Longitudinal Patterns of PTSD Symptoms Among Sexual Assault Survivors: A Latent Transition Analysis

Abstract

Objective:

The aim of this study was to identify latent classes of PTSD symptoms in a community sample of sexual assault survivors whose assaults occurred varying lengths of time in the past, and to explore patterns of transition between those latent classes over time.

Method:

Latent class analysis was used to identify naturally occurring subgroups of PTSD symptoms in a sample of sexual assault survivors who completed two mailed surveys one year apart (N = 1,271). Latent transition analysis was then used to examine individuals’ probabilities of transitioning into each latent class at Time 2 based on their latent class membership at Time 1.

Results:

A four-class model emerged as the best fitting model at both Time 1 and Time 2. Classes demonstrated overall severity and symptom cluster severity differences. Transition into a lower severity class was more common than transition into a higher severity class, though escalation was demonstrated by 6 – 20% of participants in each latent class.

Conclusions:

The substantial heterogeneity in sexual assault survivors’ PTSD symptoms highlights the variety of ways that posttraumatic stress may be experienced years after a sexual assault. Future research should explore factors that affect long-term symptoms, including cumulative lifetime trauma and social support.

Traumatic experiences are common, with over 80% of Americans reporting one or more traumatic events in their lives (Kessler, 2018). Most individuals who experience trauma report at least some psychological or physiological effects (Langton & Truman, 2014) and approximately 30% report levels of posttraumatic stress such that they can be diagnosed with posttraumatic stress disorder (PTSD) one month after the trauma occurred (Santiago et al., 2013). There is substantial variation within that figure, however, with research indicating that cumulative lifetime trauma (Solomon et al., 2012), comorbid physical and mental health issues (Dworkin et al., 2017), and experiencing a trauma that was intentionally inflicted (Santiago et al., 2013) are associated with a higher risk of developing PTSD. Crime victims are a specific subpopulation that are at high risk of developing PTSD, particularly if they experienced a violent attack, knew the person who attacked them, or blame themselves for the incident (Wohlfarth et al., 2002).

Given these risk factors, it is unsurprising that sexual assault survivors report especially high rates of post-trauma distress and PTSD (Basile & Smith, 2011; Dworkin et al., 2017). Sexual assault is a disturbingly common occurrence, with 17 – 25% of women and 1 – 3% of men assaulted in their lifetimes (Basile & Smith, 2011). Approximately three-quarters of sexual assault survivors experience moderate to severe distress after their assaults (Langton & Truman, 2014), and a recent meta-analysis found a lifetime PTSD rate of 36% among sexual assault survivors compared to 9% among those who had not experienced sexual assault (Dworkin, 2020). In an analysis of risk for PTSD based on trauma type, Kessler (2018) reported that the conditional risk of PTSD after rape was 25%, the second highest of 29 trauma types assessed. When considering both the relative risk of PTSD and the frequency with which specific traumas occur, Kessler estimated that rape and other sexual assaults account for nearly 30% of PTSD episodes in the United States. Longitudinal studies suggest that posttraumatic stress decreases over time for many sexual assault survivors (Dworkin et al., 2021) but some experience persistent distress that negatively impacts their quality of life for years to come (Pacella et al., 2013; Suris et al., 2007).

Recent years have seen an increased focus on understanding variation in posttraumatic stress experiences and trajectories. Particularly with the introduction of new symptoms and a fourth symptom cluster in the DSM-5 (APA, 2013), 636,120 unique combinations of PTSD symptoms are now possible (Galatzer-Levy & Bryant, 2013). Statistical approaches that can model this heterogeneity in clinically useful ways have been identified as an important avenue of research (Galatzer-Levy & Bryant, 2013; Pietrzak et al., 2014). One such technique is mixture modeling, in which participants’ responses to indicator variables are used to identify latent subgroups—often called latent classes—in which group members are more similar to each other than to the sample as a whole (Magnusson, 1998; Muthén & Muthén, 2000). In the context of PTSD research, these classes can communicate differences in overall severity or symptom severity, highlighting previously unidentified groupings of trauma survivors with unique experiences of posttraumatic stress.

