A Bayesian Structural Equation Modeling Factor Analysis of the UCLA Posttraumatic Stress Disorder Reaction Index for DSM-5 in a Polyvictimized Sample of Adolescents

Abstract

The revised criteria for posttraumatic stress disorder (PTSD) in the fifth edition of the Diagnostic and Statistical Manual necessitated the development of new screening tools for youth, one of the most widely used of which is the UCLA Posttraumatic Stress Disorder Reaction Index for DSM-5 (RI-5). Thus far, the few studies that have investigated the RI-5’s factor structure have supported a four-factor model. However, to date this research has been limited to youth with histories of exposure to single-event traumatic stressors, a significant limitation as evidence suggests many trauma-exposed youth report exposure to multiple types of traumatic stressors, or polyvictimization. It is imperative to determine the generalizability of previous factor models to specific populations which they are purported to represent. We investigated whether the RI-5’s four-factor model replicated in a sample of 455 polyvictimized justice-involved adolescents. Initial confirmatory factor analysis demonstrated that the four-factor model did not converge. Therefore, we utilized Bayesian Structural Equations Modeling (BSEM) to determine why the previously proposed factor structure did not converge. The BSEM model suggested that the global factor structure was acceptable and did not require addition or subtraction of any factor or cross-factor loadings. However, small and moderate residual covariances resulted in model misspecification, suggesting there may be additional associations not captured by the current DSM-5 model for polyvictimized youth. Future work should continue examining the RI-5’s factor structure in order to better understand whether the current results are unique and how measurements assessing DSM-5 PTSD symptom criteria perform in diverse trauma-exposed youth populations.

The most recent revision of the Diagnostic and Statistical Manual (DSM-5; American Psychiatric Association, 2013) included a number of changes to the diagnostic criteria for posttraumatic stress disorder (PTSD). One of these changes was to increase the number of symptom categories; whereas DSM-IV and DSM-IV-TR conceptualized PTSD as comprising three categories of symptoms, DSM-5 incorporated findings from numerous studies (see Yufik & Simms, 2010 for a review) suggesting that a four-factor model of PTSD better captures the symptom structure of the disorder. Thus, the updated criteria comprise the four symptom clusters of intrusion (Criterion B), avoidance (Criterion C), negative alterations in cognitions and mood (Criterion D), and alterations in arousal and reactivity (Criterion E), which derive from 20 individual symptoms.

One of the most widely-utilized and well-validated PTSD screening tools for children and adolescents is the UCLA Posttraumatic Stress Disorder Reaction Index (Pynoos & Steinberg, 2014; Steinberg et al., 2013). Numerous studies suggest that the most recent version of the measure designed to correspond to DSM-5 PTSD criteria, the RI-5, demonstrates high internal consistency, as well as criterion-related, convergent, and discriminant validity (Kaplow et al., 2020). For example, in the original validation study, Kaplow and colleagues (2020) found evidence for criterion-related validity using bivariate correlations among the RI-5 symptom categories and the total score of the Short Mood and Feelings Questionnaire (Angold et al., 1995), a brief measure that assesses depressive symptoms in children and adolescents. Additionally, the authors found that the RI-5 total score discriminated youth with and without diagnosable PTSD when compared to the child and adolescent version of the Clinician-Administered PTSD Scale for DSM-5 (Pynoos et al., 2015), which is considered the “gold standard” for diagnosing PTSD. Other studies (Takada et al., 2018) have established that the RI-5 demonstrates convergent validity with the subscales of the Trauma Symptom Checklist for Children (Briere, 1996), as well as divergent validity (Doric et al., 2019) in relation to symptoms of anxiety and depression assessed by the Revised Child Anxiety and Depression Scale (Chorpita et al., 2005).

In addition to establishing preliminary evidence for adequate reliability and validity, previous research has also investigated the factor structure of the RI-5. For example, in a sample of over 300 youth recruited from multiple locations across Japan, Takada and colleagues (2018) found that the four-factor DSM-5 model comprising intrusion, avoidance, negative alterations in cognitions and mood, and alterations in arousal and reactivity provided an adequate fit to their data. Similarly, using a sample of over 4,000 adolescents from eleven different countries, Doric and colleagues (2019) also found evidence for a four-factor structure. However, these studies are limited in that they include youth who either primarily reported exposure to single-event traumatic stressors (Takada et al., 2018) or youth who may not have actually experienced a traumatic event (Doric et al., 2019). Therefore, additional research using diverse samples of trauma-exposed youth is needed to replicate previous findings regarding the factor structure of the RI-5.

