Longitudinal Measurement Invariance of Posttraumatic Stress Disorder in Deployed Marines

Transcription

1 Journal of Traumatic Stress June 2017, 30, Longitudinal Measurement Invariance of Posttraumatic Stress Disorder in Deployed Marines Ateka A. Contractor, 1,2 Elisa Bolton, 1 Matthew W. Gallagher, 3,4 Charla Rhodes, 1,2 William P. Nash, 5 and Brett Litz 1,2 1 VA Boston Healthcare System, Massachusetts Veterans Epidemiology Research and Information Center (MAVERIC), Boston, Massachusetts, USA 2 Department of Psychiatry, Boston University School of Medicine, Boston, Massachusetts, USA 3 Department of Psychology, University of Houston, Houston, Texas, USA 4 Texas Institute for Measurement, Evaluation, and Statistics, Houston, Texas, USA 5 Headquarters, United States Marine Corps, Arlington, Virginia, USA The meaningful interpretation of longitudinal study findings requires temporal stability of the constructs assessed (i.e., measurement invariance). We sought to examine measurement invariance of the construct of posttraumatic stress disorder (PTSD) as based on the Diagnostic and Statistical Manual of Mental Disorders indexed by the PTSD Checklist (PCL) and the Clinician-Administered PTSD Scale (CAPS) in a sample of 834 Marines with significant combat experience. PTSD was assessed 1-month predeployment (T0), and again at 1-month (T1), 5-months (T2), and 8-months postdeployment (T3). We tested configural (pattern of item/parcel loadings), metric (item/parcel loadings on latent factors), and scalar (item/parcel-level severity) invariance and explored sources of measurement instability (partial invariance testing). The T0 best-fitting emotional numbing model factor structure informed the conceptualization of PTSD s latent factors and parcel formations. We found (1) scalar noninvariance for the construct of PTSD as measured by the PCL and the CAPS, and for PTSD symptom clusters as assessed by the CAPS; and (2) metric noninvariance for PTSD symptom clusters as measured by the PCL. Exploratory analyses revealed factor-loading and intercept differences from pre- to postdeployment for avoidance symptoms, numbing symptoms (mainly psychogenic amnesia and foreshortened future), and the item assessing startle, each of which reduced construct stability. Implications of these findings for longitudinal studies of PTSD are discussed. Longitudinal research answers questions about the causes and consequences of exposure to potentially traumatizing events (PTE). Repeated assessments of key outcomes, such as posttraumatic stress disorder (PTSD) symptoms, allow investigators to draw causal inferences about risk and protective factors and to examine patterns of change over time (e.g., Nash et al., 2015). The validity of such studies is contingent upon the statistical property of measurement invariance in the construct of interest (Borsboom, 2006; Meredith & Teresi, 2006). Ateka A. Contractor is now at the Department of Psychology, University of North Texas, Denton, TX. This study was funded by the VA Health Service Research and Development (SDR ) and by the U.S. Marine Corps and Navy Bureau of Medicine and Surgery. The authors acknowledge the Marine Resiliency Study (MRS) team who made this work possible. Correspondence concerning this article should be addressed to Brett Litz, VA Boston Healthcare System, 150 S. Huntington Avenue, 13-B74, Jamaica Plain, MA Brett.Litz@va.gov Copyright C 2017 International Society for Traumatic Stress Studies. View this article online at wileyonlinelibrary.com DOI: /jts Participants need to interpret the items in a conceptually similar manner across time (Vandenberg & Lance, 2000). Different levels of measurement invariance are tested in a progressively restrictive step-wise manner. Configural invariance requires that the construct (latent factors) be associated with identical item sets across time. Metric invariance requires that the item loadings be consistent for the latent factors, which demonstrates that the factors have the same meaning over repeated assessments. Finally, scalar or strong factorial invariance requires that item intercepts (item s point of origin) be equal across temporal groups to demonstrate construct comparability. If metric and scalar invariance are assured, scores on the measure have the same unit of measurement (factor loading) and the same origin (intercept) over each measurement interval. Consequently, inferences about patterns of change or group differences can be attributable to true score differences (Chen, 2008; Gregorich, 2006; Little & Slegers, 2005). Few published longitudinal studies of PTSD have examined construct invariance (see Table 1). The mixed study results are attributable to several factors. First, there is variability in the conceptualization of the PTSD construct across studies. It is possible to test the single overarching construct of PTSD as 259

2 260 Contractor et al. Table 1 Studies Testing Longitudinal Measurement Invariance of the Construct of Posttraumatic Stress Disorder Study Sample Timeline PTSD measure Instructional set PTSD model Invariance testing results Baschnagel et al. (2005) Undergraduate students King et al. (2009) Medical patients Krause et al. (2007) Lommen et al. (2014) Women with experience of interpersonal violence (IPV) Two samples of deployed Danish soldiers Two timepoints: 1 and 3 months post-9/11 incident Three timepoints: averaging 10 days, a month, and 5½ months post-er admission Two timepoints: within 3 months, and 1 year following IPV Two timepoints: 2 months pre- and postdeployment (1 st sample); predeployment and 5 months postdeployment (2 nd sample) Posttraumatic Diagnostic Scale (PDS) Impact of Event Scale-Revised (IES-R) Participants referenced the 9/11 incident Participants referenced the worst event experienced on the Traumatic Events Questionnaire Participants not endorsing a specific event referenced most distressing event PTSD Checklist Participants referenced IPV in past year Posttraumatic Symptom Scale-Self Report (PSS) For sample deployed to Afghanistan, PSS predeployment referenced the most aversive life event, and PSS postdeployment referenced a deployment experience Dys Metric noninvariance 4-factor (DSM-IV factors and sleep factor) Configural and partial metric invariance (unstable intrusion factor loadings) Dys Configural, metric, and phi invariance 1-factor Item threshold noninvariance Continued