Multiple mixture modeling studies have found that three latent classes, which vary according to overall severity, best represent trauma survivors’ posttraumatic stress symptoms (Ayer et al., 2012; Breslau et al., 2005; Contractor et al., 2015; Steenkamp, Nickerson, et al., 2012). These three classes have been described as: (1) no/low/mild disturbance, (2) intermediate/moderate disturbance, and (3) pervasive/severe disturbance. The plurality of participants in these studies (43 – 62%) were most likely to belong to the lowest severity group, with moderate severity as the next largest group (26 – 43%), and the smallest proportion of participants in the highest severity group (9 – 14%). Other studies have identified four (Au et al., 2013) or five (Hebenstreit et al., 2015) class solutions in which the latent classes vary by symptom cluster severity, in addition to overall severity. Minihan et al. (2018), for example, reported a four-class model in which one class showed particularly high re-experiencing and avoidance symptoms, and Hebenstreit et al. (2015) identified a five-class model in which two classes showed greater endorsement of hypervigilance symptoms than would have been expected based on the severity of the other symptoms. Still other studies have highlighted emotional numbing as a set of symptoms that differentiate otherwise similar latent classes (Hebenstreit, Madden, & Maguen, 2014; Maguen et al., 2013).

Longitudinal extensions of mixture models can also be employed to understand better how latent subgroups change over time, exploring the probability of symptom escalation, de-escalation, or other types of symptom change. One such method, growth mixture modeling (GMM), evaluates the possibility of multiple growth trajectories indicating different directions or rates of change. In their review of fifty-four studies examining trajectories of posttraumatic stress symptoms, Galatzer-Levy et al. (2018) found that the most common post-trauma response was what they termed “resilience” (n = 63 identified trajectories), in which relatively little psychological disruption was reported at each time period, followed by trajectories representing recovery (n = 49), chronicity (n = 47), and delayed onset of posttraumatic stress symptoms (n = 22). Another longitudinal mixture modeling method, latent transition analysis (LTA), models change in latent class membership, communicating the likelihood that the latent class best representing an individual’s symptoms changes over time. In their study of U.S. National Guard Members six and twelve months after return from deployment, Bohnert and colleagues (2018) found that when participants were assigned to a different latent class at twelve months than they were at six months, the transition generally signified a decrease in symptom severity. This pattern of symptom de-escalation has been reported in other LTAs (Chen & Wu, 2017; Boasso et al., 2016) but the trend is far from universal; as many as 20% of National Guard members in Bohnert’s latent classes and as many as 13% of natural disaster survivors in Chen and Wu’s (2017) latent classes transitioned into a higher severity class over time. That some individuals do show an increase in symptom severity highlights the importance of identifying and attending to these distinct groups.

To our knowledge, only two published studies have analyzed sexual assault survivors’ experiences of posttraumatic stress using longitudinal mixture modeling techniques¹. Using growth mixture modeling, Armour et al. (2012) found two latent trajectories: Low Stable (35.0%), in which PTSD symptoms were stable and comparatively low, and High Decline (65.0%), which reflected higher symptoms that significantly decreased over the course of the study. Steenkamp, Dickstein, et al. (2012) found four trajectories of PTSD symptoms using latent class growth analysis: High Chronic (6.7%), Moderate Chronic (16.0%), Moderate Recovery (47.9%), and Marked Recovery (29.4%). Both studies mark major contributions to understanding PTSD among sexual assault survivors but because their analyses relied on mean severity across PTSD symptoms, uncertainty remains regarding whether latent subgroups can be differentiated by specific symptoms or symptom clusters in addition to overall severity and rate of change. In other words, we still do not know whether sexual assault survivors can be meaningfully differentiated from one another based on the severity of specific symptoms or symptom clusters over time. Particularly in light of Hebenstreit’s (2015) findings that hypervigilance symptoms were prominent in some subgroups of intimate partner violence survivors, mixture modeling at the symptom level could indicate whether specific symptom clusters are similarly prominent in subgroups of sexual assault survivors.

There are important areas for expansion within the wider sexual assault and PTSD literature, as well. Dworkin’s (2021) meta-analysis found that 40% of U.S. studies on PTSD after sexual assault were based on samples at least two-thirds white, highlighting a need for PTSD research to be carried out with racially diverse samples. Additionally, much of the literature on sexual assault survivors’ experiences of PTSD is based on samples in the immediate aftermath of an assault (Armour et al., 2012; Dickstein et al., 2012; Dworkin et al., 2021). Sexual assault commonly affects survivors for months or years afterward (Pacella et al., 2013; Suris et al., 2007) and there is much left to learn about survivors’ posttraumatic stress symptom profiles years after an assault has taken place. Finally, most sexual assault survivors do not seek services from hospitals, rape crisis centers, or other formal helpseeking providers (Sinozich & Langton, 2014), yet those locations make up a majority of recruitment sites for studies focused on posttraumatic stress in the aftermath of sexual assault (Dworkin, 2021). There is evidence that sexual assault survivors who utilize formal helpseeking services have higher rates of PTSD than survivors who do not access formal services (Lewis et al., 2005), and recruitment from those sites may therefore oversample survivors with higher symptomology.

