Prediction and Moderation of Improvement in Cognitive-Behavioral and Psychodynamic Psychotherapy for Panic Disorder

Abstract

Objective

To identify variables predicting psychotherapy outcome for panic disorder or indicating which of two very different forms of psychotherapy - Panic-Focused Psychodynamic Psychotherapy (PFPP) or Cognitive-Behavioral Therapy (CBT) - would be more effective for particular patients.

Method

Data were from 161 adults participating in a randomized controlled trial including these psychotherapies. Patients included 104 women; 118 patients were White, 33 were Black, and 10 were of other races; 24 were Latino(a). Predictors/moderators measured at baseline or by Session 2 of treatment were used to predict change on the Panic Disorder Severity Scale.

Results

Higher expectancy for treatment gains (Credibility/Expectancy Questionnaire d= -1.05, CI_95% [-1.50, -0.60]), and later age of onset (d= -0.65, CI_95% [-0.98, -0.32]) were predictive of greater change. Both variables were also significant moderators: Patients with low expectancy of improvement improved significantly less in PFPP than their counterparts in CBT, whereas this was not the case for patients with average or high levels of expectancy. When patients had an onset of panic disorder later in life (≥27.5 yrs. old), they fared as well in PFPP as CBT. In contrast, at low and mean levels of onset age, CBT was the more effective treatment.

Conclusions

Predictive variables suggest possibly fruitful foci for improvement of treatment outcome. In terms of moderation, CBT was the more consistently effective treatment, but moderators identified some patients who would do as well in PFPP as in CBT, thereby widening empirically supported options for treatment of this disorder.

Panic disorder is associated with substantial societal and economic costs, including high rates of health service utilization and absenteeism from work (Batelaan et al., 2007; de Graaf, Tuithof, van Dorsselaer, & ten Have, 2010), poor physical health, decreased quality of life, and impairments in social and occupational functioning (e.g., Markowitz, Weissman, Ouellette, Lish, & Klerman, 1989). Fortunately, effective treatments for this condition exist. A wealth of evidence supports the efficacy of cognitive-behavioral therapy (CBT) in the treatment of panic disorder (see Mitte, 2005, for a meta-analysis and review). More recently, Milrod, Busch, Cooper, and Shapiro (1997) developed panic-focused psychodynamic therapy (PFPP), a time-limited, manualized psychodynamic treatment that specifically targets panic symptoms. In a small randomized, controlled trial (RCT), PFPP outperformed applied relaxation training (ART) in the reduction of panic symptoms and improvement in psychosocial functioning (Milrod, Leon, Busch, et al., 2007). Other supportive evidence of PFPP's efficacy comes from a study of its effectiveness in clinical practice (Beutel et al., 2013).

No single psychotherapy is effective for all patients with panic disorder. For example, in a recent large clinical trial of three psychotherapies (Milrod et al., 2016), 54% of patients failed to respond to ART, 41% to PFPP, and 37% to CBT. Thus, researchers have sought to identify which factors explain this variability in treatment response. A key question is whether there are individual differences that reliably predict better response to one treatment over another. These variables are referred to as moderators and, once identified, they can be used to guide treatment selection. A related question is whether there are certain types of patients who do not fare well in any existing treatment. Predictor variables predict treatment outcome irrespective of the type of treatment, and the identification of such variables is an important first step toward developing more effective treatments for individuals who do not benefit from existing interventions.

In the present study, we use data from the Milrod et al. (2016) trial to examine predictors and moderators of response to CBT and PFPP. Although both treatments are symptom-focused, they represent two very different approaches to treatment in theory and in execution. CBT is a very directive treatment involving psychoeducation, cognitive restructuring, and interoceptive and in vivo exposure, with copious homework designed to help the patient become less fearful by changing catastrophic thinking about panic and relinquishing avoidance behaviors. In contrast, PFPP is a short-term, transference-based psychodynamic treatment that eschews homework assignments and focuses on helping patients to uncover the unconscious meanings of their panic symptoms and to resolve core unconscious conflicts leading to panic. Frequent foci are recognizing, managing, and expressing anger and guilt and conflicts over separation and autonomy. The stark contrast between these treatments suggests it should be possible to identify patients who would benefit more from one than the other¹.

Predictors and Moderators of Response to CBT for Panic Disorder

We began variable selection for this study by conducting a systematic review and meta-analysis of the literature on predictors and moderators of response to CBT for panic disorder and/or agoraphobia (Porter & Chambless, 2015). Although the number of studies for any single variable was small, we found that higher levels of pretreatment agoraphobic avoidance, lower expectancy of change, and greater pretreatment functional impairment consistently predicted worse response to CBT in this population. There was also some indication in this review that individuals with a later age of onset and a shorter duration of disorder fared better in CBT, but these variables were conceptualized somewhat differently across studies, making the literature difficult to synthesize. Finally, this meta-analysis indicated that the presence of a comorbid Cluster C personality disorder was predictive of decreased improvement in CBT for panic disorder and/or agoraphobia. However, results from the individual studies included in the meta-analysis were mixed, with one study finding that patients with Cluster C disorders fared worse, and two studies finding no significant effects. Despite exhaustive search procedures, we located no study that examined moderators of response to a CBT treatment as compared to another form of psychotherapy.

