Impaired prediction of ongoing events in posttraumatic stress disorder

Abstract

The ability to make accurate predictions about what is going to happen in the near future is critical for comprehension of everyday activity. However, predictive processing may be disrupted in Posttraumatic Stress Disorder (PTSD). Hypervigilance may lead people with PTSD to make inaccurate predictions about the likelihood of future danger. This disruption in predictive processing may occur not only in response to threatening stimuli, but also during processing of neutral stimuli. Therefore, the current study investigated whether PTSD was associated with difficulty making predictions about near-future neutral activity. Sixty-three participants with PTSD and 63 trauma controls completed two tasks, one testing explicit prediction and the other testing implicit prediction. Higher PTSD severity was associated with greater difficulty with predictive processing on both of these tasks. These results suggest that effective treatments to improve functional outcomes for people with PTSD may work, in part, by improving predictive processing.

1. Introduction

People constantly make predictions about what will happen in the near future, anticipating how other people around them will act, what other people will say, and how objects in the environment will affect them. In fact, functions ranging from object recognition (e.g., Bar et al., 2006), to action guidance (e.g., Grush, 2004), and decision making (e.g., Doya, 2008) all require predictive processing for successful performance. However, people with certain disorders, including Posttraumatic Stress Disorder (PTSD), may experience difficulty with making accurate predictions. People with PTSD often experience hypervigilance, leading to frequent, inaccurate predictions about future danger that could disrupt predictions about true future events (e.g., Gagne et al., 2018). By definition, these hypervigilance symptoms affect people with PTSD not only in situations of objective danger, but also in everyday situations like restaurants, social gatherings, and sports stadiums (American Psychiatric Association, 2013). In addition, memory for neutral stimuli is also impaired in PTSD (e.g., Vasterling et al., 2002), suggesting that impairments generalize to non-threat-related situations. However, previous studies of prediction in PTSD have focused only on anticipation of threat-related stimuli, and there have not been any studies directly investigating prediction of naturalistic, neutral stimuli in people with PTSD. Therefore, the current study was designed as a first step toward investigating predictive processing of naturalistic, neutral activity in PTSD.

Understanding how PTSD affects processing of naturalistic, neutral stimuli is important because PTSD symptoms do not arise only in the context of aversive stimuli; hypervigilance and other PTSD symptoms also impact people’s experience of neutral or even positive experiences (American Psychiatric Association, 2013). Relatedly, differences in prediction ability may manifest in neutral or positive as well as negative events, though this has not been previously investigated. Previous studies investigating prediction of threat-related stimuli (Brinkmann et al., 2017; Grillon et al., 2009; Grupe et al., 2016) measured sequelae of anticipatory processing (startle response, brain activation), without investigating the predictive process itself. In order to determine whether people with PTSD anticipate future activity differently than healthy populations, it is necessary to directly interrogate the predictive process.

Prediction and prediction error are central to comprehending activity. In particular, prediction error monitoring has been proposed to function as a control mechanism for segmenting the continuous stream of behavior into events (Zacks et al., 2007). People spontaneously segment ongoing activity into events (Baldassano et al., 2016; Yates et al., 2021; Zacks et al., 2001a,b). Event boundaries tend to correspond to points in time when there are large changes in physical motion, objects, people, locations, goals or causes (Huff et al., 2014; Newtson et al., 1977; Zacks, 2004; Zacks et al., 2009a; Zacks et al., 2009b). At these points of change, activity becomes less predictable and prediction errors tend to increase. For example, once someone starts tying their shoes, the succession of movements is relatively predictable, but once their shoes are tied it is much more difficult to predict the next movement they will make. Models of event segmentation have proposed that cognitive systems monitor for spikes in prediction error and update event representations at those moments (Franklin et al., 2020; Gumbsch et al., 2022; Zacks et al., 2007).

Event segmentation can be measured using a behavioral task in which participants press a button to indicate boundaries between events. Studies of healthy adults have found high agreement about the locations of event boundaries both across and within participants (Zacks et al., 2006), meaning that event boundaries identified from one sample of participants can be used to investigate predictive processing around event boundaries in other samples. However, this stability may break down in people with PTSD: A previous study in a non-clinical sample found that people with more severe symptoms of PTSD agreed less on the locations of event boundaries (Eisenberg et al., 2016), and an independent analysis of event segmentation data collected from the same sample reported here found similar results (Pitts et al., 2022).

Previous research using an explicit prediction task has shown that prediction errors are larger around event boundaries (Zacks et al., 2011). For this task, participants watch movies of actors engaging in neutral, naturalistic activities. These movies are paused periodically, and during these pauses, participants make predictions about what they expect to happen 5 s later. Some pauses occur around event boundaries when prediction is most difficult and some occur within events when prediction is much easier. This task simulates situations in which people purposely make predictions about what they think will happen in the near future (e.g., someone learning how to cook a complicated meal purposely trying to predict what the chef will do next in order to better learn the correct sequence of steps). In a previous study in healthy adults, explicit prediction was less accurate when prediction occurred across event boundaries than when prediction occurred within events.

One downside to this explicit prediction task is that it interrupts viewing of the movies and is therefore not a fully naturalistic assessment of prediction ability. Eye tracking provides a mechanism for investigating prediction in a much more naturalistic manner, and it has been used to interrogate predictive processing in non-clinical populations ranging from infants to adults. For example, healthy infants make predictive looks to locations where stimuli are expected to appear (Haith and McCarty, 1990; Romberg and Saffran, 2013) and use their knowledge about objects to predictively look at the locations where they expect these objects to be used (Hunnius and Bekkering, 2010). Adults also engage in predictive looking to expected stimuli, looking ahead to where objects are expected to be moved (Flanagan and Johansson, 2003) and making predictive eye movements to locations where salient information in narrative scenes is anticipated to appear (Vig et al., 2011). In the context of sports, adults anticipate a ball’s position as it flies through the air (Hayhoe et al., 2012) and make predictive eye movements to where they believe a ball will bounce (Diaz et al., 2013; Henderson, 2017). Adults’ predictive eye movements can also be used to test implicit learning, including learning of statistical regularities in human action sequences (Monroy et al., 2018).