With these gaps in mind, the current study was undertaken with three primary aims. First, we sought to increase what is known about posttraumatic stress as it is experienced by women and racially diverse populations. Second, we aimed to expand the research literature to include a longitudinal analysis of PTSD symptoms among sexual assault survivors whose assaults occurred varying lengths of time in the past and who were recruited primarily from community, rather than clinical, settings. Finally, we set out to understand heterogeneity in PTSD symptom presentations among sexual assault survivors. In pursuit of these aims, we carried out a latent transition analysis (LTA) based on PTSD symptoms reported by a racially diverse sample of sexual assault survivors recruited from community settings varied lengths of time after their assaults had taken place. Two research questions guided our study: (1) What latent classes best represent PTSD symptom profiles in a community sample of sexual assault survivors, and (2) What are the probabilities associated with transitioning between those latent classes over time? This study was exploratory in nature and no specific hypotheses were made.

Method

Participants and Procedures

Participants were recruited with online and printed advertisements distributed in various college, community, and agency settings. The advertisements described a study open to adult women who had experienced an unwanted sexual event when they were 14 years or older and had told at least one person about the experience². Women who called study organizers in response to the recruitment flyer were mailed a hard copy of the survey. Participants who returned the survey were paid US$25 for their participation, and only participants who indicated on the Time 1 survey that they wished to receive a follow-up survey in one year were sent the second survey. All data were collected with IRB oversight and approval through the University of Illinois at Chicago.

If participants indicated having experienced more than one unwanted sexual event since age 14, they were asked to think about what they considered the most serious experience when answering questions. Based on these instructions, 86.0% of participants reported unwanted oral (8.0%), anal (2.7%), or vaginal (75.1%) sex, with the remaining participants (14.0%) reporting unwanted fondling or kissing. The referenced assault occurred an average of nearly fifteen years before the Time 1 survey (M = 14.91 years, SD = 12.21) with 13.3% of participants referencing an assault that occurred within the previous two years.

Demographic data were collected in the Time 1 survey. Comparison to 2010 U.S. census data indicated that racial identifications in the sample were relatively representative of the city in which data were collected, though the proportion of individuals who identified as Black or African American was greater in the study sample (43.5% vs. 32.9%) and the proportion of individuals who identified as white was smaller in the study sample compared to census data (37.2% vs. 45.0%). There were fewer participants in the study who identified as Hispanic or Latina than would be expected based on census data (13.1% vs. 28.9%). Additional demographic data are presented in Supplemental Table 1.

The full study included three mailed surveys with one year between each instance of data collection; only the first two surveys are utilized in this analysis. N = 1,863 women completed the first survey, but n = 182 were excluded due to not meeting the study criteria of having told at least one person about the unwanted sexual experience. This reduced the final Time 1 sample to n = 1,681 participants, 75.6% of whom (n = 1,271) completed the second survey. There was no evidence of differential attrition based on race, ethnicity, sexual identity, or education, but participants who completed both surveys were significantly older (M = 37.10, SD = 0.36) than those who attritted (M = 34.36, SD = 0.60) (t(1646) = −3.83, p < .001).

Measures

PTSD symptoms were measured with the 17-item standardized Posttraumatic Stress Diagnostic Scale (PDS; Foa et al., 1997). Psychometric testing has found high reliability and validity for the PDS. In a sample of individuals who had experienced diverse traumatic stressors, the PDS demonstrated good test-retest reliability when administered over a 2 – 3-week period (κ = .83) and excellent internal consistency (α = .92) (Foa et al., 1997). The PDS showed excellent internal consistency in the current sample, as well (α = .93). The PDS was designed based on the symptoms and symptom cluster framework of the DSM-IV (APA, 1994). Recently, an updated version of the PDS (PDS-5, Foa et al. 2016) was developed to be consistent with the DSM-5. In this manuscript, we discuss symptoms and symptom clusters through the framework laid out in the DSM-5 to maximize the applicability of our findings to current understandings of PTSD.