It seems highly likely that some of the identified predictors may not be specific to CBT, but instead might predict response to psychotherapy more broadly. For example, individuals who expect more benefit from therapy at the outset of treatment are likely to make greater gains in any form of psychotherapy (Constantino, Glass, Arnkoff, Ametrano, & Smith, 2011). Similarly, we might expect individuals with an earlier age of onset, longer duration of disorder, greater agoraphobic avoidance, and greater functional disability to do more poorly in any form of psychotherapy than patients without these pretreatment characteristics in light of the chronicity and severity of their condition. Although the literature on Cluster C diagnosis as a predictor of outcome for anxiety disorders is scant and mixed, in a 2003 review Reich concluded that, overall, personality disorders (including but not limited to Cluster C) are a predictor of poor outcome for anxiety treatments across the board, with the possible exception of selective serotonergic reuptake inhibitors.

Predictors and Moderators of Response to Psychodynamic Therapy for Panic Disorder

There is little research on predictors of response to psychodynamic treatments for panic disorder. Beutel et al. (2013) found that higher emotional awareness and lower global illness ratings at baseline predicted more change in treatment for patients with panic disorder receiving either PFPP or CBT. The preponderance of patients their study received PFPP, but the sample sizes were too small to test prediction of PFPP alone. Existing papers on moderation of response to psychodynamic treatments for panic disorder have all investigated moderators of PFPP vs. ART in a single small, exploratory study (Milrod, Leon, Busch et al., 2007), which found that PFPP was superior to ART. No statistically significant moderators were identified, which is unsurprising given the small sample size in this trial. In several other papers Milrod and colleagues sought to examine moderators in their study, but in light of their small sample appropriately did not employ statistical tests of their hypotheses (Kraemer, Wilson, Fairburn, & Agras, 2002). However, there is some suggestion from the pre-post effect sizes for different subgroups of patients that although PFPP was superior to ART overall, the difference in favor of PFPP was even greater for patients whose panic disorder onset was not precipitated by a recent interpersonal loss (Klass et al., 2009) and for those with a comorbid Cluster C diagnosis (Milrod, Leon, Barber, et al., 2007). Patients with and without a comorbid Cluster B diagnosis appeared to fare similarly well in each treatment, but the number of Cluster B patients was so small as to limit any conclusions (Milrod, Leon, Barber, et al., 2007).

Theorized Moderators of Response to CBT and PFPP

Because the empirical literature provides little guidance about which variables might moderate response to CBT and PFPP, we turned to theory and our clinical expertise in treating panic disorder to develop hypotheses about potential moderators. Suspecting that some of our hypothesized predictors of response to treatment might also serve as moderators, we predicted that patients with greater functional disability and longer durations of disorder would fare better in CBT as compared to PFPP. Because of the chronicity of their disorder and the likelihood that these patients have adapted their lives to panic disorder, they might benefit most from a more structured psychotherapy like CBT, in which the therapist is relatively directive and is prepared to break the change process down into small steps. We also predicted patients higher in pretreatment agoraphobic avoidance would do better in CBT as compared to PFPP, because CBT is highly directive in urging patients to confront feared situations and to decrease their avoidant behavior, whereas PFPP encourages such confrontation more indirectly through transference interpretation.²

A goal of PFPP and of psychodynamic therapies more broadly is to help patients to gain insight into the underlying meanings of their symptoms as they occur in real time with the goal of developing greater control through improved insight and ability to acknowledge, tolerate, and reflect on the feelings that lead to their symptoms. Some writers have theorized that patients with higher pretreatment insight may fare better in psychodynamic therapy than their low insight counterparts (see Barber, Muran, McCarthy, & Keefe, 2013). Indeed, Beutel et al. (2013) found that patients with panic disorder who were higher in emotional awareness fared better in treatment. Closely related to the concepts of insight and emotional awareness, reflective functioning (RF) is defined as the ability to understand one's own and others'; behavior in terms of mental states (desires, intentions, feelings, and beliefs) (Fonagy, Target, Steele, & Steele, 1998). When designing this study, we initially hypothesized that patients high in pretreatment RF would fare equally well in CBT and PFPP, but that those with low pretreatment RF would do better in CBT than in PFPP. However, recent data raise doubt about this hypothesis. Ekeblad, Falkenström, and Holmqvist (2016) reported that having higher scores on RF predicted better treatment outcome in a trial of CBT and interpersonal therapy (IPT) for depression in a group of patients with very impaired pre-treatment RF. Contrary to our initial hypothesis in this study, Ekeblad et al. found no moderation by type of treatment. However, IPT is less insight oriented than PFPP, and thus these results might not generalize to a more intrapsychically oriented psychodynamic therapy. Accordingly, additional research is needed: RF might prove to be a predictor of outcome across both types of psychotherapy or a moderator of treatment response.

We also suspected that patients' personalities might moderate response to CBT and PFPP, specifically that patients with comorbid Cluster C personality disorders would fare better in PFPP than in CBT. Such a finding would be consistent with results from our meta-analysis (Porter & Chambless, 2015), in which we found that a comorbid Cluster C diagnosis was predictive of less improvement in CBT, and Milrod, Leon, Barber et al.'s (2007) paper, which suggested that patients with a comorbid Cluster C diagnosis did especially well in PFPP as compared to ART. Some research suggests that the impact of Cluster C comorbidity on response to treatment may depend on the specific Cluster C disorder involved. For example, Barber and Muenz (1996) found that depressed patients with comorbid avoidant personality disorder fared better in cognitive therapy (CT) than in a more affect-focused treatment (IPT), whereas those with comorbid obsessive-compulsive personality disorder (OCPD) fared better in IPT as opposed to CT. Given the preponderance of OCPD among patients in our sample who met criteria for a Cluster C diagnosis at pre-treatment (40 of 54 Cluster C patients or 74%), these findings increase the likelihood that that patients with a Cluster C disorder will improve more in PFPP, which like IPT is strongly affect focused, than in CBT.