3 Longitudinal Measurement Invariance of PTSD 261 Table 1 Continued Study Sample Timeline PTSD measure Instructional set PTSD model Invariance testing results Meis et al. (2011) Two samples of deployed soldiers Wang et al. (2012) Adolescents who experienced earthquake in China Three timepoints: 1-month predeployment, 2 3 months and 15-months postdeployment (1 st sample)two timepoints: during deployment and 12-months postdeployment (2 nd sample) Two timepoints: 5 and 6 months after the earthquake PTSD Checklist Civilian Version (PCL-C) predeployment; PCL Military Version (PCL-M) postdeployment. Chinese version of the PTSD Checklist Participants referenced a stressful experience for the PCL-C, and stressful military experiences for the PCL-M Participants referenced the earthquake Dys Configural, metric, and phi invariance Dys, EN, and DA Configural and metric invariance; scalar and strict factorial noninvariance Note. PTSD = posttraumatic stress disorder; ER = emergency room; Dys = dysphoria model; DSM IV-TR = Diagnostic and Statistical Manual of Mental Disorders (4th ed., text rev.); IPV = interpersonal violence; PSS = Posttraumatic Symptom Self-Scale report; PCL-C = PTSD Checklist, civilian version; PCL-M = PTSD Checklist, military version; EN = emotional numbing model; DA = dysphoric arousal model; all studies assessed DSM-IV PTSD symptoms.

4 262 Contractor et al. defined by individual PTSD items or by the best-fitting factoranalytical model symptom clusters or parcels, or to investigate the invariance of the PTSD symptom clusters as defined by the best-fitting factor-analytical model. Using the former approach allows researchers to make a total score comparison, whereas use of the latter permits an examination of subcluster scores. Surprisingly, only one study has examined longitudinal invariance of the overall construct of PTSD (Lommen, Van De Schoot, & Engelhard, 2014) despite PTSD being mostly represented by a single total severity score in clinical and research contexts. No study to our knowledge has conducted a within-study comparison of stability of the overall PTSD versus PTSD symptom cluster constructs. Results of such a comparison have the potential to address the comparative benefit of both approaches, and inform future longitudinal studies assessing symptomatic change or the impact of risk/resilience factors. Second, the temporal stability has also been examined at a variety of time points. We would predict greater measurement invariance when PTSD is assessed only at intervals after the experience of a PTE (King et al., 2009; Wang, Elhai, Dai, & Yao, 2012), relative to pre- to post-pte assessment as the later timeframe might capture changes in belief structures and emotional processing difficulties post-pte (reviewed in Brewin & Holmes, 2003). More construct stability would also be predicted if the measurement of PTSD was indexed to the same event at each assessment point as references to different events may affect the meaning and relative salience of test items. Indeed, Lommen et al. (2014) found variability in the latent construct of PTSD among service members evaluated prior to and after deployment to a warzone as assessed by the Posttraumatic Symptom Scale-Self Report measure (Foa, Cashman, Jaycox, & Perry, 1997). In contrast, in a study of a large cohort of National Guardsmen, Meis, Erbes, Kaler, Arbisi, and Polusny (2011) found temporal stability in the number and the meaning of the constructs of PTSD symptom clusters as assessed by the PTSD Checklist (PCL; Weathers, Litz, Herman, Huska, & Keane, 1993). These disparate findings may be due to procedural variations or the nature of the PTEs experienced by the participants. Third, although prior studies have examined invariance using a variety of PTSD measures, no study has assessed the invariance of the PTSD construct assessed by the Clinician- Administered PTSD Scale (CAPS), an extensively validated PTSD assessment (Weathers, Keane, & Davidson, 2001). Fourth, few studies have attempted to identify the source(s) of any noninvariance (c.f., King et al., 2009; Lommen et al., 2014). Limited evidence indicates that the intrusion symptoms as defined by the Diagnostic and Statistical Manual of Mental Disorders (4th ed., text rev.; DSM-IV-TR; American Psychiatric Association [APA], 2000) vary in meaning (metric noninvariance; King et al., 2009), and most DSM-IV-TR-defined PTSD symptoms are scalar noninvariant (Lommen et al., 2014). In this article, we examine the measurement invariance of the PTSD construct and its symptom clusters, using a multimethod assessment using the PCL and the CAPS, with data collected at several timepoints, and explore the source of any identified noninvariance. PTSD symptom clusters were defined by the best-fitting model at baseline. DSM-IV-TR s 3-factor conceptualization of PTSD has limited empirical support (reviewed in Armour, Müllerová, & Elhai, 2016). The 4-factor emotional numbing (King, Leskin, King, & Weathers, 1998) and dysphoria models (Simms, Watson, & Doebbling, 2002) have substantial support. The former differentiates avoidance from numbing