What Matters More? Common or Specific Factors in Cognitive Behavioral Therapy for OCD: Therapeutic Alliance and Expectations as Predictors of Treatment Outcome

Abstract

CBT for obsessive-compulsive disorder (OCD) is a strong challenge to the contention that common factors explain most of the variance in outcomes in all therapies and all disorders, given that the treatment is focused and placebo response is low. In this study, the relative contributions of expectancy and therapeutic alliance as predictors of outcome in the treatment of OCD are examined and compared to the contribution of specific treatment effects. One hundred and eight patients with OCD were randomly assigned to two forms of CBT: exposure and response prevention (EX/RP) or stress management training (SMT). Measures of OCD symptoms, quality of life, therapist and patient expectancy and alliance were collected at several timepoints. Treatment type was a substantially stronger predictor of symptom reduction compared to alliance and expectancy. However, neither specific nor common factors predicted improvement in quality of life very well. Only in EX/RP, symptom change was associated with subsequent changes in alliance. Finally, therapist effects were estimated using Bayesian methods and were negligible.

Conclusions

In the context of CBT for OCD, the data support the specific factor model, and suggest that the relative contribution of common vs. specific factors likely varies by disorder and by treatment type.

Predictors of outcome in psychotherapy have been traditionally divided to common and specific factors (Frank & Frank, 1993; Lambert, Shapiro, & Bergin, 1986)¹. Common factors relate to therapeutic variables, which seem to be present in all forms of psychotherapy: most notably therapeutic alliance and expectations. Specific factors relate to the particular aspects of psychotherapies derived from distinct theories and models, producing unique forms of technique and intervention. The relative contribution of these factors as predictors of outcome has been long debated (e.g. Rosenzweig, 1936). Wampold and others (Luborsky et al., 2002; Wampold et al., 1997; Wampold & Imel, 2015) have argued that common factors predominantly predict outcome: any form of psychotherapy delivered in a therapeutic context, where both therapist and client expect treatment to work, should work equally for all disorders. Opponents of this view have argued for specificity of technique (e.g. DeRubeis et al 2005; Siev, Huppert, & Chambless, 2009), presenting cases where specific forms of psychotherapy explain a considerable proportion of outcome variance purportedly due to techniques that address specific problems.

The associations between common factors and outcome are well documented, though the interpretation of these associations is still debated. Several meta-analyses have attempted to estimate the alliance-outcome correlation (e.g., Martin, Garske, & Katherine, 2000), with the most recent reporting a modest correlation of .28 (Horvath, Del Re, Flückiger, & Symonds, 2011). Smaller estimations have been reported for the expectancy-outcome correlation. For example, in a meta-analysis of 46 studies, Constantino, Arnkoff, Glass, Ametrano, & Smith (2011) reported a correlation of .12 between expectancy and outcomes.