The PDS asks respondents to report whether they have experienced each symptom: (0) never or only one time, (1) once in a while, (2) half the time, or (3) almost always over the last 12 months. At Time 1, overall mean scores on the PDS were 21.06 (SD = 12.92). Overall scores were somewhat lower at Time 2 (M = 16.77, SD = 12.03). Most of the items ranged from fairly symmetrical to moderately positively skewed and none diverged to the point that the data cannot be treated as normal (Rubin, 2004). For analysis, item scores were dichotomized such that experiencing a symptom never or only one time or only once in a while was coded as “0” and experiencing a symptom at least half the time or almost always was coded as “1.” There were minimal missing data on PDS at Time 1 (0.4% – 1.7% missing) and Time 2 (0.4% - 1.3% missing), and there was no evidence of differential attrition based on PTSD symptom severity (t(1565) = 0.55, p = .58).

Analysis

LCAs were conducted independently with Time 1 and Time 2 data. The goal of an LCA is to identify the most parsimonious model (i.e., the model with the smallest number of latent classes) that adequately describes the data; the fit of a given solution is therefore usually evaluated against the fit of the solution with one fewer latent class. Statistical model fit indicators used in the class enumeration process included the Akaike information criterion (AIC), Bayesian information criterion (BIC), and sample-size adjusted Bayesian information criterion (aBIC), with lower values indicating superior model fit. Likelihood-based tests were also used as model fit indicators, specifically the Vuong-Lo-Mendell-Rubin adjusted likelihood ratio test (VLMR-LRT) and the bootstrapped likelihood ratio test (BLRT). Substantive considerations included coherence, meaningfulness, and differentiation between classes (Muthén & Muthén, 2000).

After the final latent class solution was identified at each time point, fully constrained and freely estimated joint measurement models were compared to determine appropriate invariance parameters for the longitudinal model. Comparison of invariance parameters relied on a likelihood ratio test and information criteria indicators, as well as substantive evaluation of model differences. The selected invariance parameters were then applied to the latent transition model, in which the latent variable at Time 1 was regressed onto the latent variable at Time 2. The LCAs and LTA were conducted in Mplus version 8.1 (Muthén & Muthén, 1998–2017) using the robust maximum Likelihood (MLR) estimator. Missing data was handled with Full Information Maximum Likelihood (FIML) estimation, which retains participants who lack complete data in the model.

Results

Overall Levels of Posttraumatic Stress

At Time 1, dichotomized levels of posttraumatic stress symptom endorsement ranged from 18% - 55%. Overall levels of endorsement were lower at Time 2, ranging from 9 – 40%. The proportion of the sample who endorsed dichotomized symptoms at each time point is presented in Supplemental Table 2. Through the lens of symptom clusters, the avoidance and alterations in arousal and reactivity symptom clusters were endorsed most frequently at each time point, followed by negative alternations in cognition and mood symptoms, with intrusion symptoms endorsed least frequently.

Latent Class Enumeration

At each time point, 1-class through 10-class solutions were generated and goodness of fit statistics were compared. At Time 1, the BIC continued improving through the 7-class model, but the VLMR-LRT indicated that adding classes beyond the three-class model did not significantly improve model fit (see Table 1). Conflicting model fit indicators are not uncommon, and, in those instances, it is recommended practice to evaluate candidate models for substantive meaning and coherence (Masyn, 2013; Nylund-Gibson & Choi, 2018). Closer examination of the 3 – 6 class models showed that solutions containing more than four classes produced classes similar to those that had already emerged, albeit at slightly different levels of severity. The contribution of these additional classes was determined to be too small to justify the loss of parsimony and the 4-class model was therefore selected as the final model at Time 1.

Table 1.

Fit Statistics for Latent Class Models at Time 1 and Time 2

K	LL	AIC	BIC	aBIC	LMR p-value	BLRT p-value	Entropy
Time 1
1	−17997.03	36028.07	36120.30	36066.29	--	--	--
2	−14605.42	29280.85	29470.73	29359.54	< .01	0.00	0.90
3	−13863.06	27832.12	28119.67	27951.29	<.01	0.00	0.88
4	−13623.58	27289.15	27774.35	27548.80	0.15	0.00	0.88
5	−13435.58	27049.17	27534.03	27249.28	0.07	0.00	0.83
6	−13337.35	26888.70	27469.22	27129.29	0.57	0.00	0.82
7	−13245.80	26741.60	27419.77	27022.66	0.17	0.00	0.80
Time 2
1	−11857.42	23748.84	23836.28	23782.28	--	--	--
2	−9454.22	18978.44	19158.47	19047.29	< .01	0.00	0.90
3	−8958.14	18022.29	18294.90	18126.54	< .01	0.00	0.88
4	−8793.35	17728.71	18093.90	17868.37	< .01	0.00	0.83
5	−8649.83	17747.67	17935.45	17652.75	< .01	0.00	0.82
6	−8555.25	17324.50	17874.89	17534.99	<.01	0.00	0.80
7	−8489.64	17229.28	17872.23	17475.17	0.37	0.00	0.81

