) suggest the measure is sufficiently unidimensional. Values of the Item-level Explained Common Variance (IECV) measuring between 0.5 and 1 indicate that the item reflects more of the general dimension and less of the specific dimension. If most items have IECV values above .80, this commonly implies that a unidimensional measure is preferable. Further evidence of unidimensionality is if the Percentage of Uncontaminated Correlations (PUC) is greater than .70 and the Average Relative Bias (ARB) between the general and unidimensional factor loadings is small. The Factor Determinacy index (FD) and the construct replicability (H, Hancock & Mueller, 2001) should be closely monitored to ensure stability; values of FD below 0.90 and H above 0.80 are ideal.

We sought to identify the most economical model that best fits the data by comparing the number of factors extracted from the models to those suggested by the combined results of parallel analysis and eigenvalues. Additionally, we monitored inter-factor correlation magnitudes to avoid collinearity, taking factor correlations greater than .85 (Carragher et al., 2016; Rasmussen et al., 2019) as a reference point for high collinearity and low discriminant validity of factors. Furthermore, we ensured that each factor contained sufficient items, with a minimum of three items per factor, to not only ensure model identification (Little et al., 1999) but also avoid construct underrepresentation (Kaplan & Saccuzzo, 2009).

4.	Missing data

After removing 20 cases in which the PCL-5 was skipped entirely, the sample consisted of n = 2404. In the remaining data set, in 4.83% of the subjects, the answers were incomplete (with a total of 0.6% missing data points throughout the whole data set). Little's missing completely at random test indicated that data was missing systematically (χ²(874) = 1263.7, p < .001). Therefore, the missing data treatment used the predictive mean matching method. After, we excluded non-trauma-exposed individuals arriving at our final sample size of n = 1625.

3. Results

3.1. Exploratory factor analysis

Preliminary analyses regarding the number of factors to extract resulted in multiple solutions. The analysis of eigenvalues based on an Oblimin rotation revealed the first three eigenvalues to be 12.278, 1.367 and 0.804  – hinting at one strong factor and potentially a second weak one. Parallel analysis based on the less robust WLS estimation suggests four factors based on the estimated (and simulated) eigenvalues of 11.89 (0.35), 0.97 (0.24), 0.38 (0.21), 0.30 (0.18), 0.17 (0.16). We interpreted the latter cautiously since parallel analysis based on WLSMV estimation is not available in R as there is a debate on its appropriateness with categorical data (yet Garrido et al., 2016; Weng & Cheng, 2017 make a case for applicability).

Hence, we ran and compared EFAs from a one- up to a four-factor solution. Comparing the resulting four different models (see Table 3), a four-factor solution seems to result in the best fit indices. When examining the four-factor solution closely, it is evident that items loading high on factors one and three are in fact more strongly associated with factors two and four. Similarly, in the three-factor model, several items associated with factor three have high cross-loadings (> 0.30) with factors one and two. Furthermore, all multiple-factor models show high covariances (> 0.7), suggesting more than 50% of shared variance between those factors, indicating poor discriminant validity. From a practical point of view, a two-factor model with one factor reflecting trauma-related symptoms (re-experiencing and avoidance, items 1–7) and one factor reflecting the global psychological symptoms (the trauma's higher order impact on mood, cognition, behaviour and arousal, items 8–20) seems most feasible, economical, and in line with outcomes of the Eigenvalue analyses. An agreeable fit index (CFI > .90) further supports a two-dimensional solution. Yet, high inter-factor correlations and a few high cross-loadings between the two factors impose limitations.

Table 3. Weighted Least Squares Mean- and Variance Adjusted EFA and Standardized CFA Factor Loadings and Fit Indices for Multiple Factor Solutions.