Though the general alliance-outcome association clearly exists, several considerations limit using this correlation to determine the size and extent to which alliance predicts treatment outcome. First, many previous studies tend to examine alliance as a static moderator, instead of examining how changes in alliance relate to symptomatic changes over time (Zilcha-Mano, 2017). In such studies, it is rare to account for symptom change prior to measuring the alliance, and therefore there is often a potential confound between early symptom change and alliance (Barber, 2009). Recent studies employing advanced statistical methods have stressed the importance of examining the bi-directional association of alliance-outcome (Zilcha-Mano, 2017), as well as the importance of disaggregating within and between client effects (Wang & Maxwell, 2015). Second, Horvath et al.’s (2011) estimate of the alliance-outcome correlation stems from aggregating over many disorders and treatments. Whereas this estimate is important for addressing the general association, specific disorder or disorder by treatment type might be important moderators. Siev and colleagues have demonstrated that aggregation of disorders may wash out differences among treatments (Siev, Huppert, & Chambless, 2009). Consider, as a test case OCD versus depression. OCD is chronic if untreated (Marcks, Weisberg, Dyck, & Keller, 2011), has muted response to placebo (Sugarman, Kirsch, & Huppert, 2017), but responds well to specific forms of therapy (Öst, Havnen, Hansen, & Kvale, 2015). Common factors such as alliance may have a very different influence on outcomes in comparison to depression, a disorder that is quite malleable, with high rates of spontaneous remission and has been shown to be quite responsive to placebo and to nonspecific treatments (Cuijpers et al., 2012). Indeed, in their meta-analysis, Cuipers et al. suggested that 33.3% variance of treatment of depression can be accounted for by extratherapeutic factors, 49.6% by nonspecific effects, and only 17.1% by specific effects. Third, the results of Hovarth et al. (2011) were estimated by combining outcome measures, thereby masking possible differences between symptom reduction and general quality of life (c.f., Siev et al,. 2009). Whereas a decrease in symptoms should be related to an increase quality of life, alliance may be correlated differently with each. Indeed, proponents of each side of the common vs. specific factor debate have stressed the importance of this distinction. Wampold and Imel (2015) have argued that the alliance, or the “real relationship”, should have a greater impact on quality of life than on symptom reduction per se. On the other hand, Siev et al. (2009) argue for the importance of specific outcome measures, which should be influenced most by the specific ingredients of psychotherapy. Overall, the common factor model contends that “the alliance is a central construct in the Contextual Model – if it is not associated with outcome in a robust fashion, the Contextual Model is, or should be, at risk for abandonment.” (Wampold & Imel, 2015, p. 178).

Some theorists suggest that common factors might promote change more in broader and/or more flexible treatments (e.g., supportive psychotherapy; Huppert, Fabbro, & Barlow, 2005; Kazdin 2005). Indeed, some studies have found that supportive treatments have stronger alliance-outcome relationships than structured or theory based treatments (Carroll, Nich, & Rounsaville, 1997; Cailhol et al., 2009). Similarly, the alliance was found to be related to placebo but not medication arms in a trial of the treatment of depression (Zilcha-Mano Roose, Barber, & Rutherford, 2015). An additional theory proposed by Fava, Guidi, Rafanelli, & Rickels (2017) states that it is possible that a more effective intervention may work only according to its specific mechanism and not in addition to the mechanisms that are involved in less effective treatments. However, others contend that common factors should be equivalent in all treatments, (e.g., Wampold & Imel, 2015). It is therefore important to examine whether common factors have a greater impact on broader treatments (i.e., stress management training) than on focused treatments (i.e., exposure and response prevention).

Common factors in treatment for OCD

As noted above, treatment of OCD is an important test case in examining common factors given the specific techniques used in its treatment. However, only a few studies have examined the influence of common factors, with mixed results. Hoogduin, de Haan, & Schaap, (1989) reported that mid-treatment patient alliance, but not early alliance was associated with outcome. Mid-treatment patient alliance was reported to be associated with outcome also by Vogel, Hansen, Stiles, & Götestam, (2006). On the other hand, Keijsers, Hoogduin, & Schaap (1994) found that early patient alliance (though not therapist alliance) was associated with improvement in obsessive fears, but not compulsive behaviors. In a more recent study, Simpson et al. (2011) reported the early alliance predicted outcome, but this effect was mediated by adherence. Finally, in a recent report, Wheaton, Huppert, Foa, & Simpson (2016) found that overall early alliance was not significantly related to outcome. Note that all of these studies did not take into account early symptom improvement nor did they examine temporal associations of alliance over time.

Mixed results have also been reported for the expectancy-outcome association in OCD. A number of reports found no significant effect of expectancy on outcomes (e.g. Başoğlu, Lax, Kasvikis, & Marks, 1988; Freeston et al., 1997; Steketee et al., 2011), whereas other studies reported significant associations. Expectancy was found associated with outcome in treatment of children (Lewin, Peris, Lindsey Bergman, McCracken, & Piacentini, 2011), and in adult group therapy (Vorstenbosch & Laposa, 2015), though only expectancy change during the first sessions and not pre-treatment expectancy was significantly associated with outcome in the latter study. Simpson et al. (2011) reported an association between expectancy and outcomes, but this association was no longer significant when including alliance and other variables in a multivariate model.