Note: K = number of classes; FP = Free Parameters, LL = Log-Likelihood, AIC = Akaike Information Criterion, BIC = Bayesian Information Criterion, aBIC = Sample-Size Adjusted Information Criterion, LMR = Lo-Mendell Rubin Adjusted Likelihood Ratio Test, BLRT = Bootstrapped Likelihood Ratio Test

The four classes in this solution varied both quantitatively (i.e., overall severity) and qualitatively (i.e., symptom/symptom cluster severity). One class (25.3% of participants) showed a generally low probability of endorsing any posttraumatic stress symptoms (<0.15) and was therefore named the Low Severity class. Another class (14.6% of participants) showed a generally high probability of symptom endorsement (>0.55) and was therefore named the High Severity class. The primary distinction between the two intermediate classes was that one showed a substantially higher probability of endorsing nearly all negative alterations in cognition and mood symptoms and some alterations in arousal and reactivity symptoms. This class was labeled the High Moderate—Depressed & Anxious class³ (27.7% of participants) and the remaining class, which was characterized by high probability of endorsing avoidance and reactivity items relative to other symptoms, was labeled the Low Moderate—Avoidant & Reactive class (32.4% of participants). Entropy was high (0.88), indicating that the classes were well-distinguished from one another.

At Time 2, the BIC continued improving through the 7-class solution, but the VLMR-LRT suggested that there was not significant improvement in models that contained more than 6 classes (See Table 1). Similar to the patterns identified at Time 1, closer examination of the 4 – 7 class solutions suggested that, beginning with the 5-class model, new classes demonstrated generally similar endorsement patterns to existing classes but at a different level of severity. The contribution of these additional classes was determined to be too minor to justify the loss of parsimony and the 4-class model was selected as the final Time 2 model. Similar to the final solution at Time 1, this solution also included a class that could be understood as Low Severity (45.6%), High Severity (15.4%), High Moderate—Depressed & Anxious (21.1%) and Low Moderate—Avoidant & Reactive (17.8%). Entropy was slightly lower at Time 2 (0.83) but still within a range that indicates good classification of individuals into latent classes (Collins & Lanza, 2010).

Invariance Testing and Longitudinal Findings

Both the freely estimated and fully constrained joint models showed good fit to the data (see Table 2). The log-likelihood ratio test was significant (p < .001), indicating that the freely estimated model was preferable to the fully constrained model. The AIC was superior for the freely estimated model, as well. The BIC and aBIC, however, were substantially better in the constrained model. The BIC and aBIC penalize for the number of freely estimated parameters, indicating that when parsimony is considered, the fully constrained model is superior to the freely estimated model. Substantive comparison of the latent class solutions also supported selection of the fully constrained model, as the differences between latent class solutions were deemed too minor to warrant the loss of parsimony associated with the freely estimated model. Therefore, the fully constrained model was used for the longitudinal analysis (see Figure 1).

Table 2.

Joint LCA Model Fit Indices at Different Levels of Measurement Invariance

	Free Parameters	Log Likelihood	TRD p-value	AIC	BIC	aBIC
Freely Estimated	142	−22416.93	< .001	45117.86	45888.43	45437.32
Fully Constrained	74	−22529.77	--	45207.55	45609.11	45374.02

Note. TRD p-value was calculated using the Loglikelihood formulas: cd = (p0*c0 – p1*c1)/(p0 - p1); TRd = 2*(L0 – L1)/cd. Bolded cells indicate preference for the associated model.

Figure 1. Item Probability Plot for the Final Longitudinal Model.

With the measurement invariance parameters in place, the Low Moderate—Avoidant & Reactive class was the largest at Time 1 (34.3%), followed by the High Moderate—Depressed & Anxious class (25.4%), the Low Severity class (23.9%) and finally the High Severity class (16.4%). At Time 2, the Low Severity class was the largest (40.1%), followed by the Low Moderate—Avoidant & Reactive class (31.0%), the High Moderate—Depressed & Anxious class (20.8%), and the High Severity class (8.1%). As is depicted in Table 3, PTSD scores varied substantially by latent class. The severity range associated with the PTSD measure suggests that scores of 1 – 10 be interpreted as mild, 11 – 20 be interpreted as moderate, 21 – 35 be interpreted as moderate to severe, and 36 – 51 be interpreted as severe symptomology (McCarthy, 2008). These delineations add clarity to the latent classes, indicating that the Low Severity class primarily reflects scores that would be labeled as mild, the Low Moderate—Avoidant & Reactive class primarily reflects scores that would be labeled as moderate, the High Moderate—Depressed & Anxious class primarily reflects scores that would be labeled as moderate to severe, and the High Severity class primarily reflects scores that would be labeled as severe.