Finally, we were interested in whether the experience of childhood loss might serve as a moderator. Such loss is a risk factor for separation anxiety (Silove et al., 2015), which in turn is implicated as a risk factor for a host of psychiatric disorders, including panic disorder. Indeed, in his classic work, Bowlby (1973) posited that separation and loss in childhood result in dysregulated attachment and the lack of a sense of safety, which in the extreme may then result in agoraphobia. We are unaware of any studies examining childhood loss as a predictor or moderator of response to CBT or psychodynamic therapy for panic disorder. However, Nemeroff et al. (2003) found that depressed patients who experienced childhood loss fared significantly better in the cognitive behavioral analysis system of psychotherapy, a time-limited therapy that combines techniques from CBT, IPT, and psychodynamic therapy, than on antidepressant medication. We hypothesized that patients in our trial who had experienced a childhood loss might do better in PFPP than in CBT because in PFPP dysregulated attachments, such as those that commonly occur in patients who have sustained childhood loss, are a focus of treatment.³

The Present Study

We sought to examine predictors and moderators of response to psychotherapy in the context of the Cornell-Penn Psychotherapies for Panic Disorder Study, a multisite RCT in which the benefits of CBT and PFPP for patients with a primary diagnosis of DSM-IV panic disorder with or without agoraphobia were compared (Milrod et al., 2016).

For the reasons described above, we hypothesized that, in this trial, greater expectancy for change, later age of onset, shorter duration of disorder, lower functional impairment, lower agoraphobic avoidance, and absence of a Cluster C diagnosis would be predictive of greater pre-to posttreatment improvement across the sample of patients who received CBT or PFPP. Thus, we seek to replicate prior results in the CBT literature but extend these for the first time to a psychodynamic treatment. With regard to moderation, we predicted that patients with higher levels of agoraphobic avoidance, longer duration of disorder, greater disability, and lower RF would improve more in CBT as compared to PFPP, even though within the CBT condition such patients might do worse than patients without these characteristics. Conversely, we expected patients who had experienced a childhood loss and those with Cluster C personality disorders to fare better in PFPP as compared to CBT. Our design allows us (a) to test not only whether prior findings for predictors of CBT outcome will replicate in this trial but also whether they extend to a symptom-focused psychodynamic therapy and (b) to perform the first test of moderation of these two very different forms of psychotherapy for panic disorder. Moreover, given that we have two sites in our study, our design allows us an internal test of replication of the findings across sites. This permits us to rectify what Schneider, Arch, and Wolitzky-Taylor (2015) identified as a major limitation to the moderation literature in treatment of anxiety disorders – the absence of two studies that explore the same treatments with the same populations and use the same variables to define the moderators.

Method

Participants

Participants were 161 patients with a primary diagnosis of panic disorder with (n = 127, 79%) or without (n = 34, 21%) agoraphobia drawn from the Cornell-Penn Psychotherapies for Panic Disorder trial (Milrod et al., 2016). Patients in that trial were randomly assigned to one of three forms of psychotherapy, and the present sample comprises those in two of those conditions: CBT (n = 81) and PFPP (n = 80). Of these, 104 (65%) were female; 118 (76%) patients were White, 31 (19%) were Black, 8 (5%) were Asian or Asian American, 1 (<1%) were Native American, and 3 (2%) were other or missing. Twenty-four (15%) were Latino(a). Participants ranged in age from 18-69 (M = 39.40, SD = 13.25); the duration of their current panic episode ranged from 1 to 444 months (Mdn = 15).

Inclusion criteria included a primary diagnosis of panic disorder and at least one spontaneous panic attack weekly for the month before entry. Psychotropic medication was permitted if stable for ≥ 2 months at intake and if the patient was willing for it to remain constant across the trial. Exclusion criteria included active substance dependence (< 6 months' remission), history of psychosis or bipolar disorder, acute suicidality, organic mental syndrome, evidence of medical conditions that might cause the patient's symptoms, involvement in legal or disability proceedings involving the patient's mental state, and unwillingness to forego non-study psychotherapy for the duration of the protocol. All patients provided written informed consent; the Institutional Review Boards of Weill Cornell Medical College and University of Pennsylvania approved the research.

Procedure

Patients at each site were stratified on diagnosis of agoraphobia and of major depression before randomization. Subsequent to baseline assessment and completion of 3 weeks of a panic diary, they were assigned according to availability to 1 of 8 CBT therapists or 16 PFPP therapists for twice weekly 45-minute sessions for 19-24 sessions. Therapists were trained for this study or had been trained in a prior trial to deliver one of the two treatments. Therapists treated patients in a single protocol and were supervised by an expert in their treatment modality throughout the trial.

Assessment of the primary outcome variable, the Panic Disorder Severity Scale (Shear et al., 1997), was conducted by trained diagnosticians kept uninformed as to treatment assignment and was carried out at roughly monthly intervals: Baseline, Weeks 1, 5, and 9 of treatment, and termination. Predictor and moderator variables were assessed at baseline, with the exception of treatment expectancy, which was assessed prior to or immediately after Session 2. For additional details of procedure, see Milrod et al. (2016).

Interview Measures

Anxiety Disorders Interview Schedule for DSM-IV

(ADIS; Brown, Di Nardo, & Barlow, 2004). Trained masters or doctoral level interviewers used the ADIS for diagnosis of panic disorder, agoraphobia, and possible comorbid conditions. ADIS reliability was excellent for panic disorder (κ = 1.0) and agoraphobia (κ = .94). The diagnostician also collected information about duration of the current episode of panic disorder and about the patient's age at the time of the first episode of panic disorder.

