The Temporal Course of Over-Generalized Conditioned Threat Expectancies in Posttraumatic Stress Disorder

Abstract

One key conditioning abnormality in posttraumatic stress disorder (PTSD) is heightened generalization of fear from a conditioned danger-cue (CS+) to similarly appearing safe stimuli. The present work represents the first effort to track the time-course of heightened generalization in PTSD with the prediction of heightened PTSD-related over-generalization in earlier but not later trials. This prediction derives from past discriminative fear-conditioning studies providing incidental evidence that over-generalization in PTSD may be reduced with sufficient learning trials. In the current study, we re-analyzed previously published conditioned fear-generalization data (Kaczkurkin et al., 2017) including combat veterans with PTSD (n = 15) or subthreshold PTSD (SubPTSD: n = 18), and trauma controls (TC: n = 19). This re-analysis aimed to identify the trial-by-trial course of group differences in generalized perceived risk across three classes of safe generalization stimuli (GSs) parametrically varying in similarity to a CS+ paired with shock. Those with PTSD and SubPTSD, relative to TC, displayed significantly elevated generalization to all GSs combined in early but not late generalization trials. Additionally, over-generalization in PTSD and SubPTSD persisted across trials to a greater extent for classes of GSs bearing higher resemblance to CS+. Such results suggest that PTSD-related over-generalization of conditioned threat expectancies can be reduced with sufficient exposure to unreinforced GSs and accentuate the importance of analyzing trial-by-trial changes when assessing over-generalization in clinical populations.

The landmark study by Watson and Rayner (1920) established the pathogenic potential of generalized conditioned fear: the process by which fear to a conditioned stimulus (CS+) paired with an aversive unconditioned stimulus (US) transfers to resembling, safe stimuli. In this work, Watson and Rayner showed that pairing a white rat (CS) with a loud startling noise (US) elicited a conditioned fear response that generalized to a variety of safe, furry objects in a toddler referred to as Little Albert. The excessive proliferation of conditioned fear to all things furry reported by Watson and Rayner launched interest in generalization as a mechanism by which adaptive conditioned fear to veridical threat transitions to maladaptive fear. Furthermore, this study was the first to link lab-based trauma to undue generalization, introducing over-generalization as a potential pathogenic outcome of ‘real world’ psychological trauma. Today, heightened generalization of conditioned fear is widely viewed as a core feature of posttraumatic stress disorder (PTSD; Ehlers & Clark, 2000; Foa, Steketee, & Rothbaum, 1989) through which fears of people, places, and things associated with trauma unduly extend to safe situations “resembling” the traumatic encounter (American Psychiatric Association, 2000, 2013). Over-generalization contributes to PTSD symptomatology by unnecessarily increasing the number of innocuous stimuli in the individual’s post-trauma environment that are capable of eliciting and maintaining trauma-related distress.

In support of the link between PTSD and heightened levels of generalization, a meta-analysis of 13 lab-based discriminative conditioning studies in PTSD found elevated acquisition of fear to conditioned safety-cues (CS−) bearing perceptual resemblance to CS+ among those with versus without PTSD (Duits et al., 2015), a finding consistent with over-generalization of fear from CS+ to CS− in PTSD patients. Furthermore, PTSD-related over-generalization of conditioned fear has recently been documented by studies applying systematic generalization methods to elicit generalization gradients: declines in conditioned responding as presented stimuli incrementally differentiate from CS+, with more shallow gradients indicative of over-generalization found among those with PTSD (Kaczkurkin et al., 2017; Lissek & Grillon, 2012; Morey et al., 2015). Similar effects of over-generalization have been found by generalization gradient studies in panic disorder (Lissek at al., 2010) and generalized anxiety disorder (Cha et al. 2014; Greenberg, Carlson, Cha, Hajcak, Mujica-Parodi, 2013; Lissek, Kaczkurkin et al., 2014, but also see Tinoco-Gonzalez et al., 2015), implicating generalized conditioned-fear as a transdiagnostic marker of anxiety- and trauma-related disorders.

To date, studies assessing over-generalization in clinical populations have tested patient-control differences in generalization after averaging generalization responses across trials at the subject level. While no studies tested the trial-by-trial course of disorder-related over-generalization, several past findings from studies without explicit interest in generalization are consistent with the notion that patient-control differences in generalization may reduce during later trials. For example, two discriminative fear-conditioning studies found enhanced fear to conditioned safety-cues (CS−) that perceptually resemble CS+ among those with PTSD (Grillon & Morgan, 1999) and panic disorder (Lissek et al., 2009) in early, but not late stages of conditioning. Similarly, the above mentioned meta-analysis of 13 conditioning studies in PTSD (Duits et al., 2015) found that among those with PTSD, enhanced fear to CS− bearing resemblance to CS+ was restricted to the acquisition training phase and did not emerge during the subsequent extinction phase. That is, PTSD patients displayed heightened fear reactivity to the first half (acquisition) but not the second half (extinction) of unreinforced CS− trials. Taken together, these findings suggest that over-generalization to safe stimuli resembling danger cues among those with PTSD may reduce to healthier levels given a sufficient number of learning trials. Unfortunately, no study to date has applied the generalization gradient methodology to assess the trial-by-trial course of over-generalized conditioned fear in PTSD.