Therapist effects

Advocates for a greater role of common factors have suggested that differences between therapists should account for more variance than differences among treatments (Kim et al., 2006; Wampold & Imel, 2015). Whereas a meta-analysis has shown that therapist effects are smaller in research trials than in naturalistic studies (Baldwin & Imel, 2013), others have still argued that therapists should still play important roles in any effective treatment. For example, Wampold and Imel have stated that “Therapist effects generally exceed treatment effects, which at most account for 1% of the variability in outcomes” (p. 176, Wampold & Imel, 2015). However, some research has reported little therapist variance and significant patient variance in manualized protocols (e.g., in CBT for panic disorder; Huppert et al., 2014).

The current study

Few studies have examined the relative contribution of common vs. specific factors in a given study (that is, they tend to either examine the role of common factors or specific factors). The purpose of the current study is to examine the relative contribution of expectancy and alliance in predicting outcome over and above specific ingredients of treatment for OCD. These factors are examined within the context of a randomized clinical trial that examined the difference between two cognitive behavioral treatments (CBT): exposure and response prevention (EX/RP), a treatment based on directly addressing exposure to feared situations and cessation of compulsions, and stress management training (SMT²; Simpson et al., 2008), a treatment that offers support, problem-solving, assertiveness, and relaxation skills in order to reduce stress that maintains symptoms. The current study attempts to advance our knowledge in this debate in a number of ways. First, to quantify the relative contribution of common and specific factors, we conducted model comparisons and generated effect sizes. In addition, alliance was measured at multiple time points to examine changes of alliance over time as well as disaggregating within and between effects. Furthermore, cross lag-analyses were used to examine directionality of the alliance-outcome association in both treatments. To test for differences according to outcome types, symptom severity and quality of life were examined separately. Finally, because therapists effects are included in this debate, therapist effects were calculated using Bayesian methods to yield better estimations and confidence intervals even with small samples of patients per therapist. The aim was to conduct a strong test of the common factor vs. specific factor effects.

The following hypotheses were examined: 1) Regardless of whether therapist and patient expectancies were higher in EX/RP than in SMT, we predicted that expectancy will be a significantly weaker moderator of overall outcomes in comparison with treatment condition. In addition, therapist and patient expectancies are hypothesized to be more related to outcomes in SMT than in EX/RP; 2) We predicted that early alliance will account for significantly less amount of the outcomes when compared to treatment condition. Similar to hypothesis 1, we also hypothesized that early alliance will be more associated with outcomes in SMT than in EX/RP; 3) Changes in outcomes will predict subsequent changes in alliance unidirectionally in the EX/RP group whereas in SMT changes in alliance scores will unidirectionally predict subsequent changes in outcomes. 4) There will be more variability in outcomes between patients than between therapists. 5) We predicted that for all analyses, treatment condition will be more related than common factors to symptom severity, whereas quality of life will also be more influenced by treatment condition, though less so.

Method

Participants were recruited between November 2000 and November 2005 at two sites: Philadelphia and New York. The study was approved by institutional review boards at each site. One hundred and eleven adults were randomized to EX/RP (n=56) or SMT (n=55). Three participants withdrew prior to treatment, resulting in two equal size treatment groups of 54 patients each. Study participants were required to be on a stable dose of serotonin reuptake inhibitor for at least 12 weeks prior to entry and during treatment. For complete description of the study design and CONSORT diagram, see Simpson et al. (2008).