We recently developed a new eye tracking task, called the predictive looking at action task (PLAT), to directly investigate variability in predictability of near-future activity during event comprehension without interrupting comprehension with an explicit judgment task (Eisenberg et al., 2018). In the PLAT, people passively watch movies of actors completing everyday activities while an eye tracker determines their eye gaze locations. Predictive looking is measured based on how early people look at the objects the actors are about to touch. In two studies of healthy participants (Eisenberg et al., 2018), people were more likely to predictively look at the target object as the contact time approached. In addition, predictive looking varied as a function of event boundaries, with less predictive looking early on when object contact occurred near event boundaries. In terms of the shoe-tying example, people would be expected to show better explicit prediction and better PLAT performance while watching someone in the midst of tying shoes than just as the person finishes tying them. This task therefore provides a promising method for investigating the time course of prediction of near-future activities over long, naturalistic sequences of human activity.

In the current study, the explicit prediction task and the PLAT were used to investigate predictive processing in people with PTSD. We included both types of tasks in order to investigate multiple types of naturalistic predictive processes (purposeful predictions in the explicit task and more common, implicit predictions in the PLAT). For both tasks, prediction performance around event boundaries was compared to prediction within events to determine whether PTSD impacts prediction at event boundaries, when prediction is already more difficult. Because contemporary approaches to psychopathology, including PTSD, emphasize its continuous rather than discrete nature (e.g., Kotov et al., 2018; Ruscio et al., 2002), we also included a continuous measure of PTSD symptom severity.

Before performing each task, participants listened to narratives of their traumatic event or a positive life event. Because people with PTSD often experience reminders of their traumatic events in everyday life due to generalized fear triggers (e.g., a song, a smell, or a color that is similar to something they experienced during their traumatic experience; American Psychiatric Association, 2013), these narratives simulated trauma-related triggers that might naturally occur in the participants’ environment but that might not occur in a laboratory setting. The positive life event narrative served as a control, as thinking about a positive memory would not be expected to trigger avoidance or hyperarousal symptoms in people with PTSD and therefore would not be expected to negatively impact prediction performance.

We tested the following hypotheses: 1) People in the PTSD group and people with high PTSD severity will make less accurate predictions about near-future neutral activity compared to people in the control group and people with low PTSD severity, as evidenced by decreased accuracy on the explicit prediction task and less predictive looking on the PLAT; 2) Participants across group and PTSD severity will experience more difficulty making accurate predictions around event boundaries compared to within events on both the explicit and implicit tasks; 3) Listening to the trauma narrative will reduce prediction accuracy for people in the PTSD group and people with high PTSD severity compared to listening to the positive narrative, but will have less impact on the control group and people with lower PTSD severity.

2. Materials and methods

2.1. Participants

Participants for the PTSD and the control groups were recruited from the Volunteer for Health participant registry, which is a subject pool maintained by the Washington University School of Medicine, and from advertisements posted on Craigslist. Potential participants completed a short phone screen using the Mini Neuropsychiatric Interview (Sheehan et al., 1998) to determine preliminary assignment to groups and potential exclusion from the study. PTSD diagnosis was then confirmed using the SCID during the first session of the study, and participants were assigned to groups or excluded based on the results of this assessment. Participants were paid $10/hour for their time. Sixty-three participants with PTSD and 63 trauma control participants participated in the study. Trauma control participants were recruited to match participants in the PTSD group on age (within 10 years), gender, years of education (within approximately 2 years), and ethnicity (if mixed ethnicity, at least one match). Control participants were also required to have experienced at least one criterion A traumatic event.

Exclusion criteria for the PTSD group included lack of a PTSD diagnosis, history of psychosis, current substance use disorder, and current manic episode. Exclusion criteria for the control group included more than 3 current symptoms of PTSD, any PTSD symptoms that significantly interfered with important life functioning or caused significant distress, history of psychosis, current substance use disorder, and current manic episode. Control participants were not excluded if they had current or past anxiety or depressive disorders. All participants were between 18 and 50 years old.

The decision to include people with up to 3 current PTSD symptoms in the trauma control group was made because in previous studies that used the script generation task described here (e.g., Pitman et al., 1987; Pitman et al., 1990), participants in the trauma control groups were only excluded if they had a PTSD diagnosis, which meant that they could potentially be right on the border of a diagnosis. In order to make sure that people in the trauma control group were not on the border of a diagnosis, we excluded anyone with more than half of the minimum of 6 PTSD symptoms required for a DSM-5 diagnosis of PTSD.

We powered the study to detect whether predictive processing would mediate the relationship between PTSD and event segmentation and between PTSD and memory, using power tables from Thoemmes et al. (2010). This suggested that we needed a total sample size of 92 (46 participants per group) to find medium effects for both mediation paths. We over-recruited participants for the current study because we planned to use the current sample as the sole recruitment source for a second study (Pitts et al., 2023). Because we needed to ensure that we could recruit a large enough sample for this second study, the power needed for the current study was not the rate-limiting factor in recruiting.

For the PTSD group, 103 people were recruited for the study, but 18 withdrew from the study before completing all three sessions, 6 were excluded due to a history of psychosis, 6 were excluded due to a current substance use disorder, 4 were excluded due to difficulty tracking their eyes with the eye tracker, 3 were excluded due to not meeting criteria for PTSD, 2 were excluded due to experimenter error, and 1 was not successfully matched to a participant in the control group. Therefore, 63 participants in the PTSD group were included in all analyses.

For the trauma control group, 91 people were recruited for the study, but 17 withdrew from the study before completing all three sessions, 7 were excluded due to a history of PTSD, 2 were excluded due to non-compliance with study procedures, 1 was excluded due to experimenter error, and 1 was excluded due to having completed a previous study in the laboratory that was similar to the current study. One participant was missing data for one session of the PLAT but was still included in all analyses. Therefore, 63 participants in the control group were included in all analyses. See Table 1 for demographic information for all participants. The Washington University Human Research Protection Office approved this study.

Table 1.

Descriptive statistics for participant demographics and psychological assessments.