Table 3.

Mean Total PDS Scores by Latent Class Based on the Longitudinal Model

Time 1
	n	Mean Likert PTSD Score (SD)
Low Severity	374	5.81 (4.92)
Low Mod.—Avoidant & Reactive	514	16.64 (5.14)
High Mod.—Depressed & Anxious	428	28.00 (5.45)
High Severity	251	40.99 (5.88)
Full sample with valid sum PDS score	1567	21.06 (12.92)
Time 2
	n	Mean Likert PTSD Score (SD)
Low Severity	482	5.98 (4.91)
Low Mod.—Avoidant & Reactive	356	16.91 (4.95)
High Mod.—Depressed & Anxious	271	27.59 (5.78)
High Severity	92	40.83 (5.84)
Full sample with valid sum PDS score	1201	16.77 (12.03)

Note. The Total PDS score is based on a possible range of 0 – 51. Because Mean Total PDS Score is based on individuals’ sum PDS scores, participants with missing item level data on the PDS are not included, resulting in a smaller n than was used in the latent class modeling. Values reflect actual rather than estimated means and standard deviations. Latent class groupings are based on most likely class assignment.

Table 4 presents the probabilities of participants remaining in the same class as well as the probabilities of transitioning into a different latent class over time. Because the four latent classes vary by overall severity as well as symptom cluster severity, transitions can be discussed in terms of escalation and de-escalation, with movement from the High Severity to the High Moderate—Depressed & Anxious, High Moderate—Depressed & Anxious to Low Moderate— Avoidant & Reactive, and Low Moderate—Avoidant & Reactive to Low Severity class indicating de-escalation and the reverse pattern indicating escalation. Discussion of escalation and de-escalation, however, is not meant to imply that differences between levels of severity are equal and should not obscure the qualitative differences between latent classes.

Table 4.

Latent Transition Probabilities Based on The Estimated Latent Transition Model

		Time 2
Time 1	High Severity	High Mod—Depressed & Anxious	Low Mod—Avoidant & Reactive	Low Severity
High Severity	0.34	0.26	0.23	0.18
High Mod—Depressed & Anxious	0.06	0.50	0.30	0.14
Low Mod— Avoidant & Reactive	0.03	0.10	0.46	0.42
Low Severity	<0.01	0.03	0.17	0.80

Note. Bolded cells indicate stability.

The probability of remaining in the same class over time is indicated along the diagonal of Table 4. That the greatest stability was evidenced by the Low Severity class and the least stability by the High Severity class indicates that few participants who began the study with low severity symptoms experienced escalation and that many who began the study with high severity symptoms experienced de-escalation over time. The off-diagonal cells indicate transition, with Time 1 class membership indicated by row and Time 2 class membership indicated by column. These transition probabilities support an overall pattern of de-escalation, as well. Individuals who transitioned out of the Low Moderate—Avoidant & Reactive class, for example, were more likely to transition into the Low Severity class (0.42) rather than the High Moderate—Depressed & Anxious (0.10) or High Severity (0.03) class. Similarly, individuals who transitioned out of the High Moderate—Depressed & Anxious class were more likely to transition into the Low Moderate—Avoidant & Reactive (0.30) or Low Severity (0.14) classes than the High Severity (0.06) class. Overall, these results indicate that transition between latent classes was common and generally represented movement into a lower severity class.

Discussion

The current study utilized LTA to explore heterogeneity in sexual assault survivors’ PTSD symptom profiles over time. Consistent with previous mixture modeling studies conducted with sexual assault survivors (Armour et al., 2012; Steenkamp, Dickstein et al., 2012) and survivors of other traumas (Galatzer-Levy et al., 2018), we identified multiple latent classes representing unique PTSD symptom profiles. Research Question 1 asked what latent classes best represented this sample’s PTSD symptom profiles and four classes emerged at each time point. The classes were longitudinally invariant and were characterized as Low Severity, Low Moderate—Avoidant & Reactive, High Moderate—Depressed & Anxious, and High Severity. As is implied by their names, the classes showed differences in overall severity as well as the severity of specific symptom clusters.