In the present study, we aim to fill this gap by reanalyzing previously published data (Kaczkurkin et al., 2017) with nonparametric regression models to examine the within-session temporal course of group differences in generalized conditioned threat expectancy across military veterans with PTSD, Subthreshold PTSD (SubPTSD), and no PTSD (trauma controls: TC). Though Kaczkurkin and colleagues (2017) report effects of PTSD on generalization with both behavioral and fMRI measures, the current effort looks only at behavioral indices because behavioral data, but not fMRI data, have sufficient reliability for analyses at the individual trial-level. Behavioral data included trial-by-trial online ratings of shock-US expectancy to CS+ (10 trials), two CS− (10 trials each), and three generalization stimuli (GS: 10 trials each) that together form a continuum-of-similarity across CS+, GSs, and CS−. We expected evidence for elevated levels of generalization among those with PTSD and SubPTSD, relative to TC, in the earlier and middle trials of the learning record, but a lack of evidence for such group differences in generalization during later trials when similarly low levels of generalization across groups were anticipated. Additionally, we expected a lack of evidence for group effects in the first few trials when all three groups are initially learning the signal value of GSs. While this study involved a re-analysis of recently published data, the above hypotheses were a priori, and stemmed from our longstanding views on the likely time-course of PTSD-related overgeneralization, based on past findings of elevated fear to CS− in earlier but not later phases of conditioning among those with PTSD (Duits et al., 2015, Grillon & Morgan, 1999).

Methods

Participants

The study sample consisted of participants from a previous study (Kaczkurkin et al., 2017) for which trial-by-trial conditioning data were available for analysis. Participants were 61 male, United States combat veterans from Operation Iraqi Freedom and Operation Enduring Freedom. All veterans were assessed using the Clinician-Administered-PTSD-Scale-for-the-DSM-IV (CAPS: Blake et al., 1995). If a participant did not meet criteria for PTSD, they were placed into either SubPTSD (CAPS score: 20-39) or trauma control (TC) groups (CAPS score: 0-19) based on previous recommendations (Weathers, Keane, & Davidson, 2001). Psychiatric co-morbidities were assessed with the Structured Clinical Interview for the DSM-IV (SCID-I; First, Gibbon, Spitzer, & Williams, 2001).

Exclusion criteria included: a) history of Axis I psychiatric disorders before deployment, b) history of alcohol/substance abuse or dependence within the 6 months prior to study enrollment, c) use of nicotine or caffeine on the day of testing, d) current use of mood stabilizers, anti-psychotics, anti-parkinsonian medication, anti-hypertensives, anticonvulsants, and alpha/beta adrenergic agents, e) current use of illegal substances, f) current Axis I psychiatric disorder for trauma controls, g) significant suicidal ideation/intent/behavior, h) having a medical implant, device, or condition that is not MRI safe, and i) medical conditions that interfered with study objectives. If participants were taking any medication on an “as needed” basis (e.g., stimulants, pain medications, benzodiazepines, sleep medications), they were excluded from the study unless they were able to forgo taking the medication 12 hours prior to testing without causing unnecessary symptom aggravation or impaired performance on study tasks (see Table S1 more information on abstaining participants). Nine participants meeting exclusion criteria were omitted from analyses because they failed to learn the CS+/US contingency as indicated by perceived risk of shock to the conditioned safety-cue that was equal to or higher than the conditioned danger-cue. Final analyses included 15 PTSD, 18 SubPTSD, and 19 TC participants for whom sample characteristics are displayed in the Supplement (Table S2). The study was approved by the University of Minnesota and Minneapolis VA Medical Center IRBs and all subjects completed informed consent before participating. All participants were compensated for their time.

Generalization Task

Five rings of parametrically increasing size, and one “V-shaped” stimulus (see Figure 1) served as conditioned stimuli (CS+, CS−) and generalization stimuli (GSs). This stimulus set included one CS+ and two CS−. The first CS− was either the largest or smallest ring (depending on counterbalance condition)— referred to as the oCS−, and the second CS− was a “V” shaped stimulus—referred to as the vCS−. Though all subjects were conditioned with the same vCS−, the oCS− was the largest ring for 50% of subjects and the smallest ring for the remaining half. Subjects for whom the oCS− was the largest ring were conditioned with the smallest ring as CS+ and vice versa. The three intermediately-sized rings served as GSs (i.e., GS₁, GS₂, GS₃) and formed a continuum-of-size between the CS+ and oCS− with GS₃, GS₂, and GS₁ demarcating the GS with most to least similarity to the CS+ regardless of CS+ size. The vCS− was included to test the degree to which conditioned generalization accrues to all “ringed” stimuli following reinforcement of the ring-shaped CS+.

Figure 1.

Conditioning and generalization stimuli for counterbalancing (CB) groups A and B. Half of participants were assigned to CB group A and half to B. For both CB groups A and B, GS₃ consisted of the ring closest in size to CS+, with GS₂ and GS₁ further decreasing in similarity to CS+. Ring diameters in cm (and visual angles) from smallest to largest were: 6.63 (.93°), 8.02 (1.12°), 9.38 (1.31°), 10.98 (1.54°), and 12.46 (1.75°). vCS− = v-shaped conditioned safety cue; oCS− = ring-shaped conditioned safety cue; GS₁, GS₂ and GS₃ = 3 classes of generalization stimuli; CS+ = conditioned danger cue.

All CSs and GSs were presented for 4 s on a rear-projection viewing screen mounted at the foot of the scanner with a viewing distance of 6.71 feet (204.47 cm). Inter-trial-intervals for CSs and GSs were either 2.4 or 4.8 s, during which time participants focused their gaze on crosshairs in the center of the screen. The unconditioned stimulus (US) was a 100-ms electric shock (3-5mA) delivered to the right ankle. Prior to the start of the experiment, a sample shock procedure was performed during which participants received between 1-3 sample shocks, and a level of shock rated by participants as being ‘highly uncomfortable but not painful’ was established. Further details on the task design can be found in the Supplement.

Online behavioral ratings.

Throughout testing, a behavioral task developed to maintain visual gaze at the center of the visual field (Schwartz et al., 2005) was applied. This task consists of a string of colored crosshairs (blue, yellow, red, green, purple) presented serially for a duration of 800 ms each in a quasi-random order in the center of the viewing screen during the 4 s presentations of CSs/GSs (5 crosshairs per stimulus), as well as during ITI periods lasting 2.4 s (3 crosshairs) or 4.8 s (4 crosshairs). Participants were instructed to continuously monitor the stream of colored crosshairs and rate their perceived level of risk for shock as quickly as possible, following each red crosshair, using a three button, fiber optic, response pad (Lumina LP-404 by Cedrus, San Pedro, CA), where 0 = ‘no risk’, 1 = ‘moderate risk’, and 2 = ‘high risk’.