		EFA										CFA
	items	1-factor	2-factor		3-factor			4-factor				2-factor		two-factor bifactor
		1	1	2	1	2	3	1	2	3	4	1	2	G	S1	S2
1	Unwanted upsetting memories	.84	.96	−.09	.9	.01	−.05	.52	.69	.01	−.01	.87		.85	.3
2	Nightmares	.84	.89	−.01	.9	0	.02	.51	.65	0	.06	.87		.85	.27
3	Flashbacks	.84	.77	.12	.68	.22	0	.38	.75	−.02	0	.87		.86	.18
4	Psychological cue reactivity	.84	.87	.01	.61	.33	−.13	.37	.86	.01	−.15	.87		.86	.14
5	Physiological cue reactivity	.84	.65	.25	.49	.42	−.01	.25	.73	.05	.06	.88		.88	.03
6	Avoidance of thoughts of trauma	.78	.81	.01	.02	.93	−.41	.05	.7	.38	−.08	.81		.81	−.22
7	Avoidance of reminders of trauma	.8	.75	.1	0	.95	−.34	0	.67	.39	.01	.83		.85	−.37
8	Memory impairment	.63	.29	.39	.04	.62	0	−.06	.48	.16	.16		.65	.6		.22
9	Negative beliefs	.79	.22	.63	−.06	.85	.05	−.12	.82	−.01	.03		.82	.72		.37
10	Blaming self or others	.74	.38	.42	−.07	.87	−.1	−.07	.92	.01	−.17		.77	.71		.23
11	Negative affect	.82	.34	.54	.15	.68	.07	−.01	.74	.04	.12		.85	.77		.31
12	Decreased interest in activities	.81	.1	.76	.25	.48	.31	.07	.7	−.19	.22		.84	.7		.47
13	Feeling isolated	.84	−.02	.9	−.01	.75	.28	0	.99	−.4	−.04		.86	.69		.55
14	Difficulty experiencing positive affect	.84	−.02	.89	0	.73	.28	−.02	.91	−.32	.04		.86	.68		.56
15	Irritability or aggression	.78	.05	.77	0	.7	.21	−.13	.64	−.03	.23		.8	.66		.46
16	Risky or destructive behaviour	.6	−.03	.66	.13	.37	.28	−.15	.14	.19	.53		.62	.49		.42
17	Hypervigilance	.65	.22	.48	.36	.26	.21	−.02	.02	.38	.63		.67	.6		.28
18	Heightened startle reaction	.79	.13	.72	.57	.11	.42	.04	−.01	.3	.85		.81	.69		.44
19	Difficulty concentrating	.79	−.15	.97	.55	.02	.65	0	0	.07	.91		.81	.61		.61
20	Difficulty sleeping	.75	.11	.7	.7	−.08	.5	.21	.18	−.01	.65		.78	.65		.43
	χ^2 (df)	3071.113 (170)	1526.149 (151)		968.099 (133)			613.003 (116)				1676.23 (169)		1206.21 (150)
	p-value	< .001	< .001		< .001			< .001				< .001		< .001
	CFI	.806	.899		.924			.951				.961		.973
	FCVR	-	.755		.071–.798			.009–.762				.841		-
	RMSEA	.146	.112		.104			.089				.074		.066

Note. EFA = exploratory factor analysis. CFA = confirmatory factor analysis. CFI = comparative fit index. FCVR = factor covariance range. RMSEA = root mean square error of approximation. Factor loadings of EFAs greater than 0.3 are highlighted.

It is not yet possible to rule out the possibility of a unidimensional construct. Subsequently, a CFA was conducted on the one-factor and two-factor models to examine the above findings.