Our results are consistent with prior studies that have identified intermediate severity classes in both sexual assault-specific samples (Au et al., 2013; Steenkamp, Dickstein, et al., 2012) and general trauma samples (Ayer et al., 2012; Breslau et al., 2005; Frost et al., 2020). These intermediate classes highlight that binary categorizations, in which one either has or does not have PTSD, may oversimplify the full range of posttraumatic stress experiences. The intermediate classes in our analysis were primarily distinguished from each other based on their likelihood of endorsing negative alterations in cognition and mood and arousal symptoms. Interestingly, this collection of symptoms is relatively trauma non-specific. Whereas the intrusion and avoidance items make specific reference to the traumatic event (e.g., feeling emotionally upset when reminded of the experience; trying not to think about the experience), most items assigned to the negative alterations in cognition and mood and the alterations in arousal and reactivity symptom clusters do not (e.g., feeling distant or cut off from people around you; having trouble falling or staying asleep). This suggests that there may be two distinct groups of trauma survivors at an intermediate level of severity, one whose experience of posttraumatic stress is primarily limited to reminders of the traumatic event and another whose posttraumatic stress permeates their lives more broadly.

Our findings also demonstrate a pattern, previously identified in Hebenstreit et al.’s (2015) study of intimate partner violence survivors, in which at least one class shows substantially higher endorsement probabilities for reactivity items (e.g., alert, jumpy) than arousal items (e.g., difficulty sleeping, irritability, difficulty concentrating), despite the items reflecting a shared symptom cluster. There are at least two potential explanations for this pattern. The first is that arousal and reactivity are unique constructs that may be better understood as separate symptom clusters. Simms and colleagues (2002) proposed a PTSD factor structure in which the three arousal items were grouped with those currently assigned to negative alterations in cognition and mood to create a broader symptom cluster known as dysphoria; Ullman and Long’s (2008) confirmatory factor analysis supported this model, as well. This conceptualization was considered and rejected in the DSM-5 (Armour et al., 2016), but if studies continue to identify this pattern, the dysphoria model may warrant reconsideration. An alternative explanation is that all five symptoms do represent the same underlying construct, but arousal symptoms resolve more quickly than do those related to reactivity. This is an especially compelling explanation for gender-based violence survivors who may perceive their reactivity symptoms (e.g., heightened awareness of people nearby) to be self-protective against revictimization or ongoing abuse. The actual utility of heightened reactivity may depend on the type of abuse, and regardless would not negate potential negative consequences of prolonged reactivity (e.g., high blood pressure, Smith et al., 2019). These tensions are an important area for future clinical work and research.

Research Question 2 addressed the probability of transitioning from one latent class into another over a one-year period. Consistent with Dworkin et al.’s (2021) findings that many sexual assault survivors report decreases in distress over time, our results indicated that 42 – 66% of participants in each latent class de-escalated into a lower severity class between Time 1 and Time 2. Because participants assigned to the highest severity class at Time 1 had no possibility of escalation and participants assigned to the lowest severity class at Time 1 had no possibility of de-escalation, the stability of these classes also informs the interpretation of escalation and de-escalation. As would be expected in an overall environment of de-escalation, the High Severity class showed the least stability, and the Low Severity class showed the most stability. In other words, a majority of participants assigned to the High Severity class at Time 1 de-escalated into a lower severity class at Time 2 and a majority of participants assigned to the Low Severity class at Time 1 remained in the Low Severity class. Previous research has found similar patterns of de-escalation for high severity classes and stability for low severity classes among sexual assault survivors (Armour et al., 2012), veterans (Elliott et al., 2005), and survivors of terrorist attacks (Adams et al., 2019).

Despite the relative rarity of symptom escalation, the probability was not negligible, ranging from 6 – 20% across latent classes. This pattern is also represented in the literature, with research based on veterans (Orcutt et al., 2004), terrorist attack survivors (Adams et al., 2019), and children exposed to intimate partner violence (Meijer et al., 2019) all identifying a class of participants whose symptoms worsen over time. Such escalation may be distressing above and beyond the experience of higher severity symptoms, as survivors who felt as though they were “getting better” or “recovered” may be demoralized by an increase in symptom severity. Future research should explore variables that predict transition between latent classes, including those that might predict PTSD symptom change across trauma types (e.g., cumulative lifetime trauma, social support) as well as those that might function differently across traumatic experiences (e.g., victim blame). Incorporating qualitative research into this area of study may have particular clinical utility, as it could help clinicians recognize markers of symptom escalation and normalize these transitions for clients as they occur.