symptoms, and the latter conceptualizes a nonspecific distress-laden dysphoria factor comprising the numbing symptoms and three arousal symptoms. The relatively recent 5- factor dysphoric arousal model that retains the emotional numbing model factors and differentiates arousal into dysphoric and anxious arousal symptoms (Elhai et al., 2011) has substantial support. We used data gathered from a prospective study of a large cohort of Marines over one deployment cycle. We expected to find measurement invariance, especially across all postdeployment timepoints because PTSD symptoms were indexed to a specific Criterion A event and the modal event was war-related. We additionally investigated the source of any identified measurement noninvariance without a priori predictions. Method Participants and Procedure The Marine Resiliency Study (MRS) was a multiwave longitudinal study of active-duty ground-combat Marines deployed to Iraq or Afghanistan between 2008 and For a detailed description of the data collection procedure, see Nash et al. (2015). Informed consent was obtained before enrollment at baseline. Institutional review board approval was granted by the Boston VA (Boston, MA) and the San Diego VA (San Diego, CA). For the current study, we used data collected from participants of the fourth cohort (N = 892) with a 7-month deployment to Afghanistan. We excluded data from 4 Marines who did not deploy, 4 Marines who did not respond to all measures, and 17 Marines who were killed during deployment (n = 867). Assessments were conducted 1-month predeployment (T0) and 1- (T1), 5- (T2) and 8-months postdeployment (T3). The structure of PTSD symptoms differs in people with and without prior experience to a PTE (Elhai et al., 2009). Thus, we restricted our sample to those who endorsed response options of happened to me or witnessed it for any PTE on the Life Events Checklist (LEC; Gray, Litz, Hsu, & Lombardo, 2004) measure (n = 826); or a prior deployment experience, but no PTE on the LEC (n = 9). Participants missing >30% of T0 PCL or CAPS items were excluded (n = 1), which reduced the sample size to 834. This final all-male sample had an average age of years (SD = 3.70). A majority had a high school diploma (n = 526; 63.3%), and identified as Caucasian (n = 685; 82.9%). At T0, T1, T2, and T3, the mean PCL total severity scores were (SD = 7.84), (SD = 9.68), (SD = 11.51), and (SD = 10.44), respectively. The mean CAPS

5 Table 2 Predeployment (T0) Demographic Characteristics Longitudinal Measurement Invariance of PTSD 263 Variable N M SD % Age, years Number of childhood traumatic events CAPS severity PCL severity Education 831 Some high school High school GED Some college Associates degree year college degree Master s degree Race 826 Black or African American American Indian or Alaskan Native Asian Native Hawaiian or Pacific Islander White Enlisted Marital status 828 Married Never married Divorced Separated PTSD diagnosis (CAPS diagnostic algorithm) Note. PTSD = posttraumatic stress disorder; CAPS = Clinician-Administered PTSD Scale; PCL = PTSD Checklist; GED = General Educational Development. total severity scores at T0, T2, and T3 were (SD = 14.65), (SD = 20.32), and (SD = 18.01), respectively. See Table 2 for additional demographic information. At T0, one participant was missing one CAPS item. At T1, 2 participants were missing 1 PCL item, and 301 participants were missing all 17 items on the PCL. At T2, 1 participant was missing 1 CAPS item, 523 participants were missing all CAPS items, and 522 participants were missing all PCL items. At T3, 1 participant was missing 1 CAPS item, 561 participants were missing all PCL items, and 561 participants were missing all PCL items. The sample sizes for the complete PCL measure at T1, T2, and T3 were 531, 312, and 273, respectively; the sample sizes for the complete CAPS measure at T2 and T3 were 310 and 272, respectively. Attrition was assumed to be due to refusal to participate because of competing priorities, transfer to other military facilities, new deployments, medical discharge, or retirement. Measures The Life Events Checklist (LEC; Gray, Litz, Hsu, & Lombardo, 2004) is a 17-item self-report measure assessing exposure to a wide variety of PTEs. Each item is rated using a 5-point nominal scale (1 = happened to me, 2 = witnessed it, 3 = learned about it, 4 = not sure, and 5 = does not apply). The current study utilized the LEC to identify Marines endorsing exposure to a PTE predeployment. The Clinician-Administered PTSD Scale (CAPS; Blake et al., 1995) is a clinical structured interview assessing PTSD symptoms in reference to an indexed PTE. A rating is provided for the frequency and intensity of each symptom endorsement on a 0 4 Likert-type scale; such ratings are combined to yield a total severity score for each item (range of 0 to 8; Blake et al., 1995). The CAPS has good internal consistency (α =.82 at T0 in the current study), good test-retest reliability, and strong convergent validity (Blake et al., 1995; Weathers et al., 2001). The CAPS was not administered at T1 to reduce participation burden in the early weeks postdeployment. Participants were asked to anchor their responses to the worst and most distressing PTE. The Posttraumatic Stress Disorder Checklist (PCL; Weathers et al., 1993) is a 17-item self-report questionnaire that assesses PTSD symptoms on a 5-point Likert scale from 1 (not at all bothersome) to5(extremely bothersome). The PCL has high internal consistency (α =.90 at T0 in the current study), good test-retest reliability, and strong convergent