3.2. Confirmatory factor analysis

A summary of model fit parameters across the five traditional models based on literature, the ICD-11 model and the two models derived from our exploratory factor analyses is displayed in Table 4. Overall, the ICD-11 model (χ²(6) = 26.79, p < .05, RMSEA = .046, CFI = .998, TLI = .995, SRMR = .015) has the best fit and has a number of factors that is both small and is closest to the indicated two to four recommended factors. Nonetheless, note that it is only composed of a subset of the PCL-5 items having less than three items per factor. The model using all of the items with the best model fit parameters is the Hybrid Model (χ²(149) = 839.21, p < .05, RMSEA = .053, CFI = .982, TLI = .977, SRMR = .032). Yet, we would like to note, the incremental improvement of the AFI in contrast to more parsimonious models (e.g. the classic DSM four-factor model) is relatively small. Instead, we find high factor covariances (>.85 for NA  – EB, AN  – EB, AN  – DA, NA  – AN, EB  – AA, EB  – DA, AA  – DA) which suggests collinearity and low discriminant validity between the factors. Additionally, four out of the seven factors are under-identified (< 3 items per factor), and the construct is quite complex and exceeds the recommended number of factors to extract by far. Based on fit indices, the two-factor model already shows a perfect fit (χ²(169) = 1676.23, p < .05, RMSEA = .074, CFI = .961, TLI = .956, SRMR = .048) while also satisfying the criteria of sufficient items per factor and a lower, more economic number of factors.

Table 4. CFA Summary  – Overview of Model Fit Indices for Each of the Previously Theorized Models and Models Derived from EFAs.

	WLSMV						ML
Model	χ2 (df)	CFI	TLI	RMSEA [90% CI]	SRMR	FCVR	AIC	BIC
One Factor Model	3071.11 (170)	.925	.916	.103 [.099;.106]	.068	–	66895	67110
Two Factor Model	1676.23 (169)	.961	.956	.074 [.071;.077]	.048	.841	65525	65746
2-Factor Bifactor Model	1206.21 (150)	.973	.965	.066 [.062;.069]	.037	–	64928	65252
ICD-11 Model (3-factor)	26.79 (6)	.998	.995	.046 [.029;.065]	.015	.715–.824	21923	22004
DSM Dysphoria (4-factor)	1257.25 (164)	.972	.967	.064 [.061;.067]	.043	.718–.912	65005	65253
Dysphoric Arousal (5-factor)	1202.93 (160)	.973	.968	.063 [.06;.067]	.041	.693–.921	64903	65173
Anhedonia (6-factor)	931.61 (155)	.98	.975	.056 [.052;.059]	.035	.693–.914	64561	64858
Externalizing Behaviours (6-factor)	1138.65 (155)	.975	.969	.063 [.059;.066]	.039	.645–.933	64768	65064
Hybrid Model (7-factor)	839.21 (149)	.982	.977	.053 [.05;.057]	.032	.645–.911	64409	64738

Note. EFA = exploratory factor analysis. CFA = confirmatory factor analysis. CFI = comparative fit index. FCVR = Factor covariance range. ML = maximum likelihood estimation. RMSEA = root mean square error of approximation. SRMR = standardized root mean square residual. TLI = Tucker-Lewis-Index. WLSMV = weighted least squares mean and variance adjusted estimation.

3.3. Bifactor analysis

We added a bi-factor model approach to our analysis to further explore the appropriateness of a unidimensional construct in light of (1) weak evidence for a second factor in the eigenvalue analysis, (2) high inter-factor correlations in the two-factor model and (3) consistent high item loadings of the one factor EFA solution.

The two-factor bifactor model fit the data well (χ²(150) = 1206.21, p < .001, CFI = .973, TLI = .965, RMSEA = .066, SRMR = .037) with the standardized factor loadings on the specific factors being smaller than the general factor, particularly for specific factor 1 (see Table 3). These factor-loading results suggest that while the items relate consistently to the general model, this is not necessarily the case for the specific factors. To further evaluate the bifactor model, various bifactor indices were examined based on the standards outlined earlier (see Table 5). The ECV (.79) met the minimum criteria (ECV > .70; Stucky & Edelen, 2014), with the specific factor ECVs being small. The ARB (.111) indicated a relatively small difference between the unidimensional factor solution and the general factor, yet slightly above the suggested cut-off (ARB < .10; Stucky & Edelen, 2014). The average IECV was .79 (SD = 0.15, minimum = 0.50, maximum = 1.00), with 65% of the IECV greater than .80, providing further support for the bifactor model. For the general and the second factor, the FD statistic exceeded .90 and the H statistic .70 (Rodriguez et al., 2016), which indicates trustworthiness and acceptable to perfect construct replicability. The PUC for this model was .51. The ω_H and ω_HS revealed high reliability for the general factor (ω_H = .84) but low reliability for the specific factors ω_HS. In summary, a bifactor model is more statistically appropriate than a model with two specific factors; the general factor is dominant over the specific ones.