Panic Disorder Severity Scale

(PDSS; Shear et al., 1997). The PDSS was used to assess change with treatment. Ratings on the seven PDSS questions range from 0 (none) to 4 (extreme). Questions tap frequency and severity of panic attacks, distress associated with attacks, severity of interoceptive and situational avoidance and of anticipatory anxiety, and social, familial, and vocational impairment, and form a unifactorial scale (Shear et al., 2001). Internal consistency of the measure is acceptable (α = .695 in the present sample). Prior research has demonstrated the interrater and test-retest reliability and construct validity of this measure (Shear et al., 1997, 2001). In the present study, the ADIS diagnostician, who was uninformed as to treatment assignment, completed the PDSS as well. Interrater PDSS reliability, calculated across sites with an intraclass correlation coefficient, proved excellent, r_I (2,1) = .95.

Childhood Trauma Scale: Loss

(Lizardi et al., 1995; Nemeroff et al., 2003). Using the Childhood Trauma Scale, diagnosticians interviewed patients about their experiences prior to age 15 of loss of or prolonged separation from a parent (coded as present/absent) Loss was coded when the patient reported that before the age of 15 she or he experienced the death of a parent, separation from a parent for more than 6 months, or parental divorce or separation of more than 6 months. Seventy-two patients (45%) reported such a loss.

Reflective Functioning

(Fonagy et al., 1998). Originally designed to be used with the Adult Attachment Interview and derived from its demand questions, RF ratings in the present study were based on material collected via the Brief RF Interview-Revised (Rudden, 2009) conducted by the diagnostician. Patients were asked to talk about a parent, the relationship with that parent, and the changes in that relationship across time, as well as their relationship with someone important to them at the present time. Following transcription, one of three coders rated overall RF on a -1 (negative reflective functioning/psychotic and idiosyncratic) to 9 (exceptional reflective functioning) scale. Rutimann and Meehan (2012) reported good interrater reliability (ρ_I = .79) for an earlier version of the Brief RF Interview as well as excellent concurrent validity with RF as assessed from the Adult Attachment Interview (r = .71). One coder (Marie Rudden) in the present study, who was trained in coding RF by Mary Target of the Anna Freud Centre, trained the other two coders in rating RF from the Brief RF Interview. Reliability was assessed by having pairs of coders independently code the same transcript (n = 62). Intraclass reliability coefficients among the three coders ranged from acceptable to excellent, r_I (1,1) = .63-.91.

Structured Clinical Interview for DSM-IV Axis II Personality Disorders

(SCID-II, Version 2.0;First, Spitzer, Gibbon, Williams, & Benjamin, 1996). The SCID-II was administered by the same diagnosticians as the ADIS. Cluster A (n = 24, 17%) and B diagnoses (n = 18, 11%) and traits were too infrequent for analysis of moderation. However, Cluster C diagnoses were sufficiently common for analysis (n = 54, 33.5%). Interrater reliability for presence/absence of any Cluster C personality disorder was very good, κ = .83.

Questionnaire Measures

Avoidance Alone scale of the Mobility Inventory for Agoraphobia

(Chambless, Caputo, Jasin, Gracely, & Williams, 1985) was used to assess agoraphobic avoidance. Patients are asked to rate 26 items reflecting situations patients with agoraphobia may avoid on 1 (never avoid) to 5 (always avoid) scales. Item scores are averaged to yield the total score. The internal consistency, test-retest reliability, and convergent and discriminant validity of this measure have been repeatedly supported (Chambless et al., 2011).

Credibility/Expectancy Questionnaire

(Devilly & Borkovec, 2000). The three expectancy items from this measure yielded the expectancy score. Patients are asked to indicate on 0-100% scales how much they think and feel they will improve by the end of therapy and on a 1 (not at all) to 9 (very much) scale how helpful they feel therapy will be in reducing their symptoms. Item scores are standardized and averaged for analysis. The factor analytically derived expectancy scale is internally consistent, has good test-retest reliability, and has proved useful in predicting treatment outcomes (Devilly & Borkovec, 2000).

Sheehan Disability Scale

(Sheehan, 1983) is a 3-item measure of social, family, and vocational impairment. Each item is rated on a 0 (not at all) to 10 (extremely) scale indicating the degree to which the patient's disorder disrupted functioning in that domain in the last month. Expressed as a sum, total scores range from 0-30. Leon, Shear, Portera, and Klerman (1992) have demonstrated the reliability and validity of this measure for panic disorder patients, which evinced good internal consistency in the present sample, α = .83.

Treatment

All treatments were delivered in 19-24 twice weekly sessions of 45 minutes duration. Flexibility in treatment length was permitted in CBT for cases of panic disorder without agoraphobia who did not require the last sessions of the protocol, which focused on agoraphobic avoidance, or who resolved their symptoms quickly. In contrast, the goal for PFPP was 24 sessions with pre-planned termination dates to provide a clear endpoint and framework for processing mixed feelings and distress about separation from the therapist⁴.

Cognitive-behavioral therapy

CBT followed Barlow and Craske's Panic Control Therapy protocol (Craske, Barlow, & Meadows, 2000), as modified to fit the 24-session, 45 minutes per session format of the trial. This equated CBT and PFPP in length and duration of sessions. The protocol included psychoeducation about anxiety and panic; identification and correction of maladaptive thoughts about anxiety and panic; training in slow, diaphragmatic breathing; and exposure to bodily sensations designed to mimic those experienced during panic. All sessions were followed by homework assignments and readings. In vivo exposure via homework assignments was introduced at Session 17 for those patients with significant agoraphobic avoidance. Session 24 focused on review and relapse prevention. Muscle relaxation was omitted to reduce overlap with ART.