	PTSD (N = 63)	Control (N = 63)	t	df	p
Age	34.6 (20–49)	33.9 (18–50)	0.41	123	0.67
Years of Education	14.7 (12–18)	15.1 (12–18)	1.56	123	0.12
Gender
Male	12	11
Female	51	52
Ethnicity
White	34	43
Black	14	16
Asian	2	1
Mixed Race	12	3
Unknown	1	0
Days between Sessions 1 and 2	11.2 (1–146)	8.1 (1–21)	1.27	69.94	0.20
Days between Sessions 2 and 3	11.8 (3–77)	9.2 (3–27)	1.3	90.46	0.20
PCL-C	54.6 (27–80)	20.9 (17–44)	17.83	86.85	<.001
DASS	57.2 (0–110)	9.8 (0–77)	11.56	92.43	<.001
DES	670.6 (30–2070)	184.8 (0–970)	7.36	85.57	<.001
LSAS	64.9 (13–134)	30.7 (0–96)	7.61	115.82	<.001
MSPSS	52.8 (12–84)	66.1 (14–84)	−4.48	120.57	<.001

Note: PCL-C = PTSD Checklist - Civilian; DASS = Depression Anxiety Stress Scale-42; DES = Dissociative Experiences Scale; LSAS = Liebowitz Social Anxiety Scale; MSPSS = Mulitdimensional Scale of Perceived Social Support.

2.2. Experimental procedures

Participants completed experimental tasks across three days. During session 1, which took 2–4 h, participants completed psychological assessments. During session 2, which occurred on average 9.8 days (range 1–146 days) after session 1, participants completed the explicit prediction task, the PLAT, and the event segmentation task. During session 3, which occurred on average 10.2 days after session 2 (range 3–77 days), participants completed the same tasks, but with novel movies. Sessions 2 and 3 were each 2 h long. The results of t-tests suggested no significant differences in time between sessions between the PTSD and control groups. See Table 1 for the results of the t-tests.

2.2.1. Psychological assessments

Participants completed a battery of psychological assessments, including the Structured Interview for DSM-IV-Research Version (SCID; Mood, Substance Use, PTSD, and Psychosis Modules; First et al., 2002), the PTSD Checklist for DSM-IV – Civilian version (PCL-C; Weathers et al., 1993), the Beck Depression Inventory (Beck et al., 1996), the Depression Anxiety Stress Scale-42 (Lovibond and Lovibond, 1995), the Dissociative Experiences Scale (Bernstein and Putnam, 1986), the Liebowitz Social Anxiety Scale (Liebowitz, 1987), and the Multidimensional Scale of Perceived Social Support (Zimet et al., 1988). The SCID was modified to include criteria for both DSM-IV and DSM-5, as the SCID for DSM-5 had not yet been released at the time the study began. See Table 1 for means and group differences for these measures. After completing these psychological assessments, participants were given a list of mental health resources that included therapeutic services available in the local area, including low-cost and sliding scale options.

2.2.2. Life event narratives

During session 1, participants wrote a brief narrative of their traumatic event and a brief narrative of a positive life event. The instructions for this script elicitation task were based on those used by Pitman et al. (1987) and Pitman et al. (1990), who found that this traumatic script elicited greater physiological activation and sadness, fear, anger, and disgust in people with PTSD compared to trauma exposed controls. Participants listened to their narratives before each task during sessions 2 and 3. Half of the participants heard the traumatic narrative in session 2 and the positive narrative in session 3; the other half heard the narratives in the opposite order. After listening to each recording, they rated their mood and anxiety on a 10-point Likert-type scale.

2.2.3. Explicit prediction task

Participants watched three movies of an actor performing an everyday activity. For each of eight trials per movie, the movie was paused either 2.5 s before event boundaries (across-event trials) or 2.5 s before event middles (within-event trials). Each movie contained four trials of each type. During the pause, participants were shown two movie frames, one from 5 s later in the movie they were watching, and the other from a similar but different movie. Participants chose which picture they believed represented what would occur 5 s later in the movie. Fig. 1, Panel A illustrates the sequence of trials for this task and depicts a within-events trial. Movie clips 1–4 in the supplemental materials provide examples of boundary and within-event trials for this task.

Fig. 1.

Panel A. The sequence of the explicit prediction task. The movie was paused periodically, and during the pause, participants chose the picture representing what they believed would happen 5 s later in the movie. Half of the pauses occurred around event boundaries and half occurred within-events. Panel A illustrates an example of a within event trial. Participants saw the correct answer once the movie resumed. Panel B. Example interest area for the Predictive Looking at Action Task (PLAT). An interest area (red box) was drawn around the to-be-contacted object, and the amount of time participants spent looking within the interest area was used as the dependent variable in analyses (purple dot represents eye gaze location). Many participants identified an event boundary at around the time of the interest area depicted in Panel B.

The timing of the pauses in the movies was determined in a similar manner as that described in Zacks et al. (2011). Participants recruited from Amazon Mechanical Turk for a pilot study completed the event segmentation task. The four timepoints at which participants segmented most frequently were used to define event boundaries, and the four timepoints at which participants segmented least frequently were used to define event middles.

Split-half reliability was calculated for the explicit prediction task even though the number of trials was low (8 trials for each movie, half across-event and half within-event). Mean accuracy was calculated separately for odd and even trials for each participant for each of the trial and narrative types. The Spearman-Brown prophecy formula was applied to obtain the adjusted reliability for the full-length measure. For the positive narrative condition, the across-event reliability was 0.54 and the within-event reliability was 0.63. For the traumatic narrative condition, the across-event reliability was 0.51 and the within-event reliability was 0.66. This level of reliability is suboptimal, and could potentially be improved by increasing the number of trials, but is not unacceptably poor.

2.2.4. Predictive looking at action task (PLAT) and event segmentation task

The PLAT consisted of three components: passive movie viewing, memory recall, and event segmentation. First, participants watched three movies of everyday activities while their eyes were tracked. Immediately after each movie, they had 7 minutes to type into a blank text document as much as they could remember from the movie, in the order they remembered the activity occurring. After completing the recall task for the third movie, participants watched each movie for a second time and completed an event segmentation task while watching the movies. For this task, participants were instructed to press a button whenever they believed a small meaningful unit of activity had ended and another had begun. In order to ensure that participants understood these directions and were effectively engaged in this task, they watched and segmented a practice movie before moving on to the experimental task. The results of the event segmentation and memory components of this task are reported in a separate paper (Pitts et al., 2022).