6 264 Contractor et al. validity (Blanchard, Jones-Alexander, Buckley, & Forneris, 1996; Weathers et al., 1993). Participants were asked to anchor their ratings to any lifetime PTE. The PCL was administered at all timepoints. Data Analysis We used maximum likelihood estimation with robust standard errors (MLR) in Mplus version 7 to estimate missing data, to estimate model parameters, and to address nonnormality (Muthen & Muthen, ). We identified the best-fitting PTSD factor-analytical model at baseline. We computed χ 2 values and used a p value criterion of <.010 to determine significant model fit. The MLR χ 2 statistic is equivalent to the Yuan-Bentler test statistic (Yuan & Bentler, 2000). As the χ 2 statistic is sensitive to sample size, we used additional fit indices in assessing model fit. A well-fitting (adequate) model would have a comparative fit index (CFI) and Tucker-Lewis index values.95 (.90.94), root mean square error of approximation value.06 (.07.08), and standardized root mean square residual value <.08 (.09.10; Hu & Bentler, 1999). When comparing the relative fit of the nested models, a significant χ 2 value difference indicated a better fit for the complex model. To compare the relative fit of nonnested models, a 10-point Bayesian information criterion value difference represented a 150:1 likelihood and very strong support that the model with the smaller Bayesian information criterion value fit best; a 6- to 9-point difference indicated strong support (Kass & Raftery, 1995). We examined longitudinal invariance of the constructs of PTSD and PTSD symptom clusters conceptualized according to the best-fitting model. In testing the longitudinal invariance of the construct of PTSD, items were averaged for each symptom cluster (parcels) and the obtained parcels were used as indicators for the invariance models. In examining the longitudinal invariance of each PTSD symptom cluster, each latent factor was defined by the corresponding items as identified by the best-fitting factor-analytical model. We began with the least restrictive model (Model A). Subsequent models were specified by progressively constraining additional parameters. In Model A, temporal groups varied on all parameter estimates including factor loadings, item/parcel intercepts, residual error variances, factor means, and factor variances and covariances (configural invariance) to determine if the factors were associated with the same items/parcels. For Model B, we constrained factor loadings for each PTSD item/parcel across the temporal groups to test the equivalence of factor loadings while constraining the other parameter estimates (metric invariance). In Model C, we added constrained item/parcel intercepts in addition to the constrained factor loadings across groups (scalar or strong factorial invariance) to test item/parcel-severity equivalence (Gregorich, 2006; Widaman, Ferrer, & Conger, 2010). We calculated the χ 2 value for each model and compared the values between successive models to evaluate between-model statistical significance. A nonsignificant χ 2 value difference indicates invariance (Gregorich, 2006). As the χ 2 difference test tends to reach significance in larger samples with small between-model discrepancies, we also used the CFI values to test invariance (Cheung & Rensvold, 2002). Thus, we established noninvariance of statistical parameters with a significant χ 2 difference test value (p <.005; Little, 2013), and a CFI value difference.01. Lastly, we determined the source of any identified noninvariance by testing partial invariance. We identified specific factor loadings or item/parcel intercepts that differed across the temporal groups by examining high-magnitude model modification indices (MI) greater than 10, and substantial standardized expected parameter change (EPC) values. Based on theory and empirical evidence (MI and EPC values), we sequentially freed one parameter at a time until we obtained partial invariance. Partial invariance was established based on the guidelines for invariance testing, and the determination of acceptable/good model fit as indicated above (Little, 2013; Van De Schoot, Lugtig, & Hox, 2012). Results The Confirmatory factor analysis (CFA) results indicated a relatively best-fitting emotional numbing model for the PCL and CAPS items. Results of model comparisons are available upon request. The emotional numbing model was considered as the baseline model for invariance analyses. Using both recommended benchmarks for invariance testing, we found invariance of factor loadings (metric) and nonequivalence of intercepts (scalar) for the PCL- and CAPS- assessed PTSD construct (see Tables 3 and 4). In our examination of the noninvariant parcels on the PCL, we relaxed the factor loading and intercept constraint for the T0 numbing parcel, which was associated with the largest MI and EPC values (Model Ca; MI = 27.90; standardized EPC =.03); and subsequently relaxed the factor loading and intercept constraint for the T0 avoidance parcel (Model Cb; MI = 27.73; standardized EPC =.07). For the CAPS-assessed PTSD construct, we relaxed the factor loading and intercept constraint for the T0 avoidance parcel (Model Cc; MI = 21.5; standardized EPC =.06), and subsequently relaxed the factor loading and intercept constraint for the T0 numbing parcel (Model Cd; MI = 26.88; standardized EPC =.08). Thus, the numbing and avoidance subscale scores instability from pre- to postdeployment accounted for the variability in the overarching PTSD construct. There was slightly more variability across the PCL and the CAPS when assessing the invariance of PTSD symptom clusters compared to the assessment of the overarching construct of PTSD (see Tables 5 and 6). We found noninvariance of factor loadings (metric) and of intercepts (scalar) on the PCL, and invariance of the factor loadings (metric) and nonequivalence of intercepts (scalar) on the CAPS.

