Mechanisms of Change in Exposure Therapy for Anxiety and Related Disorders: A Research Agenda

Abstract

Anxiety and related disorders are a significant public health burden with rising prevalence in the wake of the COVID-19 pandemic. As demand for effective anxiety treatment increases, so too does the need for strategies to bolster treatment outcomes. Research on the mechanisms of exposure therapy, the frontline behavioral treatment, will be critically important for optimizing clinical outcomes. We outline an initial agenda for future research on the mechanisms of change of exposure therapy, developed in collaboration with a large international team of researchers through the Exposure Therapy Consortium (ETC). Key questions and recommendations for future research focus on five priority areas: conceptualization, measurement, study design/analysis, and individual/contextual differences. Rising to the challenge of addressing these questions will require coordinated action and availability of centralized tools that can be used across trials, settings, and research groups.

As the public health impact of anxiety and its disorders continues to rise (US Preventive Services Task Force, 2020), so does the demand for intervention science. Among important aims is improving the efficacy and uptake of evidence-based treatments for these disorders. To this end, we established the Exposure Therapy Consortium (ETC; https://exposure.la.utexas.edu/), an international collaboration among academic investigators that is designed to allow investigations of exposure therapy mechanisms and outcomes to occur at scale and at relatively low cost/effort (Smits et al., in press). The scale of the ETC collaboration and potential for rapid replication of findings across sites can provide infrastructure to address research questions relevant to this first-line intervention more quickly and with more confidence.

Recognizing that mechanistic research facilitates the delivery and outcomes of interventions, we have begun to develop an agenda for research on exposure therapy mechanisms of change. The process for developing this agenda involved a series of meetings organized and attended by a subset of this paper’s authors (KB, AP, MWO, JAJS) which focused on review of empirical papers and relevant theoretical accounts. Keeping in mind clinical application and acknowledging that agendas like these cannot be complete and will evolve over time, we generated a set of key questions assigned to one of four priority areas: (1) conceptualization, (2) measurement; (3) study design and data analysis; and (4) individual and contextual differences. We then reached out to colleagues who bring relevant experience and expertise and asked them to help shape this agenda. We describe this initial agenda, hoping that it can spur initiatives for larger-scale collaborative research. We provide a brief review of extant literature for background and justification for the proposed priority areas and key research questions.

Exposure Therapy Efficacy

Exposure therapy involves repeated confrontation to feared cues or contexts without engaging in safety behaviors (i.e., avoidance, rituals). Feared cues and contexts can be external (e.g., agoraphobic situations, people, animals) or internal (e.g., emotions, bodily sensations, thoughts, images, memories) and exposure practice can vary in modality (e.g., in vivo, imaginal, virtual reality, interoceptive). Likewise, safety and defensive behaviors vary substantially (e.g., in relation to threat imminence, Hamm, 2020). Exposure is a primary component of established cognitive behavioral interventions for anxiety, obsessive-compulsive, and trauma and stressor-related disorders (Abramowitz, Deacon, and Whiteside, 2019). Exposure-based treatments have consistently fared well in clinical trials, outperforming placebo and active psychotherapy controls, and rivaling or outperforming pharmacological interventions (Carpenter et al., 2018).

The observation that a significant minority of patients receiving exposure-based treatment do not respond, remain symptomatic, or experience return of fear has promoted research on combination and augmentation strategies. Approaches to developing and testing these strategies have varied and included adding a single anxiolytic strategy (e.g., benzodiazepine) and combining exposure therapy with strategies to further engage putative mechanisms (e.g., cognitive enhancers). These efforts have yielded mixed successes (Tuerk, 2014) and often have seen initial promising findings weaken or disappear in follow-up studies (Rosenfield et al., 2019). One reaction to this pattern is increasing emphasis on mechanisms of change, or renewed focus on understanding how exposure works. Exposure therapy benefits greatly from a rich body of experimental work with animals and humans (McNally, 2007), and studies have increasingly focused on translating mechanistic knowledge from these settings into clinical treatment. This has produced important advances, but gaps remain and efforts to leverage mechanistic knowledge have not realized full potential for improving clinical outcomes.

1. Conceptual Considerations

In general, a mechanism of change can be defined as the processes through which treatment (e.g., exposure) produces the change (see Kazdin, 2007). Theories of the maintenance and recovery of fear, anxiety and their disorders have identified multiple potential mechanisms of exposure therapy efficacy (e.g., emotional processing theory [EPT]; (Foa and Kozak, 1986), inhibitory learning model [IL]; (Craske et al., 2008), acceptance and commitment model; (Twohig et al., 2015), self-efficacy theory; Bandura, 1988). Table 1 provides brief descriptions of common candidate mechanisms in research to date (see Knowles and Tolin, 2022, for a recent review). Keeping with the aims of this paper, we will not provide a comprehensive review of these theories and mechanisms, but instead share some conceptual observations that can provide direction for building upon earlier work. Table 2 provides an overview of identified challenges related to this and other priority areas, along with related recommendations for a coordinated research agenda and action steps for individual researchers.

Table 1.

Candidate mechanisms of exposure therapy

Candidate Mechanism	Brief description	Clinical example
Fear activation/Emotional engagement	Elevation of distress, fear, or anxiety; typically at the start or peak of an individual exposure exercise	Initial exposure to a phobic stimulus (e.g., a spider, bodily sensations, thoughts, images, trauma memories) results in elevation of fear
Counter-conditioning	Pairing a fear-eliciting stimulus with a positive stimulus to replace the fear response with a positive response	Receiving sweets whenever a phobic stimulus is presented
Within-session extinction	Decline of fear response within a single exposure exercise (often referred to as within-session “habituation”)	A visit to a grocery store (e.g., agoraphobia) is continued until the patient experiences a decrease of physiological and/or subjective fear
Between-session extinction	Decline of peak fear response across multiple exposure exercises and/or sessions (often referred to as between-session “habituation”)	Repeated exposure to the same grocery store until peak fear levels start to decline
Change in maladaptive fear cognition	Learning that a feared stimulus or situation is safer than previously believed	The visit to the grocery store is continued until the patient learns that they will not experience a heart attack
Expectancy violation	Exposure exercises should violate threat expectancy (anticipation of a specific threat) through learning that the threat does not actually occur during exposure	A patient expects that a panic attack will result in a heart attack during exposure practice in a grocery store, and experiences “surprise” when this does not actually occur
Expectancy change	Reduced threat expectancy across multiple exposure exercises and/or sessions	After exposure, a patient no longer expects that a heart attack will occur when having a panic attack in a grocery store
Distress tolerance	Repeated exposure results in higher distress tolerance (capacity to experience and withstand negative psychological and physiological experiences)	Repeated exposure to an aversive stimulus such as shortness of breath (in interoceptive exposure for panic sensations) increases a patient’s ability to tolerate the discomfort associated with those sensations
Instrumental learning	Exposure practice changes behavioral responses from avoidance to approach behavior through positive and negative reinforcement	During exposure practice a patient approaches a grocery store, which is followed by social reinforcement, feelings of pride, and/or absence of a feared outcome
Experiential avoidance	Exposure reduces avoidance/escape of aversive internal states and associated contexts, thus facilitating a more flexible and approach-oriented stance	Following exposure to crowds, a patient is better able to attend social events (e.g., parties), thus better able to discern the behavioral choices they have in such situations, rather than simply avoid them
Self-efficacy	Exposure strengthens specific beliefs related to the ability to cope with anxiety-provoking stimuli	After public speaking exposure practice, a patient realizes they are more capable of giving a speech and coping with distress during the speech than they had imagined
Meta-cognition	Exposure weakens beliefs about the benefits of dysfunctional strategies (e.g., repetitive negative thinking) to prevent threats, and also weakens beliefs about repetitive negative thinking being uncontrollable and dangerous	A patient believes that their worry helps them to be prepared and to prevent bad things from happening, but then worries that their habitual worrying itself is harmful and uncontrollable; these beliefs shift over time with exposure practice
Therapeutic relationship	Conducting therapist-guided exposure strengthens the therapeutic relationship, in turn improving treatment outcome	By repeatedly doing exposures together, and having the therapist model them first, the patient feels more connected and trusting toward the therapist

Table 2.