Panic-Focused Psychodynamic Psychotherapy

is a manual-based treatment (Milrod et al., 1997) rooted in the assumption that panic symptoms have a psychological meaning and that uncovering these unconscious meanings will lead to relief. The therapist explores the circumstances and feelings surrounding panic onset, the personal meaning of panic symptoms, and the feelings and content of panic episodes. PFPP aims to lessen vulnerability to panic by helping patients understand and alter core unconscious conflicts, which are often identified and understood through their emergence in the transference. Frequent themes include conflicts over separation and autonomy; recognizing, managing, and expressing anger; and guilt. Termination, addressed prominently in the last third of treatment, permits patients to re-experience conflicts directly with the therapist so that underlying feelings can be articulated and rendered less frightening.

Adherence

to the therapy manuals was assessed by teams of coders for each of the three treatments. As reported in the main outcome paper (Milrod et al., 2016), we confirmed that adherence was acceptable for both treatments, and that therapists did not mix interventions from the other two protocols with their assigned protocol.

Statistical Analyses

Change during treatment

Standard approaches to the longitudinal analysis of treatment data assume that missing data are missing at random, but this was not the case with these data. For this reason, we analyzed change in treatment over time with a shared parameters approach (Ten Have, Pulkstenis, Kunselman, & Landis, 1998), rather than the typical multilevel modeling (MLM) approach, which ignores the missing data mechanism. Each patient was simultaneously modeled with (a) MLM for change over time on the PDSS and (b) a survival model for attrition. The two processes share a common random effect, creating a quantifiable correlation between change in the PDSS outcome and dropout processes. In this sample this correlation was substantial (r = -.44, p = .03), demonstrating outcome-missing data dependency: Patients who were improving at a slower rate were more likely to drop out of treatment more quickly.

At Level 1 of the MLM, PDSS scores varied within patients over time, expressed as weeks from baseline. Level 2 captured between person variability as a function of treatment assignment.⁵ The attrition process was assessed with a discrete time survival model: At each evaluation point, an individual was classified a dropout, treatment completer, or continuing in treatment. By jointly modeling outcome and attrition processes, we were able to evaluate outcome while adjusting for the association between the outcome and attrition models.

A priori covariates were use of anxiolytics and of antidepressants, treatment condition, site, and Treatment × Site interaction effects. In addition, because there were Treatment × Site differences in age and gender in the original study (Milrod et al., 2016), these variables were included as covariates. To preserve power, predictors or moderators of interest were introduced one per equation. Prediction and moderation of rate of change (the slope) are the critical tests. Additionally, we also included the three-way interaction of Moderator × Treatment Condition × Site to determine whether effects differed across sites. Continuous predictors/moderators were centered for analysis. When interactions involving a continuous predictor/moderator were significant, we probed these by estimating and contrasting slope estimates for the treatments with the continuous predictor set at the mean and 1 SD above and below the mean.

A number of predictors were substantially correlated with PDSS baseline scores, thus risking a spurious relationship with slope because of the steeper PDSS slope evinced by patients who started with higher PDSS scores. To reduce the likelihood of artefactual findings, we regressed these predictors on baseline PDSS scores and used the residualized variables in the place of the original form of the predictors in our analyses.

The main results of this trial were complicated by significant Site × Treatment interaction effects: Among patients treated at Cornell, there were no differences between the two treatments on the primary outcome measure, change in symptom severity on the Panic Disorder Severity Scale (PDSS; Shear et al., 1997), whereas among patients treated at Penn, those who received CBT showed significantly greater improvements in PDSS change than those who received PFPP. Thus, before conducting shared parameters analyses, we ascertained whether there were significant Site × Treatment differences on the predictors/moderators. If this had been the case, the effects of the predictor/moderator might have been confounded with Site × Treatment differences in treatment response. However, none of these interactions were significant, all ps ≥ .11.

SAS Version 9.4 (2015) was used for all analyses. Degrees of freedom were estimated with the Kenward-Roger's (1997) approximation.

Power analyses

To preserve power, we set alpha at .05 for each analysis. Due to participant or interviewer error in collecting data or recording failure, sample sizes for predictors and moderators ranged from 93-161. Power for tests of prediction was estimated as .80 for a medium d of 0.48. Power for tests of moderation depends on the form of any interaction obtained (ordinal vs. disordinal), the relative within-treatment condition effect sizes, and the sample sizes of each treatment condition. Based on hypotheses, we assumed any interactions would be ordinal, indicating the effect was in the same direction across the intervention arms but of different magnitudes, and calculated power under a number of scenarios of larger and smaller effect sizes across the two treatment conditions. Sample size for our study was determined on the basis of power for our randomized test of treatment outcome differences and not for moderation. Thus, we only had adequate power (> .80) when we assumed large differences (e.g., d of 0.9) in the efficacy of one of the treatments at low and high levels of the moderator and very small effects for the second. Accordingly, tests of moderation may well be underpowered and should be taken as exploratory (Kraemer et al., 2002).

Results

Descriptive data for the sample and a correlation matrix for predictor, moderator, and dependent variables may be found in Supplemental Tables 1 and 2. Additional details of results from the shared parameters model are reported in Supplemental Table 3.