Design.

The generalization paradigm included three phases: 1) pre-acquisition— consisting of 20 trials of each stimulus type (CS+, GS₁, GS₂, GS₃, oCS−, vCS−) all presented in the absence of any shock US; 2) acquisition— including 15 CS+, 15 oCS−, and 15 vCS−, with 12 of 15 CS+ co-terminating with shock (80% reinforcement schedule as done in our previous studies: e.g., Lissek, Bradford et al., 2014); and 3) generalization test— including 20 trials of each stimulus type (unreinforced CS+, GS₁, GS₂, GS₃, oCS−, vCS−), and an additional 10 CS+ co-terminating with shock (33% reinforcement schedule) to prevent extinction of the conditioned response during the generalization sequence. Trials for all 3 phases of the study were arranged in quasi-random order such that no more than two stimuli of the same class occurred consecutively. An additional constraint for the generalization sequence was the arrangement of trials into 10 blocks of 13 trials (2 unreinforced CS+, 1 reinforced CS+, 2 oCS−, 2 vCS−, 2 GS₁, 2 GS₂, 2 GS₃) to ensure an even distribution of trial types throughout runs. Importantly, trials were presented in the same order for every participant, rendering data conducive to trial-by-trial analyses.

Procedure.

Participants were not instructed of the CS/US or GS/US contingencies but were told they might learn to predict the shock if they attended to the presented stimuli. Next, participants practiced using the button box to respond to red crosshairs appearing at the center of CSs and GSs. Shock electrodes were then attached and a shock workup procedure was completed, during which 3-4 sample shocks were administered, participants rated the aversiveness of shocks, and the experimenter adjusted shock intensity to achieve a level that was highly uncomfortable but not painful. Finally, participants were placed in the MRI magnet, and structural scans were acquired, followed by pre-acquisition, acquisition, and generalization phases of the study.

Data Analysis

Nonparametric regression.

The trial-by-trial time course of group effects on acquisition and generalization of perceived risk were analyzed with nonparametric mixed-effects regression models (e.g., Gu & Ma, 2005; Helwig, 2016) using a three-way (Group x Stimulus-type x Trial) smoothing spline analysis of variance (SSANOVA: Gu, 2013; Wabha, 1990). While parametric regression assumes an exact mathematical model of relationships between predictor variables and response variables and aims to test hypotheses about model parameters (e.g., Faraway, 2014), nonparametric regression provides data driven models of relationships among variables of interest (e.g., Gu, 2013; Helwig & Ruprecht, 2017). Because the mathematical function that best describes PTSD-related differences in rates of learning across generalization trials is unknown, a nonparametric approach was selected. Because SSANOVA veers from the typical parametric ANOVA models applied in the fear-conditioning literature, a more complete description of the SSANOVA approach is provided.

SSANOVA model.

Predictor variables entered into the SSANOVA model included: 1) trial (8 levels: trials 1-8), group (3 levels: PTSD, Sub-PTSD, TC), and stimulus-type (3 levels: vCS−, oCS−, CS+) for the acquisition analysis; and 2) trial (10 levels: trials 1-10), group (3 levels: PTSD, Sub-PTSD, TC), and stimulus-type (6 levels: vCS−, oCS−, GS₁, GS₂, GS₃, CS+) for the generalization analysis.

Typical linear mixed-effects models assume that the response changes as a linear function of each predictor, and the goal is to estimate the unknown parameters (i.e., the β_j coefficients) that describe the linear relationships. Instead of assuming linear relationships, the SSANOVA model assumes that the expected response changes as an unknown function, f(·), of the predictors, and the goal is to estimate the unknown function, f(·), from the data. Thus, our mixed-effects SSANOVA model has the form,

$$y_{its}=f(trial,grou{p}_i,stim)+u_i+{\epsilon }_{its}$$

where y_its is the observed risk rating for the i-th subject for the t-th trial and s-th stimulus, u_i denotes a random intercept (i.e., baseline risk appraisal) that is unique to each subject (with group specific variance terms for the present analysis), and ε_its denotes the model error term, which is assumed to be independent, identically distributed, and independent of u_i terms. The model mean function, f(·), can be decomposed as

$$f(trial,grou{p}_i,stim)=f_0+f_T(trial)+f_G(grou{p}_i)+f_S(stim)+f_{TG}(trial,grou{p}_i)+f_{TS}(trial,stim)+f_{GS}(grou{p}_i,stim)+f_{TGS}(trial,grou{p}_i,stim)$$

where f₀ is the intercept term, f_T is the main effect of trial, f_G is the main effect of group, f_S is the main effect of stimulus, f_TG is the trial by group interaction effect, f_TS is the trial by stimulus interaction effect, f_GS is the group by stimulus interaction effect, and f_TGS is the three-way interaction effect among trial, group, and stimulus. Of note,

$$f(trial,grou{p}_i,stim)={\widehat{y}}_{gts}$$

where ${\widehat{y}}_{\mathit{gts}}$ is the expected risk rating for any given trial, group, and stimulus combination.

The model was fit using the bigssp function in the “bigsplines” package (Helwig, 2018) in R (R Core Team, 2018), which estimates variance and smoothing parameters using the two-stage approach described in Helwig (2016). The two-stage approach estimates the variance parameters via restricted maximum likelihood estimation (Patterson & Thompson, 1971) and then estimates smoothing parameters via generalized cross-validation (Craven & Wahba, 1978), which reduces model overfitting. With our fitted SSANOVA model, we can obtain estimates of the model mean function, f(·), as well as standard errors of the estimates. Together, the estimates and standard errors can be used to form confidence intervals for expected risk appraisals at any combination of input predictors. This allows us to make inferences on the expected risk appraisals, or any linear combination of the expected risk appraisals, such as differences in expected risk appraisals (i.e., contrasts) between groups and/or stimuli.