Priority areas for research on mechanisms of change in exposure therapy

Priority Area	Current Challenges	Research Agenda Areas of focus for the field	Action Steps Concrete actions for individual researchers
1. Conceptual Considerations
	• Focus on theoretical differences obscures similarities across overlapping constructs • Different findings likely arise from emphasis on different stages (e.g., initial learning, consolidation, retrieval) or aspects (e.g., cognitive, emotional, behavioral) of the same underlying processes • The distinctions between mechanisms of change versus a) processes that moderate or facilitate them and b) read-out measures that signal them are often unclear • There may be multiple read-out measures for the same mechanism, which could differ across individuals	• Work toward integrated mechanistic models that account for overlapping constructs and distinguish mechanisms versus their facilitators/moderators and read-out measures • As described in 2.1: development of consistent terminology, definitions and methods, test construct validity against conceptually overlapping measures • Examine fine-grained temporal changes of constructs in relation to one another	• Include substantive discussion of conceptual overlap in all manuscripts (e.g., implications for study design, interpreting results) • To the fullest extent possible, measure an array of distinct, plausible mechanisms at multiple timepoints (see also 3.4) in each study • Design studies with the understanding that mechanisms and strategies that engage them (facilitators) may not always match as theorized • Collaborate with investigators who have expertise in “competing” theories
2. Measurement
2.1 Psychometric properties of mechanistic measures	• The most commonly used mechanistic measures have very limited psychometric support • Construct definitions, measurement procedures, and calculation methods are highly variable, making it difficult to compare findings across studies • Measurement error from multiple sources is likely and can greatly reduce power	• Work toward consistent operational definitions for mechanistic constructs • Establish reliability and construct validity of mechanistic measures • Enhance the ontological study of extinction with investigations of relationships between measures and the construct of extinction learning • Focus efforts on practical measures to support forward translation • Determine optimal measurement timing and calculation methods • Establish best practices for reducing or accounting for measurement error introduced by therapy content	• Include psychometric data (reliability, correlations with other measures, predictive validity) for mechanistic measures in publications and/or public repositories • Create and publicly share detailed, reproducible, written procedures related to mechanism measurement (e.g., SOPs), including rater training/reliability procedures • Deploy a priori and/or post-hoc methods to reduce the impact of data pollution (see 2.1) • Increase measurement frequency to reduce error • Follow recommendations for the reporting of mechanistic behavioral research (Birk et al., 2023)
2.2 Multimodal measurement	• Self-report, physiological, and behavioral measures typically do not converge; degree of convergence may vary by measure, context, and individual differences (e.g., age, insight)	• Design studies to explicitly test measure convergence and predictive validity • Examine individual and contextual factors that influence measure convergence and predictive validity • Establish best practices for multimodal measurement that minimizes burden with maximum potency	• Incorporate multimodal measurement into clinical trials • Include psychometric data (reliability, correlations with other measures, predictive validity) for multimodal measures in publications and/or public repositories
2.3 Psychometric properties of laboratory models	• Commonly used stimuli have weak face validity • Recent fear acquisition used in lab models may differ from longstanding, complex fear responses associated with clinical symptoms • Limited data to support predictive validity with treatment outcomes	• Develop and test more “lifelike” stimuli • Develop and test models that incorporate avoidance behavior • Conduct larger tests of predictive validity in clinical trials • Examine subgroup differences in validity (e.g., based on level of impairment, symptom duration, or exposure to chronic stressors)	• Develop close collaborations between clinical researchers and experimental researchers; doing so early and often will enhance validity of lab models
2.4 Advances from models with non-human animals	• Early translational work with animals can inform novel approaches by identifying promising windows for measurement/intervention, novel mechanisms, and new measurement/treatment approaches	• Innovative paradigms with non-human animals should continue to inform measurement strategies and initial trials with humans • Human clinical studies are needed to examine promising interventions that target promising mechanisms from animal studies (e.g., retrieval-extinction targeting reconsolidation blockade	• Develop close collaborations between non-human animal researchers and clinical researchers; do so early and often to enhance validity of animal models and speed translation into clinical trials
3. Study Design and Analysis
3.1 Experiment and clinical trial design	• Laboratory experiments that manipulate mechanism using brief exposure protocols are efficient and well-suited for isolating mechanisms of interest • Full clinical trials are needed to test mechanisms in conjunction with efficacy and examine generalizability of earlier translational findings (e.g., to treatment-seeking clinical populations, multi-session exposure) • Limited evidence is available about mechanism-relevant strategies that may have been used in original exposure efficacy trials • It is difficult to design trials that compare mechanistic theories given conceptual overlap	• Laboratory experiments should continue to be used for initial mechanism tests and to identify the most promising strategies for engaging mechanisms • Rigorously designed clinical trials should specify, test, and confirm mechanisms and clinical efficacy before making recommendations for clinical practice • Secondary analysis of exposure delivery in past trials is needed to inform design of “exposure as usual” comparison conditions for future trials • Non-traditional (e.g., full factorial in MOST framework) and adaptive designs (e.g., SMART) can be used to test mechanisms and tailored treatment decisions related to mechanisms • Innovative single-subject designs (e.g., multiple baseline, alternating treatment, knock-out) should be used to isolate the contribution of specific exposure strategies to mechanistic change (see also 3.5)	• Avoid making recommendations for practice change without rigorous clinical efficacy data (e.g., from masked RCTs with clinical, treatment-seeking populations receiving a full course of treatment) • Share detailed treatment manuals, fidelity measures, and related therapist training procedures in supplemental materials and/or public repositories • Design trials to test promising strategies under an integrated theoretical framework, rather than to compare models against one another (see also section 1)
3.2 Measuring and Accounting for Exposure Delivery Variation	• Exposure delivery approach and dose varies widely, across and even within individual studies; measuring this is critical for understanding proximal “triggers” of mechanistic change, interpreting the validity of mechanistic measures (see 2.1), and statistical modeling • Adherence/fidelity measures are commonly used in clinical trials but rarely have psychometric support, lack predictive value, and fail to capture nuances relevant for mechanism	• Develop and test detailed measures of exposure dose as well as specific aspects of setup/selection, delivery, and post-processing; these should use clear operational definitions for replicability • Studies of exposure in group format should incorporate multi-level modeling and other novel approaches (e.g., social relation, one-with-many modeling) to disentangle group-level dynamics • Trials should incorporate detailed measurement of homework exposure and deploy novel methods (e.g., CACE) to account for individual differences in homework completion	• Report detailed descriptions of methods in manuscripts or extended supplementary materials • Create and publicly share detailed treatment manuals • Deploy detailed measures of exposure dose and delivery in all clinical trials; report related data (including psychometric data) in manuscripts and/or supplementary materials • Use clear operational definitions when describing delivery techniques; avoid using labels that infer mechanism or fail to convey what the technique “looks like”
3.3 Rigor and reproducibilit y	• Variation in exposure delivery, quality monitoring, measurement procedures, and/or allegiance to a particular mechanism likely contributes to discrepant mechanistic findings across studies • Subjective measures of mechanism (e.g., self-report) are commonly used and least burdensome but increase risk for bias; this is exacerbated without masking	• Develop and validate low-burden objective measures (e.g., using wearables or brief computerized tasks) • Optimize the utility of publicly available platforms for quickly locating and sharing “tools”	• Develop and publicly share detailed “tools” that support measurement and treatment procedures, including clinician training/monitoring (e.g., rating manuals, SOPs) • Pre-register hypotheses when possible • Increase sample size in clinical trials to the fullest extent possible (e.g., by participating in multisite research consortiums) to offset decreased power resulting from measurement error • Design clinical trials to be masked and use independently-rated primary outcomes, particularly if using subjective measures of mechanism
3.4 Outcome measurement/analysis in clinical trials	• In traditional RCTs the mediator is not randomized, leading to overestimation of a causal effect on outcome • Relationships between mediators and outcomes can be driven by a third variable (i.e. other plausible mediators)	• Deploy enhanced cross-lagged panel models (e.g., random-intercept, latent change score) to most accurately assess the causal effect of a mediator	• Assess candidate mediators and outcomes frequently • Power trials to investigate all possible mechanisms and their interactions, rather than only focusing on outcomes. • Assess delivery and/or contextual factors that may impact outcomes (e.g., significant life events, see also 3.2). • Collaborate with researchers bringing relevant statistical expertise, during both design and analysis
3.5 Machine learning applications	• Assumptions about mechanistic processes are typically made on the within-subject level, but most analyses focus on the between-subject level • Mediation is often tested in linear models; non-linear relationships may be present but traditional modeling risks model misspecification • Exposure use/techniques are not always randomized (e.g., homework exposures) • Machine Learning (ML) models are more flexible, make fewer assumptions, and better suited for handling multi-collinearity vs. enhanced cross-lag panel models (see 3.4)	• Deploy supervised ML to simultaneously integrate many predictors and nonlinear variable relationships • Deploy unsupervised ML to identify patient “clusters” (e.g., those that respond best to specific exposure conditions)	· Consider ML augmented inverse propensity score weighting with doubly robust estimation to accurately estimate causal effects • Consider ML methods based on mixture modeling, which are designed a priori to address measurement error • Collaborate with researchers bringing relevant statistical expertise, during both design and analysis
4. Individual and Contextual Differences
4.1 Magnitude/type of emotion elicited during exposure	• Exposure often targets emotions other than or in addition to fear (e.g., disgust, incompleteness); pattern of change and clinical outcomes can differ for these • Emotional distress can be amplified or qualitatively altered by individual factors (e.g., anxiety sensitivity)	• Studies should measure multiple affective states to enable comparison of change patterns and outcomes by type of emotion • Develop targeted augmentation and/or exposure delivery strategies for different emotions	• Design studies to measure a range of negative emotions across time; also consider measuring positive emotions. • Include measures of individual factors relevant for elicited emotions (e.g., anxiety sensitivity) • Increase sample sizes to the fullest extent possible (e.g., by participating in multisite research consortiums) to power subgroup analyses related to emotion
4.2 Developmental considerations	• Exposure is efficacious for adolescents, but response may be less robust and enduring vs. that for younger children and adults; adolescents experience normative deficits in extinction learning • It is not known whether the higher response rate in younger children relates to true developmental differences or compensatory strategies (e.g., greater involvement of caregivers) • Some treatment packages for youth include less exposure in favor of other CBT ingredients (e.g., relaxation); higher proportion of exposure strongly predicts better outcomes • Younger patients have more difficulty articulating fear cognitions and/or understanding differences between mechanistic constructs (e.g., fear vs. expectancy)	• Longitudinal studies are needed to understand changes in extinction learning across development • Compare patterns of mechanistic change and clinical outcomes across age cohorts in clinical trials, particularly when holding treatment components constant • Conduct psychometric examination of mechanistic measures across age cohorts • Establish “best practices” for streamlined use of self-report measures by age • For adolescents, test intervention strategies that circumvent reliance on the prefrontal cortex (e.g., targeted exposure augmentation strategies; emphasize rewarding stimuli)	• When possible, include objective measures (e.g., psychophysiology, observer ratings) as part of multimodal mechanism measurement for youth • Include detailed measures of treatment delivery, including developmental adaptations and caregiver/peer involvement • Increase sample size in clinical trials to the fullest extent possible (e.g., by participating in multisite research consortiums) to power tests across age cohorts
4.3 Consideration s for culturally diverse populations	• Exposure appears to be effective for diverse populations, but studies have been limited by small sample sizes and inability to test for subgroup effects (e.g., based on race, ethnicity, intersectionality) • Very few studies have examined mechanistic differences in marginalized individuals • Factors such as discrimination and race can influence key mechanistic constructs (e.g., experiential avoidance, response to threat) • Patient report of symptoms and mechanistic changes differ across cultures (e.g., endorsement of somatic vs. cognitive distress)	• Design lab-based and clinical studies to explicitly test mechanisms among marginalized individuals • Conduct psychometric examination of mechanistic and symptom measures across cultural groups and languages • Identify optimal strategies for providing culturally responsive exposure therapy (i.e. tailored for cultural context) • Examine elements of culturally responsive care in relation to exposure mechanisms	• All studies should include measures that characterize historically marginalized groups (e.g., based on race, gender, sexual orientation, disability, economic status) and their experiences (e.g., discrimination, negative life events) • Increase sample size in clinical trials to the fullest extent possible (e.g., by participating in multisite research consortiums), to detect subgroup differences for minoritized individuals

1.1. Overlap among theorized mechanisms

Most research to date has aimed to identify single theory-based mechanisms and/or compare theoretical approaches against one another. However, discussing and testing different theory-based mechanisms as “competitors” risks obscuring the considerable overlap across theories, relationships among different mechanisms, and links with earlier translational work-- and does not align with the idea that exposure most likely works via parallel and successive mechanisms (Himle, 2015). First, variable use of terminology may foster artificial boundaries between different approaches and hinder translational progress. For example, “habituation” is often discussed as a mechanism, referring to the reduction of fear responses within or between exposure exercises. However, in basic research, habituation is strictly defined as a short-term, implicit adaptation in the form of diminished responding to an inconsequential stimulus not due to fatigue (a non-associative learning process; Rankin et al., 2009). This definition does not align with the idea of associative learning presumed to underlie exposure according to multiple theories, which is more accurately described as extinction (i.e. decrement of fear behaviors when associative relationships underlying original fear learning have changed; Lovibond, 2004); Table 1). The field will benefit from establishing consistent, precise terminology and concrete operational definitions for both mechanistic constructs and related exposure optimization strategies (see 2.1 and 3.3).