In particular, studies investigating the factor structure of the RI-5 would benefit from including youth who endorse polyvictimization, defined by exposure to diverse types of interpersonal traumas and other traumatic events (Finkelhor et al., 2007b). Research shows that a significant proportion of trauma-exposed adolescents can be characterized as polyvictims (Finkelhor et al., 2007a). Additionally, evidence suggests that polyvictimized youth are at elevated risk for displaying more severe PTSD symptoms (Soler et al., 2012) over longer periods of time (Finkelhor et al., 2007a) when compared to traumatized youth who have not experienced polyvictimization. Furthermore, polyvictimized youth are more likely to experience other behavioral health problems, such as depression, anxiety, aggression, and suicidality (Charak et al., 2019; Grasso et al., 2016), as well as lower levels of family and peer support (Turner et al., 2017), all of which may affect the way that youth express and report PTSD symptoms. Therefore, because polyvictimized youth likely make up a substantial proportion of those who are assessed with PTSD symptom screening measures and may express PTSD symptoms differently than youth who report exposure to a single type of traumatic event, more research is needed to determine the factor structure of the RI-5 in this population.

One group of adolescents that has evidenced disproportionately high rates of polyvictimization are those youth who are involved in the juvenile justice system. Previous studies have found that justice-involved youth report particularly high rates of polyvictimization compared to other cohorts of victimized adolescents (Ford et al., 2010; Ford, Grasso, Hawke, et al., 2013). As might be expected with such high rates of trauma exposure, the prevalence of PTSD among justice-involved youth is also elevated in comparison to their peers in the community (Wood et al., 2002). Furthermore, youth involved in the juvenile justice system who report polyvictimization are also more likely than those who are not polyvictims to evidence severe emotional and behavioral difficulties, such as anxiety, depression, and suicide ideation (Ford, Charak, et al., 2018; Ford, Grasso, Hawke, et al., 2013). Thus, given that justice-involved youth often experience exposure to diverse types of traumatic events and evidence high levels of posttraumatic stress symptoms, this population provides an important context in which to investigate the psychometric properties of measures commonly used to assess PTSD, such as the RI-5.

To this end, the present study sought to investigate the structure of the RI-5 in a sample of youth with exposure to diverse traumatic events. Under the current replication crisis, it is imperative to not just assume that factor structures replicate, but instead examine whether previous findings replicate in diverse populations. This is especially true for populations that may have been under-represented in initial investigations of a measure’s factor structure. Thus, we investigated whether the DSM-5 four-factor model of the RI-5 replicated in this sample of polyvictimized, justice-involved youth. When the four-factor model did not replicate in the current sample, we utilized advanced Bayesian modeling techniques to determine the sources of model misfit and implied model modifications.

Method

Participants

Participants included 455 youth (341 boys, 112 girls, 2 transgender youth) recruited from a juvenile detention center in the Western United States. Participants were between the ages of 12 and 19 years old (M = 15.95, SD = 1.33). Consistent with the demographics of detention centers in this geographic region, 44.6% of the sample identified as White, 33.3% as Latino(a)/Hispanic, 7.1% as Biracial/Multiracial, 5.4% as Black/African American, 3.9% as Native American, 2.7% identified as Pacific Islander/Native Hawaiian, 2.4% as an Other race.

Measures

Posttraumatic stress symptoms.

The University of California at Los Angeles (UCLA) Posttraumatic Stress Disorder Reaction Index for DSM-5 (RI-5; Pynoos & Steinberg, 2014) is a self-report screening measure that assesses lifetime exposure to traumatic events and past-month symptoms of PTSD in youth. In the first part of the measure, the RI-5 screened for exposure to 14 potentially traumatic events according to Criterion A. In addition to Criterion A events, youth were also screened for exposure to other types of potentially traumatic events that are developmentally salient, including parental substance use, neglect, parental incarceration, and emotional abuse, given this was part of the larger study’s aims. Youth who endorsed exposure to diverse traumatic events were asked to choose the event they perceived as the “worst,” which was used as the index trauma to assess PTSD symptoms. In the second part of the measure, youth indicated how frequently (from 0 = none of the time, to 4 = most of the time) they experienced past-month PTSD symptoms. The RI-5 includes 27 items assessing the 20 DSM-5 PTSD symptoms, as well as four additional items that assess whether respondents meet criteria for the dissociative subtype. Because the purpose was to investigate the structure of the four DSM-5 PTSD symptom clusters, and in line with previous work investigating the factor structure of the RI-5 (Doric et al., 2019), the current study included only the 27 primary symptom items. Previous psychometric research suggests that score of 35 or more on the DSM-5 symptom total score indicates probable PTSD (Kaplow et al., 2019). In the current study, the Cronbach’s alphas were .88 for Criterion B (Intrusion), .68 for Criterion C (Avoidance), .87 for Criterion D (Changes in Cognition and Mood), and .74 for Criterion E (Arousal).

Procedure

All study procedures were approved by the institutional review boards of the University of Utah and the Utah Department of Human Services. Informed consent was obtained from all individual participants included in the study. Specifically, youths’ legal guardians provided informed consent during visiting hours at the juvenile detention center, after which youth were approached by research staff to determine if they were interested in participating in the study. If youth were interested in participating, they provided informed assent and completed self-report measures on a laptop in a private visiting room at the detention center. No compensation was offered to youth for their participation.