This study’s findings should be understood in the context of its limitations, one of which is that the PTSD measure used as the basis for the latent classes reflects the DSM-IV rather than the updated DSM-5. This limitation, while noteworthy, is unlikely to substantially impact the results of the study. Although data were collected with a DSM-IV measure, the PTSD items were interpreted and discussed according to the DSM-5 symptom structure to inform current research and practice. Additionally, sixteen of the seventeen PTSD symptoms reflected in the DSM-IV remain in the revised DSM-5, and there is little reason to think that endorsement patterns on those items would change with the addition of the new DSM-5 symptoms. There are also limitations associated with mixture modeling as an analytic method. LCA describes latent classes but cannot prove that they exist (Vermunt & Magidson, 2002) and classes must therefore be evaluated against extant theory and research. Furthermore, latent classes are categorical and many, if not most, complex concepts are continuous at their core. Identifying categorical groupings along a continuum may still provide a useful framework for understanding heterogeneity (Masyn, 2013), so long as latent classes are understood as those that best represent these data and not as reified categories.

There are limitations associated with the study’s sample, as well. Participants were recruited through community advertisements describing a study for women with unwanted sexual experiences and there may be a variety of factors that differentiate women who chose to participate from those who did not. It is possible that more symptomatic individuals did not volunteer for the study due to their heightened distress but also possible that less symptomatic individuals did not volunteer because they perceived the study as less relevant to them. Despite the limitation associated with these unknown differences, the community nature of the sample can be considered a strength, as well. Whereas previous mixture modeling research with sexual assault survivors has been conducted with help-seeking samples (e.g., Armour et al., 2012; Steenkamp, Dickstein et al., 2012), this study helps to establish the variety of symptom profiles that emerge among survivors who are not necessarily seeking formal support.

In spite of the study’s limitations, the findings can inform clinical practice and research. For decades, practitioners have criticized the binary framing in which one either “has” or “does not have” PTSD (Clark et al., 2017) and the results of this study underscore the validity of those concerns. Differentiation of latent classes based on symptom cluster, in addition to overall, severity also suggests potential utility in diagnostic tests whose results communicate the relative spread of trauma survivors’ PTSD symptoms across symptom clusters. Such diagnostic tests would not necessarily change treatment approaches but could provide a shared language for treatment providers and patients to talk about the relative weight of specific symptoms in patients’ experiences of posttraumatic stress. Finally, the prominence of reactivity symptoms suggests that these markers of hypervigilance could be especially important to address with sexual assault survivors.

Regarding implications for research, the current findings emphasize the importance of modeling PTSD heterogeneity at the symptom or symptom cluster level. Longitudinal mixture models that include only indicators of overall severity have provided crucial insight into PTSD trajectories but have been unable to clarify temporal resolution of symptom clusters. In their latent growth model with survivors of motor vehicle accidents, Wu and Cheung (2006) determined that avoidance symptoms may take longer to resolve than intrusion or hyperarousal symptoms. Although our analysis cannot speak to this directly due to participants’ varying lengths of time post-assault, the high levels of avoidance that remained in both intermediate classes support the uniqueness of the avoidance symptom cluster and encourage additional research in this domain. Additionally, that we did not find previously identified subtypes of PTSD (e.g., internalizing/externalizing) prompts consideration of the conditions under which subtypes of PTSD may emerge. This subtype has been identified in samples recruited on the basis of high PTSD symptoms (Miller, 2004; Wolf et al., 2012), and it is possible, even probable, that different subtypes are likely to emerge depending on the symptomology required for study participation. This possibility further highlights the importance of expanding research samples to include participants with a wide diversity of posttraumatic stress experiences.

In sum, this study adds to a growing literature that demonstrates meaningful heterogeneity in trauma survivors’ posttraumatic stress responses. That multiple latent classes of PTSD symptoms remain years post-assault for many sexual assault survivors highlights the need to expand populations of study to include trauma survivors whose traumatic events may have occurred years prior. Regardless of whether trauma survivors report diagnosable levels of PTSD, many experience ongoing PTSD symptoms that warrant research and clinical attention.

Clinical Impact Statement:

The substantial variation in sexual assault survivors’ posttraumatic stress symptoms, as well as the frequency with which transition occurred between latent classes, indicates that change in symptoms continues to occur years, even decades, after a sexual assault. The varied prominence of symptom clusters across latent classes encourages attention to how specific symptoms may impact individuals’ experiences of posttraumatic stress over time. Consistent elevation of reactivity symptoms suggests that these symptoms may warrant particular clinical attention for gender-based violence survivors.