Second, much of the research on mechanisms of change has been narrow in that studies have typically focused on testing one phase of treatment (e.g., before, during, or after in-session exposure, between-session exposure homework). This misses an important opportunity to understand relations among optimization strategies used across different phases of mechanistic learning, which include initial learning (during exposure; Benito et al., 2018), consolidation (in the hours after exposure; Kredlow, Eichenbaum, and Otto, 2018), retrieval (during repeated exposures; Craske et al., 2014), and generalization (when encountering new stimuli and/or contexts; Richter et al., 2021). When considering all phases, we have an opportunity to fully optimize outcomes by deploying the most promising strategies targeting each.

Third, many constructs typically associated with different theories (e.g., EPT, ILT, ACT) share a high degree of conceptual overlap and may not be distinct (i.e., represent measured variables of a shared latent construct). A case in point is extinction and expectancy violation. The reduction of fear is incongruent with the idea that fear will be of infinite intensity or duration (a common prediction in anxiety). How does this differ from expectancy violation or change? It seems plausible that changes in fear and expectancy represent emotional and cognitive aspects of the same underlying mechanism of change (e.g., Craske et al., 2022; Hofmann, 2008). Similar concerns have been raised regarding overlap between self-efficacy and outcome expectations (Breuninger, Tuschen-Caffier, and Svaldi, 2019) and between distress tolerance, experiential avoidance, and emotional engagement (Schloss and Haaga, 2011). Additionally, many of these models are consistent with neurocognitive, computational models of reinforcement learning contending that prediction errors (online predictions followed by feedback) are the basis of most learning, including fear extinction (Koban et al., 2017; Song et al., 2022). In some cases, constructs with different labels share near identical definitions (e.g., change in maladaptive fear cognition, expectancy change, threat reappraisal) but have become associated with seemingly different clinical approaches (Himle, 2015). A more comprehensive, integrated theoretical conceptualization is needed to account for shared core features across existing theories--and will aid in accelerating optimization research with the greatest potential to produce clinical gains. Understanding shared and unique core features may also help to pinpoint whether these features qualify as common mechanism of change or whether some are unique to specific phenomena of anxiety disorders (e.g., see discussion on fear vs. anxiety, Perusini et al., 2015; Ohman, 2008 ).

1.2. Mechanism or facilitator

Some candidate mechanisms appear to be important for therapeutic change but are not causes of symptom improvement. Consider, for example, emotional engagement/activation. Patients show emotional activation in everyday life without lasting change in anxiety symptoms., Brief emotional activation does not qualify as a mechanism of change because it does not directly cause anxiety reduction. However, it may be a “facilitator” (i.e. strategy that facilitates mechanism engagement) or a moderator, in that it is necessary but not sufficient for mechanism engagement and/or strengthens the influence of another mechanism. A similar argument can be made about therapeutic alliance. Disentangling mechanisms from moderators and facilitators will expedite future research on proximal “levers” that cause mechanistic change—a critically important area of work that will aid in identifying the most promising strategies for treatment optimization. As part of this work, it will be necessary to consider that facilitators may trigger change in more than one mechanism, including those they are not theoretically paired with (e.g., exposures conducted with focus on fear extinction likely produce change in expectancy or vice-versa). As such, individual studies should include a comprehensive plan for measuring all plausible mechanisms—a necessary step for definitive mechanism testing (see 3.4). This will also speed identification of the most potent optimization strategies for various mechanisms, and may help identify alternate optimization strategies when tailoring is needed (e.g., for youth who have difficulty articulating fear cognitions, see 4.2).

1.3. Mechanism or read-out measure

Testing mechanisms is inherently dependent on imperfect “signals” or read-out measures of those mechanisms, i.e., observable responses used to measure the mechanism.. For example, human fear extinction research is at risk of failing to distinguish between read-out measures and the underlying latent process of fear extinction (see Chronbach and Meehl (1955). Typical read-out measures in fear extinction research are assessed on the subjective (e.g., threat expectancy), physiological (e.g., skin conductance responses), and behavioral level (e.g., safety behavior). Read-out measures can thus be understood as the observable expression of a latent mechanism. Despite the assumption that they all measure the same underlying process, these read-outs often do not converge (see 2.2.) and subtle changes in measurement may differentially influence read-outs (see Lonsdorf et al., 2017 ).

The distinction between mechanism and read-out also applies to exposure theories. In the original description of EPT, within-session fear reduction is a read-out that indicates successful emotional processing. Findings that within-session fear reduction is not consistently associated with treatment success (Craske et al., 2008) could thus be an issue with the read-out measure, not the mechanism itself. Conversely, a read-out measure might signal change in an underlying mechanism other than the one with which it is commonly associated. For example, within-session extinction could be consistent with inhibitory learning if it serves as a signal of initial learning (followed later by consolidation and retrieval). The distinction between mechanism and read-out has long been discussed in behaviorism, for example, by demonstrating that learning can occur without changes in observable behavior (“latent learning”,Tolman & Honzik, 1930). In addition to having implications for measurement error and power in clinical trials (see section 2), this issue is of notable clinical importance, as therapists depend on observable read-outs to guide exposure. Going forward, it will be important to distinguish between mechanisms and read-out measures, which may be influenced by multiple factors apart from the mechanism itself. This calls for more basic research on the link between specific read-out measures and a proposed underlying mechanism and consistent operational definitions and measurements for mechanisms of change.

2. Measurement

Mechanism measurement has largely followed traditional approaches used in experimental laboratory research. This has commonly included assessment of self-reported distress (e.g., Subjective Units of Distress Scale; SUDS) or psychophysiological indices (e.g., skin conductance) during exposure. More recent studies have included a broader array of measurement approaches in response to advances in mechanistic theory, new efficacy data, and emphasis on the Research Domain Criteria (RDoC) for studies funded by the U.S. National Institute of Mental Health (Insel et al., 2010). Despite a long tradition of incorporating mechanistic measures in both analog and clinical studies of exposure, relatively little psychometric data exists to guide selection and/or administration timing. As a result, researchers have developed a variety of “home grown” measures and supporting procedures (e.g., for rater training). This presents a challenge for replicability and likely contributes to inconsistent findings across trials. This problem may become even more apparent during forward translation, as resources decrease and exposure delivery shifts to fit various “real world” contexts. It will be important to develop and follow a coordinated research agenda that emphasizes both psychometric rigor and pragmatism in the measurement and reporting of exposure mechanisms, an issue receiving ongoing attention in the NIH Science of Behavior Change (SOBC) program, including recommendations for standards in the reporting of mechanistic behavior change research (Birk et al., 2023).

2.1. Psychometric Properties of Measures

Future work will need to evaluate mechanistic measures in a rigorous way, with emphasis on establishing reliability, construct validity, and predictive validity (i.e. with treatment outcomes). This is particularly true for psychotherapy process-based constructs (i.e. events or interactions that occur during psychotherapy sessions) traditionally assessed at the exposure practice- or session-level (e.g., expectancy, extinction, distress tolerance). It will be important to establish inter-rater reliability for measures rated by therapists or other observers, and to detail the procedures needed for training them to a reliable criterion and maintaining it over time. These procedures might include developing a detailed manual to guide ratings, double-rated observations, and monitoring to prevent drift. While time-consuming, such procedures reduce measurement error, boost power, and support rigor and replicability across research groups, particularly when made publicly available. Innovations in advanced computing (e.g., artificial intelligence [AI] using natural language processing) might also be tested to replace human reporters, reducing or eliminating the need for ongoing monitoring (Ewbank et al., 2020). It will also be essential to establish construct validity through comparison with established measures of similar constructs (convergent validity) and those of different constructs (divergent validity). It is not yet clear whether measures commonly used to assess different theorized mechanisms will separate from one another or whether they represent overlapping aspects of the same underlying process (e.g., cognitive vs. emotional; see section 1). This conceptual ambiguity has often resulted in studies using different construct labels despite identical measurement approaches (Cooper, Clifton, and Feeny, 2017). Whereas a number of studies have measured more than one mechanistic construct, virtually none have deployed them in a way that facilitates examination of construct validity (i.e. with the same timing of administration during exposure). It will be important for the field to develop consistent definitions and measurement approaches for mechanism-relevant constructs that can be adopted across research groups and through forward translation. In contrast, numerous studies have examined predictive validity of mechanistic measures with clinical outcomes after treatment. This is particularly true for constructs based on SUDS (e.g., fear activation, extinction, variability), though studies have increasingly focused on others (e.g., expectancy, experiential avoidance, self-efficacy). However, few measures have consistently predicted outcomes across studies and research groups. At present, it is not clear whether inconsistent findings relate to differences in study procedures and/or samples, weak relations of these constructs with outcomes, or one or more common measurement challenges (as described below, e.g., limited reliability, different measurement approaches and timing).

Limited reliability and construct heterogeneity can drastically impact findings in mechanistic studies. Wilcox et al. (2013) present a seemingly benign example where 10% of analyzed data departs from the main sample data pattern. This scenario could readily be observed in measurements of exposure and related symptom reduction. In this case, between-group effect sizes are reduced by over 70% and statistical power is reduced by more than 60%, which increases the required sample size more than 10-fold (Faul et al., 2009). These problems are compounded in mechanistic research designs because there are multiple stages of analysis that sequentially worsen the problem. A key design in exposure research involves mediation with two parts, for example, where an intervention first predicts level of exposure engagement, and then exposure engagement predicts symptom response. In this case, commonly accepted levels for reliability (e.g., .80) would be expected to reduce standardized mediation effects by more than 50%. These examples fall under the umbrella of “data pollution” (De Nadai, Hu, and Thompson, 2022)--unintentional errors in data that have substantial impact on results. Data pollution can be addressed post-hoc through analysis (e.g., robust estimation, latent variable estimation) and a priori through research design (e.g., selection of measures that exceed conventional norms for reliability). Once elements of data pollution are addressed, improved effect sizes and model precision substantially reduce the number of participants needed, which will contribute to accelerated scientific progress.