Data Analysis

We conducted all analyses using Mplus version 8.0 (Muthén & Muthén, 2012–2018). Missing data ranged from 0.9 to 2.9% for items assessing past-month PTSD symptoms. Data were missing because youth were given the option of “choose to skip” for every question on the RI-5.

Initial attempts to fit a confirmatory factor analysis (CFA) based on the four-factor model resulted in convergence issues due to a PSI-Matrix error. Additional exploratory diagnostics indicated that the PSI-Matrix error was a result of the Criterion C (Avoidance) factor consisting of only two indicators, as per the DSM-5 criteria, which commonly results in model identification and convergence issues (Schmitt et al., 2018). The PSI-Matrix error rendered the model uninterpretable, and this issue was persistent across alternative factor models suggested by the literature (details available from the authors by request). Therefore, Bayesian structural equation modeling (BSEM) was used to determine whether the factor structure proposed by the DSM-5 was an acceptable model for the RI-5 in the current sample. BSEM is an exploratory modeling technique that can be used to identify a better fitting factor structure, as this technique can suggest modifications to the tested factor structure and can assist in determining whether the assumptions made by the measurement model are being met. Additionally, BSEM can be used to determine if the factor structure is a good representation of the construct, if items cross-load (i.e., represent multiple factors simultaneously), and if there are important correlations between residual variances. Bayesian methods differ from frequentist methodologies (e.g., maximum likelihood) by treating parameters as variables with distributions, rather than as constants. Bayesian methods function by specifying priors, defined as the distributions of parameters, which can be diffuse and non-informative, or informative by specifying distributional properties, specified by theory, pilot studies, or revision of hypotheses (Muthén & Asparouhov, 2012). Additionally, whereas both cross-loadings and residual covariances are assumed to be zero in traditional frequentist CFAs, the BSEM approach tests the frequentist assumption that these values should be fixed to zero by specifying informative priors that allow these estimates to vary slightly around zero (see Muthén & Asparouhov, 2012 for a more expansive review). This feature is particularly relevant for the present study given both the paucity of research validating the DSM-5 factor structure in polyvictimized youth, as well as evidence that youth with a history of polyvictimization may experience more complex or intense symptoms. the assumption that the cross-loadings and residual covariances are zero is likely too strong, further extending the utility of applying the BSEM approach in the present study.

To fit the BSEM models, a uniform metric was first created by standardizing the variables for the scale free model. Latent variables (i.e., the various PTSD symptom clusters) were allowed to correlate with one another, error covariances were fixed to zero, and factor variances were fixed at one. A CFA model without cross-loadings and residual covariances was conducted first. Next, an acceptable prior variance for cross-loadings was selected to replicate the 95% confidence interval for the difference between the observed and replicated chi-square values from the CFA model without cross-loadings. We then increased the cross-loading prior variance, and identified the prior variance value which converged quickly enough and minimized the confidence interval limits for the prior variances. Finally, we computed the inverse Wishart priors distribution using the residual variances from the CFA model to specify the priors for the residual variances and the covariances between residuals.

Model fit of the exploratory BSEM model was assessed using the posterior predictive p-value (PPP) computed for the 95% confidence interval for the difference between the observed and replicated chi-square values. Low PPP indicates poor fit. A 95% confidence interval for the difference in the f statistic for the real and replicated data was generated, with a positive lower limit in line with a low PPP and poor fit, and an excellent-fitting model having a PPP around .5 and an f statistic difference of zero being near the center of the confidence interval (Asparouhov & Muthén, 2010; Muthén & Asparouhov, 2012). After conducting the BSEM analyses, we tested the frequentist (i.e., non-Bayesian) CFA implied by our BSEM model. We utilized a number of indices to evaluate the final model’s overall fit, including the comparative fit index (CFI; Hu & Bentler, 1999), the Tucker-Lewis index (TLI; Hu & Bentler, 1999), the root-mean square error of approximation (RMSEA; Steiger, 1990), and the standardized root mean square residual (SRMR; Hu & Bentler, 1999). Based on the guidelines proposed by Hu and Bentler (1999), an excellent fit is obtained when SRMR ≤ .08, CFI and TLI ≥ .95, and RMSEA ≤ .06; an adequate fit is obtained when CFI and TLI ≥ .90, and when SRMR and RMSEA ≤ .08.