7 Longitudinal Measurement Invariance of PTSD 265 Table 3 Fit Statistics of Each Model to Test Invariance of the Posttraumatic Stress Disorder (PTSD) Construct a Model χ 2 df RMSEA 90% CI CFI TLI SRMR BIC PCL Configural invariance **.04 [.03,.05] Metric invariance **.04 [.03,.05] Scalar invariance **.05 [.05,.06] Partial scalar invariance (Ca) **.05 [.04,.06] Partial scalar invariance (Cb) **.04 [.04,.05] CAPS Configural invariance **.05 [.04,.06] Metric invariance **.05 [.04,.06] Scalar invariance **.06 [.05,.06] Partial scalar invariance (Cc) **.05 [.04,.06] Partial scalar invariance (Cd) **.05 [.04,.05] Note. Model Ca removed the intercept and factor loading constraint on numbing at T0; Model Cb removed the intercept and factor loading constraint on numbing and avoidance (T0); Model Cc removed intercept and factor loading constraint on avoidance (T0); Model Cd removed intercept and factor loading constraints on avoidance and numbing (T0). df = degrees of freedom; RMSEA = root mean square error of approximation; CI = confidence interval; CFI = comparative fit index; TLI = Tucker-Lewis index; SRMR = standardized root mean square residual; BIC = Bayesian information criterion; PCL = PTSD Checklist; CAPS = Clinician-Administered PTSD Scale. a Parcels according to emotional numbing model conceptualization. *p <.005. **p <.001. Table 4 Fit Statistics of Model Comparisons to Test Invariance of the Posttraumatic Stress Disorder (PTSD) Construct a Model comparisons χ 2difference df CFI value difference PCL Configural vs. metric invariance Metric vs. scalar invariance ** Metric vs. partial scalar invariance (Ca) ** Metric vs. partial scalar invariance (Cb) ** CAPS Configural vs. metric invariance Metric vs. scalar invariance ** Metric vs. partial scalar invariance (Cc) ** Metric vs. partial scalar invariance (Cd) ** Note. Model Ca removed the intercept and factor loading constraint on numbing at T0; Model Cb removed the intercept and factor loading constraint on numbing and avoidance (T0); Model Cc removed intercept and factor loading constraint on avoidance (T0); Model Cd removed intercept and factor loading constraints on avoidance and numbing (T0). df = degrees of freedom; CFI = comparative fit index; PCL = PTSD Checklist; CAPS = CAPS = Clinician-Administered PTSD Scale. a Parcels according to emotional numbing model conceptualization. *p <.005. **p <.001. For the PCL-assessed symptom clusters, we relaxed the factor-loading constraint on the T0 item assessing startle response (Model Ba; MI = 32.70; standardized EPC =.07), and additionally relaxed the factor-loading constraint on the TI item assessing amnesia (Model Bb; MI = 36.41; standardized EPC =.18). These items had the largest MI and EPC values. To examine partial scalar invariance, we additionally relaxed the factor-loading and intercept constraint for the T0 item assessing foreshortened future (Model Ce; MI = 22.57; standardized EPC =.08). For the CAPS-assessed PTSD symptom clusters, we relaxed the intercept and factor-loading constraint on the T0 item assessing startle response (Model Cf; MI = 58.11; standardized EPC =.12), and subsequently relaxed the intercept and factor-loading constraint on the T0 item referencing foreshortened future (Model Cg; MI = 28.17; standardized EPC =.21). These items had the largest MI and EPC values. Thus, the PCL and CAPS items assessing startle response and foreshortened future pre- to postdeployment, and the PCL-assessed item assessing amnesia for parts of the traumatic event immediately

8 266 Contractor et al. Table 5 Fit Statistics of Each Model to Test Invariance of Posttraumatic Stress Disorder (PTSD) Symptom Clusters (Emotional Numbing Model) Model χ 2 (df) df RMSEA 90% CI CFI TLI SRMR BIC PCL emotional numbing model factors Configural invariance **.03 [.03,.03] Metric invariance **.03 [.03,.04] Partial metric invariance (Ba) **.03 [.03,.04] Partial metric invariance (Bb) **.03 [.03,.03] Scalar invariance **.03 [.03,.04] Partial scalar invariance (Ce) **.03 [.03,.04] CAPS emotional numbing model factors Configural invariance **.02 [.02,.02] Metric invariance **.02 [.02,.02] Scalar invariance **.02 [.02,.03] Partial scalar invariance (Cf) **.02 [.02,.03] Partial scalar invariance (Cg) **.02 [.02,.02] Note. Model Ba removed constraints on factor loading for item 17 (T0); Model Bb removed constraints on factor loading for items 17 (T0) and 8 (T1); Model Ce removed constraints on factor loading and intercept for items 17 (T0), 8 (T1), and 12 (T0); Model Cf removed constraints on factor loading and intercept for item 17 (T0); Model Cg removed constraints on factor loading and intercept for items 17 and 12 (T0). df = degrees of freedom; RMSEA = root mean square error of approximation; CFI = comparative fit index; TLI = Tucker-Lewis index; SRMR = standardized root mean square residual; BIC = Bayesian information criterion; PCL = PTSD Checklist; CAPS = Clinician-Administered PTSD Scale. *p <.005. **p <.001. Table 6 Fit Statistics of Model Comparisons to Test Invariance of PTSD Symptom Clusters (Emotional Numbing Model) Model comparisons χ 2difference df CFI value difference PCL numbing model Configural vs. metric invariance ** Configural vs. partial metric invariance (Ba) ** Configural vs. partial metric invariance (Bb) * Partial metric (Bb) vs. scalar invariance ** Partial metric (Bb) vs. partial scalar invariance (Ce) ** CAPS numbing model