It will also be critical to establish “best practices” for handling common challenges that arise in process-based mechanism measurement. First, optimal assessment timing and variable calculation are unknown. Calculation of the same construct varies widely across studies, for example, using one time point (van Minnen and Hagenaars, 2002), average of time points (e.g., (Jaycox, Foa, and Morral, 1998), change scores that sum all time points (Benito et al., 2018), all time points in a multi-level model to compare effects across time (e.g., (Sripada and Rauch, 2015), or calculation of reliable change indices (Cooper, Clifton, and Feeny, 2017). Calculations based on more frequent assessment within and across sessions will facilitate better understanding of the time course of change and reduce measurement error (thus improving power; Faul et al., 2009). Some constructs are traditionally sampled more frequently than others (e.g., within-session variability vs. extinction); to ensure equally robust measurement/power across constructs it may be important to use similar time points in the calculation of each. Finally, a majority of process-based measures focus on acute changes (i.e. within-session) or longer-term (i.e. between-session) changes during exposure tasks. From a learning perspective, these may be most relevant for understanding initial learning and later retrieval but miss the opportunity to understand processes related to consolidation (e.g., during post-exposure processing). Understanding these processes provides an opportunity for memory editing (Phelps & Hofmann, 2019). Consolidation is thought to occur within a 6-hour period, leaving open the possibility that memory may be modified even well after the exposure session (Kredlow, Eichenbaum, and Otto, 2018). Similarly, process-based measures have largely missed the opportunity to assess generalization of learning to new stimuli and/or contexts. Future studies should enhance efforts to assess consolidation (e.g., by following up with participants during the 6 hours following exposure) and generalization (e.g., with regular assessment of potential experiences with new stimuli/contexts). For assessing these and other constructs outside of treatment sessions, studies might deploy low-burden Ecological Momentary Assessment (EMA) strategies to capture detailed temporal changes (e.g., Walz et al., 2014). Researchers might also deploy methods for controlled observation. For example, Richter et al. (2021) used a post-treatment Behavioral Approach Task (BAT) to demonstrate modest generalization of exposure learning to new stimuli. Future studies should also work to identify the optimal window for measuring both learning generalization and change during the consolidation period, which may vary across development, individuals, and/or context.

A second challenge is that therapy content itself may create bias and measurement error. For example, psychoeducation could introduce expectancy or social desirability bias in later self-report ratings of mechanistic constructs (e.g., patients may be more likely to report changes in expectancy vs. other constructs after psychoeducation focused on expectancy violation as a goal). Similarly, measures of mechanism depend on the quality of the exposure context, which is influenced by a wide variety of potential external factors (e.g., changes in stimulus, therapist actions) and patient behaviors (e.g., avoidance). Successful measures may need to account for the occurrence of these as well as their timing in relation to mechanism measurement (e.g., decrease in expectancy may not reflect mechanistic learning if preceded by avoidance). Selection of the exposure task may also influence mechanistic ratings, for example, if “easy” tasks produce a floor effect in ratings. A longstanding issue in the calculation of between-session changes relates to the introduction of increasingly difficult exposure tasks across treatment, making it difficult to determine whether mechanistic changes are occurring. This remains problematic even when variable task selection is an explicit goal (Craske et al., 2014), as many patients become more willing to try challenging exposures later in treatment. See section 3.2 for additional discussion about delivery variation. Measuring and accounting for these variations is likely to be complex, particularly as delivery becomes less controlled and patient presentation more variable through forward translation. Future studies will need to explore these possibilities in greater detail and consider innovative solutions for handling complexity in a practical manner (e.g., with AI approaches; Ewbank et al., 2020).

2.2. Multimodal Measurement and Convergence Across Measurement Levels

One central measurement issue is whether multimodal assessment across measurement levels (e.g., self-report, behavioral, physiological, neural) and/or internal process (e.g., cognitive, emotional) is needed to ensure an accurate accounting of mechanistic change. Threat conditioning paradigms provide a useful perspective on the challenges inherent in this issue. Although some studies have indicated correspondence between physiologic (e.g., SCL) and expectancy ratings (Fanselow and Pennington, 2018), many do not. Lubin and associates (2023) found that these measures shared less than 6% of variance, and Constantinou et al. (2021) reported shared variance ranging from a low of 0.5% during extinction phase assessment to near 11% when early and late responding was parsed. These results caution against the assumption that these measures accurately assess the same construct, even in the context of tightly controlled laboratory studies. However, the distinction between mechanistic construct (e.g., extinction learning) and observable read-out measures (e.g., SCRs, expectancy ratings) also needs to be considered. Read-out measures are influenced by a variety of factors and may thus indicate the same mechanism despite low overlap between measures (see section 1).

Desynchrony among measures of fear is an old problem, with some indication that degree of desynchrony among measures–e.g., reflecting self-report (verbal/cognitive), motor, and physiologic response systems (Lang, 1971)--may be moderated by degree of fear, external demands, type of intervention employed, time passed since improvement, and measure used (Hodgson and Rachman, 1974). Each spells trouble for the assessment of mechanism in clinical trials, raising questions about whether the study of mechanism should attend to what Hodges and Rachman proposed five decades ago, that [mechanistic] “effectiveness of a therapeutic technique should be assessed across all response systems” (p. 322). Indeed, examination of separate measures in the verbal/cognitive realm–self report of fear (SUDS) vs. expectancy–reveals that both have had predictive successes and failures (e.g., Smits, Rosenfield, Otto, Marques, et al. 2013; Pittig, van den Berg, and Vervliet, 2016). This might be explained by individual differences in interoception accuracy (Khalsa et al., 2018) or reliance on expectancies or habit in information processing (Fradkin et al., 2020) that influence the meaning/accuracy of self-reported fears or expectancies, respectively. These factors may be even more relevant earlier in development (see 4.2), raising questions about whether/how multimodal assessment might be streamlined for youth while optimizing predictive value. Further, the link between these measures and the core concept of fear continues to be debated in relation to whether verbal/cognitive, motor, and physiologic response systems arise from the same systems or represent separate pathways to fear and fear-related behaviors (LeDoux and Pine, 2016). Indeed, the distinction between higher and lower order processes can aid the understanding of discordance among measures and associated processes (e.g., forecasted action vs. conditioned action, self-report of fear or expectancies vs. physiologic reactivity; Taschereau-Dumouchel et al., 2022), and encourages investigators to pause and conceptually map linkages that might underlie a given measure of mechanism.

Some have argued that the essence of the emotional experience of fear and anxiety is the subjective experience and that the associated behaviors and physiological responses are, at best, indirect indicators of those inner experiences (LeDoux & Hofmann, 2018; Taschereau-Dumouchel et al., 2022). The subjective experience can only be assessed through verbal report, which constitutes a methodological barrier to studying conscious feelings in animals as well as humans who are nonverbal or earlier in development (Hofmann, 2008). Researchers often ignore desynchrony between these measures or arbitrarily choose one measure over others (often favoring objective over subjective measures). However, different outcome measures do not necessarily measure the same construct or may measure the construct at a different level of cognitive awareness (e.g., verbal report at the conscious level and physiological measures at the automatic and subconscious level). Similarly, some measures might provide a meaningful comparison between animal and humans (e.g., physiologic) but others may be more valuable for understanding the human experience (e.g., self-report). Thus, each measure may contribute unique information that can be seen as informative rather than a problem that needs rectifying (e.g., through error correction; Guolo, 2008). Moving forward, extinction research should focus greater attention on ontological issues, including explicit examination of relationships among readout measures and development of broader frameworks that incorporate measurement of extinction learning relative to other indices of declarative and associative learning and encompass related concepts (e.g., perseveration and behavioral persistence; Eisenberg et al., 2019).

2.3. Psychometric Properties of Laboratory Models with Humans

A fair amount of mechanistic research is based on laboratory models used as proxies for exposure therapy. Fear and avoidance conditioning and extinction are among the most prominent (Krypotos, Vervliet, and Engelhard, 2018; Pittig et al., 2020). Briefly, fear conditioning involves repeated presentations of a neutral stimulus (designated conditioned threat stimulus, CS+) with an aversive unconditioned stimulus (US), resulting in the CS+ triggering anticipatory fear responses (e.g., elevated threat expectancy or skin conductance, Lonsdorf et al., 2017). Fear extinction, as a proxy for exposure tasks, subsequently involves repeated unreinforced presentations of the CS+, leading to a reduction in fear responses to that CS+. In avoidance conditioning, the aversive US associated with the CS+ can be prevented by a predefined avoidance response to the CS+. During avoidance extinction, the CS+ is no longer followed by a US and changes in avoidance responses are examined (Pittig et al., 2020).

Fear and avoidance extinction has been studied across species, enabling insights into theoretical (Bouton, 2002) and neurobiological (Milad and Quirk, 2012) underpinnings (see 2.4 for additional detail about animal models). Despite being highly influential, only a handful of studies have assessed the test-retest reliability of fear response measures during fear extinction, such as self-report, psychophysiological, or neural indices (e.g., (Klingelhöfer-Jens et al., 2022). Results indicate limited reliability across measures, although floor effects may contribute, as fear responses decreased during extinction and reliability was better during fear acquisition. Even less is known about psychometric properties of avoidance conditioning and other laboratory models (e.g., emotion regulation tasks to model experiential avoidance, laboratory stress challenges to model distress tolerance, false feedback to manipulate self-efficacy beliefs). The psychometric properties of responses within these models have, to the best of our knowledge, not been investigated.

Further, less is known about the external validity of laboratory models of exposure therapy. To what extent do methods and findings of laboratory models match clinical phenomena and translate to practice? Conceptually, validity criteria for laboratory models include face validity, construct validity, content validity, and predictive validity (Scheveneels et al., 2016; Krypotos, Vervliet, and Engelhard, 2018; Haynes et al., 1995). However, few studies have tested these in a robust manner (Scheveneels et al., 2016; Pittig et al., 2020). For example, face validity of stimuli used is relatively weak. More complex, multi-sensory, and life-like stimuli will better simulate the naturalistic situations confronted during exposure (Scheveneels et al., 2016). Moreover, recently acquired fear or avoidance responses or brief inductions of emotion regulation or self-efficacy may be substantially different from patient responses that have been elaborated, generalized, and socially reinforced over years. Recent studies have made efforts to improve relevance of laboratory models for patient populations and exposure therapy, with promising results (e.g., inclusion of overnight consolidation after fear acquisition, use of explicit instructions to ensure fear responses are reliably acquired; Hollandt et al., 2020). With respect to predictive validity, only a handful of studies have investigated whether/how individual differences in laboratory responses predict clinical response to exposure therapy (Scheveneels, Boddez, and Hermans, 2021). This is critical, as pre-treatment laboratory assessments could have prognostic value or guide treatment tailoring. In summary, despite the importance of laboratory research on mechanisms of exposure, the psychometric properties and validity of models is still largely unknown.