2.3. Materials

A total of 12 movies of actors completing everyday activities were used in the study. Six of these movies (putting together Legos: 245 s; putting books on shelf: 311 s; washing car: 431 s; putting up a tent: 377 s; changing tire: 342 s; and washing clothes: 300 s) were used in the explicit prediction task and six (making breakfast: 329 s; preparing for a party: 376 s; planting plants: 354 s; walking through a library: 249 s; sweeping: 263 s; doing dishes: 327 s) were used for the PLAT. The sweeping movie was played at an incorrect frame rate (30 instead of 24 frames per second) due to experimenter error and was therefore sped up slightly when presented to participants (should have been 328 s). In addition, one of the movies (library) used for the PLAT was not suitable for the eye tracking analyses necessary to determine predictive looking, as the camera was not stationary, and is therefore not included in the analyses reported here. All other movies were filmed from a fixed, head-height perspective, with no pan or zoom.

2.4. Eye tracking

During both prediction tasks, gaze location from the participants’ right eye was tracked using an eye tracker (EyeLink 1000; SR Research Ltd, Mississauga, ON, Canada) that sampled at 1000Hz. The eye tracker camera was mounted on the SR Research Desktop Mount. Participants were instructed to keep their heads in an SR Research chin and forehead rest to minimize head movement. The camera was positioned 52 cm from the top of the forehead rest. The movies were presented on a 19 in (74 cm) monitor (1400×900 resolution, viewing distance of 58 cm from the forehead rest, viewing angle of 38.6°). Data was exported from Data Viewer software (SR Research Ltd., Mississauga, ON, Canada) into text files, which were then imported into R (R Core Team, 2014) for analysis.

Calibration of the eye tracker was conducted before beginning the first task. Participants were instructed to look at each of five to nine dots presented serially across the central and peripheral visual field. Following calibration, the measurements were validated by having the participants look at each of these dots again as they appeared on the screen. Validation was considered good when there was an average error of 0.50 degrees of visual angle or less and when the maximum error for any given dot was 1.00° or less. Calibration and validation were repeated until errors were under or as close to these cut-offs as possible. Due to technical errors, validation values from 27 eye tracking sessions could not be extracted. For the other 226 sessions, the mean average error was 0.36° (range = 0.13–0.89), and the mean maximum error was 0.80 (range = 0.37–2.32).

2.5. Data analysis

All data analysis was conducted in R (R Core Team, 2014), and all figures were created using the ggplot2 package (Wickham, 2016). Linear regression models were conducted in base R and beta values were calculated using the lm beta package (Behrendt, 2023). Mixed effects models were run using the lmerTest package (Kuznetsova et al., 2017), confidence intervals for main effects were calculated using the confint function’s profile method, and post-hoc tests for interaction contrasts were run using the emmeans package (Lenth, 2019), using the Tukey HSD method to adjust for multiple comparisons.

2.5.1. Predictive looking at action task

For the PLAT, analyses were similar to those described in Eisenberg et al. (2018). The time course of anticipatory looking was determined by calculating how much time participants spent looking at the objects the actor was about to touch during the 3000 ms before the actor contacted those objects. For each movie, time points at which the actor came into contact with an object were identified. Interest areas were drawn around target objects from 3000 ms before contact through 1000 ms after contact. Fig. 1, Panel B shows an example interest area. Interest areas were placed using the same rules as described in Eisenberg et al. (2018; see Section 1 of supplemental materials). There were 37, 60, 29, 46, and 35 interest areas for the breakfast, party, planting plants, sweeping, and dishes movies, respectively. To determine whether there were group differences in predictive looking near event boundaries, we coded whether each object contact happened within 1000 ms of an event boundary. This was done separately for each participant, based on their own segmentation data. For the PTSD group, 24.2% of the interest areas were near event boundaries and 75.8% were within-event (SD = 0.12). For the control group, 25.5% of the interest areas were near event boundaries and 74.5% were within-event (SD = 0.13). Many participants identified an event boundary at around the time of the interest area depicted in Panel B. Movies 5–12 in the supplemental materials provide examples of other event boundary and within-event trials for the PLAT.

Once interest areas were identified for each movie, eye tracking data from the 3 s before contact were divided into six 500 ms bins, as described in Eisenberg et al. (2018). For each subject, we calculated how long their gaze fell within the interest area during each of the six time bins. This variable was the dependent measure for the mixed-effects models.

To determine whether a similar pattern of effects would hold for predictive looking around normative event boundaries, we conducted the same analyses using event boundaries identified by a separate sample of healthy participants. See supplemental materials for more information about this sample.

Split-half reliability was calculated for the PLAT by determining the linear slope for predictive looking between the bin representing predictive looking from 3000 to 2500 ms before contact and the bin representing predictive looking from 500 to 0 ms before contact. This was done separately for the positive and traumatic narrative conditions. Each interest period was treated as a trial. Mean slopes for odd and even trials were calculated for each participant. The Spearman-Brown prophecy formula was applied to obtain the adjusted reliability for the full-length measure. The resulting split-half reliability was 0.89 for the positive narrative condition and 0.91 for the traumatic narrative condition, indicating very high reliability for this task.

3. Results

3.1. Life event narratives

For the life event narrative analysis, separate regression models predicting mood and anxiety were created. For both models, independent variables were trial number, group, and narrative condition (traumatic vs. positive). All possible interactions were also entered into the model. For the mood model, the two-way interactions between group and narrative condition (β = −0.17, p < .001) and between trial number and narrative condition (β = 0.12, p < .02) were significant. The shape of the interaction between group and narrative condition indicated that while both the control and the PTSD group experienced lower mood in the traumatic compared to the positive narrative condition, the PTSD group experienced a greater reduction in mood in the traumatic narrative condition (see Fig. 2, panel A). The two-way interaction between trial number and narrative condition suggested that the difference in mood ratings for the positive and traumatic narrative conditions became closer together over the course of the trials (see Fig. 2, panel C). This suggests that the narrative manipulation may have had less of an impact on mood over the course of the experimental sessions. For the anxiety model, only the main effects of group (β = .27, p < .001) and narrative type (β = 0.24, p < .001) were significant (see Fig. 2, panels B and D). The results were very similar when PTSD severity was used as an independent variable instead of group (see Section 2 of supplemental materials, including Table S1). Overall, these results suggest that the narrative condition impacted mood and anxiety, as seen in previous studies using this manipulation. See Table 2 for the results of these linear models.

