Moderators of outcome for youth anxiety treatments: Current findings and future directions

Abstract

Although multiple treatments for youth anxiety have established efficacy, there is considerable heterogeneity in treatment response. To improve outcomes and create more personalized interventions, the field has sought to identify moderators of treatment response (variables that specify which treatments work for whom and under what conditions). Despite research effort, few consistent moderators have emerged. The majority of studies to date have examined variables of convenience rather than theory-driven variables, including demographics (age, sex, race, ethnicity, socioeconomic status), pretreatment youth characteristics (symptom severity, primary diagnosis, comorbidity) and pretreatment parent variables (parent psychopathology, parenting). Findings from these studies are reviewed and future directions for the identification of meaningful moderator variables are discussed.

Introduction

Several treatments have shown efficacy in the treatment of child and adolescent (referred to hereafter as youth) anxiety. An evidence-based update identified six treatments that met criteria for a well-established treatment: cognitive behavioral therapy (CBT), exposure, modeling, CBT including parents, education, and CBT plus medication (Higa-McMillan, Francis, Rith-Najarian, & Chorpita, 2016). Other treatments have been identified as probably or possibly efficacious. Despite the availability of these treatments, a meaningful portion of youth do not respond or do not respond optimally. Current thinking suggests that a more personalized approach to treatment, within the treatments found to be efficacious, may improve treatment response.

To move toward the goal of personalized intervention, moderators will need to be identified, and the successful identification of moderators requires adoption of common definitions across studies. Moderators are pre-randomization characteristics that identify which treatments work for whom and under what conditions (Baron & Kenny, 1986; Holmbeck, 1997; Kendall & Comer, 2011; Kraemer, Wilson, Fairburn, & Agras, 2002). These variables do not explain treatment main effects. Instead, moderators interact with treatment condition on treatment outcome measures, such that treatment response for a single individual depends on that individual’s level of the moderator present at baseline (Kraemer et al., 2002). Pre-randomization characteristics that show a main, but not interactive, effect on outcome are referred to as non-specific predictors of outcome. Predictors identify for whom treatments work more broadly, whereas moderators go a step further to facilitate evidence-based treatment decision making (i.e., which treatment to implement for which client)¹.

To date, few variables have been identified as consistent moderators for youth anxiety treatments, although occasional studies have found that certain variables moderated specific outcomes. The variables that have been examined most frequently include readily available demographics (age, sex, race, ethnicity, socioeconomic status), pre-treatment youth clinical characteristics (symptom severity, primary diagnosis, comorbidity) and pre-treatment parent variables (parental psychopathology, parenting). Although there are theoretical constructs that are intriguing as potential moderators, few studies have examined them. Findings from these studies are discussed below² and future directions are outlined.

Demographic variables

Chronological age

Age is perhaps the most frequently examined potential moderator, with some significant findings. Results from early studies examining the efficacy of CBT with additional family components (i.e., family management and parental anxiety management) suggested that younger (ages 7–10) but not older (ages 11–14) youth were more likely to respond to CBT plus family management at posttreatment and 12-month follow-up compared to CBT alone (Barrett, Dadds, & Rapee, 1996), although there were no differential treatment effects across age groups when CBT was compared to CBT plus parental anxiety management (Cobham, Dadds, & Spence, 1998) or CBT plus cognitive parent training (Nauta, Scholing, Emmelkamp, & Minderaa, 2003). Another study found that older children (dichotomized using a median split of 13.08 years) self-reported greater reductions in social anxiety following attention bias modification treatment (ABMT) compared to an attentional control training (ACT; Pergamin-Hight, Pine, Fox, & Bar-Haim, 2016). In the School-Based Treatment for Anxiety Research Study (STARS), which examined effectiveness of school-based clinician delivery of modular CBT (M-CBT), M-CBT had stronger effects relative to treatment as usual (TAU) for children who were older (1 SD above the mean age of 14.1), although this difference was not maintained at 1-year follow-up (Ginsburg, Pella, Pikulski, Tein, & Drake, 2019).

Of the many studies that have examined age, the majority have found that age does not significantly moderate outcome across various treatment types, including individual CBT (ICBT; Kendall et al., 1997b; Silk et al., 2018), group CBT (GCBT; Silverman et al., 1999b) family-based CBT (FCBT; Bodden et al., 2008; Jongerden & Bögels, 2015; Kendall, Hudson, Gosch, Flannery-Schroeder, & Suveg, 2008; Shortt, Barrett, & Fox, 2001, Suveg et al., 2009), one session treatment (OST) for specific phobias (Ollendick et al., 2009; Ollendick et al., 2015), social effectiveness therapy for children (SET-C; Alfano et al., 2009; Beidel, Turner, & Morris, 2000), metacognitive therapy (MCT-c; Walczak, Breinholst, Ollendick, & Esbjørn, 2019), fluvoxamine (Vitiello & GROUP, 2003), and cognitive bias modification of interpretations treatment (CBM-I; Krebs et al., 2018). Age was also not a moderator for any of the treatment conditions examined in the child/adolescent anxiety multimodal study (CAMS; Walkup et al., 2008), which included CBT (i.e., Coping cat), medication (i.e., sertraline; SRT), combination of CBT and medication (COMB) and placebo (PBO; Albano et al., 2018; Compton et al., 2014; Sanchez et al., 2019), nor were CBT effects found to be different in a review of CBT for anxious adolescents (Kendall & Peterman, 2015).