Table 5. Bifactor Models and their Indices.

	ECV1	ECV2	ω/ωS	ωH/ωHS	H	FD
General	.79	.79	.96	.84	.97	.98
SF1	.07	.03	.93	0	.30	.83
SF2	.30	.18	.92	.21	.76	.91

Note. ECV = explained common variance; measurement of the strength of a specific factor relative to all explained variance of all items. ECV1 = includes loadings of all specific factors. ECV2 = only includes loadings of the specific factor of interest. H = construct replicability. FD = factor determinacy.

3.4. Model misspecification

An alternative to the goodness-of-fit test is to investigate whether misspecifications are present in the model using the method suggested by Saris et al. (2009, p. 571). To implement the Saris-Satorra-van der Veld procedure, Jrule was used to examine the power and significance of potential cross-loadings by setting the misspecification cut-off, λ, to .20 and .40, alpha level rate to .05, and power to .80. We tested the PCL-5 models from Table 4 for model misspecification at the factor loading and residual covariance levels. Results from Table 6 show that an increase in the number of factors leads to a rise in model misspecifications. Our findings indicate that the one-factor and two-bifactor models are probably the most suitable. Only zero or two modifications may be necessary depending on the magnitude of λ to optimize the models. As Schmitt et al. (2018) found, probably, the second dimension (related to the second eigenvalue) is likely a consequence of an item pair (items 6 and 7) exhibiting shared variance that is not associated with the factor, which is essentially a manifestation of dimensionality.

Table 6. Number of Model Misspecifications at the Factor Loading, Correlated Residual and Overall (Factor Loading plus Correlated Residuals) Levels.

	λ = 0.2			λ = 0.4
	Factor loadings	Residual covariance	Overall	Factor loadings	Residual covariance	Overall
One Factor Model	0	2	2	0	0	0
Two Factor Model	8	0	8	0	0	0
Two Bi-Factor Model	0	2	2	0	1	1
ICD-11 Model (3-factor)	0	0	0	0	0	0
DSM Dysphoria (4-factor)	24	0	24	5	0	5
Dysphoric Arousal (5-factor)	35	0	35	8	0	8
Anhedonia (6-factor)	40	0	40	10	0	10
Externalizing Behaviours (6-factor)	46	0	46	14	0	14
Hybrid Model (7-factor)	46	0	46	10	0	10

Note. The number of model misspecifications is based on the following criteria: misspecification of at least .20 or .40, a Type I error of .05 and power equal to .80.

Based on the results of confirmatory factor analyses and misspecification indices (Tables 4 and 6), a two-factor bifactor model seems statistically and practically most appropriate. Therefore, research should consider using the total PCL-5 score Figure 1.

Figure 1. The Factor Model and Standardized Factor Loadings for the Two-Factor Bifactor Model.

3.5. General psychometric properties of the PCL-5

Participants reported an average sum score of 8.19 (SD = 11.37) on the PCL-5. According to the cut-off scoring method, 115 individuals, constituting 4.78% of the total sample and 7.07% of the trauma-exposed individuals, met the DSM-5 criteria for PTSD. Further, they reported an average of 1.56 traumatic events experienced (SD = 0.49). We evaluated the psychometric properties of the PCL-5 item and scale characteristics of the uni-, two- and four-dimensional PCL-5-conception based on the total trauma-exposed sample (n = 1625). As shown in Table 7, item-total correlations are in the upper range. Internal consistency as a measure of the reliability of the scale can be considered to be very good (Cronbach's α = .94 for the total scale, α = .91 for both two-factor model subscales, α = .9 for the re-experiencing cluster, α = .82 for the avoidance cluster, α = .87 for the NACM cluster and α = .83 for the AAR cluster). Results indicate better internal consistencies for a one- or two-dimensional model.