2.4. Advances from Models with Non-human Animals

Early translational studies using non-human animals have long been important for understanding exposure mechanisms and will continue to be a key avenue for future research. Animal literature has informed a variety of paradigms/behaviors relevant for laboratory models with humans and exposure therapy. These include fear potentiated startle, where fear is initially conditioned by pairing a CS with mild shock (the US), then a sudden loud noise is presented during the CS. Conditioned fear results in an exaggerated startle response (Davis, 2006). Passive vs. active avoidance can also be assessed using tasks where an individual is trained to inhibit the natural inclination to avoid negative consequences. A simple distinction between passive and active avoidance is that in the former, individuals choose to not engage in a response that will increase discomfort, and in the latter, they explicitly choose to use behaviors that decrease discomfort (Krypotos, Vervliet, and Engelhard, 2018). Conditioned suppression of reward seeking can also be assessed using paradigms where rat is first trained to carry out a task for reward (e.g., press a bar) and then fear conditioned (perhaps also with extinction). Willingness to engage in bar pressing is assessed during CS presentation. Conditioned responding provides insight about willingness to engage in pleasurable activities following an anxiety-inducing experience. For example, a rat may show a significant reduction in fear response (freezing) following extinction, but fail to resume reward seeking (Shumake and Monfils, 2015). Likewise, a patient may report distress reduction after treatment but refrain from activities they used to enjoy—a distinction that likely has important ramifications for long-term functional outcomes.

In fear extinction paradigms (a proxy for exposure therapy, see 2.3), individuals are susceptible to the return of fear. This is evident through the phenomena of spontaneous recovery, reinstatement, and renewal, where spontaneous recovery occurs after the passage of time, reinstatement is brought on by a stressful event, and renewal occurs in new contexts. Animal research has provided important insights into these phenomena and potential mechanisms through which novel treatment approaches could improve them. Nader proposed that once retrieved, memories become susceptible to disruption, and blocking the molecular cascade engaged at retrieval can attenuate fear expression and prevent return of fear—known as reconsolidation blockade. In an initial proof-of-concept study (Nader, Schafe, and Le Doux, 2000) and many that followed, reconsolidation blockade was achieved with protein synthesis inhibitors that cannot be used in humans. Since then, studies have explored pharmacological approaches to target fear memories with varying degrees of success. In a first study with humans, administration of propranolol after retrieval of a traumatic memory reduced physiological responding compared to placebo (Brunet et al., 2008). Additional studies of propranolol post memory reactivation demonstrated some success (e.g., (Kindt, Soeter, and Vervliet, 2009; Elsey and Kindt, 2017). Since then, additional drugs have been tested for reconsolidation blockade (e.g., Sirolimus, Surís et al., 2013; mifepristone, Pitman et al., 2011) . Despite promising results, this approach is not universally effective (e.g., Wood et al., 2015). Future studies should continue to examine whether targeting mechanisms that underlie reconsolidation blockade—first identified in animals–will effectively translate into improved outcomes in humans.

Focus on reconsolidation offered an opportunistic window for targeting fear memories. While this area has received substantial focus to date, our discussion of it here serves to emphasize a recognizable example of cross-translation between basic neuroscience and the clinic. Monfils et al. (2009) proposed retrieval-extinction, where extinction training is used during reconsolidation—a finite period when memories are susceptible to disruption. In this paradigm, an isolated retrieval trial is presented; then, after sufficient time has elapsed for the memory to destabilize, extinction training is introduced. This has been successful in some studies with rodents and humans (e.g., Agren et al., 2012; Rao-Ruiz et al., 2011), and translated with some success in clinical and analog samples (e.g., Lancaster, Monfils, and Telch 2020). One advantage of reconsolidation-based approaches is that they are theoretically less susceptible to return of fear (Monfils and Holmes, 2018)—and unlike extinction, need not engage the prefrontal cortex (Agren et al., 2012). They could thus prove effective in treating populations in which engaging the prefrontal cortex proves challenging (e.g., individuals with traumatic brain injury, or youth). The above approaches are in their infancy, and future research should continue to test whether vulnerability opened by memory retrieval will prove an effective therapeutic window on a larger scale in humans.

Ultimately, animal-to-human translation should be viewed as an iterative process or “feedback loop” that also serves to improve basic models and enable continued treatment refinement. This will facilitate ongoing consideration of species-specific methods and measures, feeding back practical information about what can reasonably be translated across domains and where future adjustments should be made (e.g., Haaker et al., 2019). Importantly, rodent study successes over the years suggest that most in-roads have evolved through feedback from behavioral developments and approaches. Human behavioral studies can serve to inspire rodent researchers to further develop novel assays and improve understanding of the broad repertoire of rodent fear expression and processing. Though more homogenous than humans, rodents still display vast individual differences in fear responding. Rodent research is uniquely suited for explaining individual variance given the ability to isolate individual factors of interest under tightly controlled conditions. Ultimately, close collaboration between basic and clinical researchers will facilitate development of methods with a balance of internal and external validity, and will speed treatment and study development.

3. Study Design and Analysis

3.1. Experiment and Clinical Trial Design

In a seminal review of mechanisms of change in psychotherapy research, Kazdin, (2007) highlighted the advantages of using an experimental approach for testing therapeutic change mechanisms. This emphasis is complemented by calls to redesign clinical trials with specific tests of how treatment elements are linked to mechanism(s) as a prelude to evaluating clinical significance (Nielsen et al., 2018). As such, progress in understanding mechanisms relies on well-designed experiments that specify putative mechanisms as well as clinical trials that can show the effects of targeting specific mechanisms. Across both preclinical experiments and clinical trials, there is a need for traditional and innovative randomized group-based designs, as well as single-subject designs (e.g., multiple baseline, alternating treatment, knock-out). Small N designs capture within-subject change that is often missed in group-based designs and are ideally suited for isolating the contribution of specific exposure strategies. They are also valuable as part of a study series that follows the experimental therapeutics framework, by testing initial mechanism engagement before moving on to a larger pilot trial. For example, one recent study used a multiple baseline design to test a therapist training tool for optimizing exposure delivery in community treatment settings (Benito et al., 2021). Results demonstrated that therapists met delivery benchmarks and patients showed highest levels of within-session extinction only after the training tool was introduced–demonstrating that successful manipulation of exposure delivery via the training tool likely had a causal effect on a theorized mechanism.

Experimental designs

Experimental designs in laboratory settings fill an important translational gap between basic laboratory models (e.g., for fear and avoidance conditioning/extinction) and full clinical trials and will continue to be part of a meaningful research agenda. These studies aim to manipulate putative mechanisms using a brief exposure protocol and examine how changes in these relate to subsequent reduction in fear and/or anxiety (or other dimensional symptom outcomes). Samples are often composed of healthy participants and/or those with elevated symptoms (i.e. a clinical diagnosis is not required), which facilitates speedy recruitment. The ability to isolate mechanisms of interest and rapidly determine which “levers” engage those mechanisms best is a key advantage of these designs. Because they require fewer resources than full treatment trials, these studies are also highly efficient. For example, in one study focused on testing threat reappraisal as a core mechanism of change, Telch and colleagues, (2000) examined the independent and joint effects of distraction and threat focus with disconfirmation on fear reduction during exposure. Participants with severe claustrophobia symptoms were randomly assigned to exposure (a) with guided threat reappraisal, (b) with a cognitive load distractor task, (c) with both or (d) without either. The cognitive load task had a marked detrimental effect on threat reappraisal and fear reduction, while guided threat reappraisal (regardless of cognitive load) had a facilitative effect. The greatest level of threat reappraisal and fear reduction and the lowest level of return of fear were observed in the guided threat reappraisal without cognitive-load condition—providing a promising avenue for future clinical trials.

Clinical Trial Designs

Full clinical treatment trials will continue to be important for testing the efficacy of exposure optimization strategies (and corresponding mechanisms) before making recommendations for clinical practice, and are essential for ensuring that promising strategies hold up under increasingly generalizable conditions (e.g., with treatment-seeking clinical populations, using a multi-session course of treatment). For example, optimization strategies could function differently when moving from experimental protocols with a limited number of sessions into a full course of exposure therapy (e.g., with repeated consolidation and retrieval of learning over a period of months). Additionally, individuals with anxiety disorders show impaired extinction learning; this is even more pronounced among those with more than one anxiety diagnosis (Marin et al., 2017). Optimal strategies for engaging exposure mechanisms may differ when moving from a population with subclinical/mild symptoms into a treatment-seeking clinical population. For example, it is possible that those with more severe/impairing symptoms need additional compensatory strategies, are less able to experience long-term benefit from exposure in the presence of factors that interfere with learning (e.g., safety behaviors), and/or need a higher “dose” of various strategies.

Where possible, secondary studies that examine exposure delivery in past clinical trials will be important for advancing knowledge about which mechanism-related strategies were actually used (intentionally or unintentionally). One common assumption is that older trials were conducted using a “habituation” model guided by EPT, yet few treatment manuals explicitly called for this and the limited available evidence suggests considerable variation in delivery, with exposures more commonly including features consistent with other models (e.g., IL; Benito et al., 2018). Determining whether clinical outcomes from prior trials can be improved with mechanism-informed exposure techniques is perhaps the highest priority goal for future research. To that end, information about exposure delivery in older trials will be needed to design realistic “exposure as usual” conditions in future trials, thus enhancing significance. Additionally, given the conceptual overlap of mechanistic theories (section 1), it will be challenging–if not impossible–to design trials that directly compare various models (e.g., extinction vs. inhibitory learning vs. cognitive) without inadvertently creating artificial “straw man” comparison conditions. Instead, it may be more valuable to follow an integrated theoretical framework with focus on testing the most promising mechanisms and corresponding strategies (e.g., for before, during, and after exposure) identified in earlier translational work.

Innovative clinical trial designs, such as sequential multiple assignment randomized trial (SMART) designs, are likely to be helpful for testing multiple exposure optimization strategies and/or related mechanisms. SMARTs test adaptive interventions via multiple randomization points based on time and/or tailoring factors (Almirall et al., 2014). Information about mechanisms might be used as a tailoring factor that triggers a subsequent treatment decision (i.e., re-randomization; Sauer-Zavala et al., 2023; Southward and Sauer-Zavala, 2020). To the degree that change in mechanism precedes clinical improvement, mechanisms can serve as a proximal indicator of treatment response and guide relevant decision-making earlier in treatment. For example, patients could be initially randomized to a condition (e.g., different sets of exposure optimization strategies vs. exposure as usual), followed by a second randomization after a specified number of weeks. At that time, those who show evidence of mechanism engagement might continue within their assigned treatment arm while those that do not are randomized again (e.g., to continue vs. switch arms). This example design would accommodate mechanistic heterogeneity across individuals (e.g., with respect to relevant mechanisms and/or optimization strategies) and explicitly test alternate treatment options for those that do not show initial change in a given mechanism. Alternatively, individuals that do show evidence of mechanism engagement early in treatment could be randomized to continue or terminate treatment (Sauer-Zavala et al., 2023). This type of design would test whether those showing evidence of mechanism engagement could end treatment early, reducing excess treatment and allowing providers to treat additional patients. These are just a few examples of the ways in which SMARTs or other adaptive designs might use information about mechanisms to make complex treatment and/or dosing decisions and optimize outcomes.