Of note when considering age, samples were frequently divided into “younger” and “older” age groups using inconsistent grouping methods (i.e., varying pre-determined cut-offs, median split). Although age does not appear to moderate CBT response when assessed either dimensionally or categorically (Bennett et al., 2013), use of inconsistent cut-offs limits cross-study comparisons and dichotomizing samples increases risk for Type I error (Kraemer, Frank, & Kupfer, 2006). In addition, arbitrary age groupings, which are presumably used as a proxy for child versus adolescent developmental stage, do not allow for a nuanced examination of how individual variation in pubertal timing might influence treatment response. The theory behind examining developmental stage is sound (e.g., that for peer-group focused adolescents the role of the parent in treatment may be of less importance), but chronological age is likely a less-than-ideal measure to test this theory.

Sex

Some evidence suggests that biological sex moderates outcome. Studies have shown that females respond better to CBT with increased family involvement, including additional family management (Barrett et al., 1996) and parent anxiety management when caregivers had high anxiety (Cobham et al., 1998). Although sex was not a moderator in CAMS (Compton et al., 2014), secondary analyses indicated that males showed decreased parent-reported school impairment compared to females when randomized to medication treatments (SRT and COMB) relative to PBO, suggesting that medication may be a particularly helpful treatment to target anxiety-related school impairment for males (Sanchez et al., 2019).

Some studies reported that sex moderated follow-up, but not posttreatment, response, particularly when outcome measures beyond symptom reduction were examined. For example, females randomized to complete OST were more likely than males to be diagnosis-free at 6-month follow-up (Ollendick et al., 2009), although sex did not moderate outcome in another trial examining parent-augmented OST (Ollendick et al., 2015). Sex did not moderate response in a trial comparing CBT, FCBT and a family-based educational control (Kendall et al., 2008), but secondary analyses of the dataset found that females showed more improvement from posttreatment to 1-year follow-up on mother-reported (but not father-reported) social competence when randomized to CBT compared to FCBT (Suveg et al., 2009). Consistent with the importance of examining anxiety-related school impairment as an outcome for males, Suveg et al. (2009) also found that males showed significantly more positive change on mother-reported school performance at 1-year follow-up in both CBT conditions compared to the active control treatment.

Although some results point to sex as a potential moderator of follow-up outcome and functional treatment response (i.e., social competence, school impairment), the majority of studies have reported that sex does not moderate outcomes. Specifically, sex did not moderate outcomes for CBT (Kendall et al., 1997a; Silk et al., 2018), ICBT plus cognitive parent training (Nauta et al., 2003), family-based group CBT (Shortt et al., 2001), GCBT (Manassis et al., 2002; Silverman et al., 1999b), fluvoxamine (Vitiello & GROUP, 2003), SET-C (Beidel et al., 2000), MCT-c (Walczak et al., 2019), CBM-I (Krebs et al., 2018) or in the STARS trial (Ginsburg et al., 2019). Again, questions emerge regarding the theory behind examining sex as a potential moderator, and whether sex or gender is the construct of interest. If sex is a proxy for identification with stereotypical gender roles (i.e., a comparatively higher focus on relationships among females, a decreased emphasis on sharing emotional experiences among males), then theory-driven continuous, rather than binary, measures should be examined.

Race/ethnicity

Few studies have examined whether race or ethnicity are moderators. Although race/ethnicity did not moderate response in the CAMS trial (Compton et al., 2014), a secondary analysis using receiver operating curves found that Hispanic ethnicity was associated with more severe anxiety symptoms after CBT per independent evaluator (IE) report, but more severe anxiety after SRT per parent report (Taylor et al., 2018). That said, most studies have found that treatments are equivalent across racial and ethnic groups, including for ICBT (Pina, Silverman, Fuentes, Kurtines, & Weems, 2003), GCBT (Silverman et al., 1999b), fluvoxamine (Vitiello & GROUP, 2003), SET-C (Beidel et al., 2000) and in STARS (Ginsburg et al., 2019).