Table 7. Psychometric Properties of the German Adaptation of the PCL-5.

					1- factor model			2-factor model			DSM model
	Item	M	SD	Skew	α	rit			α	rit		α	rit
1	Unwanted upsetting memories	0.86	0.99	1.16	.94	.72		Trauma-related symptoms	.91	.81	R	.9	.83
2	Nightmares	0.6	0.9	1.61		.73				.81			.82
3	Flashbacks	0.47	0.84	2.01		.74				.77			.78
4	Psychological cue reactivity	0.92	1.08	1.09		.74				.82			.82
5	Physiological cue reactivity	0.52	0.91	1.87		.76				.76			.75
6	Avoidance of thoughts of trauma	0.82	1.06	1.21		.68				.75	AV	.82	.77
7	Avoidance of reminders of trauma	0.63	0.99	1.61		.69				.73			.77
8	Memory impairment	0.38	0.75	2.24		.53		Global psychological symptoms	.91	.51	NACM	.87	.49
9	Negative beliefs	0.4	0.81	2.27		.7				.72			.74
10	Blaming self or others	0.45	0.89	2.21		.63				.61			.65
11	Negative affect	0.54	0.89	1.83		.75				.74			.74
12	Decreased interest in activities	0.43	0.85	2.26		.71				.73			.73
13	Feeling isolated	0.38	0.79	2.36		.73				.77			.81
14	Difficulty experiencing positive affect	0.34	0.76	2.62		.71				.75			.76
15	Irritability or aggression	0.34	0.73	2.54		.67				.7	AAR	.83	.64
16	Risky or destructive behaviour	0.29	0.64	2.6		.48				.52			.54
17	Hypervigilance	0.73	0.99	1.29		.58				.57			.6
18	Heightened startle reaction	0.46	0.78	1.85		.71				.73			.78
19	Difficulty concentrating	0.44	0.79	2.06		.68				.73			.78
20	Difficulty sleeping	0.69	1.02	1.54		.67				.67			.68

Note. M = mean; SD = standard deviation; rit = part-whole corrected item-total correlation (related to the total score).

4. Discussion

Our study investigates the dimensionality, psychometric properties, and conceptual determination of the PCL-5 in a large-scale population-based sample, addressing the ongoing debate about the PCL-5′s dimensionality. Regarding the general factor analytical procedure, all tested models (standard and common alternative DSM-5 models, the ICD-11 model, a unidimensional model, a two-dimensional model and the 2-bifactorial model) showed at least satisfactory to good fit indices. In particular, we found the ICD-11 and Hybrid models to fit excellently with the data. However, only the one-, two-factor models and the two-factor bifactor model included all items and additionally (1) had a number of factors supported by the combined outcomes of parallel analysis and eigenvalues, (2) showed sufficient discriminant validity between factors, (3) had sufficient items per factor and (4) a low number of model misspecifications. Based on our data, we find a 2-bifactor model to have the best overall fit. It consists of a solid general PTSD factor and two weaker rather descriptive specific factors – one containing traumatic stimuli-related behaviour (re-experiencing and avoidance) and one describing the trauma's higher-order impact on mood, cognition and arousal. It provides a model that fits well, has easily interpretable parameters and approximates reality parsimoniously (as recommended by Preacher & Merkle, 2012). This result advocates for the use of the cut-off scoring method. Not least, the internal consistencies of the total scale and two-factorial subscales were higher than those of the DSM model's subscales, providing further evidence to support the reduced dimensional structure of the PCL-5.