3.2. Measuring and Accounting for Exposure Delivery Variation

Exposure delivery can vary widely even in tightly controlled clinical trials with highly adherent therapists (Benito et al., 2018; Benito, Machan, et al., 2021)–and this can have profound implications for conclusions that might be drawn about mechanism. Most trials use some form of adherence or fidelity monitoring, yet these are often trial specific, lack psychometric support, and designed to capture only basic information about “ingredients” used in each session (e.g., whether exposure was used, whether proscribed elements from another treatment arm were used). Moving forward, studies should routinely and rigorously assess nuanced aspects of delivery with direct relevance to theorized mechanisms (e.g., exposure “dose,” targeted core fear, use of compound exposures, process of task selection, key elements of post-exposure processing, and many more). Researchers should also seek to develop psychometric support for measures of exposure delivery, including inter-rater reliability and validity. As part of this process, it will be essential to develop clear operational definitions (i.e. concrete descriptions of what a delivery strategy “looks like”) to facilitate replication across studies and translation into service settings (e.g., through clinician training). Much like problems with labeling and defining mechanisms (i.e. different mechanistic labels often used for the same measurement approach and vice-versa; see sections 1 and 2.1), specific delivery techniques are likely labeled and defined differently by different research groups. Importantly, some clinical trials have failed to find group differences when comparing different exposure approaches (e.g., Cooper, Clifton & Feeney, 2017), leading some to question whether the robust effect of exposure simply overshadows relatively smaller effects associated with variation in delivery. However, other well-powered studies have found modest to meaningful differences (e.g.,Gloster et al., 2011; Pittig et al., 2021; Cooper, Clifton & Feeney, 2017) and/or association between delivery variation and clinical outcomes (e.g., Benito et al., 2021). Incorporating detailed measurement of a full range of exposure delivery variables will be important for identifying/ruling out reasons that some trials find differences while others do not (e.g., whether other, uncontrolled delivery factors may have interfered).

Group-Based Exposure

Group-based exposure has many desirable elements that are unavailable in individual therapy, including peer modeling/encouragement and efficiency (i.e., treating multiple patients at once). However, it is difficult to separate specific causes of exposure effects due to added variables introduced by group-dynamics (e.g., personality characteristics of group members at the subject level, cohesion/morale at the group level). In addition to traditional multilevel modeling approaches, social relation modeling (Kenny et al., 2006) can explicitly separate effects between individual participants and the broader group, and one-with-many modeling (Brinberg et al., 2022) can address the therapist role (the therapist as the “one” and individual group members as the “many”). The one-with-many design is easier to implement in practice, where patients simply rate their therapist. Social relations modeling requires group members to rate all other group members, which presents additional logistical challenges. The one-with-many design has been used in a number of clinical settings, including evaluation of clinician effects in primary care and adolescent substance abuse treatment (Marcus et al., 2011; Anagnostopoulos et al., 2012). These group-based designs allow researchers to separate specific effects of exposure and provide an opportunity to identify specific elements that enhance exposure effects at both individual- and group-levels. For example, it is possible to quantify how effects of a particularly resilient group member shape more positive outcomes for the entire group, or how a suboptimal experience of one group member can affect outcomes of other individual group members and overall group results.

Variation in Exposure “Dose”

There is strong theoretical rationale for the importance of a high “dose” of in-session and out-of-session exposure, and a higher number of in-session exposures robustly predicts improved treatment outcomes (Peris et al., 2017). Nevertheless, treatment protocols vary in the amount of exposure called for (e.g., number of sessions with exposure) and even highly adherent therapists deliver different amounts (e.g., number/duration of exposures) while remaining within protocol parameters (Benito et al., 2018; Benito et al., 2021). It will be important for treatment studies to measure and report these exposure “dose” variables, which may explain variation in outcomes across patients and facilitate comparison of findings across trials. Given that homework completion can be highly variable (Kazantzis et al., 2017), a major problem is that patients can never be fully randomized with regard to out-of-session exposure dose. As a result, traditional intent-to-treat approaches lose the ability to draw causal inferences and can underestimate the magnitude of intervention effects (Connell, 2009). A creative analytic approach designed to address this problem is complier average causal effects modeling (CACE; Yau and Little, 2001). The CACE approach relies on latent variable modeling and has been applied in multiple domains, including substance abuse (e.g., Cordovilla-Guardia et al., 2017). Data generated during exposure therapy research fits well into CACE modeling strategies but requires careful measurement of exposure dose and outside exposure practice. The CACE modeling approach can also be used to address participants who prematurely discontinue treatment, which is common (Swift et al., 2017).

3.3. Rigor and Reproducibility in Clinical Trials

Although a full discussion is beyond the scope of this paper, several issues that relate specifically to exposure therapy mechanisms are worth highlighting here. As described in 2.1 and 3.2, researchers can work to enhance both rigor and reproducibility by using well-validated measures to assess mechanisms and clearly articulating detailed measurement methods for mechanisms and aspects of exposure delivery (e.g., through rating manuals, standard operating procedures for training/monitoring). Ideally, these could be publicly shared–for example, as part of a larger exposure research “toolbox” available for researchers. Transparency and consistency of mechanism measurement and treatment delivery will accelerate work in the field by reducing discrepant findings across studies (and contributing to knowledge gained when findings do differ). Another important area for future work relates to the need for objective mechanistic measures (e.g., psychophysiological, observed behavior) and clinical outcomes (e.g., rated by masked independent evaluator), which will greatly enhance rigor. Relying primarily on subjective measures of mechanism (e.g., self-report) introduces bias; in trials that also lack double-masking, the risk for bias related to subjective measures increases further (e.g., to 22% as reported in Savović et al., 2012). As described above (2.1), this degree of error has a profound impact on power and dramatically alters findings (Wilcox et al., 2013). This issue is likely compounded in trials that use subjective ratings of both mechanism and clinical outcome, particularly if both rely on the same rater where shared forms of bias can inflate the risk of type I error. At present, objective measures of mechanism are infeasible in many trials; it will be particularly important that these trials use masking and objective clinical outcomes. Future studies should work to improve feasibility of objective measures (e.g., psychophysiologic or behavioral indices assessed through wearables) for use in clinical trials, for example, functional clinical outcomes might be assessed based on geographic range of location in daily life during and after exposure for agoraphobia. Finally, many exposure trials to date have suffered from modest sample sizes which, coupled with measurement error, lead to underpowered studies. Leveraging multi-site research consortia can dramatically increase total sample size, thereby producing well-powered studies of putative exposure mechanisms that are more likely to be replicated (Smits et al., in press).

3.4. Outcome Measurement/Analysis in Clinical Trials

Perhaps guided by seminal papers on research on the mechanisms of change of psychosocial treatments (e.g., Kazdin, 2007), many studies that have aimed to test mechanisms have done so with secondary analyses of clinical trial data. In these studies, repeated measures of symptom severity and candidate mechanisms have been modeled to test mediation (degree to which change in mechanistic variables accounts for clinical change). However, the mediator is not randomized, which introduces confounding that limits causal interpretations of associations between mediator and outcome. Hence, mediation analyses that are correlational (e.g., parallel process models that relate the slope of the mediator to the slope of the outcome; or use of time-varying mediators to relate to time-varying outcomes (even lagged outcomes) without controlling for previous outcomes) will often overestimate the causal effect of the mediator on outcome. Researchers have developed a number of new approaches using variations of cross lag panel models that more accurately assess the causal effect of a mediator on outcome, e.g., random-intercept cross lag panel models, (Hamaker, Kuiper, and Grasman, 2015); latent change score models (McArdle and Hamagami, 2001); integrated cross lag and fixed effects models (Falkenström, Solomonov, and Rubel, 2022); difference in differences. To enhance accuracy of causal estimates in these cross-lag models, RCTs can include the following methods.

First, assessment of mediators and outcomes should be frequent. In any cross-lag model, the lag must accurately reflect the time that it takes for an effect to be observed in the dependent variable. Many studies assess mediators and outcomes infrequently (e.g., every 4 weeks). However, if the effect of the mediator is relatively rapid, any effect on the outcome may be washed out by the time the lagged outcome is measured. Similarly, if the effect is longer than the time between assessments, then no effect will be found. It is best to assess measures too often vs. not often enough, since time-series analysis can initially determine the appropriate lag time for cross-lag models. Further, these methods assume that data are missing at random (MAR). Frequent assessments also enhance the likelihood that data are MAR since, if dropout is related to scores on the outcome, recent assessments of the outcome should be related to the missingness, thereby meeting MAR.

Second, clinical trials should measure many possible mediators and include them simultaneously in the model; the vast majority of trials measure just one or two potential mediators. Any relation between these mediators and outcome may not reflect true causal relations but rather a correlation of measured mediators with true mediators (i.e., third variable confounds). This problem is present even with advanced cross-lag panel analyses which examine within-person relations between mediators and outcomes. Thus, if there are multiple plausible mechanisms through which the intervention affects the outcome, whether those are specific (e.g., expectancy violation) or non-specific (e.g., therapeutic alliance), all should be measured and included in analysis to ensure that relations between mediators and outcomes are not over-estimated due to omitted third variables. Sensitivity analyses can be conducted for models with and without covariates to examine model robustness.

3.5. Machine Learning Applications

An agenda that proposes increasingly complex theories about exposure mechanisms has a better chance of succeeding if it embraces methodological advances in statistical approaches designed to rigorously test such complex theories. Although the enhanced cross lag panel models described above can be generalized to multiple predictors, these models are very complex and often require multiple constraints to converge to a solution even when considering only one predictor of outcome. Thus, they may be impractical for teasing apart the exact mechanisms by which exposure impacts outcome. Machine learning (ML) methods provide a toolkit of flexible algorithms that may overcome some of the aforementioned limitations. Although ML has become increasingly mainstream in psychological science, the focus of application has been on prediction and classification (as opposed to causal inference). Yet many of the tools of ML can be applied to test novel mechanistic hypotheses in a robust manner designed to protect against pitfalls that often impact replicability and generalizability. In this section we provide an overview of how ML can be applied to examine both mediation and moderation.