Contrasts.

The SSANOVA results were used to form two types of contrasts (Helwig, Shorter, Ma, and Hsiao-Wecksler, 2016): stimulus contrasts and group-stimulus contrasts.

Stimulus contrasts.

A stimulus contrast takes the estimated risk appraisal for a given group, trial, and stimulus (e.g., PTSD group’s estimated risk appraisal at trial 1 for vCS−) and subtracts it from the estimated risk appraisal for the same group at the same trial, but for a different stimulus (e.g., PTSD group’s estimated risk appraisal at trial 1 for GS₃). The stimulus contrasts (SC) take on the form

$$SC=f(Trial,Group,StimulusA)-f(Trial,Group,StimulusB).$$

The vCS− was used as a non-circular control stimulus to compare against all other generalization stimuli (i.e., the vCS− was Stimulus B in the above equation for all generalization phase contrasts); in other words, the vCS− was used to control for fear unrelated to the circular shape. Increased generalization was defined by increases in stimulus contrasts between the three circular GSs (GS₃, GS₂, and/or GS₁), relative to the non-circular vCS−. Increases in discrimination between CSs was defined by increases in stimulus contrasts between the CS+, relative to the oCS− and non-circular vCS−. The goal of the stimulus contrast analyses was to model the degree to which each separate group generalized the GSs across the course of the generalization phase. We also examined how discrimination of the CS+ vs. vCS−, CS+ vs. oCS−, and oCS− versus vCS− changed throughout the course of the acquisition phase.

We computed stimulus-type contrasts using risk appraisal estimates from our fitted SSANOVA model to find contrast estimates of: 1) CS+ minus vCS−, CS+ minus oCS−, and oCS− minus vCS− at each of 8 trials for each of the 3 groups during the acquisition phase, and 2) all three GS estimates averaged minus vCS− and each separate stimulus type (CS+, GS₃, GS₂, GS₁, oCS−) minus vCS− at each of 10 trials for each of the 3 groups during the generalization phase. While all GSs averaged minus vCS− captured the time-course of overall generalization, each separate GS minus vCS− reflected levels of generalization for GSs with high (GS₃), moderate (GS₂), and low (GS₁) resemblance to CS+. We also found contrast estimates of CS+ minus oCS− and CS+ minus vCS− for the generalization phase.

Group-stimulus contrasts.

The group-stimulus contrasts are an extension of the stimulus contrasts. Group-stimulus contrasts tested group differences in trial by trial stimulus contrasts to compare levels of generalization and discriminative conditioning across each pair of groups. For these contrasts, over-generalization was defined as stronger generalization, captured by the size of stimulus contrasts, in the PTSD or SubPTSD group versus the TC group. The group-stimulus contrasts (GSC) take on the form

$$GSC=f(Trial,GroupX,StimulusA)-f(Trial,GroupY,StimulusB).$$

As an example of how group-stimulus contrast inferences work, suppose we want to find whether or not the PTSD group over-generalizes GS₃ at trial 1. We would find the group-stimulus contrast by first obtaining the stimulus contrast of GS₃ at trial 1 for the TC group and the stimulus contrast of GS₃ at trial 1 for the PTSD group. We would then subtract the TC stimulus-contrast estimate from the PTSD stimulus-contrast estimate to examine the degree to which the PTSD group generalized the GS₃, relative to the TC group, at trial 1. As we define over-generalization as an increase in generalization in the PTSD (and SubPTSD) group compared to the healthy, TC group, we always subtracted the TC group’s stimulus contrast estimates from the PTSD (and SubPTSD) group’s corresponding stimulus contrast estimates. We did the same when comparing discrimination differences between groups. To compare the PTSD versus the SubPTSD group in generalization and discrimination, we subtracted the SubPTSD group’s stimulus contrast estimates from the PTSD group’s stimulus contrast estimates. Group-stimulus contrasts were computed for each trial/stimulus contrast combination across the course of both acquisition and generalization phases.

Statistical inference.

We tested differences in generalization between groups using 95% Bayesian “confidence intervals” (CIs; Nychka, 1988; Wahba, 1983). Note that our approach is not fully Bayesian, given that we used cross-validation to select the smoothing parameter. As a result, the intervals that we use are typically referred to as Bayesian “confidence intervals” (in quotes) to reflect that they are standard (frequentist) CIs using the Bayesian estimate of the standard errors. To avoid confusions, we simply refer to them as CIs throughout our discussion of the results. Note that conclusions obtained from these 95% CIs directly correspond to conclusions drawn from p-values set at α = .05. More specifically, if zero is contained within the upper and lower bound 95% CI for a contrast estimate, then p > .05; on the other hand, if zero is not contained within the upper and lower bound 95% CI for a contrast estimate, then p < .05. Of note, we did not correct for multiple comparisons because the SSANOVA confidence intervals tend to have “across the function” coverage (Gu & Wahba, 1993; Nychka, 1988; Wahba, 1983). In other words, rather than computing separate CIs for each combination of input predictors, we form a 95% CI around the function estimate, $\widehat{f}(\cdot )$, which is expected to contain about 95% of the true function realizations across the combinations of input predictors. Effect sizes (i.e., size of differences between groups) as well as corresponding standard errors and 95% CIs for each trial/stimulus combination are reported in tables to provide a fuller picture of results. Additional statistical/mathematical details of the SSANOVA model and contrasts can be found in the online supplemental materials.

Results

Findings during the pre-acquisition and acquisition phases of the study are of secondary interest and can be found in the Supplement (text, Tables S3–S5, Figures S1–S2).