Our findings align with Schmitt et al. (2018), endorsing a bifactorial model characterized by a robust general factor and two subordinate specific factors. Despite the excellent model fit exhibited by the traditional models (Hybrid (Armour et al., 2015), DSM Dysphoria (Simms et al., 2002), Dysphoric Arousal (Elhai et al., 2011), Anhedonia (Liu et al., 2014), Externalizing Behaviours (Tsai et al., 2015)) on our PCL-5 data, the increase in the number of factors leads to higher inter-factor correlations, increased model-misspecifications, and more underidentified factors. This undermines the models’ statistical robustness and diminishes their clinical utility through overly intricate diagnostic algorithms and less distinguishable factors. In summary, our research challenges the prevalent trend of factor over-extraction, demonstrating that simpler models provide good fit, statistical robustness, and enhanced clinical utility. In the conceptualization of the two facets, we find a resemblance to the differentiation between PTSD (core symptoms) and complex PTSD (general dysphoria and anxiety symptoms) of the ICD-11.

Even though our study is based on a large-scale population-based sample, and we have implemented a thorough and precise statistical and psychometric methodology, some shortcomings and future challenges must be mentioned. Noteably, this study refrains from reporting prevalences. We acknowledge minor deviations in the analysis sample from expected sociodemographic group norms, such as the underrepresentation of individuals older than 70 and a gender bias favouring more females in the working-age sample. However, exact representativeness is not essential for structural modelling (Rothman et al., 2013). Additionally, individuals residing in institutions, including older individuals with World War-related traumas and those with severe trauma leading to mental and physical impairment, are not represented in our study since data collection took place in private households. The broad definition of the A-criterion, encompassing indirect exposure (‘learned about it’), was chosen to construct a sizable sample for factor structure validation analyses, potentially inflating the base rate of trauma. The observed low rate of PTSD in the analysis sample poses a limitation to the interpretation of our findings on the factorial structure of PTSD. Yet this corresponds to the rarity of this mental disorder in the general population (Hauffa et al., 2011; Maercker et al., 2008).

The presentation of questions in questionnaire form by non-medical/psychological interviewers from a survey institute could lead to uncertain compliance from participants in providing information about their psychological experiences. Likewise, it is imperative to consider memory bias when attempting to capture events from the distant past. In standard statistical procedures, caution is recommended when conducting both exploratory and confirmatory factor analyses on a single dataset, as this practice may introduce bias by inflating fit statistics through chance discoveries in the exploratory phase. However, adhering to Schmitt et al. (2018) structured approach, we advocate a judicious combination of CFAs with EFAs to pinpoint model misspecifications and gain a comprehensive understanding of empirical data. We recommend implementing the proposed two-bifactor model on independent samples in subsequent studies. Due to the lack of access to other instruments and a parallel clinical interview, general psychometric properties in this study are limited to internal consistencies. We cannot provide a validated cut-off score for the general German population. However, Krüger-Gottschalk et al. (2017) describe the psychometric properties in a German clinical sample. Additionally, no statements can be made about the impact of trauma type, frequency of exposure and time passed since exposure on the factorial structure, as they have not been included in the present analysis. The factorial structure of the measure is contingent on the sample population; our findings can be generalized to the German general population and the German version of the PCL-5. However, further research is required to determine the stability of this factorial structure in independent samples, in clinical samples with confirmed PTSD diagnosis, across diverse cultural backgrounds and other language versions of the PCL-5.

In conclusion, our research demonstrates that the PCL-5 is a reliable and effective screening tool for assessing PTSD in clinical and research settings. Its economic structure with a general factor and two specific facets (trauma related symptoms and global psychological symptoms) renders the diagnostical algorithm unnecessary. Given these findings we suggest using the cut-off scoring method.

Disclosure statement

During the revision of this work, the author used ChatGPT-3.5 in order to refine language. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication. The authors reported no potential conflict of interest.

Data availability statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to the informed consent given by the participants.

Additional information

Funding

This work is funded by the Open Access Publishing Fund of Leipzig University supported by the German Research Foundation within the program Open Access Publication Funding.