Results

Descriptive Sample Statistics

On average, participants reported exposure to five (SD = 2.59; range 1–13) different types of potentially traumatic events according to DSM-5 Criterion A. There were thirty-five youth (7.69%) who reported exposure to only one Criterion A traumatic event but also endorsed exposure to at least one other developmentally-salient traumatic event, such as witnessing parental substance use or parental incarceration, and as such were retained for the current analyses. In regards to the traumatic experiences participants identified as the worst (index) event, the most common index events endorsed included the death of someone they knew (10.1%), seeing parental substance use (7.7%), hearing about the death or injury of a loved one (7.3%), and sexual abuse or assault (6.6%). Using the 20 DSM-5 PTSD symptom items, the average total symptom score for the RI-5 was 27.82 (SD = 17.28, range 0–74). One-hundred and sixty youth scored 35 or higher on the RI-5, suggesting that 35.2% of the sample likely met criteria for PTSD.

Results of Exploratory BSEM Model

The BSEM model was estimated with all cross-loadings such that all of the indicators were allowed to load onto all latent variables rather than only those latent variables specified in the DSM-5 PTSD criteria (i.e., the Criterion B-E symptom clusters). Additionally, the cross-loadings were specified to have a mean of zero and a small prior variance. In standard CFA these cross-loadings would be assumed to be exactly zero, whereas measurement theory would suggest that we are unlikely to create measures that are this accurate, especially across samples that may differ from those which measures were initially validated (Asparouhov et al., 2015). Thus, by utilizing BSEM, we can specify priors, which allow these cross loadings to have distributional properties, and thus vary slightly around zero. Consistent with the guidelines suggested by Asparouhov, Muthén, and Morin (2015), the final selected BSEM model demonstrated adequate fit, PPP = .297, 95% CI [−76.0, 90.0], and is presented in Figure 1. The results of the full BSEM model are presented in the supplemental materials.

Figure 1.

Four-Factor CFA Model with All Significant Residual Covariances. This figure shows depicts the four-factor model with corresponding residual covariances. The separate residual covariances between Items 15 and 27, Items 19 and 22, Items 25 and 26, Items 4 and 27, Item 26 and 27, and Items 7 and 12 were included in the constrained CFA model. NACM = Negative Alterations in Cognition and Mood.

Our exploration identified small cross-loading priors variances of .01. Additionally, none of the cross-loadings were significant. Significant cross-loadings can indicate the need to add or drop a latent factor or can signal where the measurement model assumptions failed. The lack of significant cross-loadings suggests that the factors are a reasonable characterization of the construct underlying the indicators, with the addition of cross-loadings that allowed all relevant information to be capitalized on. Sensitivity analyses indicated that when these cross-loadings were fixed to zero, the model fit was less good, but was still adequate. Thus, the four-factor model confirmed in previous studies (e.g., Doric et al., 2019; Takada et al., 2018) may be a good representation of the RI-5 data in the current sample; however, without additional considerations, it does not produce an interpretable model due to convergence issues and additional model misspecification in the residual covariances.

The addition of correlations between residual variances significantly improved the model fit. The degrees-of-freedom parameter for the priors specification for the residual covariances was selected to be 200 after three iterations (Muthén & Asparouhov, 2012); this yielded a model with fast convergence and a PPP > .05, consistent with suggested benchmarks (Asparouhov et al., 2015). There were many small significant correlations, most which were among items that loaded onto the same factor. The two items that loaded onto Criterion C, which contributed to the initial PSI-Matrix error in the maximum likelihood CFA, had only one additional significant residual correlation outside of the Criterion C items. This, in conjunction with the lack of significant cross-loadings, suggest that the Criterion C items should not be re-distributed to a different factor, since there is no clear factor to which these items belong, and no clear conceptual factor which the pattern of correlations would suggest to be missing.

Determining to Retain or Drop a Factor in Final Model

BSEM can also be used to help determine whether there is a missing factor, or too many factors, in a given model and also provides clarification about which measurement assumptions are not being met by the data. Results indicated that the covariances between factors were moderate to relatively high, ranging between .439 and .778, but not high enough to suggest that reducing the number of factors would be recommended. The presence of many large residual covariances (~r = .25 – r = .4) can indicate that there is a missing factor. Whereas our model produced many small residual covariances, in addition to the few moderate ones, these do not suggest that there is missing factor. Rather, the many small residual correlations suggest that the failure of the CFA models may be due to small differences between the sample and the population, and that a major change to the factor structure of the CFA is not called for. These small residuals, in conjunction with only two indicators loading onto a factor, likely resulted in the model convergence issues and poor CFA fit.