Fig. 2.

Participants listened to either their traumatic or positive narrative before completing each task in the study and rated their mood and anxiety after each time they listened to these narratives. For the mood ratings, there were significant two-way interactions between group and narrative condition and between trial number and narrative condition. For the anxiety ratings, there were significant main effects of group and narrative condition but no significant interactions. For panels A and B, the x-axis displays the narrative condition and the y-axis displays mood (panel A) and anxiety (Panel B) ratings. For panels C and D, the x-axis displays the trial number and the y-axis displays mood (panel C) and anxiety ratings (panel D). The error bars are 95% confidence intervals.

Table 2.

Results from linear models for mood and anxiety ratings.

Independent Variables	Mood		Anxiety
	β	p	β	p
Number	−0.06	0.070	−0.04	0.281
Group	−0.13	0.005	0.27	<.001
Narrative Condition	−0.46	<.001	0.24	<.001
Number x Group	0.01	0.840	−0.003	0.954
Number x Narrative Condition	0.12	0.030	−0.09	0.13
Group x Narrative Condition	−0.17	0.003	0.10	0.095
Number x Group x Narrative Condition	−0.05	0.400	0.06	0.322

3.2. Explicit prediction task

To determine whether participants with PTSD performed worse on the explicit prediction task, mixed models with accuracy as the dependent variable were created. For one model, group (PTSD vs. control), narrative type (traumatic vs. positive), and condition (across-event vs. within-event) were entered as the fixed effects. The second model was identical except that PCL-C PTSD severity scores were entered into the model instead of group. For both models, the intercepts for subject and movie were entered into the model as random effects. For both models, the fixed effect of condition was significant, with lower accuracy for across-event compared to within-event trials (Model 1: β = 0.13, p < .001, confidence interval [0.11, 0.15]; Model 2: β = 0.13, p < .001, confidence interval [0.11, 0.15]). In Model 1, the fixed effect of group indicated that those with PTSD performed non-significantly worse than controls (β = −.02, p = .097 [−0.04, 0.004]). Model 2 provided stronger evidence for an effect of PTSD severity on prediction performance: higher PCL-C scores significantly predicted reduced accuracy (β = −0.00082, p = .007, confidence interval [−0.001, −0.0002]). The fixed effect of narrative type was not significant for either model (Model 1: β = 0.01, p = .27, confidence interval [−0.01, 0.04]; Model 2: β = 0.01, p = .27, confidence interval [−0.01, 0.04]). Fig. 3 displays these results.

Fig. 3.

The y-axis displays the accuracy (proportion correct) on the explicit prediction task. The x-axis displays PTSD Checklist (PCL-C) PTSD severity. The dark red dots and lines represent participants in the control group and the light blue dots and lines represent participants in the PTSD group. The shaded areas are 95% confidence intervals. Participants were more accurate when predictions were within-event versus across-event, and higher PTSD severity was associated with decreased accuracy. Participants in the PTSD group tended to make less accurate predictions than participants in the control group, but this effect was not significant.

3.3. Predictive looking at action task (PLAT)

To determine whether predictive looking differed between the groups and depended on whether there was an event boundary nearby, mixed-effects models were tested. Time spent looking at the contacted object during each 500 ms bin leading up to contact was the dependent variable. Two models were tested: one with group (PTSD vs. control) and the other with PTSD severity on the PCL-C as a fixed effect. The other fixed effects for both of these models were bin, group, narrative type, and boundary condition. Two-way interactions for bin by PTSD, boundary condition by PTSD, bin by boundary condition, and narrative type by PTSD were also included. Finally, the three-way interaction between PTSD, boundary condition, and bin was included. For both models, intercepts for subject, movie, and object were random effects.

For both the group (PTSD vs. control) model and the PTSD severity model, the model with the three-way interaction between PTSD group/severity, boundary condition, and bin explained significant additional variance compared to the model with all of the two-way interactions (group model: χ² = 18.55, df = 5, p = .002; PTSD severity model: χ² = 17.66, df = 5, p = .003). Therefore, the conditional effects for the two-way interactions and main effects are interpreted below. Non-significant results that are not reported below are reported in Section 3 of the supplemental materials, including Tables S2–S5. Fig. 4 displays the form of the three-way interaction for each model and Table 3 displays the results for the fixed effects for both models.

Fig. 4.

Predictive Looking at Action Task examining interactions with boundary condition. For both panels, the y-axis displays the time spent looking at the contacted object within each 500 ms time bin. This represents the extent of anticipatory looking participants engaged in during each time bin. The x-axes display the 3000 ms before object contact, broken up into six 500 ms bins. Error bars represent 95% confidence intervals. People with PTSD (Panel A) and people with greater PTSD severity (Panel B) showed reduced predictive looking in the 500 ms preceding object contact, especially for objects contacted near an event boundary. The colored lines represent participants who reported PTSD severity scores on the PTSD Checklist (PCL-C) within each of three severity bins (higher scores mean greater PTSD severity). These severity bins represent PTSD scores below the mean (38), up to one standard deviation above the mean (58) and two or more standard deviations above the mean. These bins were used only for the purposes of visualization; continuous PTSD severity scores were entered into the models.

Table 3.

Results from linear mixed effects models for group and PTSD severity.

Fixed Effect	Group Analysis		Severity Analysis
	F	p	F	p
Time Bin	4284.30	<.001	1364.58	<.001
Boundary Condition	7.00	0.008	10.81	0.001
PTSDa	1.50	0.220	4.11	0.044
Narrative Condition	9.10	0.002	13.98	<.001
Time Bin x PTSDa	14.00	<.001	21.66	<.001
Boundary Condition x PTSDa	0.70	0.410	4.96	0.02
Time Bin x Boundary Condition	30.00	<.001	17.25	<.001
PTSDa x Narrative Condition	6.40	0.110	6.78	0.009
Time Bin x Boundary Condition x PTSDa	3.70	0.002	3.53	0.003

^aNote: For the group analysis, PTSD refers to the fixed effect of group, and for the PTSD severity analysis, PTSD refers to the fixed effect of PTSD severity.

We did not have any a priori hypotheses about the direction of potential interactions between narrative condition, boundary condition, and bin, but the results of models that included these interactions are reported in Section 4 of the supplemental materials (Fig. S2 and Table S6). When symptom cluster (reexperiencing, avoidance, and increased arousal) were used in the models instead of total PTSD severity scores, the pattern of results was the same. Results from these models are reported in Section 5 of the supplemental materials (Tables S7–S9).