A frequently noted, but infrequently addressed, limitation of most RCTs is limited sample diversity (e.g., the mode proportion of Hispanic/Latinx participants in 16 youth anxiety trials was zero; Pina et al., 2003), although cross-cultural investigations are currently being conducted (e.g., Wong et al., 2019). As a result of this limitation, multiple studies have collapsed race/ethnicity categories into “white” versus “other,” thus losing within-group cultural variation (i.e., cultural beliefs regarding the use of mental health services, individualistic versus collectivistic orientation) and centering whiteness. In one of the few studies to date to recruit a more diverse sample (40.9% Caucasian, 59.1% Hispanic/Latinx), Latinx ethnicity or Spanish language did not moderate outcomes for a culturally-sensitive prevention/early intervention CBT approach with and without parent involvement, although within-group heterogeneity for the Hispanic/Latinx group was not examined due to sample size (Pina, Zerr, Villalta, & Gonzales, 2012). Thus, it appears that race and ethnicity are not moderators of outcome, although issues of low power and resulting type II error (false negatives) rates are pronounced when examining the impact of multiple, intersecting identities (Crenshaw, 1989; Rosenthal, 2016) on response to treatment.

Socioeconomic status

Two studies examined whether socioeconomic status (SES), assessed using yearly family income and parental highest education level, identifies which treatments work for whom. In CAMS, membership in a “high” or “low” SES group did not moderate treatment response, although a measure of caregiver strain that included items assessing financial stress was a predictor of poorer outcomes (Compton et al., 2014). There were no significant moderation effects of SES in another trial examining fluvoxamine efficacy (Vitiello & GROUP, 2003). Taken together, both studies suggest that SES is not associated with differential treatment response. However, future studies should adopt a more nuanced approach to assessment of the complex SES construct (e.g., disposable income, social capital available for use in childcare, neighborhood, family distress associated with financial difficulties). The recruitment, and retention, of more socioeconomically diverse samples is also recommended.

Characteristics of participant youth

Global anxiety severity

Global anxiety symptom severity was identified as a predictor of treatment response in one review (Compton et al., 2014) but not in others (Knight, McLellan, Jones & Hudson; Nilsen, Elsemann & Kvernmo, 2013). However, this variable does not emerge as a consistent moderator. One study found that youth with moderate levels of fear and avoidance showed stronger response to intensive CBT for adolescent panic disorder compared to youth with more severe fear and avoidance patterns (Elkins, Gallo, Pincus & Comer, 2016). Anxiety severity also moderated outcomes in the STARS trial, such that M-CBT was more effective for youth with more severe baseline anxiety, whereas those with less severe anxiety were more likely to respond to TAU (Ginsburg et al., 2019). However, this difference was not maintained at 1-year follow-up and other studies have not found that symptom severity significantly moderates response for CAMS (Compton et al., 2014), CBT (Silk et al., 2018), fluvoxamine (Vitiello & GROUP, 2003) or MCT-c, although there was a trend such that youth with more severe symptoms responded better to CBT than MCT-c (Walczak et al., 2019). Duration of anxiety symptoms, a potential proxy for anxiety severity, was also not a moderator of response to ICBT versus ICBT including cognitive parent training (Nauta et al., 2003).

Principal diagnosis

Youth principal diagnosis, particularly social anxiety disorder (SoP), is one of the more consistent moderators of treatment response. Principal diagnosis moderated outcome in CAMS, although only on one of two outcome measures (Compton et al., 2014) and not for parent and self-reported outcome at posttreatment or 36-week follow-up (Albano et al., 2018). Results from initial moderator analyses found that youth with generalized anxiety disorder (GAD) and separation anxiety disorder (SAD) responded best to COMB (GAD: COMB>CBT>SRT>PBO; SAD: COMB > SRT = CBT = PBO), whereas youth with SoP responded best to treatments including medication (COMB = SRT > CBT = PBO; Compton et al., 2014). Consistent with results from CAMS, another study suggested that medication was a particularly efficacious treatment for SoP (Birmaher et al., 2003). In this study, individuals with SoP, but not GAD and SAD, showed better clinical and functional response to medication (fluoxetine) compared to placebo, although sample sizes were small and comorbidity was high. SoP has also been shown to moderate response to individual and group treatments, but the direction of the relationship is unclear, with one study suggesting that youth with high social anxiety self-reported a better response to ICBT (Manassis et al., 2002) and results from other studies suggesting that a group format was optimal per some, but not all, measures (Liber et al., 2008; Wergeland et al., 2014). This differed from findings for other primary diagnoses (GAD: ICBT > GCBT; SAD: ICBT = GCBT; Wergeland et al., 2014). Finally, youth with SoP were more likely to respond to M-CBT than treatment as usual in the STARS trial at posttreatment (but not 1-year follow-up), whereas youth without SoP were more likely to respond to treatment as usual (Ginsburg et al., 2019). Presence of GAD or SAD did not moderate outcomes in this study. It is notable that the pattern of findings examining SoP as a moderator mirrored the findings for symptom severity in this study, which raises the possibility that SoP may be a proxy for overall symptom severity. Empirical evaluation of this question is warranted.