Configural vs. metric invariance Metric vs. scalar invariance ** Metric vs. partial scalar invariance (Cf) ** Metric vs. partial scalar invariance (Cg) Note. Model Ba removed constraints on factor loading for item 17 (T0); Model Bb removed constraints on factor loading for items 17 (T0) and 8 (T1); Model Ce removed constraints on factor loading and intercept for items 17 (T0), 8 (T1), and 12 (T0); Model Cf removed constraints on factor loading and intercept for item 17 (T0); Model Cg removed constraints on factor loading and intercept for items 17 and 12 (T0). df = degrees of freedom; CFI = comparative fit index; PCL = PTSD Checklist; CAPS = Clinician-Administered PTSD Scale. *p <.005. **p <.001. postdeployment, accounted for variability in the measurement of the PTSD symptom clusters. Discussion We examined the measurement invariance of the construct of PTSD in a large sample of Marines deployed to war and studied prospectively. To the best of our knowledge, this is the first multimethod examination of measurement invariance of PTSD and its symptom clusters. Our analyses demonstrated partial measurement invariance in the latent constructs of PTSD and its four symptom clusters of reexperiencing, avoidance, numbing, and arousal. For the most part, we found longitudinal consistency in the pattern of item mappings onto latent factors and the strength of factor loadings for the constructs of PTSD

9 Longitudinal Measurement Invariance of PTSD 267 and its symptom clusters. Thus, we can infer consistency in meaning of these latent factors across time in our sample (Chen, 2008; Gregorich, 2006). In contrast, we observed that metric invariance was not supported for the latent symptom clusters assessed by the PCL. One possibility is that the structured clinical interview measure (CAPS) captures the PTSD construct in a more stable manner longitudinally specifically because, as intended, trained interviewers provide consistent definitions of symptoms to interviewees (Weathers et al., 2001). The fact that we found scalar noninvariance for the latent constructs of PTSD and its symptom clusters is consistent with prior research (e.g., Krause, Kaltman, Goodman, & Dutton, 2007; Wang et al., 2012). This suggests that our findings may not be unique to our sample, this population, or the measures employed. For the construct of PTSD, variability in factorloading and intercept values of the numbing and avoidance symptom clusters pre- to postdeployment contributed to measurement instability. For the symptom clusters, inconsistency in the factor-loading and intercept values of items assessing startle responses, amnesia for traumatic events, and foreshortened future mainly pre- to postdeployment added to the measurement instability. The latter two items are components of the numbing symptom cluster and each have very low base rates among service members, which could have contributed to instability. Thus, changes in these specific item-level and subscale-level scores are only partially associated with changes in the construct of PTSD. Item or subscale-level variability may also be due to unspecified response biases, a desire of respondents to portray themselves differently at different times, a tendency to respond more strongly to items at one time point than at a different time point irrespective of PTSD severity, or difficulty in the reliable measurement of these low base-rate aspects of the PTSD construct. It may be that the uncertainty of preparing to deploy to combat at the predeployment interval would cause a number of participants to respond more strongly to these items. Nevertheless, we argue that this does not invalidate the measurement of the construct of PTSD as the majority of the indicators were consistent across time. Further, the difference between pre- and postdeployment factor loadings for the item on traumatic amnesia is consistent with the prior findings (Baschnagel, O Connor, Colder, & Hawk, 2005; Palmieri, Weathers, Difede, & King, 2007). This suggests that the amnesia item may be problematic for the reasons stated above, chiefly its low base rate. It is possible that removal of these items could result in a more valid measure of the construct of PTSD and enhance the likelihood of capturing responses truly influenced by one s experiences of PTSD rather than measurement artifacts or other time- and context-varying influences. Most of the measurement variability in this sample occurred between the pre- and the immediate postdeployment time points. Lommen et al. (2014) concluded that in their sample the construct of baseline symptoms prior to the experience of a PTE is different than the construct of PTSD assessed post-pte. Our study results do not completely align with these findings, which may reflect sample-specific differences, including the fact that our sample was composed solely of individuals who reported exposure to a PTE prior to deployment. It is possible that an additional experience of a PTE may alter how the person responds to items assessing startle responses, and items comprising the numbing and avoidance symptom clusters. We thus conclude that in this sample, the overall consistency in measurement indicates that the same constructs were assessed accurately and reliably pre- and postdeployment in this sample. The results were also, for the most part, replicated across two different measures of PTSD and we found invariance in the