Machine Learning in Mediation Analysis

Mediation is often tested in (linear) structural equation models, which can be adapted to incorporate multiple mediators, different types of confounding, and repeated measures (Moerkerke, Loeys, and Vansteelandt, 2015). Increasing the complexity of mediation models may facilitate the identification of the causal effects of interest; however, this also increases the potential for statistical misspecification (e.g., assuming linear associations when they are actually non-linear (Valente et al., 2017). To address this, many tools of machine learning (ML) that have been honed in the predictive modeling framework are increasingly applied in the causal inference literature (Blakely et al., 2019). ML methods (vs. traditional regression) are more flexible and make fewer assumptions. This means that if the true relations between treatments, mediators, and outcomes are non-linear, estimates from a data-adaptive ML algorithm are more likely to reflect reality, especially as sample sizes increase.

Modern approaches to specifying causal models and evaluating them statistically are heavily influenced by the counterfactual framework (Pearl, 2019). A key feature of this framework is that cause-and-effect questions are mathematically translated to methods for the estimation of potential outcomes at the individual level (Broadbent, 2015). Models are developed that use relevant pre-treatment variables (e.g., participant and contextual characteristics) to estimate outcomes after experimental treatment, and after control treatment. These outcome models can be applied to each individual to get an estimate of outcome after their actual treatment and after their counterfactual treatment. This is different from standard analysis where treatment effect is estimated through a comparison of group averages. Standard regression approaches can be used to develop outcome models, but make restrictive parametric assumptions that may be unrealistic. As an alternative, outcome models can be developed using data-adaptive machine learning algorithms that model complex non-parametric associations, including high-dimensional interactions (van der Laan and Rose, 2011).

A variety of machine learning methods have been developed that can incorporate estimates from outcome models along with estimates from treatment propensity models to enhance estimation of the treatment effect of interest (e.g., targeted maximum likelihood estimation and augmented inverse probability weighting; (van der Laan and Rose, 2011). Although adoption of these methods has been slow in the behavioral sciences, the combination of open-source machine learning software, increased access to high-performance computing, and larger sample sizes will facilitate increased application of modern causal inference approaches in future research on mechanisms of exposure therapy. Researchers interested in applying ML methods to test causal hypotheses could use a super learner (stacked ensembling) approach that includes a diverse set of ML algorithms which allows researchers to make fewer assumptions about which algorithm may be best suited for their data by incorporating estimates from multiple, diverse algorithms. The potential for overfitting is reduced by applying cross-estimation (or cross-fitting) procedures (Zivich and Breskin, 2021) that provide outcome estimates for each individual that are based on models that were developed in datasets that excluded that individual (analogous to cross-validation in the predictive modeling framework).

Machine Learning in Moderation Analysis

Exposure outcomes vary by patient, yet we lack a clear information about predictive patient profiles and/or mechanisms through which variation occurs. Work in this area has highlighted relevant factors such as motivation and insight (Abramowitz, Deacon, and Whiteside, 2019). To complement this variable-focused work (which has been based on traditional statistics or supervised ML), unsupervised machine learning (ML) takes a person-centered analytic approach. It can provide a more comprehensive data-driven profile of patients, especially with regard to how exposure mechanisms differentially affect individuals. Unsupervised ML encompasses multiple approaches that find homogeneous participant subgroup clusters within a set of heterogeneous variable responses. In the case of exposure, unsupervised ML can identify groups of patients who are more or less likely to respond to specific strategies and the specific conditions in which a number of candidate exposure mechanisms have effects. For example, De Nadai et al., (2023) investigated how fMRI-based neural mechanisms differentiated response to exposure-based treatment for OCD. By applying unsupervised ML to results across many OCD-relevant brain areas, they found three groups of patients. One group showed a neuroimaging pattern consistent with healthy controls and exhibited greater reduction in OCD symptoms with exposure therapy vs. general stress management training.

There are many options for implementing unsupervised ML and some are better suited for exposure research. Compared to areas where ML methods are often developed (e.g., computer vision research), behavioral research has substantially more measurement error, and ML methods developed in other fields often do not accommodate the resulting bias. Methods based on mixture modeling fit explicitly account for measurement error and are more likely to be robust in the face of such error. Examples include Gaussian mixture modeling (for continuous variable responses; also called latent profile analysis) and latent class analysis (when applied to categorical responses). Other unsupervised ML approaches (e.g., k-means, hierarchical clustering) rely on post-hoc approaches to model calibration (e.g., cross-validation), which can reduce robustness. Post-hoc calibration is problematic for clinical research, which often cannot readily provide the steady stream of data needed to continually refine and test the model. Mixture modeling also provides a more precise framework for identifying causes and consequences of identified patient clusters, through partial assignment of patients to each cluster in a way that acknowledges that imperfect cluster assignment and accounts for this in subsequent covariate analyses (Asparouhov and Muthén, 2014).

Promise and Pitfalls of Machine Learning

The above approaches must be considered in the context of potential tradeoffs. Many ML-based models are “black box” models where the specific contributing role of each individual variable remains unknown, thus provide little insight into theoretical underpinnings related to who will benefit from what treatment. This property of ML-based models can make it difficult to correct models that behave incorrectly, especially related to bias against various demographic groups (Pearl, 2019). Representative samples and humility with regard to individual predictors are essential. Out-of-sample cross-validation, while rare in the behavioral sciences, can mitigate some of these problems. Data analysis also does not occur in isolation, and many innovative methods require careful pre-trial design in to function properly.

Application of these methods also requires appropriate lags and assessment of relevant confounds (i.e., other possible mediators) as recommended for RCT design above (3.4). Further, ML enhanced propensity score approaches have been developed for dichotomous predictors/mediators, though few are dichotomous and transformation into such leads to loss of accuracy and power. Fortunately, researchers have recently proposed extensions of the propensity score approach to include continuous predictors. These methods, once fully developed, will provide an alternative analytical framework for causal mediation analysis.

4. Individual and contextual differences.

The most relevant question for future research may not be “Which mechanism matters most?” but “Which mechanism matters most for whom, and under what circumstances?” It is highly likely that successful exposure depends on multiple interrelated mechanisms, and that these mechanisms (and optimal methods/measures) will vary by individual, context, and/or timing (Rothman and Sheeran, 2021). As just one example, earlier translational work points to several contextual features that may underlie return of fear/avoidance following within-session extinction (Bouton, 2002). Contexts moderating extinction retention in animal models include internal affective state (e.g., degree of arousal), external contexts (e.g., smells, cage design), and passage of time. Human studies have replicated these effects with location (room), imagined location, and degree of arousal (e.g., Mystkowski et al., 2006). Accordingly, assessment of overall treatment outcome and session-by-session retention of extinction effects may need to consider context as a moderating variable in exposure efficacy. Similarly, failure to find meaningful differences in trials that compare exposure delivery approaches could be explained by moderation of treatment effects by individual or contextual differences. For example, trials may enroll individuals with varying deficits in fear extinction learning (e.g., some with deficits in initial learning while others experience problems with consolidation); the delivery variations being tested may not target the relevant deficit for all patients. Little work has been done to elucidate important contextual, individual, or timing variations in mechanisms over a full course of clinical treatment; below we explore potentially important factors to consider for future studies in this area.

4.1. Magnitude and Type of Emotion Elicited During Exposure

Factors that influence the magnitude and/or quality of distress during exposure may have implications for understanding mechanisms. For example, anxiety sensitivity (AS) has been conceptualized as an important amplifier of emotional/somatic distress (Otto et al., 2016), and this amplification may have important effects on safety learning during exposure. Patients high in AS may simultaneously face two types of fear cues: e.g., external cues arranged by the therapist and internal cues amplified by AS. Threat expectancies could be tied to both, and while one type may be disconfirmed (e.g., actual harm in the presence of the external cue) another is inadvertently confirmed (e.g., feeling overwhelmed by “dangerous” levels of anxiety). When this occurs, there is potential for the exposure to appear successful relative to external dangers but constitute a sensitization experience relative to internal dangers.

Furthermore, exposure classically targets emotions of fear and anxiety, but other types of distress can be prominent in individual presentations and/or symptom clusters, and are often targeted with exposure (e.g., disgust, incompleteness, shame, guilt). Disgust has been associated with contamination-related OCD, blood-injection-injury phobia, spider phobia (Cisler, Olatunji, and Lohr, 2009), and some forms of PTSD (Badour and Feldner, 2018). Different emotional targets may show different response patterns or have unique implications for treatment strategies or outcomes. Disgust has distinct neural and physiological substrates and is less closely tied to cognitions about negative outcomes, instead characterized by visceral discomfort (Cisler, Olatunji, and Lohr, 2009). In fact, disgust has been linked to evaluative conditioning, in which a CS acquires the valence (e.g., dislike) of a paired US. Disgust also has shown greater resistance to extinction than fear in some studies (e.g., Olatunji, Forsyth, and Cherian, 2007).

Incompleteness is a sensory-affective experience that things are “not just right,” involved in some presentations of OCD and OC-spectrum disorders (Summerfeldt, 2004). Like disgust, incompleteness is less tied to specific negative cognitions, and relevant exposures are self-confirming: uncomfortable feelings occur and individuals learn to tolerate them. Also like disgust, incompleteness may show distinct response patterns, such as a shallower decline or even a flat trajectory during exposure (Milgram et al., 2022). In one study of pediatric OCD, greater baseline disgust and incompleteness, but not fear, predicted poorer response to treatment that included exposure (Cervin and Perrin, 2021).

Shame and guilt have been associated with PTSD. Shame is a negative affective-cognitive state that entails judgment of the self, whereas guilt entails judgment of one’s actions (Saraiya and Lopez-Castro, 2016). In contrast to the present- or future-oriented focus of fear and anxiety, shame and guilt are past-focused emotions (Pugh, Taylor, and Berry, 2015). There is a small, mixed literature on the role of shame and guilt in the treatment of PTSD. Some studies suggest that traditional or prolonged exposure (PE) is less effective at targeting shame and guilt. These findings have facilitated development of tailored treatments, such as PE with imagery rescripting or cognitive processing therapy, to more directly target related cognitions (e.g., Saraiya and Lopez-Castro, 2016). In sum, extant studies suggest that several unique processes are involved in “non-fear” exposure targets, with potential implications for exposure strategies and outcomes. These emotions may decline more slowly during exposure or benefit from augmentation strategies. Future laboratory and clinical studies will benefit from concurrent measurement of multiple affective states, which will enable direct comparisons.