Expectancy

Unlike other predictors, expectancy was assessed once treatment had begun (at Session 2 of the first week of treatment). Accordingly, we calculated the slope effect beginning from Week 1 of treatment rather than at baseline. Expectancy was a significant predictor of change in treatment, t(84) = -4.85, p < .0001, d = -1.05, CI95% [-1.50, -0.60]. However, this effect was moderated by an interaction with treatment condition, t(83) = -2.19, p = .03. Probes of the interaction indicated that patients with low levels of expectancy improved more in CBT than in PFPP (t(83) = 2.39, p = .02, d = 0.52, CI_95% [0.09, 0.95]), whereas at average and high levels of expectancy the two treatment conditions did not differ significantly, ds ≤ 0.28, ps ≥ .20. The expectancy effect was not modified by interactions with Site or Site × Treatment, ps > .97.

Age of Onset and Duration of Episode

Also consistent with prediction, age of onset (log transformed to achieve normality) was significantly predictive of slope: Patients who were older when they had their first episode of panic disorder changed more in treatment, t(147) = -3.95, p = .0001, d = -0.65, CI_95% [-1.50, -0.60]. However, this effect was moderated by treatment, t(146) = -2.61, p = .01. In the case of patients who were relatively young when they had their first episode of panic disorder (< 14.32 years old), PFPP patients improved significantly less than patients in CBT, t(146) = 3.77, p = .0002, d = 0.62, CI_95% [0.29, 0.95]. The same was true at the average age of onset (23.68⁶ years of age): CBT patients improved more than PFPP patients, t(146) = 2.58, p = .01, d = 0.42, CI_95% [0.10, 0.75]. Among patients with later onset (> 39.17 years of age), there was no significant difference between treatments, t(146) = -0.06, p = .95, d = -0.01, CI_95% [-0.33, 0.31]. To better pinpoint the age of onset at which treatment differences in CBT and PFPP no longer approached significance, we tested differences between these two treatments at 0.1 SD increments above the mean. At 0.3 SD above the mean (27.53 years old), p for the contrast exceeded .05. Using this cut point, we find that for the top 37% of the age of onset distribution the two treatments were comparable in outcome. These effects were not moderated by interactions with Site or Site × Treatment, ps ≥ .35.

Contrary to hypothesis, duration of present episode (log transformed to achieve normality) was not associated with change in treatment, t(150) = 0.23, p = .82, d = 0.04, CI_95% [-0.28, 0.35]. The test of the hypothesis that patients who had a longer duration of panic disorder would fare relatively better in CBT than in PFPP indicated that this was not the case, t(149) = -0.38, p = .70, d = -0.06, CI_95% [-0.38, 0.26]. The effect of duration was not modified by Site or Site × Treatment interactions, ps ≥.62.

Sheehan Disability Scale

Because baseline Sheehan and PDSS scores were highly correlated (r = .54, p < .001), residualized Sheehan scores were used for analysis. The Sheehan was neither an overall predictor of change (t(147) = 1.67, p = .10, d = 0.27, CI_95% [-0.05, 0.60]) nor a moderator of treatment outcome, t(146) = -1.91, p = .06, d = -0.31, CI_95% [-0.64, 0.01]. We probed the interaction for heuristic purposes (see Supplemental Table 3). CBT proved to be more effective than PFPP at low and moderate levels of disability but not at high levels of disability. The effect of disability was not modified by Site or Site × Treatment interactions, ps ≥ .41.

Agoraphobic Avoidance

Because avoidance was highly correlated with baseline PDSS scores (r = .57, p < .001), residualized avoidance scores were used for these analyses. As hypothesized, patients higher in agoraphobic avoidance did worse in treatment, t(138) = 3.72, p = .001, d = 0.55, CI_95% [0.22, 0.89]. However, this effect was moderated by site. Although the pattern was the same at both sites, the effect was significant only at Cornell: Cornell t(137) = 3.66, p = .0004, d = 0.62, CI_95% = 0.28, 0.96; Penn t(137) = 1.38, p = .17, d = 0.23, CI_95%[-0.10, 0.57]. The interaction with Treatment and the Avoidance × Site × Treatment interaction were not significant, ps ≥ .15.

Cluster C Personality Disorder Diagnosis

Cluster C did not significantly predict outcome, t(152) = 1.84, p = .07, d = 0.30, CI_95% = -0.02, 0.61. The effect size was in the predicted direction but small. Contrary to prediction, the Cluster C × Treatment interaction was not significant, t(151) = 1.21, p = .23, d = 0.20. There was no significant Cluster C × Site or Cluster C × Site × Treatment interaction, ps > .74.⁷

Reflective Functioning

Contrary to hypothesis, patients higher in RF fared no better than those with lower RF. Indeed the small effect size was in the opposite direction, t(147) = 1.73, p =.09, d = 0.28, CI_95% = -0.04, 0.61. Moreover, the Treatment × Reflective Functioning test of moderation was not significant, t(146) = -1.90, p = .06, d = -0.31, CI_95% [-0.64, 0.01]. In light of our a priori hypothesis, we probed the interaction for heuristic purposes (see Supplemental Table 3). Consistent with prediction, patients who were low or average on RF improved more in CBT than in PFPP, whereas there was no significant difference between treatments for those high in RF.

Childhood Trauma Scale: Loss

Loss was not a predictor of treatment outcome, t(152) = -1.68, p = .095, d = -0.27, CI_95% [-0.59, 0.05]. Contrary to prediction, the Loss × Treatment interaction was not significant, t(151) = 0.90, p= .37, d = 0.15. There were no significant interactions of Loss with Site or Site × Treatment, ps ≥.13.