Generalization Phase

Predictor variables within our SSANOVA model accounted for approximately 40% of the variance in risk appraisals (R² = .40). A graph of fitted values (i.e., estimated risk appraisals) for each group, stimulus-type, and trial combination can be found in Figure S3 of the Supplement. Using the SSANOVA model, estimated stimulus contrasts averaged across trials for each group revealed increased risk appraisals to CS+ versus both oCS and CS+ versus vCS−, indicating that conditioning persisted during generalization for each group (statistical results reported in the Supplement). Additionally, risk appraisals, as predicted by the stimulus-type main effect term only, fell along a downward gradient of generalization, with increasing stimulus similarity to CS+ corresponding to increases in risk appraisals (see Figure S4).

Group-stimulus contrasts.

Conditioned stimuli.

Full trial-by-trial results (i.e., effect size estimates, associated 95% CIs, and statistical significance) for group differences in perceived risk to CS+ versus both oCS− and vCS− can be found in Table S6 and Figure S5 of the Supplement. As can be seen, significantly elevated perceived risk to CS+, relative to vCS− and oCS−, was found in PTSD versus TC at trials 2-10. Additionally, significant elevations in SubPTSD versus TC were found at trials 3, 4, 9, and 10 for CS+ versus vCS−, and at trials 2, 3, 9 and 10 for CS+ versus oCS−. Finally, those with PTSD, compared to SubPTSD, displayed significantly elevated perceptions of risk to CS+ versus oCS− and CS+ versus vCS− during trials 5-8 and 5-7, respectively.

Overall generalization.

Full trial-by-trial results for group differences in overall generalization (all GSs averaged) can be found in Table 1 and Figure 2. While each of three groups showed declining levels of overall generalization across the 10 trials, reflective of discrimination learning, larger group differences in generalization emerged at particular points in the learning record. Specifically, overall generalization was significantly elevated in PTSD versus TC at trials 1-7, but not at trials 8-10 (see Figure 2). That is, over-generalization of conditioned risk appraisals in PTSD emerged for early and middle trials but resolved to levels that were non-significantly different from trauma controls later in the learning record. The SubPTSD group showed a time-course of group differences similar to that of PTSD, with significantly elevated overall generalization in SubPTSD versus TC at trials 2-5, but not at trial 1 or trials 6-10. Thus, like PTSD, those with SubPTSD displayed over-generalization of conditioned risk appraisals in early and middle trials that was no longer significant toward the end of the learning record. Despite similarities across PTSD and SubPTSD, over-generalization was slower to resolve in PTSD versus SubPTSD, with significant over-generalization extending through trial 7 in PTSD but only through trial 5 in SubPTSD. No significant differences in overall generalization were found between PTSD and SubPTSD at any trial.

Table 1.

Trial-by-trial results for group-stimulus contrast statistics reflecting group differences in overall generalization.

	PTSD vs. TC					SubPTSD vs. TC					PTSD vs. SubPTSD
Trial	Est.	Std. Error	t	95% CIs	p	Est.	Std. Error	t	95% CIs	p	Est.	Std. Error	t	95% CIs	p
1	.299	.134	2.23	[.04, .56]	.026*	.235	.130	1.81	[−.02, .49]	.070	.064	.135	0.47	[−.20, .33]	.636
2	.324	.097	3.33	[.13, .51]	.001*	.238	.094	2.53	[.05, .42]	.011*	.086	.098	0.88	[−.11, .28]	.381
3	.340	.089	3.82	[.17, .51]	<.001*	.237	.086	2.75	[.07, .41]	.006*	.103	.090	1.15	[−.07, .28]	.252
4	.334	.090	3.71	[.16, .51]	<.001*	.219	.087	2.52	[.05, .39]	.012*	.115	.090	1.27	[−.06, .29]	.203
5	.297	.092	3.25	[.12, .48]	.001*	.180	.088	2.05	[.01, .35]	.041*	.118	.092	1.28	[−.06, .30]	.201
6	.244	.092	2.64	[.06, .43]	.008*	.135	.088	1.53	[−.04, .31]	.125	.109	.093	1.18	[−.07, .29]	.239
7	.191	.092	2.07	[.01, .37]	.038*	.107	.087	1.22	[−.06, .28]	.222	.085	.092	0.92	[−.10, .27]	.359
8	.161	.092	1.75	[−.02, .34]	.080	.111	.087	1.27	[−.06, .28]	.202	.051	.092	0.55	[−.13, .23]	.583
9	.162	.102	1.59	[−.04, .36]	.113	.137	.095	1.45	[−.05, .32]	.148	.024	.102	0.24	[−.18, .22]	.810
10	.181	.140	1.29	[−.09, .45]	.196	.170	.131	1.3	[−.09, .43]	.194	.011	.140	0.08	[−.26, .28]	.939

Overall generalization is defined by risk appraisals to all generalization stimuli (GS₁, GS₂, GS₃), averaged, minus risk appraisals to the V-shaped conditioned safety-cue (vCS−). Estimates (Est.) reflect unstandardized effects sizes for group differences in overall generalization, with more positive estimates indicating greater overall generalization in PTSD versus trauma controls (TC), subthreshold PTSD (SubPTSD) versus TC, or PTSD versus SubPTSD. t-values were obtained by taking the estimate divided by the standard error of the estimate (SE).

^*p < .05.

Figure 2.

Trial-by-trial levels of overall generalization in (A) PTSD and (B) subthreshold PTSD (SubPTSD). Overall generalization is defined by stimulus contrasts assessing differences in risk appraisals to all three generalization stimuli (GS₁, GS₂, GS₃) averaged versus risk appraisals to the V-shaped conditioned safety-cue (vCS−). The top row of graphs plots levels of overall generalization and standard error bars across trials for each group. The bottom row displays trial-by-trial results for group-stimulus contrasts reflecting (A) PTSD versus Trauma Controls (TC) and (B) SubPTSD versus TC differences in overall generalization, with higher values indicating greater overall generalization in PTSD or SubPTSD relative to TC. The shaded regions reflect 95% CIs, with lower-bound CIs that do not cross 0.00 indicating significant group effects.