The results supported our hypothesis that people with PTSD would display worse prediction accuracy and that this difficulty would be exacerbated around event boundaries compared to within events. Post-hoc tests for these conditional effects found that the two-way interaction between group and boundary condition was only significant for the time bin at 500 ms before contact (F = 13.57, p < .001; See Table S2 for results for all time bins), suggesting that this difficulty in prediction ability was most evident right before object contact. While both people with PTSD and controls spent less time looking at the target object within events than around boundaries (PTSD: 10.8 ms, z = −4.04, p = < .001; Controls: 24.6 ms, z = −9.32, p < .001), people with PTSD spent less time than controls looking at the target object near boundaries (23.5 ms, z = 4.88, p < .001), though not within events (9.7 ms, z = 2.41, p = .07).

As hypothesized, similar results were also seen in the PTSD severity analysis: Higher PTSD severity was associated with less looking at the target object at 1000 ms and 500 ms before contact, both within events and near boundaries (1000 ms bin, within events: slope = −0.21, z = 2.09, p = .04; 500 ms bin, within events: slope = −0.38, z = −3.75, p < .001; 1000 ms bin, boundary: slope = −0.41, z = −3.39, p = < .001; 500 ms bin, boundary: slope = −0.77, z = −6.41, p < .001; See Table S3 for results for all time bins). In addition, for the 500 ms before contact, higher PTSD severity was associated with particularly impaired predictive looking around event boundaries compared to within events (slope difference = .40, z = 4.21, p = .002; see Table S4 for results for all time bins).

The results did not support our hypotheses that listening to a trauma narrative would negatively impact prediction accuracy for people with PTSD and people with higher PTSD severity scores (see Fig. 5). Although the two-way interaction between narrative type and group/PTSD severity was significant for both models (group model: F = 6.43, p = .01; severity model: F = 6.78, p = .009), further interrogation of the results suggested that controls increased their looking to the target during the trauma narrative compared to the positive narrative (−3.8 ms, z = −3.92, p < .001) and people with PTSD were not significantly impacted by listening to the trauma narrative (−0.3 ms, z = −0.34, p = .98). The results were similar for the PTSD severity model. Although there was a significant slope difference between the trauma narrative and positive narrative conditions (slope difference = 0.09, z = 2.60, p = .009), when this result was interrogated at specific values of PTSD severity (−1 SD, mean, +1 SD), we found that only people with lower PTSD severity scores displayed a performance difference between the trauma and positive narrative conditions, with better performance in the trauma narrative condition (PTSD severity = 18, −1 SD: 3.88 ms, z = −4.02, p < .001; PTSD severity = 37.9, mean: 2.10 ms, z = −3.08, p = .002; PTSD severity = 57.8, +1 SD: 0.34 ms, z = −0.35, p = .73).

Fig. 5.

Predictive Looking at Action Task examining interactions with narrative condition. The y-axes display the time spent looking at the contacted object within each 500 ms time bin. Error bars represent 95% confidence intervals. Panel A: Two-way interaction between narrative condition and group predicting time spent looking at the target object during the 3 s before contact. People in the control and PTSD groups spent similar amounts of time looking at the target object during the positive narrative condition. However, in the trauma narrative condition, people in the control group (brown) spent more time looking at the target object than people in the PTSD group (aqua). Panel B: Two-way interaction between narrative condition and PTSD Checklist (PCL-C) severity scores predicting time spent looking at the target object during the 3 s before contact. People with low to moderate PTSD severity scores spent more time looking at the target object during the positive (green) compared to the traumatic (dark purple) narrative condition, with no significant effect of narrative condition at higher PTSD severity scores.

In addition to these primary hypotheses, we also replicated two effects reported in Eisenberg et al. (2018). First, predictive looking increased continuously during the 3000 ms before contact (main effect for time bin for group analysis: F = 4284.30, p < .001; severity analysis: F = 1364.58, p < .001). Second, there was a cross-over interaction between time and boundary condition, such that controls spent less time looking at the target object near event boundaries vs. within events at 3000 ms before contact (8.4 ms, z = 3.18, p = .008) but more time looking at the target object near event boundaries vs. within events at 500 ms before contact (−24.63 ms, z = −9.32, p < .001). The same pattern was visible for the PTSD group, but the difference at 3000 ms before contact was not significant (3000 ms before contact: 3.15, z = 1.18, p = .64; 500 ms before contact: 10.82 ms, z = −4.04, p < .001; Fig. S1 displays the cross-over interaction and Table S5 shows all of the comparisons).

Though not an a priori hypothesis, we also investigated the relationship between the two prediction tasks to determine whether people’s performance on the explicit prediction task predicted their performance on the PLAT. Mixed-effects models predicting performance on the PLAT were created, with explicit prediction performance, boundary condition, group or PTSD severity, time bin, and narrative type as fixed effects. Subject was entered as a random effect. Models with only the main effects were compared against models that included all of the possible two-way and three-way interactions with explicit prediction performance. For both the group and PTSD severity analyses, the models that included the interaction terms did not explain additional significant variance in performance on the PLAT (group model: χ² = 0, df = 6, p = 1; PTSD severity model: χ² = 0, df = 6, p = 1). For both models, the significant main effects were explicit prediction performance (group model: F = 341.65, p < .001; PTSD severity model: F = 340.87, p < .001), time bin (group model: F = 2698.6, p < .001; PTSD severity model: F = 2698.61, p < .001), and boundary condition (group model: F = 35.09, p < .001; PTSD severity model: F = 35.02, p < .001). The main effects of group/PTSD severity (group model: F = 1.12, p = .29; PTSD severity model: F = 2.25, p = .14) and narrative type (group model: F = 2.52, p = .11; PTSD severity model: F = 2.52, p = .11) were not significant for either model. These results indicate that people who performed better on the explicit prediction task also performed better on the PLAT irrespective of group or any of the other variables.