Although a pattern of support for social anxiety as a potential moderator variable is evident, other studies have not found a moderating role for principal disorder and thus significant results should be interpreted with caution. For example, studies examining individual CBT (Kendall et al., 1997a; Silk et al., 2018), fluvoxamine (Vitiello & GROUP, 2003), OST (Ollendick et al., 2009) and CBT with parent anxiety management training (Cobham et al., 1998) and family management (Barrett et al., 1996) have not shown that principal disorder moderated response. In addition, comorbid SoP did not moderate response to MCT-c for children with primary GAD, although there was a trend in favor of CBT (Walczak et al., 2019), and severity of social anxiety symptoms was not a moderator for ABMT across most measures, although some findings suggested that adolescents with less severe social anxiety responded better to ABMT than control at 3-month follow-up (Ollendick et al., 2019).

Comorbidity

Comorbidity (both within anxiety disorders and across other disorders) is the norm (Kendall, Brady, & Verduin, 2001), which may limit the utility of examining “principal” anxiety disorder as a potential moderator. Nevertheless, several studies have examined whether general comorbidity (i.e., presence or absence of a comorbid diagnosis) moderates outcomes and found limited evidence (for a review, see Ollendick, Jarrett, Grills-Taquechel, Hovey & Wolff, 2008), including in trials of ICBT (Kendall et al., 1997a; Kendall, Brady & Verduin, 2001; Silk et al., 2018), GCBT (Silverman et al., 1999b), family-based group CBT (Shortt et al., 2001) and SET-C (Beidel et al., 2000).

Other studies have examined whether specific comorbidities, particularly externalizing symptoms, are moderators; findings are again inconsistent. One study found that youth with higher levels of ADHD symptoms benefited more from FCBT than CBT, although this was only at 1-year follow-up and findings were inconsistent across measures (Maric, van Steensel, & Bogels, 2018). Externalizing disorders did not moderate CAMS outcomes (Compton et al., 2014), but a secondary analysis of CAMS data reported that youth with comorbid ADHD had worse posttreatment outcome in CBT, but not other conditions, compared to those without comorbid ADHD (Halldorsdottir et al., 2015). This difference was not maintained at 6-month follow-up, and there were no differences for youth with comorbid ODD, although another secondary analysis of CAMS reported that more rule breaking behaviors were associated with worse parent-reported outcome in CBT (Taylor et al., 2018). Results from other studies have shown that hyperactive symptoms do not moderate response to GCBT compared to CBT (Manassis et al., 2002), that comorbid ADHD did not moderate response to fluvoxamine (Vitiello & GROUP, 2003), and that externalizing symptoms did not moderate response to OST or parent-augmented OST (Ollendick et al., 2015).

With regards to comorbid depression, a secondary analysis of CAMS data found that more severe youth-reported depressive symptoms were associated with less severe youth-reported anxiety symptoms after both therapy treatments (Taylor et al., 2018). Another study reported that youth with lower baseline depressive symptoms unexpectedly showed a more marked response to fluvoxamine compared to placebo, which was driven by lower improvement in placebo within the low severity group, although it is notable that a diagnosis of major depressive disorder (MDD) was an exclusion criterion for the study (Vitiello & GROUP, 2003). Interestingly, there is some suggestion that comorbid affective disorder (i.e., major depressive disorder and/or dysthymia) may account for differential treatment effects within SoP samples (Crawley, Beidas, Benjamin, Martin, & Kendall, 2008), although depressive symptoms did not moderate outcomes for SET-C treatment for youth with SoP (Alfano et al., 2009). In addition, comorbid depression was not a moderator for GCBT (Berman, Weems, Silverman & Kurtines, 2000).

Only one study examined comorbid autism symptoms as a potential moderator (Puleo & Kendall, 2011), although comorbid anxiety is common among youth with autism spectrum disorder (ASD; White, Mazefsky, Dichter, Chiu, Richey & Ollendick, 2014) and research has shown that CBT is an efficacious treatment for anxiety for youth with comorbid ASD and anxiety (Wood et al., 2020). In this study, children with moderate autism symptoms were less likely to improve in ICBT than typically developing children, even after controlling for anxiety severity and social anxiety symptoms, although this was not the case for FCBT. Another study found that comorbid selective mutism moderated CAMS outcomes, such that IE-rated anxiety was more severe after COMB and parent-rated anxiety was more severe after CBT for youth with comorbid selective mutism (Taylor et al., 2018). The same study found that (a) increased separation anxiety was associated with lower IE- and parent-reported anxiety after COMB, but higher parent-reported anxiety after SRT and (b) increased panic symptoms were associated with lower youth-reported anxiety after COMB, but higher parent-reported anxiety after CBT.