Participants

Participants who began treatment were adults (age: M = 39.2, SD= 13.9; sex: 43% female) who received a DSM-IV diagnosis of OCD with a duration of at least one year (average duration: 22 years), and reported at least minimal improvement from an adequate SRI trial but remained symptomatic (YBOCS ≥ 16; average YBOCS = 25.82). Patients were excluded for mania, psychosis, prominent suicidal ideation, substance abuse or dependence, an unstable medical condition, pregnancy or nursing, or prior CBT while receiving an adequate SRI trial. Other comorbid diagnoses were permitted if secondary to OCD (44% received comorbid diagnoses). Diagnoses were confirmed by the Structured Clinical Interview for DSM-IV (First, Spitzer, Gibbon, & Williams, 1996). All participants provided written informed consent prior to entry and randomization in the study. In the consent, a brief overview of each treatment was provided. EX/RP was presented as the standard treatment for OCD and SMT was presented as an alternative therapy which has been proven to reduce anxiety.

Treatments

Although different in content, the format of both CBT treatments was identical, and a manual was written for each one: 17 twice-weekly sessions (each 90–120 minutes), daily homework assignments, and between-session phone calls (twice per week, each <20 minutes). Expert supervisors in each approach led weekly group supervision teleconferences after reviewing videotaped sessions. Four experienced CBT clinicians, blind to outcome, assessed the use of prescribed treatment procedures in 25 randomly selected sessions of EX/RP, and 28 randomly selected sessions of SMT. Therapists exhibited excellent protocol adherence: for EX/RP cases, 83% of EX/RP (versus 0% of SMT) procedures were used; for SMT cases, 79% of SMT (versus 6% of EX/RP) procedures were used³.

Exposure and response prevention (EX/RP)

The protocol for EX/RP followed the procedures of Kozak and Foa (1997). It included two treatment planning sessions and 15 sessions using in vivo and/or imaginal exposures, at least two of which occurred in the participant’s home environment to promote generalization. Patients were asked to stop ritualizing after the first exposure session. The rationale provided to patients was that by experiencing exposure without rituals, they would learn that anxiety decreases with time alone (“habituation”) and that feared consequences do not occur. As homework, patients were asked to record any rituals and spend at least 1 hour per day conducting self-guided exposures.

Stress management training (SMT)

Stress management training included two introductory sessions and 15 active treatment sessions including breathing retraining, progressive muscle relaxation, assertiveness training, and problem solving. The rationale for the treatment was that life stressors can trigger and exacerbate OCD symptoms and that stress management skills would reduce stress and thereby reduce OCD symptoms and that in addition, reducing stress would also reduce their symptoms via facilitation of medication effects. As homework, patients were asked to monitor daily stressors and practice the stress management skills for at least 1 hour each day.

Therapists

Ten therapists, nine psychologists and one psychiatrist, provided both forms of therapy in study (5 therapists at each study site). The average number of patients per therapist was 10.8 (range: 3–24). Therapists at each study site received training and supervision from faculty from Philadelphia who were experts in either EX/RP or SMT and who had no contact with study patients. Training included reviewing each manual and completion of at least one training case of each treatment type under supervision.

Measures

Symptom severity was evaluated using the clinician rated Yale-Brown Obsessive Compulsive Scale for OCD (YBOCS; Goodman, Price, Rasmussen, Mazure, Delgado et al., 1989; Goodman, Price, Rasmussen, Mazure, Fleischmann et al., 1989). YBOCS was administered by independent evaluators blind to treatment assignment at three timepoints: baseline (week 0), mid-treatment (week 4) and at the end of treatment (week 8). To examine interrater reliability for these assessments, a second independent evaluator listened to 30 taped diagnostic interviews; intraclass correlations were high (r = .96, p < .001).

Quality of life was assessed using Quality of Life Enjoyment and Satisfaction Questionnaire (QLESQ; Endicott, Nee, Harrison, & Blumenthal, 1993) a self-report measure administered also at weeks 0, 4, and 8. Alliance was measured using the Working Alliance Inventory – Short form, Revised (WAI; Tracey & Kokotovic, 1989) for clients (WAI-C) and therapists (WAI-T). Alliance was assessed at four timepoints: twice early in treatment (after the first and second introductory sessions), once in mid-treatment (after the session 10) and at the end of treatment (after session 17). Patients expectancy ratings (PER) and therapists expectancy ratings (TER) were obtained after the second session using 3 Likert scales (0–8) assessing expectancies regarding improvement in treatment on obsessions, compulsions and general distress⁴. Scores were summed to create one expectancy score for patients and another for therapists. Both measures showed high internal consistency (PER: α = 0.79, TER: α = 0.99).