Generalization to specific GSs.

Full trial-by-trial results for group differences in generalization to GSs with high (GS₃), moderate (GS₂), and low (GS₁) resemblance to the conditioned threat-cue (CS+) can be found in Tables 2–3 and Figure 3. Results for PTSD versus SubPTSD can be found in Table S7.

Table 2.

Trial-by-trial results for group-stimulus contrasts reflecting PTSD versus trauma control differences in generalization to stimuli with high (GS₃), medium (GS₂), and low (GS₁) resemblance to the conditioned danger-cue.

	GS3					GS2					GS1
Trial	Est.	Std. Error	t	95% CIs	p	Est.	Std. Error	t	95% CIs	p	Est.	Std. Error	t	95% CIs	p
1	.387	.158	2.44	[.08, .70]	.015*	.338	.153	2.20	[.04, .64]	.028*	.174	.145	1.20	[−.11, .46]	.230
2	.439	.115	3.83	[.21, .66]	<.001*	.362	.112	3.24	[.14, .58]	.001*	.172	.106	1.61	[−.04, .38]	.107
3	.482	.105	4.57	[.27, .69]	<.001*	.374	.102	3.69	[.18, .57]	<.001*	.163	.096	1.71	[−.02, .35]	.088
4	.498	.106	4.68	[.29, .71]	<.001*	.362	.102	3.53	[.16, .56]	<.001*	.141	.096	1.47	[−.05, .33]	.143
5	.476	.108	4.40	[.26, .69]	<.001*	.316	.105	3.02	[.11, .52]	.003*	.100	.099	1.02	[−.09, .29]	.309
6	.430	.109	3.93	[.22, .64]	<.001*	.252	.106	2.39	[.05, .46]	.017*	.050	.099	0.51	[−.14, .24]	.613
7	.381	.109	3.48	[.17, .59]	.001*	.187	.105	1.79	[−.02, .39]	.074	.005	.098	0.05	[−.19, .20]	.957
8	.352	.110	3.22	[.14, .57]	.001*	.149	.105	1.42	[−.06, .36]	.156	−.018	.098	−0.18	[−.21, .18]	.858
9	.354	.120	2.96	[.12, .59]	.003*	.145	.116	1.24	[−.08, .37]	.214	−.015	.110	−0.13	[−.23, .20]	.895
10	.374	.165	2.27	[.05, .70]	.023*	.162	.159	1.02	[−.15, .47]	.308	.005	.149	0.04	[−.29, .30]	.971

Generalization is defined as levels of risk to each generalization stimulus (GS) minus the V-shaped conditioned safety-cue (vCS−). Estimates (Est.) reflect unstandardized effects sizes for PTSD versus trauma control (TC) differences in generalization to GS₃, GS₂, and GS₁, with more positive estimates indicating greater generalization in PTSD versus TC. t-values were obtained by taking the estimate divided by the standard error of the estimate.

^*p < .05.

Table 3:

Trial-by-trial results for group-stimulus contrasts reflecting Subthreshold PTSD versus trauma control differences in generalization to stimuli with high (GS₃), medium (GS₂), and low (GS₁) resemblance to the conditioned danger-cue.

	GS3					GS2					GS1
Trial	Est.	Std. Error	t	95% CIs	p	Est.	Std. Error	t	95% CIs	p	Est.	Std. Error	t	95% CIs	p
1	.325	.153	2.11	[.02, .63]	.034*	.254	.149	1.70	[−.04, .55]	.088	.127	.141	0.90	[−.15, .40]	.366
2	.334	.111	3.01	[.12, .55]	.003*	.252	.108	2.32	[.04, .46]	.020*	.129	.103	1.25	[−.07, .33]	.213
3	.338	.102	3.30	[.14, .54]	.001*	.245	.099	2.48	[.05, .44]	.013*	.128	.093	1.38	[−.05, .31]	.167
4	.317	.103	3.08	[.12, .52]	.002*	.220	.099	2.22	[.03, .41]	.026*	.119	.093	1.28	[−.06, .30]	.201
5	.267	.104	2.58	[.06, .47]	.010*	.174	.101	1.73	[−.02, .37]	.083	.098	.095	1.03	[−.09, .28]	.305
6	.212	.104	2.04	[.01, .42]	.041*	.123	.101	1.22	[−.07, .32]	.221	.070	.095	0.73	[−.12, .26]	.465
7	.180	.103	1.74	[−.02, .38]	.082	.091	.100	0.91	[−.10, .29]	.362	.049	.094	0.52	[−.14, .23]	.603
8	.194	.103	1.88	[−.01, .40]	.060	.093	.099	0.93	[−.10, .29]	.350	.046	.094	0.49	[−.14, .23]	.624
9	.238	.112	2.13	[.02, .46]	.033*	.118	.109	1.08	[−.10, .33]	.281	.056	.104	0.53	[−.15, .26]	.594
10	.290	.154	1.88	[−.01, .59]	.060	.150	.150	1.00	[−.14, .44]	.318	.070	.142	0.49	[−.21, .35]	.621

Generalization is defined as levels of risk to each generalization stimulus (GS) minus the V-shaped conditioned safety-cue (vCS−). Estimates (Est.) reflect unstandardized effects sizes for subthreshold PTSD (SubPTSD) versus trauma control (TC) differences in generalization to GS₃, GS₂, and GS₁, with more positive estimates indicating greater generalization in SubPTSD versus TC. t-values were obtained by taking the estimate divided by the standard error of the estimate.

^*p < .05.

Figure 3.