Characteristics of participant parents

Several parent variables have been examined as potential moderators, particularly parent anxiety. Although parent state and trait anxiety did not moderate outcomes in CAMS (Compton et al., 2014), secondary latent growth curve analyses showed unexpectedly that higher levels of parental anxiety were associated with a faster and larger symptom reduction compared to youth with low anxious parents only within SRT (Gonzalez et al., 2015). Another study found that older youth (ages 11–14) with at least one anxious parent were less likely to respond to individual CBT (Cobham et al., 1998). In a trial examining ICBT and FCBT, father anxiety moderated response when 1-year follow-up outcomes were examined, such that children randomized to ICBT who had an anxious father were less likely to show improvements in follow-up than children randomized to FCBT or control; this was not the case for youth with an anxious mother (Kendall et al., 2008). In another trial of FCBT, ICBT was more efficacious than FCBT on some measures when children had a parent with an anxiety disorder themselves (Bodden et al., 2008). However, parent anxiety was not a moderator for parent-augmented OST (Ollendick et al., 2015).

Some findings suggest that global parental psychopathology may be an important factor to consider for young children and for individual compared to group treatment (Berman, Weems, Silverman, & Kurtines, 2000). Although there was no support for global parental psychopathology or family mental health treatment history as a moderator in CAMS (Compton et al., 2014), a secondary analysis found that more severe global parental psychopathology predicted lower IE-, child-, and parent-rated anxiety severity after SRT, along with higher IE-reported anxiety after CBT and lower parent-reported anxiety after COMB (Taylor et al., 2018), consistent with unexpected previous findings (Gonzalez et al., 2015). Other parenting variables have been examined, including parent overprotection (Ollendick et al., 2015), caregiver strain and family dysfunction (Compton et al., 2014), but have not been shown to moderate outcomes.

The lack of consistent findings for parenting variables is surprising given the role that parental factors play in the maintenance of youth anxiety disorders and the fact that parental psychopathology, family stress and parenting behaviors have been shown to predict treatment response (e.g., Compton et al., 2014; Cooper, Gallop, Willetts & Creswell, 2008; Crawford & Manassis, 2001; Keeton et al., 2013; Schleider et al., 2015; Settipani, O’Neil, Podell, Beidas, & Kendall, 2013). Counterintuitive findings may be due to the current focus on broad parenting variables rather than more specific behaviors that are more directly targeted in treatment, like parental accommodation (Kagan, Peterman, Carper, & Kendall, 2016). It may also be the case that parents are still key players even in the context of individual treatments for youth. For example, within the individual CBT Coping Cat protocol, parents meet with clinicians at the close of each session and there are two parent sessions. This represents a potential complication for identifying who would respond optimally to individual versus family treatments.

Other variables

Although the majority of studies have focused on the variables discussed above, studies have explored other potential moderators, including emotion dysregulation (Suveg et al., 2018), treatment expectancies (Compton et al., 2014; Norris et al., 2019b), referral source (Compton et al., 2014), baseline intellectual level (Vitiello & GROUP, 2003) and whether treatment type moderated the relationship between therapeutic relationship and outcome (Shirk & Karver, 2003), with results indicating nonsignificant moderation. A secondary analysis of CAMS data found that child-rated coping skills at baseline was a moderator, such that increased coping skills were associated with better child-reported outcomes after CBT (Taylor et al., 2018). This study also reported that increased thought problems were associated with better child-reported response to SRT, while school phobia was associated with worse parent-reported response to SRT and better parent-reported response to CBT, although findings were not consistent across informants.

Two studies examined informant response styles. Results from one study showed that youth who disclosed high distress in an initial evaluation made greater treatment gains, suggesting that “under-reporting” youth may respond less optimally to CBT (Panichelli-Mindel, Flannery-Schroeder, Kendall, & Angelosante, 2005). Similarly, youth who self-reported lower symptoms compared to parent-report were less likely to remit when randomized to CBT, but not other CAMS treatments (Becker-Haimes, Jensen-Doss, Birmaher, Kendall, & Ginsburg, 2018). It is possible that “under-reporting” reflects a lack of treatment “buy in” on the part of the youth or perhaps a lack of insight into symptoms that may be required for psychological treatments. It has also been suggested that this effect may be explained in part by increased parental psychopathology (Becker-Haimes, Jensen-Doss, Birmaher, Kendall, & Ginsburg, 2018).