overall PTSD construct and in subclusters (factors). This means that it is valid to interpret the differences in the means in the measures of PTSD employed in this study and that the obtained consistent pattern of partial invariance is reasonable to interpret (Meredith & Teresi, 2006; Sass, 2011). These results are sample- and study-specific. Despite the high level of combat experienced by this cohort, this was an epidemiological study of highly healthy individuals who reported relatively low PTSD severity over time. The Marines lost to follow-up had greater T1 functional impairment and endorsed more events on the LEC (Nash et al., 2015). The impact of missing data and the correlates of missing data on invariance is uncertain and is a question for future research. Further, the PTSD assessments were indexed to the most currently distressing event in the Marine s lifetime. Of participating Marines, 53% had served in deployments prior to this deployment; 99% reported experiencing or witnessing at least one PTE on the LEC. Respondents may have anchored their responses to different events at each time and some Marines may have anchored their responses to PTEs experienced in the past. It is beyond the scope of this study to examine if such variability impacted construct invariance. However, despite variation in the timing and type of index events, we found support for partial invariance for the construct of PTSD. Finally, it is unclear how our results affect DSM-5- (APA, 2014) based PTSD assessments. Changes in PTSD s diagnostic criteria from DSM-IV-TR to DSM-5 include the separation of the avoidance and numbing symptoms, conceptualization of the PTSD construct as encompassing four symptom clusters, conceptualization of the numbing symptoms as those encompassing the negative alterations in cognitions and mood cluster, and the addition of several new items (Friedman, 2013). Despite these changes, there is considerable similarity between the two latest DSM versions, and the numbing model used in the current study is similar to the DSM-5 conceptualization of PTSD. In addition, the findings regarding the intrusion and avoidance clusters, and certain specific items such as traumatic amnesia are applicable to DSM-5 PTSD criteria. In summary, we demonstrated procedurally that, for the large part, the constructs of PTSD and its symptom clusters, as assessed by two widely used measures, have longitudinal stability in the assessment of deployed service members over time. However, responses to items assessing avoidance and numbing symptoms (specifically foreshortened future, and amnesia) may be influenced by other factors rather than PTSD severity. This

10 268 Contractor et al. small noninvariance suggests that in this kind of study context, these items may not be specific to PTSD (Simms et al., 2002). It needs to be emphasized that measurement invariance is a study-specific and measure-specific phenomenon. In this case, our results are generalizable mainly in terms of procedures as well as the broader implications of the statistical technique for the interpretation of results. References American Psychiatric Association. (2000). Diagnostic and statistical manual of mental disorders (4th ed., text rev.). Washington, DC: Author. American Psychiatric Association. (2013). Diagnostic and statistical manual of mental disorders (5th ed.). Arlington, VA: Author. Armour, C., Müllerová, J. & Elhai, J. D. (2016). A systematic literature review of PTSD s latent structure in the Diagnostic and Statistical Manual of Mental Disorders: DSM-IV to DSM-5. Clinical Psychology Review, 44, doi: /j.cpr Baker, D. G., Nash, W. P., Litz, B. T., Geyer, M. A., Risbrough, V. B., Nievergelt, C. M.,... the M.R.S. Team (2012). Predictors of risk and resilience for posttraumatic stress disorder among ground combat marines: Methods of the marine resiliency study.preventing Chronic Disease, 9,E97. doi: /pcd Baschnagel, J. S., O Connor, R. M., Colder, C. R. & Hawk, L. W. (2005). Factor structure of posttraumatic stress among Western New York undergraduates following the September 11th terrorist attack on the World Trade Center. Journal of Traumatic Stress, 18, doi: /jts Blake, D. D., Weathers, F. W., Nagy, L. M., Kaloupek, D. G., Gusman, F. D., Charney, D. S. & Keane, T. M. (1995). The development of a clinician-administered PTSD scale. Journal of Traumatic Stress, 8, doi: /jts Blanchard, E. B., Jones-Alexander, J., Buckley, T. C. & Forneris, C. A. (1996). Psychometric properties of the PTSD Checklist (PCL). Behaviour Research and Therapy, 34, doi: / (96) Borsboom, D. (2006). When does measurement invariance matter? Medical Care, 44, doi: /01.mlr cc Brewin, C. R. & Holmes, E. A. (2003). Psychological theories of posttraumatic stress disorder. Clinical Psychology Review, 23, doi: /s (03) Chen, F. F. (2008). What happens if we compare chopsticks with forks? The impact of making inappropriate comparisons in cross-cultural research. Journal of Personality and Social Psychology, 95, doi: /a Cheung, G. W. & Rensvold, R. B. (2002). Evaluating goodness-of-fit indexes for testing measurement invariance. Structural Equation Modeling, 9, doi: /s sem0902_5 Elhai, J. D., Biehn, T. L., Armour, C. A., Klopper, J. J., Frueh, B. C. & Palmieri, P. A. (2011). Evidence for a unique PTSD construct represented by PTSD s D1 D3 symptoms. Journal of Anxiety Disorders, 25, doi: /j.janxdis