Data Preparation and Data Analytic Approach

Data were analyzed using full intent-to-treat longitudinal mixed effects models (LMLM). Data analyses were conducted on the full sample, though particular analyses which estimated common and specific factor effects included fewer patients when there were no data for a given patient (sample sizes of given analyses range from 43 to 51 per group). Hypotheses 1 and 2 examine to what extent common factors (expectancy and alliance) moderate outcome change. First, we examined the interactions of factors with time via LMLM adjusted for repeated measures using restricted maximum likelihood estimation methods using ‘nlme’ R package in R (Pinheiro, Bates, DebRoy, Sarkar, & R Core Team, 2016). This method has been shown to be robust when there are data that are missing at random. In addition, for direct comparison of the specific vs. common factor models we employed a model comparison approach and computed model comparison statistics. The common factor model was compared to the specific factor model, and both models were compared to a combined model. All model formulas are specified in the supplemental material. The combined model was a 9 parameter model, including effects for time (Time; level one predictor), treatment group (Tx; level two predictor), a common factor (CF - patient/therapist early alliance or expectancies; level two predictors), and two interactions: time by treatment group (Time X Tx), and time by common factor (Time X CF). The remaining parameters included the intercept, and three variance parameters: intercept variance, time slope variance and residual variance. The common factor model was modeled similar to the combined model, with the exclusion of the time by treatment interaction. The specific factor model was modeled in the same manner, by removing the time by common factor interaction from the combined model. This method ensured that both the common factor model and the specific factor model were nested within the combined model, and therefore can be compared to the same model. Moreover, effects of treatment and common factor at baseline were included in all of these analyses to compare models that differed only in the parameter of interest (factor by time). Time was coded based on weeks in therapy (0–8) and centered around 0 (first measurement). Early alliance was computed by averaging alliance measured after the first and second session. All common factors were group mean centered. Dummy coding was used to code treatment group: 0 for SMT and 1 for EX/RP. For estimating slopes for SMT was set at 1 instead. When comparing models which are not nested, we used 3 model comparison statistics: the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC) and the Nakagawa & Schielzeth R² (Johnson, 2014; Nakagawa & Schielzeth, 2013). Both BIC and AIC are used for model selection, with lower values indicating better fit. Both calculate the fit of the model to the data while controlling for overfitting by introducing a penalty term for the number of parameters in the model. Nakagawa & Schielzeth R²s were computed using package ‘r2glmm’ in R (Jaeger, 2016). When comparing two nested models, likelihood ratio test were used in addition to the other statistics.

Hypothesis 3 examined the cross-lagged time-varying relationship between alliance and outcome. Following Wang and Maxwell’s (2015) recommendations, we did not control for linear time effects in our data. Three models were fit separately for each alliance (patient/therapist) and outcome (YBOCS/QLESQ) pair. In the first model, we examined the prediction of outcome from alliance at the same timepoint (simultaneous model). In the second model, we examined the prediction of outcome from alliance in the previous timepoint. In the third model, we reversed the roles and examined the prediction of alliance from outcome in the previous timepoint (cross-lag models). This approach can potentially facilitate a better understanding of the direction of the relationship and also establish its temporal sequence. Following the recommendations of Falkenström, Finkel, Sandell, Rubel, and Holmqvist (2017), we did not include the lagged dependent variable as a predictor due to the problem of endogeneity.