Within the context of attention-based treatments, attentional control and threat-attention bias have been examined as moderators, with mixed results. Attention bias has not been shown to moderate outcomes (Ollendick et al., 2019; Pergamin-Hight et al., 2016; Pettit et al., 2020), although findings for attentional control have been more limited. One study found that children with lower attentional control benefitted more from ABMT relative to control (Pergamin-Hight et al., 2016). However, other studies have found both that higher attentional control was associated with better response to ABMT (Ollendick et al., 2019) and that attentional control was not a moderator (Pettit et al., 2020).

An emerging body of literature has examined biomarkers of treatment response using both genetic and neuroimaging measures. These studies have potential for elucidating the mechanisms of action underlying youth anxiety treatments, but likely do not have utility in most real-world clinics. In one study, a baseline polygenic score indexing environmental sensitivity was found to moderate outcome, with individuals who were high in environmental sensitivity showing better response to CBT compared to group or parent-led CBT (Keers et al., 2016). Moderation analyses examining neural emotional face processing have not yielded significant findings (Bunford et al., 2017; Burkhouse et al., 2017; Kujawa et al., 2016). Of note, these neuroimaging studies used small samples (N≤41) and allowed self-selection into treatment conditions. Thus, the search for moderator biomarkers is still in its infancy.

Future directions

Group to individual generalizability and new analytic approaches

A critical question facing those searching for moderators, and social science researchers more broadly, is the degree to which group-level data are generalizable to the individual and stable across time (i.e., whether moderation processes are ergodic; Molenaar, 2004; Molenaar 2013). The majority of moderation studies to date have tested group-level moderators at a single timepoint, resting on the assumption that results will have direct applicability to an individual client at the time that they present to a clinic (i.e., the ecological fallacy; Robinson, 1950) and will thus facilitate clinical decision making. The validity of this assumption has gone largely unchecked in the moderation literature, but empirical demonstrations in other contexts indicate that aggregated estimates at the group-level are inconsistent with individual-level estimates (e.g., Borkenau & Ostendorf, 1998; Fisher, Medaglia & Jeronimus, 2018; Hamaker, Nesselroade & Molenaar, 2007; Molenaar, Huizenga & Nesselroade, 2003; Molenaar, 2004; Na et al., 2010). For example, Fisher and colleagues (2018) compared intra- and inter-individual variation across six datasets and found that, while mean estimates were generally consistent across levels, variance estimates were two to four times larger at the individual level. Such findings call into question the clinical utility of group-level moderation analyses, even if consistent group-level moderators are ultimately identified. If the spirit behind the search for moderators is to help practitioners answer the question “which treatments work for which client,” then concerns about ergodicity suggest that group-level approaches may not be providing a complete answer.

Assuming low group to individual generalizability, an argument can be made for moving entirely away from group-level approaches, bringing the focus instead towards person-level moderators derived from group-level moderators or fully idiographic methods (Fisher, 2015; Hamaker, 2012; Molenaar, 2004). We agree that group-level moderation analyses are insufficient on their own to meaningfully inform personalized intervention efforts. However, group-level moderators may be informative in some cases, depending on where moderation processes fall along the continuum of ergodicity (Adolf, Schuurman, Borkenau, Borsboom & Dolan, 2014; Adolf & Fried, 2019; Brose, Voelkle, Lovden, Lindenberger & Schmiedek, 2015; Voelkle, Brose, Schmiedek & Lindenberger, 2014; Voelkle, Gische, Driver & Lindenberger, 2019; for an opposing argument see Medaglia, Jeronimus & Fisher, 2019). In addition, both within- and between-person analyses have limitations, and address different, often complementary, questions. Thus, consistent with Hayes et al. (2019), we argue for a hybrid approach, emphasizing that group- and individual-level analyses should be interpreted and reported within the context of the question each approach is designed to address (Hamaker & Ryan, 2019).

To the extent that current approaches continue, the ergodic assumption should be assessed and evaluated, assuming limited group- to individual-generalizability until demonstrated otherwise (Medaglia et al., 2019). This could involve use of statistical approaches that better account for individual differences (e.g., multi-level models) or instituting norms requiring reports of ergodicity (e.g., Fisher et al., 2018; Kievit, Frankenhuis, Waldorp & Borsboom, 2013). Such norms might also require adherence to a consistent analytic approach more broadly, as there is considerable cross-study variation in reported analyses [e.g., chi square group comparisons rather than examining interactions as specified by Kraemer and colleagues (2002)]. Concurrent with these efforts, increasing emphasis should be placed on person-centered methods. Such methods are beginning to be implemented. For example, Wallace and colleagues (2017) applied an optimal combined moderator statistical approach (Kraemer, 2013; Wallace, Frank & Kraemer, 2013) within the Child Anxiety Treatment Study (CATS) sample (Silk et al., 2018) in which a single score was generated for each individual client using multiple potential moderating variables. This approach found that total number of anxiety diagnoses, several sleep variables (sleep quality, ease of waking, sleep efficiency), parental college education and negative interpersonal concerns together differentiated response to client-centered treatment and CBT. Another study used the probability of treatment benefit (PTB) method, which comprehensibly summarizes the probability that an individual will benefit from treatment given baseline stratification variables (Beidas et al., 2014). Results suggested that youth in the CAMS sample with moderate to severe anxiety symptoms did best in COMB treatment relative to other conditions. Receiver operating characteristics (ROC; Taylor et al., 2018) and latent profile analyses (Norris et al., 2019a; Phillips, Norris & Kendall, 2020) have also been implemented to identify prognostic sub-groups, rather than examining individual variables in isolation.