Results of Final Confirmatory Factor Analysis Model

Based on the results from the model identified through the exploratory BSEM model fitting process, we re-fit the maximum likelihood CFA and included all the significant residual covariances. This yielded a model with acceptable fit, RMSEA = .039, SRMR = .038, and CFI = .956 (see Figure 1). We also conducted an additional sensitivity analysis of a maximum likelihood CFA where the small residual correlations implied by the BSEM model were constrained to zero, which is in line with only including meaningful residual correlations to avoid over-fitting of the CFA model (Asparouhov, Muthén, & Morin, 2015). This model, which included only the six moderate residual correlations, also had an adequate fit (RMSEA = .049, SRMR = .048, and CFI = .914), and the magnitude and direction of the factor structure did not change when the majority of the residual correlations were constrained. It is important to note that this model included many small but significant correlations between residual variances. This is not consistent with assumptions made by traditional measurement models. Ideally, if the theoretical construct had been correctly parsed, there would have been few or no correlations between residuals. Instead, the pattern in the current results shows that there are relationships between individual indicators (i.e., the items on the RI-5) that are not accounted for by the covariances between factors. Measurement theory frequently attributes these relationships to non-construct-related relationships, such as parallel formats in questions. For example, the residual correlations between items “I have thoughts like ‘I am bad.”” and “I have thoughts like, ‘The world is really dangerous.’” could be due to the parallel wording of the questions. This may account for some of the small correlations between residuals, yet this may not fully account for the number of significant moderate correlations in the current results. Instead, these moderate correlations likely indicate additional relations among individual indicators not captured by the measurement model. For example, the residual correlations between “I don’t feel like doing things with my family or friends or other things that I liked to do” and “I feel alone even when I am around other people” could be due to some unique component of Negative Alterations in Cognition and Mood Factor (e.g., impacts on relationship quality) not shared by the remainder of the Negative Alterations in Cognition and Mood items. These moderate residual correlations may potentially be due to high rates of co-occurring PTSD symptoms.

Discussion

The revised diagnostic criteria for PTSD in DSM-5 proposed a four-factor conceptualization of the disorder, prompting the development of updated PTSD screening tools. Thus far, few published studies have confirmed the factor structure of one of the most widely-used screening tools for child and adolescent populations, the UCLA PTSD-RI for DSM-5 (RI-5; Pynoos & Steinberg, 2014). Furthermore, those studies that have investigated the factor structure of the RI-5 have primarily included youth who reported exposure to single-event traumatic stressors. This is a limitation, as evidence suggests that many youth who report trauma exposure endorse polyvictimization, or exposure to multiple types of traumatic events. The current study attempted to replicate previous findings by testing a four-factor structure of the RI-5 in a sample of polyvictimized adolescents.

Results of the initial CFA model in which we followed the traditional assumptions of CFA (i.e., not allowing the residual covariances to vary from zero) was uninterpretable due to a non-positive definite PSI-Matrix error. When a model evidences significant misfit, it is important to examine the ways in which the model might be failing to meet the underlying measurement assumptions. Because the initial CFA model did not converge, the BSEM technique was used to determine if addition or subtraction of factors was necessary in order to improve the model, if individual items significantly predicted other factors, and if residual correlations were present. The BSEM model indicated the presence of both trivial and moderate residual correlations and the maximum likelihood CFA model suggested that the residual correlations may account for both the estimation error and significant model misfit. Ultimately, the identified BSEM model suggested that the four-factor model was the best fit, but only when the assumption that residual covariances were zero was relaxed. Furthermore, results also demonstrated high co-occurrence of symptoms among youth in this polyvictimized sample that have not been present in previous CFA studies performed on samples of youth who report exposure to single-event traumatic stressors.

The poor model fit and convergence issues, in conjunction with the many small residual correlations (Muthén & Asparouhov, 2012) and additional associations among individual indicators, suggests that the current sample likely differs from the sample in which the RI-5 was initially tested. For example, although previous research confirming the factor structure of the RI-5 also followed the DSM-5 by defining the Criterion C symptom cluster with only two items (Doric et al., 2019; Takada et al., 2018), these studies did not report convergence issues with the frequentist CFA model. Additionally, previous studies have not reported additional associations among the individual indicators outside of the specified symptom clusters, which in the present study suggested a high co-occurrence among the individual symptoms in their samples. In the present analyses, these additional associations needed to be accounted for through the residuals in order to produce acceptable model fit. It may be that the current sample, or the population the sample represents, experiences more symptoms of PTSD at greater intensities, and thus is more likely to report a high level of co-occurring symptoms than youth exposed to single-event traumas. Additionally, it is possible that polyvictimized youth report more diverse and severe PTSD symptoms due to substantial levels of comorbid emotional and behavioral difficulties, which have been reported in other research (Dierkhising et al., 2019). Future research should continue to investigate the extent to which youth with a history of polyvictimization experience and report PTSD symptoms differently when compared to youth with exposure to acute traumatic stressors, as this could have important implications for how PTSD screening tools are constructed.