Discussion

In a large sample of panic disorder patients treated with one of two different forms of psychotherapy, we examined prediction of treatment outcome across both PFPP and CBT by variables found in our systematic review (Porter & Chambless, 2015) to predict outcome of CBT: agoraphobic avoidance, functional impairment, and expectancy of improvement. Of the three replicated predictors Porter and Chambless identified, only one proved to predict outcome across treatments and sites in this study - expectancy of change. Another, agoraphobia avoidance, predicted poorer outcome at one of the two sites. Of three other variables Porter and Chambless identified as possible predictors, earlier age of panic disorder onset proved to be significant in the present sample, but duration of disorder and Cluster C personality disorders did not. We further tested whether lower reflective functioning, higher functional impairment, more severe agoraphobic avoidance, longer duration of the disorder, childhood loss of a parent, and diagnosis of a Cluster C personality disorder might serve as moderators of outcome, indicating which treatment might be preferable for a given patient. Our moderation hypotheses were not confirmed, but two unpredicted significant moderators of treatment were identified and proved to replicate across the two study sites – expectancy and age of onset.

Consistent with two other studies of CBT for panic disorder (see Porter & Chambless, 2015) and with the results of a meta-analysis of expectancy across disorders and types of treatment (Constantino et al., 2011), treatment outcome expectancy as assessed at Session 2, on the whole, predicted better treatment response in the present sample for both treatments with a medium-large effect size. We assessed expectancy after the patient had one or two meetings with the therapist to learn the rationale for the treatment. Thus, it is conceivable that the expectancy measure could reflect not only the treatment rationale but also the patient's initial reaction to the therapist, that is, not only will this treatment help me, but also will this person help me? Research on why expectancy is related to outcome is not extensive, but findings to date (Constantino et al.) suggest that expectancy may affect outcome through its effects on completion of homework and on the therapeutic alliance. Huppert et al. (2014) have demonstrated the positive relationship of working alliance at Session 3 with treatment outcome on panic disorder symptoms. Thus, examining whether the working alliance mediates the relationship of expectancy to outcome would be a worthwhile pursuit. Moreover, in studies of CBT for various disorders, both the quantity and quality of homework completion is associated with better outcome (see meta-analysis by Kazantzis et al., 2016). It seems likely that expectancy might also affect the level of engagement in challenging activities in session such as interoceptive exposure in CBT and interpretation of intrapsychic conflict and the transference in PFPP. Why low expectancy predicted worse outcome for PFPP than CBT is not clear. However, this might be the result of CB therapists' use of cognitive restructuring for low expectations of change – an approach readily available in their armamentarium. That is, had we had measures of expectancy at later sessions, we might have found that in CBT initially low expectancy patients had a more positive view of their prospects for improvement later in therapy. Unfortunately, such measures are lacking in our sample.

Constantino et al. noted the poor methodological quality of many of the 46 studies they included in their meta-analysis. Thus, the present study makes a meaningful contribution to the literature by assessing expectancy in a reliable and valid fashion and by measuring change with treatment subsequent to expectancy assessment, thus avoiding confounding the effects of early panic disorder improvement and expectancy on outcome. Although correlational, these findings argue for the importance of therapists' efforts to increase early treatment expectancy. Constantino et al. suggested a number of strategies including explicit discussions of patients' expectancies during which the therapist is respectful of any doubts patients hold, providing information on the known efficacy of an intervention plus reasons the therapist believes the intervention will be beneficial for this particular person, and the like.

Age of onset of the first episode of panic disorder was both a predictor of poorer treatment outcome and a moderator of outcome, in that regardless of treatment condition patients did worse when they had an earlier age of onset, but this was especially marked for PFPP. For roughly 63% of the sample, CBT was likely to be more effective than PFPP, but for patients in the upper 37% of the distribution (≥ 27.5 years old at first onset), there was no advantage to CBT over PFPP. These data solidify the uncertain findings reported in Porter and Chambless (2015) as to age of onset's predictive utility and are consistent with reports that age of onset is related to worse outcome in psychotherapy for social anxiety disorder as well (e.g., Borge, Hoffart, & Sexton, 2010). Interpreting why socially anxious patients with earlier onset did worse in both CBT and IPT, Borge et al. hypothesized that their findings might reflect greater personality problems in the younger onset patients, such as higher neuroticism. We do not have measures of neuroticism but note that Cluster C diagnosis was not related to worse outcome in the present study. In our clinical experience the younger the onset the more normal development is disrupted by panic disorder, and the less patients have a compelling image of the life they could have if they confronted their fears in treatment. Progress may seem punishing in that it may confront the patient with the challenge of delayed developmental experiences, for example, seeking further education or employment or establishing an independent residence. Longer term treatment with support to develop life skills may be required. PFPP may fare worse than CBT with patients of younger onset because CB therapists are more likely to provide such concrete coaching.

Cluster C was not a predictor of poor treatment outcome overall, thus adding another negative study to the column of those reported by Porter and Chambless (2015) and making it less likely that this is a negative prognostic sign for CBT. In light of the high preponderance of patients with OCPD in our Cluster C patients (74% of those with a Cluster C diagnosis), we predicted that Cluster C patients would improve more if in PFPP than in CBT (cf. Barber & Muenz, 1996). We found no statistically significant moderation by Cluster C diagnosis. However, in analyses conducted for heuristic purposes, we found that, contrary to prediction, CBT was superior to PFPP for patients with OCPD. Thus, these results are inconsistent with Milrod et al.'s (Milrod, Leon, Barber, et al., 2007) small study in which the pattern of results suggested PFPP was particularly more effective than ART for Cluster C patients (albeit that ART is a different form of cognitive-behavioral treatment than Panic Control Therapy). Under these circumstances, no clear statement can be made about moderation other than to say that the findings do not suggest that a psychodynamic therapy should be selected over CBT when a panic disorder patient has a Cluster C diagnosis. Given the small number of patients with Cluster A and B disorders in our study, our results do not speak to treatment selection where other personality disorder types are concerned.