(A) Trial-by-trial levels of generalization to stimuli with high (GS₃), medium (GS₂), and low (GS₁) resemblance to the conditioned danger-cue across groups. Generalization is assessed by stimulus contrasts comparing risk appraisals to each generalization stimulus versus the V-shaped conditioned safety-cue (vCS−). Standard error bars accompany the estimates. (B-C) Trial-by-trial group differences in generalization across (B) PTSD versus Trauma Controls (TC) and (C) Subthreshold PTSD (SubPTSD) versus TC for each of three generalization stimuli. Higher values indicate greater generalization in PTSD or SubPTSD relative to TC. The shaded regions reflect 95% CIs, with lower-bound CIs that do not cross 0.00 indicating significant group effects.

GS₃.

Those with PTSD, relative to TC, displayed significantly elevated generalization to GS₃ that endured across all 10 trials, whereas significantly elevated generalization in those with SubPTSD, relative to TC, was limited to trials 1-6 and 9. Additionally, elevated generalization to GS₃ was observed in PTSD versus SubPTSD at trial 6 (see Table S7). Such findings indicate that over-generalization to the GS most closely resembling CS+ was both more persistent across trials and somewhat more robust in PTSD versus SubPTSD.

GS₂.

Those with PTSD, relative to TC, displayed significantly elevated generalization to GS₂ at trials 1-6, but not trials 7-10. Additionally, those with SubPTSD, relative to TC, displayed significantly elevated generalization to GS₂ at trials 2-4, but not trials 1 or 5-10. This pattern of results reflects greater evidence of over-generalization to the GS with moderate resemblance to CS+ among those with PTSD and SubPTSD toward the beginning of trials, with less evidence of over-generalization at later trials. Though group effects for PTSD and SubPTSD at GS₂ were characterized by a similar time-course, elevated generalization to GS₂ was slower to resolve for PTSD versus SubPTSD, with significant over-generalization extending through trial 6 in PTSD but only through trial 4 in SubPTSD. Levels of generalization to GS₂ in PTSD versus SubPTSD did not differ at any trial (see Table S7).

GS₁.

No significant group differences in generalization to GS₁ emerged at any of the trials for PTSD versus TC, SubPTSD versus TC, or PTSD versus SubPTSD (Figure 3, Table S7).

Discussion

The present study represents the first effort to model the trial-by-trial course of PTSD-related over-generalization of threat expectancies to test the prediction that the expression of such over-generalization among those with PTSD can be reduced with sufficient exposure to unreinforced generalization stimuli (GSs). Results largely support this hypothesis, with greater overall generalization among those with PTSD versus trauma controls (TC) found in beginning and middle trials but not toward the end of the learning record. Those with subthreshold PTSD (SubPTSD) showed a similar temporal course, with increases in overall generalization, relative to TC, found in the first but not second half of the generalization sequence. Despite these similarities across PTSD and SubPTSD, heightened overall generalization persisted further into the learning record in PTSD versus SubPTSD suggesting that those with greater PTSD symptomatology require more learning trials to achieve levels of generalization that are more similar to those of healthy TCs.

Additionally, results for individual classes of GSs revealed unique time-courses for PTSD-related over-generalization to GSs with high (GS₃), moderate (GS₂), and low (GS₁) resemblance to the conditioned threat-cue (CS+). In PTSD, significant over-generalization persisted through all 10 trials for GS₃, but was not found toward the end of the learning record for GS₂, and was not present at any of 10 trials for GS₁. That overgeneralization to GS₃ in the PTSD group endured across all trials is contrary to predictions, and future work is needed to determine whether such over-generalization might resolve to levels more similar to TCs with additional GS₃ learning trials. In SubPTSD, significant over-generalization to GS₃ and GS₂ was no longer present by the middle and end of the learning record, respectively, and was not present at any trial for GS₁. This pattern of findings across PTSD and SubPTSD suggests that PTSD-related over-generalization is more persistent when evoked by stimuli with higher resemblance to a CS+ (GS₃), declines with sufficient trials when evoked by stimuli with moderate CS+ resemblance (GS2), and is not present at any trial in response to stimuli with low resemblance to CS+ (GS₁). Though similar patterns of results were found for PTSD and SubPTSD, over-generalization to safe stimuli bearing high (GS₃) and moderate (GS₂) resemblance to CS+ among those with SubPTSD versus PTSD required fewer trials before reducing to levels that were non-significantly different from TCs.

Findings are generally consistent with the notion that expression of over-generalized conditioned fear can be reduced with sufficient exposure to unreinforced generalization stimuli. While this is the first study to demonstrate reductions in PTSD-related over-generalization with repeated learning trials, several past fear-conditioning studies in anxiety and trauma-related disorders yield relatable findings. Specifically, elevated fear responding to safety cues resembling CS+ among those with PTSD and panic disorder have been shown to resolve during later stages of conditioning (Grillon & Morgan, 1999; Lissek et al., 2009; but see Jovanovic et al., 2010; Orr et al., 2000). Additionally, meta-analytic findings reflecting lab-based conditioning results across multiple case-control studies in PTSD found heightened fear reactivity to safety cues (CS−) resembling danger cues (CS+) among those with PTSD during the first half (acquisition training) but not the second half (extinction test) of unreinforced CS− trials (Duits et al., 2015). Finally, studies assessing block-by-block extinction of conditioned fear in traumatized samples have found elevated fear to (unreinforced) CS+ during early and middle, but not later extinction trials among those with PTSD (Fani et al., 2012; Norrholm et al., 2011; Norrholm et al., 2013, but see Orr et al., 2000) and trauma survivors with more severe re-experiencing symptoms (Norrholm et al., 2015). That is, PTSD symptoms were associated with slowed rates of safety learning to unreinforced CS+ presentations. Taken together, current and past findings suggest that those with PTSD are able to learn the safety value of both safe GSs and CS+ no longer associated with aversive outcomes, but doing so requires increased exposure to unreinforced GSs and CS+ relative to healthy individuals.