The results of the current study also suggested that, despite moderate to high covariances among the four factors, reducing the number of factors was not recommended. Additionally, the small residual covariances produced by the model did not suggest that adding another factor would improve model fit. However, the final CFA model identified by BSEM produced only an adequate fit, thereby indicating that there was still some model misfit even after making the suggested adjustments. In sum, although the current results indicated that the RI-5 data in the present sample were best represented by four factors, the RI-5 items may not be adequately capturing the way in which polyvictimized youth experience or express PTSD symptoms. Namely, it is possible that the RI-5 is missing some key items that may be applicable to the way in which polyictimized youth experience posttraumatic stress. For example, scholars have previously posited that youth who experience chronic exposure to interpersonal trauma and other types of traumatic events may evidence a broader and more diverse set of symptoms, including disturbances in typical bodily functioning, dysfunctional close relationships, and a persistent negative sense of self (van der Kolk et al., 2009). Therefore, including such items on future measures aimed at assessing posttraumatic reactions in polyvictimized youth may provide more insight into the way in which this population experiences posttraumatic stress. It is also possible that specific items function differently for this population. Thus, in addition to considering how various groups of traumatized youth respond to current PTSD screening tools, it will also be important to continue assessing whether current PTSD screening tools are adequately capturing the breadth of posttraumatic symptoms trauma-exposed youth experience.

The implication that PTSD symptoms manifest in unique ways in polyvictimized youth aligns with theory suggesting that youth who experience repeated, chronic, and diverse traumatic events exhibit a different symptom presentation than youth with exposure to single-event stressors. For example, Terr (1991) differentiated between Type I traumas, or acute traumatic stressors, and Type II traumas, or traumatic experiences that are pervasive and repeated; she posited that exposure to Type II traumatic events would be associated with more complex symptom presentations and higher levels of overall dysregulation. More recent developmental models include dysregulation across multiple domains as a key component of post-trauma sequelae for those children and adolescents who are exposed to diverse traumatic events (van der Kolk, 2005). In line with this, scholars have proposed that a new diagnosis, Developmental Trauma Disorder (DTD), should be included in our diagnostic compendia (van der Kolk et al., 2009). This proposed diagnosis highlights the developmental consequences of exposure to pervasive trauma and victimization over the course of childhood and adolescence, which are characterized by profound levels of physiological, emotional, cognitive, and behavioral dysregulation. Although DTD ultimately was not included in the DSM-5, evidence has accumulated indicating that the diagnosis is considered clinically useful by treatment providers (Ford, Grasso, Greene, et al., 2013) and can be reliably assessed by standardized diagnostic tools (Ford, Spinazzola, et al., 2018). Given the current findings and previous research suggesting that youth exposed to repeated and diverse traumatic events may evidence a pattern of symptoms that is not accurately captured in the diagnostic criteria for PTSD in the DSM-5 (D’Andrea et al., 2012; Pynoos et al., 2009), further investigations are warranted into the links between specific trauma exposure typologies and distinct symptom presentations.

Alternatively, it is worth noting that the model convergence issues of the initial CFA, as well as the necessity of inclusion of residual covariances in future models, could be due to inherent flaws in the DSM-5’s four-factor model of PTSD. Indeed, after the DSM-5’s revisions to the PTSD criteria, some scholars have criticized the inclusion of the new symptoms due to the conceptual overlap with other DSM diagnoses, such as depression and anxiety disorders (Pai et al., 2017). Furthermore, previous research has found support for other conceptual models of PTSD among both adults and youth, including a five-factor model (Armour et al., 2012; Bennett et al., 2014), a six-factor model (Tsai et al., 2015), and a seven-factor model (Armour et al., 2015), though the six- and seven-factor models have been critiqued due to statistical problems and lack of construct validity among the factors (Rasmussen et al., 2019; Silverstein et al., 2018). Ultimately, future studies should investigate alternative conceptual models of PTSD. Additionally, future research would benefit from more extensive measurement explorations, such as BSEM, in order to determine whether and how frequentist CFA assumptions are being violated in the current DSM-5 measurement structure of PTSD.

Implications for Future Research

The findings of the current study have a number of implications for future research. First, future studies should continue to examine the factor structure of the RI-5, especially utilizing techniques such as BSEM, in order to explore not just if the CFA is of adequate fit, but also where the measurement model is failing assumptions and if there are implied modifications to the measurement model. Research continuing to investigate the psychometric properties of the RI-5, including internal consistency, discriminant validity, and convergent validity, with other trauma screening tools designed for use in children and adolescents is also necessary. This work also highlights the potential pitfalls of mono-method and self-report methods, which allow us to analyze statistically where misfit is originating, but which do not allow us to discern the reason that misfit is occurring. Future work should capitalize on multi-method studies to be better able to address validity concerns and disentangle which variance is accounted for by the measurement method, rather than the phenomenon of interest. Furthermore, more research is needed to better understand the ways in which polyvictimized youth might express posttraumatic stress symptoms in ways that differ from other populations. In addition to comparing the expressions of symptom evidenced by youth who have experienced polyvictimization versus those who have encountered single-incident events, it also will be important to further investigate the heterogeneity of posttraumatic stress symptom expression among polyvictims. For example, it is possible that certain aspects of polyvictimization exposure differentially influence the psychometric properties and factor structure of the RI-5. Indeed, previous research has suggested that the developmental timing (Dierkhising et al., 2019), chronicity (Grasso et al., 2016), and type (Charak et al., 2019) of polyvictimization influence the manifestation of posttraumatic stress. Thus, future research may benefit from using mixture modeling or integrative data analysis techniques (e.g., moderated nonlinear factor analysis) to examine whether various characteristics of polyvictimization exposure, such as the pattern or severity of exposure, influence the way in which the RI-5 performs, as this could better inform assessment and intervention strategies for this high-risk group of youth.