Elhai, J. D., Engdahl, R. M., Palmieri, P. A., Naifeh, J. A., Schweinle, A. & Jacobs, G. A. (2009). Assessing posttraumatic stress disorder with or without reference to a single, worst traumatic event: Examining differences in factor structure. Psychological Assessment, 21, doi: /a Foa, E. B., Cashman, L., Jaycox, L. & Perry, K. (1997). The validation of a self-report measure of posttraumatic stress disorder: The Posttraumatic Diagnostic Scale. Psychological Assessment, 9, doi: / Friedman, M. J. (2013). Finalizing PTSD in DSM-5: Getting here from there and where to go next. Journal of Traumatic Stress, 26, doi: /jts Gray, M., Litz, B. T., Hsu, J. & Lombardo, T. (2004). Psychometric properties of the Life Events Checklist. Assessment, 11, doi: / Gregorich, S. E. (2006). Do self-report instruments allow meaningful comparison across diverse population groups? Testing measurement invariance using the confirmatory factor analysis framework. Medical Care, 44, doi: /01.mlr f Hu, L. T. & Bentler, P. M. (1999). Cutoff criteria for fit indexes in covariance structure analysis: Conventional criteria versus new alternatives. Structural Equation Modeling: A Multidisciplinary Journal, 6, doi: / Kass, R. E. & Raftery, A. E. (1995). Bayes factors. Journal of the American Statistical Association, 90, doi: / King, D. W., Leskin, G. A., King, L. A. & Weathers, F. W. (1998). Confirmatory factor analysis of the clinician-administered PTSD Scale: Evidence for the dimensionality of posttraumatic stress disorder. Psychological Assessment, 10, doi: / King, D. W., Orazem, R. J., King, L. A., Lauterbach, D., Hebenstreit, C. L. & Shalev, A. Y. (2009). Factor structure of posttraumatic stress disorder as measured by the Impact of Event Scale Revised: Stability across cultures and time. Psychological Trauma: Theory, Research, Practice, and Policy, 1, doi: /a Krause, E. D., Kaltman, S., Goodman, L. A. & Dutton, M. A. (2007). Longitudinal factor structure of posttraumatic stress symptoms related to intimate partner violence. Psychological Assessment, 19, doi: / Little, T. D. (2013). Longitudinal CFA model. In: T. D. Little & N. A. Card (Eds.), Longitudinal structural equation modeling (pp ). New York, NY: Guilford Press. Little, T. D. & Slegers, D. W. (2005). Factor analysis: Multiple groups. In: B. S. Everitt & D. C. Howell (Eds.), Encyclopedia of statistics in behavioral science (pp ). Chichester, England: Wiley. Lommen, M. J. J., Van De Schoot, R. & Engelhard, I. M. (2014). The experience of traumatic events disrupts the measurement invariance of a posttraumatic stress scale. Frontiers in Psychology, 5, doi: /fpsyg Meis, L. A., Erbes, C. R., Kaler, M. E., Arbisi, P. A. & Polusny, M. A. (2011). The structure of PTSD among two cohorts of returning soldiers: Before, during, and following deployment to Iraq. Journal of Abnormal Psychology, 120, doi: /a Meredith, W. & Teresi, J. A. (2006). An essay on measurement and factorial invariance. Medical Care, 44, doi: / 01.mlr Muthen, L. K. & Muthen, B. O. ( ). Mplus user s guide. (6thed.). Los Angeles, CA: Author. Nash, W. P., Boasso, A. M., Steenkamp, M. M., Larson, J. L., Lubin, R. E. & Litz, B. T. (2015). Posttraumatic stress in deployed Marines: Prospective trajectories of early adaptation. Journal of Abnormal Psychology, 124, doi: /abn Palmieri, P. A., Weathers, F. W., Difede, J. & King, D. W. (2007). Confirmatory factor analysis of the PTSD Checklist and the Clinician-Administered PTSD Scale in disaster workers exposed to the World Trade Center Ground Zero. Journal of Abnormal Psychology, 116, doi: / x

11 Longitudinal Measurement Invariance of PTSD 269 Sass, D. (2011). Testing measurement invariance and comparing latent factor means within a confirmatory factor analysis framework. Journal of Psychoeducational Assessment, 29, doi: / Simms, L. J., Watson, D. & Doebbling, B. N. (2002). Confirmatory factor analyses of posttraumatic stress symptoms in deployed and non-deployed veterans of the Gulf War. Journal of Abnormal Psychology, 111, doi: / x Van De Schoot, R., Lugtig, P. & Hox, J. (2012). A checklist for testing measurement invariance. European Journal of Developmental Psychology, 9, doi: / Vandenberg, R. J. & Lance, C. E. (2000). A review and synthesis of the measurement invariance literature: Suggestions, practices, and recommendations for organizational research. Organizational Research Methods, 3, doi: / Wang, M., Elhai, J. D., Dai, X. & Yao, S. (2012). Longitudinal invariance of posttraumatic stress disorder symptoms in adolescent earthquake survivors. Journal of Anxiety Disorders, 26, doi: /j.janxdis Weathers, F. W., Keane, T. M. & Davidson, J. R. (2001). Clinician- Administered PTSD Scale: A review of the first ten years of research. Depression and Anxiety, 13, doi: /da.1029 Weathers, F. W., Litz, B. T., Herman, D., Huska, J. & Keane, T. M. (1993, October). The PTSD Checklist (PCL): Reliability, validity, and diagnostic utility. Paper presented at the Annual Convention of the International Society for Traumatic Stress Studies, San Antonio, TX. Widaman, K. F., Ferrer, E. & Conger, R. D. (2010). Factorial invariance within longitudinal structural equation models: measuring the same construct across time. Child Development Perspectives, 4,10 18.doi: /j x Yuan, K.-H. & Bentler, P. M. (2000). Three likelihood-based methods for mean and covariance structure analysis with nonnormal missing data. Sociological Methodology, 30, doi: /