Prediction Error as a Putative Mechanism for Prolonged Over-Generalization in PTSD

Prominent theories of classical conditioning (Pearce & Hall 1980; Rescorla & Wagner 1972) implicate prediction error as a key determinant of the associative strength between a conditioned stimulus (CS) and an unconditioned stimulus (US). Through classical conditioning, the CS comes to signal the US, and subsequent presentations of the CS in the absence of the US create a prediction error, or a discrepancy between what was expected and what occurred. Such prediction errors promote learning by updating expectations of the US in the presence of the CS to increase the match between the expected and actual US outcome. In the current study, GSs elicit expectations of the US as a function of their perceptual similarity to CS+. This is particularly true for the first GS trial when participants have yet to experience GSs in the absence of the shock-US. The non-occurrence of shock during the first GS trial should thus elicit a prediction error, leading to a reduced expectancy of shock during the second GS trial. A similar effect should occur following each subsequent non-reinforced GS trial, leading to incremental decreases in perceived risk of shock across GS trials of the kind found in the current study (e.g., Figure 2). The heightened maintenance of shock expectancy across unreinforced GS trials found among those with PTSD and SubPTSD versus TC may therefore reflect a PTSD-related deficit in the efficient use of prediction errors to update GS-US associations.

Treatment Implications

The link between PTSD and slowed reductions in generalization of perceived risk prescribes a therapeutic approach aimed at reducing expectations of harm elicited by innocuous encounters that resemble aspects of the traumatic event (i.e., generalization stimuli). This could be achieved through in vivo and imaginal exposures to a fear hierarchy of trauma cues, with exposures to stimuli resembling trauma cues added at each level of the hierarchy. Importantly, such exposures should aim to maximize prediction error by providing patients repeated contact with trauma cues and generalization stimuli in the absence of feared outcomes, while directing patients’ attention to the resulting violation of expectations (see Craske, Hermans, & Vervliet, 2018). Additionally, generalization stimuli at each level of the fear hierarchy might best elicit prediction errors if they are presented randomly to patients, rather than according to their degree of similarity to the CS+, at each level of the fear hierarchy (Scheveneels, Boddez, Vervliet, and Hermans, 2019).

Take, for example, an individual with combat-related PTSD who fears safe, roadside objects in their post-deployment environment that resemble roadside bombs encountered during combat. This patient may benefit from randomly ordered exposures to different kinds of safe roadside objects, which vary in similarity to the encountered roadside bombs, in the absence of the predicted harm. The resulting prediction error should serve to disconfirm expectancies for threatening outcomes in the presence of safe roadside objects. In order to direct attention to exposure-elicited prediction errors, this patient would be asked to articulate what was predicted and what occurred, and the degree of “surprise” experienced following each exposure.

These kinds of exposure-based prediction errors should be repeatedly elicited until erroneous expectancies for aversive outcomes are reduced to a minimum, an approach shown to create more robust safety learning than standard exposure techniques aimed at fear-reduction (e.g., Deacon et al., 2013; Salkovskis, Hackmann, Wells, Gelder, & Clark, 2007). Present results suggest that reductions in aversive expectancies for safe stimulus events resembling features of the trauma should be achievable in those with PTSD given a sufficient number of exposures, and that protracted exposure regimens may be needed to reduce threat expectancies to innocuous stimuli with high resemblance to features of the trauma. Of note, exposures to trauma cues themselves may reduce threat-related responding to resembling stimuli, as lab-based extinction to a CS+ has been found to decrease threat expectancies to generalization stimuli (Vervliet, Vansteenwegen, Eelen, 2004). Thus, including exposures to trauma cues in generalization-focused exposure may serve to decrease threat expectancies to both trauma cues and resembling stimuli.

Limitations

Present results were derived from a fairly small sample of combat veterans. Though the applied SSANOVA protects against problems of overfitting associated with small sample sizes, the heterogeneity of symptoms across individuals diagnosed with PTSD may reduce replicability of results across different samples, particularly when sample sizes are small. This brings the generalizability of current results into question, which should be addressed through replication studies with larger Ns. Our small sample size may also have resulted in underpowered statistical tests. Specifically, non-significant group differences in generalization during later trials may have resulted from insufficiently powered analyses rather than the absence of true group effects. Non-significant group effects on generalization in the present study thus reflect tentative evidence for the absence of PTSD-related elevations in generalization and await replication with larger samples. Additionally, our small sample size accentuates the importance of attending to effect sizes and their corresponding confidence intervals when interpreting current findings.

An additional weakness of the present study is the inability to discern whether prolonged generalization across trials among those with PTSD reflects a premorbid risk factor for PTSD, or the PTSD ‘disease process’ itself. Longitudinal work in at risk samples is needed to assess for this generalization abnormality both before and after psychological trauma. Such work could clarify whether prolonged generalization of innocuous stimuli can be used as a prospective predictor of risk for PTSD or a diagnostic indicator and treatment target for the disorder.

Conclusions

Current findings suggest that heightened generalization of threat expectancies in PTSD and subthreshold PTSD can be reduced to levels closer to those displayed by healthy trauma controls with sufficient learning trials, and that more learning trials are required to reduce PTSD-related over-generalization of threat expectancies to safe stimuli bearing higher resemblance to the conditioned danger-cue. Such results support the use of extended courses of exposure therapy in PTSD that maximize violations of threat-related expectancies for safe stimulus-events resembling the traumatic encounter. Finally, findings highlight the importance of assessing the trial-by-trial course of generalization when testing for over-generalization in anxiety- and trauma-related disorders.

Highlights for Hammell et al.

• Threat overgeneralization in PTSD was found in early but not late generalization trials.

• Reductions in PTSD-related overgeneralization to stimuli more similar to the danger cue required more generalization trials.

• Overgeneralized conditioned fear in PTSD can be reduced with sufficient generalization trials.

Abbreviations:

CS+

conditioned danger-cue

CS−

conditioned safety-cue

US

unconditioned stimulus

GS

generalization stimulus

TC

trauma controls

SubPTSD

subthreshold PTSD