Other analytic approaches warrant consideration. For example, machine learning (ML) may be a fruitful approach (Jordan & Mitchell, 2015), assuming appropriate emphasis is placed on model validation in new samples. ML represents an appropriately flexible framework for the identification of predictions at the individual level, as the focus is on predictive fit rather than explanatory inference or effect interpretability. ML models also allow for inclusion of complex relationships among features with small effects, which may be particularly relevant for treatment moderators (Kraemer, 2013). Other analytic approaches have also been suggested, including a move towards more process-focused methods like time series and network analyses rather than focusing on baseline variables (Hayes et al., 2019), use of the dynamic p-technique to flexibly generalize individual-level inferences to the group (Kurtz, Johnson, Kellum & Wilson, 2019) and the personalized advantage index (Cohen & DeRubeis, 2018; DeRubeis et al., 2014). Importantly, if these approaches are implemented post-hoc, results should be reported as exploratory, rather than confirmatory (Kraemer et al., 2002). In addition, efforts will need to be made to disseminate actuarial decision making into clinically useful language and tools (e.g., implementation of machine learning algorithms into “phone app” format for clinician use).

Addressing power concerns

Another pervasive problem facing researchers interested in identifying moderator variables lies in the statistical power of their analytic tests. This problem is particularly pronounced when examining interactions (Brookes et al., 2004), and yet it is likely that complex interactions at multiple levels of assessment, rather than simple main effects, will be most informative in determining which treatments work best for whom (e.g., Tiemens, Bocker & Kloos, 2016). One resource- and time-intensive solution to this problem is to conduct additional multi-site RCTs that are adequately powered to detect moderation (e.g., the Genes for Treatment study; Hudson et al., 2015), although inconsistent findings across moderation analyses hamper evidence-driven decisions on measure selection. Another option that we propose, in line with the move toward big data in psychology (Harlow & Oswald, 2016), is a youth anxiety disorders treatment equivalent of the national database for autism research (NDAR), RDoC database (RDoCdb), national database for clinical trials related to mental illness (NDCT) and the NIH pediatric MRI repository (PedsMRI). A team science approach to the identification of moderators via development and ongoing maintenance of an aggregated data repository may make group-level moderation tests more informative. In addition, use of data repositories could accelerate research progress by allowing for more adequately powered secondary analyses that can generate hypotheses subsequently tested via RCTs. Dataset aggregation has already been implemented successfully for some youth anxiety treatment trials, including a project using individual patient data meta-analysis (IPD-MA) to examine age as a moderator of CBT response (Bennett et al., 2013) and another project implementing integrative data analysis to examine predictors of youth anxiety treatments (Skriner et al., 2019). Several limitations of data repositories warrant consideration (e.g., lack of overlapping measures, inconsistent outcome measures, difficulties associated with data harmonization, site differences, concerns related to Simpson’s Paradox) and ongoing development and maintenance of such a repository will require considerable structural shifts (e.g., a standardized set of pre- and post-treatment measures administered across trials, adherence to data sharing agreements, changes to informed consent procedures, use of open access practices). However, given the sizable number of trials examining anxiety treatments that have already been conducted, presence of overlapping measures administered across datasets (Bennett et al., 2013, Skriner et al., 2019) and the rise of open access resources, this may be a particularly fruitful area for exploratory research and recommended infrastructure to facilitate such efforts is beginning to be outlined (e.g., Creswell et al., 2020).