4. Discussion

Events in modern life are often neutral in valence and rarely are truly dangerous. However, the results of this study suggest that people with PTSD have difficulty making adaptive predictions about future experiences even in innocuous settings. The frequent prediction errors people with PTSD experience when their predictions are incorrect may result in everyday experiences feeling subjectively less predictable, leading to more efforts to increase the predictability of everyday life, including increased avoidance and hypervigilance. This could result in a vicious cycle leading to even less subjective predictability, leading to increased symptoms, and so on.

In this study, participants completed two prediction tasks: an explicit prediction task and an implicit prediction task (PLAT). Consistent with our first hypothesis, people with greater PTSD severity made less accurate predictions in the explicit prediction task. For the PLAT, participants in the PTSD group spent significantly less time looking at the to-be-contacted object during the 500 ms before contact than the control group, and this effect was greater for the boundary condition than the within-event condition. Similarly, people with greater PTSD severity made fewer anticipatory looks to the to-be-contacted objects on the PLAT during the second before contact. This study is the first to demonstrate difficulty with prediction of near-future, neutral activity in people with PTSD.

In addition, consistent with our second hypothesis, event boundaries modulated prediction for both tasks. For the explicit prediction task, people in both groups were more accurate when making predictions within events compared to across events. For the PLAT, participants in both groups engaged in less anticipatory looking to the to-be-contacted object when contact occurred near an event boundary compared to within-event. This replicated the results found in Zacks et al. (2011). Moreover, there were significant cross-over two-way interactions between time bin and boundary condition: Participants looked at to-be-contacted objects earlier when contact occurred within-event, and later when contact occurred near an event boundary. A very similar result was reported by Eisenberg et al. (2018). As discussed there, this cross-over interaction could occur because when object contact occurs near an event boundary, participants make less accurate predictions about which object will be contacted. This makes them slower to look to the object that will be contacted−resulting in both reduced early looking and increased late looking. In other words, when viewers are predicting more accurately, the viewer’s eye may already have moved on to a subsequent prediction by the time the actor’s hand gets close to the object, but when viewers are predicting less well, their eyes may arrive at the target object more slowly.

There were also three-way interactions between time bin, boundary condition, and group/PTSD severity. Close to object contact, participants with PTSD spent less time looking at the to-be-contacted object than participants in the control group when contact occurred near an event boundary compared to within-event. A similar interaction was present for the model with PTSD severity, with participants who reported the most severe levels of PTSD displaying less anticipatory looking when contact occurred around an event boundary compared to within-event. These results suggest that the PTSD-related difference in anticipatory looking is only present right before object contact. Although we originally hypothesized that there would be main effects of group and PTSD severity on anticipatory looking behavior, this interaction does provide evidence that PTSD affects people’s ability to make accurate predictions about near future activity. This interpretation is strengthened by the results of the explicit prediction task, which also found reduced prediction accuracy for people with higher PTSD severity.

Participants’ own event boundaries were used to test the primary hypotheses for the PLAT because prediction performance was expected to be worse at the times that each participant experienced greater prediction error, i.e., at their own event boundaries. However, an exploratory analysis was conducted to determine whether the same pattern of predictive looking held when normative event boundaries were used for the PLAT instead (see Section 6 of the supplemental material, including Figs. S3 and S4 and Tables S10–S12). Overall, the pattern of results was similar, but the effects were weaker, as might be expected. For both the group and PTSD severity analyses, participants engaged in greater predictive looking within-event versus around event boundaries early on before object-contact. However, for the group analysis, there was no difference between the two groups in predictive looking around event boundaries. Only the participants with the most severe PTSD symptoms experienced prediction difficulty regardless of whether there was a normative event boundary nearby.

The narrative condition had the expected effect on mood and anxiety, with participants reporting worse mood and anxiety in the traumatic narrative condition compared to the positive narrative condition. In addition, the traumatic narrative condition had a stronger impact on mood for people in the PTSD group compared to controls. While mood and anxiety were less affected by the manipulation in later trials of the study, the order of the experimental tasks in this study were counter-balanced, meaning that this decrease in the manipulation’s effect over time was unlikely to explain this study’s findings. These results suggested that the narrative manipulation had the desired impact on affect and allowed for investigation of the impact of narrative on prediction performance.

Consistent with our third hypothesis, having participants listen to a narrative of their traumatic event before completing each task affected performance on the PLAT, though not on the explicit prediction task. However, contrary to our hypothesis, the traumatic narrative condition impacted people in the control group rather than people in the PTSD group. People in the control group spent more time looking at the target object during the 3 s before contact when they had just listened to a traumatic narrative. People in the PTSD group showed a different pattern, performing similarly in the positive and traumatic narrative conditions. The analysis based on PTSD severity revealed similar effects, with people with low to moderate PTSD displaying increased predictive looking for the traumatic versus the positive condition, with people with higher PTSD severity displaying similar performance across the two narrative conditions.

While these results were surprising, there are a few potential explanations. First, it is possible that for people without PTSD, listening to their traumatic narrative led to greater alertness and therefore better predictive looking. On the other hand, given that people with PTSD generally report higher levels of arousal even without specific triggers in their environment, it is possible that they were already at ceiling and that the traumatic narrative did not add to their already high level of arousal. These results are consistent with results from a study of event segmentation (Sherrill et al., 2018), which found that greater state anxiety was not associated with reduced segmentation agreement in a group of healthy undergraduates without severe PTSD. If this were the case, we would have expected better prediction performance in people with PTSD under the positive narrative condition, given that mood was higher and anxiety was lower for people with PTSD in the positive versus traumatic narrative conditions. However, prediction performance did not improve for people with PTSD in the positive versus the traumatic narrative condition, providing evidence against this type of ceiling effect explanation. Another potential explanation is that because people with PTSD often attempt to avoid thinking about their traumatic experiences, they may have engaged in numbing or suppressing of the negative affect from the trauma narrative manipulation, leading to less impact of the trauma narrative on prediction performance. People in the control group may have been less likely to engage in this type of avoidance behavior and may have been more affected by the negative affect elicited by the trauma narrative as a result. However, this explanation appears unlikely based on the pattern of mood ratings made by participants after listening to their traumatic and positive narratives, as mood and anxiety for both groups was negatively impacted by the traumatic narrative compared to the positive narrative and this effect was even stronger for mood in the PTSD group. Finally, the lack of a narrative effect at all in the explicit task may be due to this task being easier overall, meaning that performance may be harder to influence through interventions.