Another potential area for future research involves investigating whether the inclusion of dissociative symptoms alters the factor structure of the RI-5. The updated diagnostic criteria for PTSD in the DSM-5 included a new dissociative subtype specifier for individuals evidencing depersonalization and/or derealization (American Psychiatric Association, 2013). Although the RI-5 and other screening tools for PTSD include items that assess for the presence of dissociation, previous factor analytic studies of the RI-5 (Doric et al., 2019, Takada et al., 2018) have not included these symptoms given that they represent a subtype and are not considered primary symptoms of the diagnosis. However, previous research in adults comparing factor analytic models comprised of only primary DSM-5 PTSD symptoms versus models containing additional dissociative items has indicated that models that incorporated both primary PTSD symptoms and additional dissociation items provided the best fit to the data (Armour et al., 2014). Including dissociative symptoms in factor analytic studies involving polyvictimized youth may be particularly relevant, given that prior work has found disproportionately high levels of dissociation among youth who report exposure to diverse types of traumatic events (Kisiel et al., 2014), including youth in the juvenile justice system (Bennett et al., 2015; Plattner et al., 2003). Therefore, investigating models that include both primary PTSD and dissociative symptoms, particularly using newer statistical techniques for analyzing item interrelations, such as network analysis (Cramer et al., 2020; Ross et al., 2020), may ultimately provide a better understanding of how dissociative symptoms are related to other PTSD symptom clusters and how patterns of symptom endorsement differ among youth exposed to unitary stressors versus diverse traumatic events.

Strengths and Limitations

Among its strengths, this study used advanced statistical techniques to assess the measurement fit and misfit associated with the four-factor model of the RI-5 in a sample of polyvictimized adolescents. Additionally, the study used data drawn from a racially diverse sample of justice-involved youth, a population that is particularly vulnerable to experiencing posttraumatic stress symptoms and one that is typically under-served. However, there are a number of limitations that should also be acknowledged. For example, participants for the present study were drawn from a different population than the original sample of youth in the initial validation study of the RI-5. Namely, the current sample was comprised of youth involved in the juvenile justice system, a population that is routinely found to have experienced particularly high rates of polyvictimization. Therefore, the current findings may not generalize to other populations of youth who have experienced only discrete types of traumatic events or who are not involved in the justice system. Additionally, this study does not represent a full exploration of the RI-5 measurement system. Instead, it serves as an exploration into the reason the proposed factor structure of the RI-5 fails to converge in a sample for which measurement of PTSD is crucial. Explorations such as these are essential to preserving the replicability of previous research, as well as provide important information on the functionality of scales in subpopulations which the measurement structure should represent. Furthermore, it should be noted that for those youth who identified their index traumatic event as exposure to the death of someone close to them, the RI-5 does not inquire as to whether that death was sudden or accidental, and so it cannot be determined whether the death about which youth were reporting would be considered a Criterion A event. However, previous research has found that bereaved youth who have lost a caregiver due to an anticipated death, such as the result of a prolonged medical illness, exhibit higher levels of posttraumatic stress than did youth whose loved one died unexpectedly (Kaplow et al., 2014), suggesting it may be developmentally appropriate to assess for posttraumatic stress symptoms in youth who report exposure to deaths that are not necessarily violent, unexpected, or accidental. Despite these limitations, this study contributes novel findings to the current literature that may have greater implications in regards to how PTSD symptoms may manifest in polyvictimized youth and how current PTSD screening measures function in this population.

In summary, this study utilized a sample of adolescents with diverse trauma histories to explore the modifications necessary to the proposed DSM-5 structural model of PTSD symptoms as assessed by the RI-5. The BSEM technique was used to determine if addition or subtraction of factors was necessary in order to improve the model, if items were significantly predicting other factors, and if residual correlations were present. The identified BSEM model suggests that the four-factor model was the best fit, but only when the assumption that residual covariances were zero was relaxed. Results also implied high co-occurrence of symptoms, suggesting that PTSD may manifest differently in youth with exposure to diverse traumatic events, and thus highlighting the possibility that current PTSD screening tools may not accurately capture the unique way in which polyvictimized youth experience and report PTSD symptoms. The current findings speak to the importance of continuing to investigate posttraumatic stress in polyvictimized youth, as this could facilitate the development of more effective methods for assessing and treating the negative sequelae of polyvictimization in this population.

Table 1.