Study design

When designing studies to test moderation, continued emphasis should be placed on randomized design, as randomization can help bolster claims of conditional ergodicity (Adolf et al., 2019; for an opposing argument see Medaglia et al., 2019). Novel study designs also merit consideration, particularly adaptive treatment methods like the sequential multiple assignment randomized trial (SMART; Almirall, Nahum-Shani, Sherwood, & Murphy, 2014). In a SMART design, participants are randomized, assessed regularly, and re-randomized at specified time points. This type of design might allow for a more efficient evaluation of moderators (Almirall & Chronis-Tuscano, 2016). Stepped-care designs have also been leveraged to identify predictors of youth treatment response (Legerstee et al., 2008; Legerstee et al., 2009; Legerstee et al., 2010) and could be used to examine moderation. In addition, intra-individual analyses typically require repeated sampling of the individual over time, even when the focus is not on longitudinal change, and thus repeated observation of the individual (e.g., via ecological momentary assessment methods) should be emphasized to facilitate quality examinations of moderation (Fisher et al., 2018; Hayes et al., 2019). Finally, careful attention should be paid to features of a recruited study sample (i.e., inclusion/exclusion criteria and the population from which samples are recruited). Overly restrictive inclusion and exclusion criteria decrease the range of a particular moderator variable that can be examined (Cohen et al., 2018) and few studies have been conducted in community clinics (for an exception, see Villabø, Narayanan, Compton, Kendall, & Neumer, 2018). It is possible that a less restricted range within examined moderator variables and assessment in a real-world environment may yield different results than those reported in the context of an RCT.

Theory-driven measure selection

To date, the majority of variables examined in moderator analyses are primarily measures of convenience (i.e., demographic variables and youth functioning), perhaps because early RCTs were not designed with moderation in mind. Future studies would benefit from a theory-driven approach to moderation, in conjunction with data-driven exploratory efforts. These theory-driven studies should examine variables that can test hypotheses about why certain individuals may benefit from certain treatments more than others, including explicit theories regarding variable interactions. For example, depressive symptoms have been examined as a moderator of outcome, but it is possible that more specific features of the disorder are associated with differential outcome. Low responsiveness to reward, rather than symptom severity more broadly, may make it difficult to motivate youth to complete homework or exposure tasks, both of which are critical components of efficacious treatments. Such theory-driven examinations of moderators would then facilitate targeted updates to treatments more readily. For example, should low reward responsiveness moderate response to behavioral treatments, it may be necessary to include additional sessions of motivational interviewing at pre-treatment or begin with a positive affect treatment (PAT) that addresses low reward sensitivity (Craske et al., 2019). Other theory-driven variables have been suggested, including intolerance of uncertainty and parental accommodation (Kendall, Norris, Rabner, Crane, & Rifkin, 2020), that warrant future investigation.

A clearer understanding of the mechanisms of action of youth anxiety treatment may also help to inform theory-driven selection of potential moderators (Baron & Kenny, 1986), although to date few consistent mediators of treatment response have been identified (Carper, Makover, & Kendall, 2018). For example, the theory of inhibitory learning posits that exposure efficacy rests in part on expectancy violations (Craske, Treanor, Conway, Zbozinek, & Vervliet, 2014). If this is the case, then contrast avoidance (Newman & Llera, 2011) may be a potential moderator of interest, if a tendency towards avoiding contrasts hampers the occurrence of expectancy violations during exposure treatments. Process variables (e.g., alliance) may also be important mediators to consider when generating hypotheses regarding moderator variables.

Other measurement considerations warrant attention. Moderation findings were often inconsistent across informants and different outcome measures (i.e., functional outcomes compared to symptom measures). Decisions about informant selection should be clearly discussed in moderation analyses. For example, youth depressive symptoms are a fairly consistent predictor of treatment response (Berman et al., 2000; Crawley et al., 2008), although it appears that self-reported depressive symptoms may be more informative in predicting response than other informant reports (O’Neil & Kendall, 2012). Moderators were also more frequently identified when examining functional outcomes compared to symptom reduction. Future studies should be sure to examine the relationship between potential moderators and these kinds of real-world outcomes, and to explicitly outline theories underlying informant and outcome measure selections. Pre-registration will also be helpful in this case, to resist the temptation towards “outcome/informant switching” in the frustrating search for significant moderators.

Conclusion

There are few consistent conclusions that can be drawn from the literature on group-level moderators of youth anxiety treatments, despite decades of research on the topic. To date, a clear and replicable pattern of moderator variables has not emerged, with inconsistent results across informants, measures and studies. Even if such a pattern had emerged, it is unclear whether findings at the group-level would have clinical utility for individual clients (i.e., if moderation processes are ergodic). The variables examined have primarily focused on readily available demographic and symptom information, with a limited focus on underlying theory. Just as often, moderators have been examined in isolation, which belies the reality that any single individual presents with multiple characteristics that likely interact to influence treatment response. Consistent with the reported findings, it is possible that results suggest comparable efficacy of youth anxiety treatments across a range of demographic and symptom presentations. However, given the fact that a portion of youth do not respond to current anxiety treatments, additional work is needed to better clarify which treatments work for whom and to move the field closer to personalized treatment assignment decisions.