4.2. Developmental Considerations

Although exposure-based CBT is a first-line treatment with strong efficacy for youth with anxiety disorders, OCD, or PTSD (Freeman et al., 2018; Higa-McMillan et al., 2016), not all youth experience symptom remission and many fail to maintain gains over time (Kodal et al., 2018). Adolescents in particular appear to experience less benefit from exposure than both younger children and adults (Turner et al., 2018), and may also be less likely to maintain symptom remission through long-term follow-up–particularly when receiving exposure alone (i.e. without medication; 20–50% remission rate at 6 years; (Ginsburg et al., 2018). These clinical outcomes are paralleled in the laboratory by marked differences in extinction learning. Adolescent rodents and humans show developmentally normative deficits in the retention of extinction learning as indicated by higher threat responding at later extinction recall and in some cases poorer initial extinction learning (e.g., Pattwell et al., 2012). One benefit of animal models is the ease of testing species-homologous modifications in exposure therapy (via fear extinction protocols) and other interventions while taking into account discrete age boundaries. Such evidence can systematically track the influence of stages of brain development on responses to treatment. In contrast, despite apparent mechanistic and clinical differences across development, the vast majority of clinical trials include youth of all ages and have not been powered to examine differences based on age (Seligman and Ollendick, 2011). It will be critically important for future clinical studies to be sufficiently powered to test the most relevant mechanisms and examine treatment outcome differences across different phases of development.

Continued maturation of the prefrontal cortex and associated regulatory behaviors across adolescence has emerged as a potential explanation for why treatments that rely on prefrontal functioning may be less effective in pediatric populations. Extinction is largely dependent on the prefrontal cortex, development of which continues into the early 20s (Giedd et al., 1999). In contrast, the amygdala matures on a much faster timescale, reaching structural maturity during adolescence (Hu et al., 2013) while simultaneously exhibiting greater functional activation relative to adults in response to fear-associated stimuli (Lau et al., 2011). As structural and functional connectivity between prefrontal control regions and the amygdala strengthens into early adulthood, approaching a greater degree of top-down inhibition of the amygdala, the capacity for affective control increases. Clinical populations of youth (i.e. those with disorders treated by exposure) appear to experience even greater deficits during this developmental process (e.g., Liu, Bilek, and Fitzgerald, 2016; Zugman, Winkler, and Pine, 2021), leading researchers to suggest that some youth may benefit from targeted exposure augmentation strategies to improve top-down control (Fitzgerald, Schroder, and Marsh, 2021). Available evidence also suggests the potential benefit of intervention strategies that emphasize rewarding stimuli and circumvent the prefrontal cortex (e.g., Meyer et al., 2023). Importantly, while continued maturation of prefrontal circuitry among youth may contribute to deficits in exposure learning, it also presents a unique opportunity for intervention during a developmental window in which neural mechanisms for cognitive control are most amenable to modulation (Fitzgerald and Taylor, 2015). Novel augmentation strategies that bolster functioning in key prefrontal circuitry during this time (e.g., non-invasive neuromodulation, Garnaat et al., 2019); behavioral strategies; Fitzgerald, Schroder, and Marsh, 2021) have potential to produce potent and enduring effects into adulthood.

In addition to neurobiological changes, the socioemotional landscape shifts dramatically across development, particularly during adolescence when school and social activities reorient towards same-aged peers and away from caregivers and immediate family. Another potential explanation for attenuated exposure outcomes during adolescence is that caregivers are less involved in treatment (vs. for younger children), but adolescents have not yet independently mastered skills deployed consistently by most adults. Additional research is needed to determine whether age-related effects relate to true developmental differences, use of compensatory treatment strategies (e.g., caregiver involvement), or both. Longitudinal studies will also be essential for future work in this area, as work to date has largely been cross-sectional. This will facilitate clearer understanding of the developmental trajectory underpinning mechanistic changes for youth and may aid in identifying critical or sensitive windows for intervention.

Another potential challenge for understanding mechanisms in exposure-based treatments for youth relates to the common inclusion of various developmental adaptations in treatment packages. In general, packages for youth (vs. adults) are more likely to include tailored delivery strategies (e.g., psychoeducational content, community/school-based delivery) and greater emphasis on other components in addition to exposure (e.g., relaxation, cognitive restructuring, caregiver training; Chorpita and Daleiden, 2009). Although there is clear empirical support specifically for the inclusion of caregiver/family-based treatment components (e.g., Freeman et al., 2018), it is important to note that a higher proportion of in-session exposure robustly predicts clinical improvement among youth (Peris, et al., 2017)--suggesting that exposure should continue to be emphasized as a primary and necessary treatment component in CBT packages for youth. Notably, findings in the pediatric OCD literature (where protocols typically include a higher proportion of sessions with exposure compared with those for anxiety disorders, Franklin, Freeman, and March 2019; Freeman and Garcia, 2008) seem to show better long-term maintenance of gains (e.g., 73% remission at 3 years; (Melin et al., 2020); 61% at 7–9 years; (Fatori et al., 2020). Although these trials cannot be directly compared for a variety of reasons, future studies may wish to explore whether a higher “dose” of exposure can mitigate against deficits in extinction learning experienced by youth.

Finally, additional consideration should be given to measurement differences that arise across development. Younger children are not yet able to articulate cognitions regarding threat and distress in the same way as adults or even adolescents (Freeman, Flessner, and Garcia, 2011). Youth also exhibit less insight, manifesting as difficulty with cognitive processing and effectively articulating symptoms (Geller et al., 2001). This likely makes it more difficult to leverage cognitive constructs (e.g., expectancy violation) as a clinical tool. It also presents challenges for measuring multiple constructs via self-report, as it may be difficult for youth to reliably report on differential changes across those constructs (e.g., distress vs. expectancy vs. stimulus valence). Although enhanced training (e.g., through psychoeducation) could be considered for bolstering this skill in youth, it will likely be difficult to do so without inadvertently contributing to response bias (see 2.1). Burden associated with multiple measures is also likely to be higher among youth than adults (Norris et al., 2023), particularly through forward translation into clinical settings. Going forward, it will be important to empirically identify optimal streamlined approaches to self-report measurement of mechanism in youth of various ages. Future studies may also wish to place greater emphasis on objective measures (e.g., psychophysiology, observer ratings) as part of multimodal mechanism measurement for youth.

4.3. Considerations for Culturally Diverse Populations

Thus far, findings from randomized clinical trials examining exposure-based interventions have found mostly comparable outcomes for individuals from diverse backgrounds on measures of clinical response, remission, symptom severity and overall functioning (e.g., McLean et al. 2022; Kline, Feeny, and Zoellner, 2020). However, many of these have been limited by small sample sizes, inadequate analyses examining race and ethnicity, and lack of consideration of intersectionality or sociocultural context (beyond a checkbox identity). Few studies have examined whether hypothesized mechanisms are differentially important for individuals from different historically marginalized communities (e.g., based on race, ethnicity, sexual/gender identity, income). It is likely that sociocultural processes experienced by these groups have an impact on exposure mechanisms. As an example, discrimination is an adverse experience associated with physiological, cognitive, and behavioral changes that increase vulnerability for experiencing anxious symptoms (Lara-Cinisomo et al., 2016; Pascoe and Smart Richman, 2009). According to models of stigmatized identity, experiences of discrimination are associated with increased vigilance to negative cues in the environment (Helzer, Connor-Smith, and Reed, 2009; Pachankis, 2007), a process associated with heightened levels of anxiety and increased levels of avoidance (Cisler and Koster, 2010). Such cognitive vigilance and physiological reactivity may negatively impact mechanisms relevant to exposure therapy, such as experiential avoidance and threat expectancies. Findings from a primary care sample of Latine adults suggest that discrimination has an indirect effect on depression, social anxiety, and anxious arousal symptoms through experiential avoidance (Zvolensky et al., 2022). Indeed, overreliance on such strategies may contribute to overall psychological distress and impact willingness to experience distressing emotions during exposure. Neurobiological findings also suggest that exposure to negative life experiences (inclusive of violence exposure, low family income and neighborhood disadvantage) impact mental health and treatment through influence on brain regions that support threat-related emotional functioning. In a prospective study of adolescents who identified as Black-Americans and White Americans, which included self-report questionnaires and a Pavlovian fear conditioning task during fMRI, racial differences in neural (fMRI activity), behavioral (threat expectancy), and psychophysiological responses to threat (Harnett et al., 2019) were found. Importantly, racial differences in brain activity to threat were smaller after accounting for negative life experiences. These findings highlight the impact that such experiences may have on symptoms and exposure mechanisms. Sociocultural background may also influence the reliability and validity of self-report measures used to assess exposure mechanisms. For example, individuals from a wide variety of cultures worldwide may be most likely to endorse anxiety using somatic or culture-specific terms (Lewis-Fernández et al., 2010) and men may be more likely than women to describe feelings of anger or irritability when experiencing anxiety and/or sadness (Genovese et al., 2017). Reliance on assessment of limited cognitive or emotional experiences during exposure may miss important information about the experience of mechanistic change for these groups. Significant work remains for the field to understand exposure mechanisms, best methods for measuring them, and treatment response across different sociocultural groups, particularly for those that have been historically marginalized. This information will be critically important for future efforts to optimize culturally-responsive assessment and care in the context of exposure therapy.

Discussion

Coordinated action is needed to accelerate research on exposure therapy mechanisms, particularly related to key issues of conceptualization, measurement, design, analysis, and individual differences. This paper outlines an initial agenda for addressing specific questions in each of these priority areas. In addition to specific recommendations for future work in each area (see Table 2), several overarching themes emerge. First, considerable work is needed to conceptually integrate overlapping mechanistic constructs, disentangle these from facilitators or read-out measures, and establish consistent mechanistic terminology with clear operational definitions. As has been initiated for other psychological constructs with clear ties to important clinical outcomes (e.g., Eisenberg et al., 2019), extinction research can benefit from focused ontological research to provide greater specification of the link between the construct of extinction and its observable measurement. This will improve replicability, aid in identifying the most potent mechanisms, and is likely to require ongoing collaborative work across research groups. Similarly, there is a pressing need for open-source, empirically supported tools that might be deployed across trials, settings, and/or research groups (e.g., reliable and valid measures of mechanism, detailed fidelity measures, clinician/rater training and monitoring protocols, treatment manuals, analytic code, variable calculation procedures, other standard operating procedures). Such a “toolbox” could improve rigor and replicability across nearly all priority areas outlined in this paper and—as empirically-supported tools become available— might be hosted by the ETC (see Smits et al., in press) or other groups with a similar mission. A final overarching theme relates to the need for larger sample sizes in mechanistic trials, which will facilitate examination of individual differences, bolster power in the presence of measurement error, and enable use of advanced analytic techniques. A central goal of the ETC is to facilitate research collaborations across sites, leading to larger sample sizes through enhanced recruitment and/or data sharing (Smits et al., in press). As research in these areas progresses and new priorities emerge, we anticipate that the present agenda will continue to evolve. We anticipate that this will eventually lead to new or improved theoretical models of fear and anxiety, which in turn will produce new research questions for the future.