We tested whether RF would predict better treatment outcome across therapies, as was the case in treatment of depressed patients with very low RF (Ekeblad et al., 2015), or would moderate outcome such that those low in RF would do better in CBT than in PFPP due to their low capacity for reflection on mental states. The effects for prediction and moderation both failed to reach significance. For heuristic purposes we probed the moderation effect and found that, consistent with the hypothesis, CBT was superior to PRPP at low and average levels of RF but not at high levels. We note that average RF scores were substantially higher in our study than in that of Ekeblad et al., and this may have made it less likely that we would find this variable to be a predictor of outcome in our sample.

Contrary to hypothesis, we did not find that patients who had suffered a significant loss in childhood benefited more from PFPP than from CBT. This is not to say that a panic disorder patient who is suffering from such a loss and needs to talk about it might not do better in a treatment that has space for such a discussion (which would have been difficult in our highly structured CBT unless therapists went off protocol). However, the mere fact of such a loss does not indicate that the patient needs to focus on this in treatment: Childhood loss neither predicted poorer outcome nor moderated results of treatment.

Limitations

There are several limitations to our tests of moderation: The obtained moderation effects, although replicated across sites, were not those we predicted, our study had adequate power only for large moderation effects thus risking Type II error, and to maintain power we conducted a number of tests without correcting for Type I error. Moreover, we acknowledge that the substantial majority of our patients were White and non-Latino. Whether our results would generalize to a more racially and ethnically diverse sample remains to be determined. In addition, we conducted CBT on a slower schedule than is often the case, spreading a treatment often delivered in 11-12 sessions (e.g., Barlow, Gorman, Shear, & Woods, 2000) over 19-24 sessions. It is impossible to assess how this might have affected the results, although we note the consistency of our three predictors of outcome with the prior literature.

CBT does not work for all patients (e.g., Milrod et al., 2016), and we had hoped to detect moderators identifying patients who might benefit particularly from a relatively new treatment with empirical support of efficacy – PFPP. These hopes were not fulfilled. At best, we identified patients for whom PFPP was a reasonable alternative to CBT, the treatment with the most empirical support: These were patients whose panic disorder began later in life (age 27.5 or later) and patients who were average or above average on expectancy for improvement with treatment. Given that psychologists are enjoined to consider patients' preferences as well as empirical evidence in selecting a treatment approach (American Psychological Association Presdential Task Force on Evidence-Based Practice, 2006), these results are important for practitioners treating patients with panic disorder who would prefer a psychodynamic approach.

We failed to identify patients who would benefit preferentially from a psychotherapy other than CBT using the set of moderators we had available. However, there were important potential moderators that we did not assess or had inadequate sample size to test, such as the effect of Cluster B diagnoses and of separation anxiety. In light of a recent report by Milrod et al. (2014) of preliminary data indicating separation anxiety moderated the outcome of PFPP vs. ART in treatment of panic disorder (with PFPP being superior for patients with separation anxiety) and Aaronson et al. (2008)'s study finding that adult separation anxiety predicted worse treatment outcome for CBT for panic disorder, it is unfortunate that we did not have a measure of separation anxiety in our data set. Finally, we note that our measure of childhood loss was heterogeneous, and it is highly likely that some kinds of loss (e.g., death of a parent) have greater implications than others (e.g., parental separation for 6 months or more). It is possible that had we had the sample size to test the effects of the more severe forms of loss, our results would have been different, and we would have found that PFPP would be advantageous for such patients.

Conclusions and Clinical Implications

We identified two treatment moderators – expectancy and age of onset – that indicate when patients are significantly more likely to benefit from CBT than from PFPP. These effects were replicated across the two study sites, enhancing confidence in their reliability. In addition, for heuristic purposes we have reported two cases where moderation effects were not statistically significant in the overall tests of moderation but where the tests of simple effects were significant. These were reflective functioning and OCPD. Noting the difficulty of amassing sufficient statistical power to test medium but nonetheless potentially meaningful moderators in a single study, Schneider et al. (2015) advocated reporting such results to make them available for meta-analysis, and in that spirit we highlight them as potential targets for future research.

Finally, we have identified two predictors of poor treatment outcome (again, expectancy and age of onset) and, have suggested how these might guide therapists' efforts to improve outcomes for patients with these characteristics. An important next step is to elucidate the mechanisms behind these prediction effects, and we plan both qualitative and quantitative studies to test such mechanisms. Such research, along with the present study, will provide direction to continued efforts to improve treatment outcome for panic disorder.

Public Health Significance.

Panic disorder is associated with substantial societal and economic costs. With current psychotherapies, as many as a third of patients do not benefit substantially from treatment. Identifying predictors of better outcome may point to ways to improve treatment delivery, whereas identifying moderators of response to particular psychotherapies may allow a more personalized approach to treatment. In this study expectancy of improvement and later age of onset both predicted better outcome and moderated response to cognitive-behavioral therapy (CBT) vs. panic-focused psychodynamic therapy (PFPP) such that CBT was the more beneficial treatment when patients were low in expectation of improvement and when patients' panic disorder began before age 27.5.