Factor Correlations from Unconstrained Four-Factor CFA Model

	Estimate	S.E.	Two-Tailed P-Value
Criterion B correlated with
Criterion C	0.819	0.043	0.000
Criterion D	0.814	0.025	0.000
Criterion E	0.724	0.034	0.000
Criterion C correlated with
Criterion D	0.773	0.026	0.000
Criterion E	0.608	0.043	0.000
Criterion D correlated with
Criterion E	0.812	0.029	0.000

Table 2.

Factor Loadings from Unconstrained Four-Factor CFA Model

	Estimate	S.E.	Two-Tailed P-Value
Criterion B regressed on
Item 18	0.835	0.032	0.000
Item 10	0.713	0.040	0.000
Item 5	0.715	0.040	0.000
Item 11	0.782	0.035	0.000
Item 14	0.822	0.033	0.000
Criterion C regressed on
Item 13	0.863	0.042	0.000
Item 3	0.849	0.025	0.000
Criterion D regressed on
Item 23	0.518	0.051	0.000
Item 2	0.653	0.040	0.000
Item 9	0.537	0.046	0.000
Item 16	0.628	0.043	0.000
Item 15	0.677	0.041	0.000
Item 19	0.620	0.041	0.000
Item 6	0.571	0.047	0.000
Item 22	0.594	0.046	0.000
Item 25	0.644	0.047	0.000
Item 27	0.448	0.051	0.000
Item 7	0.542	0.050	0.000
Item 17	0.642	0.042	0.000
Item 12	0.595	0.045	0.000
Criterion E regressed on
Item 4	0.541	0.045	0.000
Item 20	0.379	0.063	0.000
Item 26	0.610	0.049	0.000
Item 1	0.514	0.046	0.000
Item 24	0.627	0.045	0.000
Item 8	0.637	0.040	0.000
Item 21	0.618	0.042	0.000

Table 3.

Residual Covariance Estimates from Unconstrained Four-Factor CFA Model

	Estimate	S.E.	Two-Tailed P-Value
Item 1 correlated with
Item 9	0.094	0.038	0.013
Item 20	−0.100	0.043	0.017
Item 27	0.120	0.041	0.004
Item 2 correlated with
Item 6	−0.107	0.034	0.002
Item 7	−0.039	0.034	0.251
Item 9	0.097	0.039	0.013
Item 23	−0.056	0.038	0.135
Item 24	−0.070	0.035	0.045
Item 3 correlated with
Item 13	−0.206	0.053	0.000
Item 4 correlated with
Item 15	0.086	0.038	0.023
Item 17	−0.064	0.031	0.040
Item 20	−0.101	0.042	0.017
Item 24	−0.098	0.043	0.021
Item 27	0.166	0.043	0.000
Item 5 correlated with
Item 6	−0.066	0.032	0.041
Item 6 correlated with
Item 12	−0.044	0.029	0.124
Item 22	0.176	0.040	0.000
Item 24	0.052	0.039	0.179
Item 7 correlated with
Item 9	−0.076	0.037	0.039
Item 12	0.182	0.044	0.000
Item 15	−0.045	0.035	0.196
Item 17	0.227	0.046	0.000
Item 20	0.114	0.056	0.041
Item 25	−0.024	0.035	0.498
Item 9 correlated with
Item 12	−0.069	0.035	0.047
Item 16	0.055	0.036	0.130
Item 19	−0.083	0.038	0.032
Item 22	−0.095	0.037	0.010
Item 23	−0.076	0.040	0.060
Item 10 correlated with
Item 11	−0.054	0.029	0.060
Item 12 correlated with
Item 13	0.098	0.033	0.003
Item 15	−0.067	0.033	0.042
Item 17	0.183	0.043	0.000
Item 19	−0.035	0.030	0.231
Item 25	−0.008	0.031	0.797
Item 13 correlated with
Item 24	0.048	0.034	0.155
Item 14 correlated with
Item 18	−0.084	0.022	0.000
Item 15 correlated with
Item 17	−0.033	0.034	0.339
Item 27	0.113	0.039	0.004
Item 16 correlated with
Item 19	−0.135	0.035	0.000
Item 22	−0.141	0.035	0.000
Item 24	−0.076	0.035	0.029
Item 25	−0.094	0.036	0.009
Item 17 correlated with
Item 20	0.124	0.041	0.002
Item 22	0.002	0.027	0.954
Item 19 correlated with
Item 22	0.172	0.039	0.000
Item 27	−0.063	0.035	0.071
Item 21 correlated with
Item 26	−0.107	0.034	0.002
Item 27	−0.055	0.038	0.146
Item 22 correlated with
Item 25	0.073	0.034	0.034
Item 24 correlated with
Item 26	−0.091	0.040	0.021
Item 25 correlated with
Item 26	−0.160	0.034	0.000
Item 27	−0.085	0.038	0.023
Item 26 correlated with
Item 27	0.165	0.043	0.000