Although to our knowledge there are no previous studies investigating anticipatory processing of neutral, everyday activities in people with PTSD, previous studies have found differences in how people with PTSD anticipate negatively valanced stimuli. People with PTSD displayed a stronger startle response to unpredictable versus predictable aversive stimuli than people with generalized anxiety disorder and healthy controls (Grillon et al., 2009) and individual differences in hyperarousal symptoms were associated with decreased ventromedial prefrontal cortex activity (Grupe et al., 2016). Increased activation of other brain regions often implicated in PTSD has also been found during anticipation of unpredictable aversive stimuli (Brinkmann et al., 2017). These studies provide further evidence that people with PTSD experience the anticipation of future stimuli differently than healthy populations.

What, mechanistically, is responsible for impairments in predictive processing in people with PTSD? Though this study was not designed to directly discriminate between potential mechanisms, a reasonable hypothesis is that this difficulty is related to symptoms of increased arousal and hypervigilance. One of the diagnostic criteria for a PTSD diagnosis is hypervigilance, which can result from a fear that something dangerous will happen and a desire to prevent that dangerous thing from happening. It follows that hypervigilance can be conceptualized as frequent predictions that danger will occur, leading to behaviors aimed at preventing that danger from actually occurring. A similar conceptualization has also been suggested by Gagne et al. (2018). Interpreting the results of the current study using this framework suggests that people with PTSD performed worse on the prediction tasks because their hypervigilance led to more globally focused attention to the scene to ensure that nothing dangerous was about to happen, keeping them from attending to and accurately anticipating the neutral activity that was actually unfolding. If so, this phenomenon could have significant implications for the everyday functioning of people with PTSD, especially given that most everyday experiences are neutral in valence.

An important question for future research is which specific steps in the cascade of predictive processing are affected in PTSD. Effective prediction requires maintaining a task set, ongoing attention to that task (vigilance), accurate sensory and perceptual processing, the ability to form and maintain representations of anticipated events, and the ability to detect when these representations no longer match the current event. PTSD pathophysiology is not known to be associated with basic sensory or perceptual impairments (American Psychiatric Association, 2013), but further work is required to establish whether the impairments observed here result specifically from processes involved in forming and maintaining representations of anticipated activity, or from more general deficits in task set maintenance or vigilance. However, some features of the current task suggest that impairments in vigilance alone are unlikely to explain these effects. Throughout both tasks, participants kept their heads in a headrest, making it difficult to look anywhere other than the screen. In addition, participants knew that experimenters could tell whether they closed their eyes, and during breaks between tasks, experimenters reminded participants to keep their eyes open during tasks. Future studies should include explicit measures of task set maintenance and vigilance to ensure that deficits in these abilities cannot explain the current results.

In addition, there are some limitations of the narrative approach used here. Because state affect was only measured after participants listened to their narrative before completing each task, it is not possible to know how long the impacts of the narrative condition lasted. In addition, this study did not include a neutral narrative condition, meaning that it is not possible to know whether both positive and traumatic narrative conditions impacted predictive processing differently than if there were no narrative condition included. Future research that measures state affect after each task and that includes a neutral condition would further elucidate the impact of emotional processing on prediction performance in people with PTSD. Furthermore, this study did not assess trait levels of mood and anxiety, instead focusing on measuring changes in state mood and anxiety related to the narrative manipulation. However, the narrative manipulation did have an impact on mood and anxiety in both groups, suggesting that differences in trait mood and anxiety alone cannot fully explain these results. Assessing for trait levels of mood and anxiety in future research could help to clarify the impacts of state versus trait mood and anxiety on predictive processing in people with PTSD. It is also important to note that while the effects of narrative condition on predictive processing were significant, the effects were small (differences of under 7 ms). Additional research investigating the effects of other methods for simulating trauma-related triggers may yield larger effects on performance on both tasks.

Eye tracking research investigating hypervigilance to threat has found that people with PTSD experience sustained attention to threat-related stimuli but do not evidence increased threat detection (Lazarov et al., 2019). This sustained attention to threat-related stimuli could have an impact on predictive processing, and could vary depending on the threat level depicted in stimuli. Therefore, it is possible that sustained attention to threat-provoking or trauma-related stimuli would further reduce predictive processing ability in people with PTSD. In addition, people’s familiarity with and knowledge about certain settings and stimuli may impact hypervigilance and therefore predictive processing, suggesting that future studies would benefit from assessing familiarity and prior knowledge of the settings depicted in the stimuli in addition to including a range of neutral to threat-related stimuli.

This study was not designed to determine whether difficulties with predictive processing are unique to the PTSD clinical population. Because of the high level of comorbidity of anxiety and depressive disorders in the PTSD population (Spinhoven et al., 2014), participants were not excluded if they had current or past anxiety or depressive disorders. Due to this high degree of comorbidity, attempting to control for a secondary diagnosis of an anxiety and/or depressive disorder would likely have resulted in a high level of multicolinearity between this control variable and the predictor of interest (PTSD) in the model. Future studies could interrogate whether people with other anxiety and/or mood disorders also experience difficulty with predictive processing.

Understanding predictive processing in PTSD may also lead to insights into how existing gold standard interventions for PTSD, such as Cognitive Processing Therapy and Prolonged Exposure, work to reduce symptomatology in people with PTSD. It is possible that by helping people gain a more accurate and complete understanding of the context of their traumatic experiences, these treatments also help improve the accuracy of people’s predictions about what will happen in future experiences. Future research investigating predictive processing before and after people engage in PTSD treatment would provide evidence for this hypothesis and provide insights into the mechanisms driving the effectiveness of existing treatments.

5. Conclusion

The ability to adaptively anticipate what will happen in the near future is essential for successful everyday functioning. The current results indicate that people with PTSD have difficulty making accurate predictions about the near future during processing of neutral, everyday activities. This anticipatory processing difficulty may be associated with symptoms of hypervigilance due to a perception that everyday events are unpredictable. It is possible that existing interventions for PTSD work in part by helping people improve their predictive processing, and future research could be greatly beneficial for understanding the mechanisms by which existing interventions improve functional outcomes for people with PTSD.

Data availability

Raw eye tracking data will be made available on request. All other data is available in the Open Science Framework respository at https://osf.io/5g36r/?view_only=50eecee665b04517b6853c331368e60b.