Hypothesis 4 examined therapist effects. Following Baldwin & Imel’s (2013) review, ICCs were computed by dividing the therapist or patient variance component by sum of the therapists and patients variance components. Estimates were calculated following the procedure in Kim, Wampold, & Bolt (2006) and Wampold & Bolt (2006), which recommend using a two level model, and controlling for symptom severity by adding pre-treatment outcome scores as a predictor of post-treatment outcome scores. Whereas therapists treated patients in both EX/RP and SMT, the treatment effect was controlled for as well by entering the treatment group as a predictor. The model formula is presented in the supplemental material. Using this model, 4 variance components were estimated, 3 for the therapist level and one for the patient level. For the therapist level: a) variance attributed to differences between therapists in treatment outcome (outcome variance) b) variance attributed to differences between therapist in the effect of symptom severity on treatment outcome (severity variance), c) variance attributed to differences between therapist regarding the effect of treatment group on treatment outcome (treatment group variance). For the patient level, only one variance component was estimated: the variance attributed to differences in outcomes among patients within therapists⁵. To account for all four sources of variance we computed four ICCs, corresponding to the relative proportion of variance between the four variance components. Because variance estimations may be close to zero, and the due to the asymmetric nature of the ICCs likelihood function, maximum likelihood estimations may be biased (Zilcha-Mano et al., 2015). Bayesian estimates, using Markov Chain Monte Carlo (MCMC) procedures seem to be more equipped for this analysis (Betancourt & Girolami, 2015; McElreath, 2016) and were therefore employed for examining the first hypothesis. Analysis was executed using R (version 3.4) and Stan via ‘rstanarm’ package (Stan Development Team, 2016). Rstanarm by default takes a conservative approach using weakly informative priors which provide moderate regularization and stabilize computation. Due to the asymmetric shape of the posterior distribution, median estimates were used when referring to point estimates, accompanied with 95% confidence intervals (highest posterior density intervals). Modes and means are provided as well, in the appropriate table.

Results

Therapist effects

As seen in Table 1, the median proportion of variance attributed to differences between therapists in outcomes on the YBOCS (therapist ICC for outcome) was estimated at 0.12 % with a 95% confidence interval between 0% and 3.06%. As can be seen in Table 1, similarly negligible effects were found for the median estimation of ICCs representing the differences between therapists regarding the effect of patient severity and treatment group. However, the median for estimating the patients ICC, representing the proportion of variance attributed to differences in YBOCS outcomes among patients within therapists was estimated at 99.62% with 95% confidence interval between 94.29% and 100%. Similar effects were found for the QLESQ, as can be seen in Table 1. In sum, therapist effects in this sample were extremely small, accounting for very little variance in outcomes in accordance with hypothesis 4. Therefore, we did not continue to take therapist effects into account in further analyses.

Table 1.

Posterior distributions for variance components and ICC for therapist and patient level on the YBOCS and QLESQ, obtained using MCMC procedure.

				Mode	Median	Mean	Lower 95%	Upper 95%
YBOCS	Variance component	Therapists	Outcome	<0.0001	0.0370	0.2002	<0.0001	0.9699
			Effect of severity	<0.0001	0.0262	0.0667	<0.0001	0.2589
			Effect of treatment group	<0.0001	0.0405	0.1561	<0.0001	0.6477
		Patients	Outcome	27.4323	30.4362	30.9285	22.5225	40.294
	ICC	Therapists	Outcome	<0.0001	0.0012	0.0062	<0.0001	0.0306
			Effect of severity	<0.0001	0.0008	0.0021	<0.0001	0.0085
			Effect of treatment group	<0.0001	0.0013	0.0049	<0.0001	0.0204
		Patients	Outcome	>0.9999	0.9962	0.9874	0.9429	>0.9999
QLESQ	Variance component	Therapists	Outcome	<0.0001	0.0436	0.8005	<0.0001	3.8495
			Effect of severity	<0.0001	0.0188	0.0456	<0.0001	0.1764
			Effect of treatment group	<0.0001	0.0372	0.4756	<0.0001	1.8838
		Patients	Outcome	143.398	149.6915	152.1805	109.8615	199.4425
	ICC	Therapists	Outcome	<0.0001	0.0003	0.0050	<0.0001	0.0247
			Effect of severity	<0.0001	0.0001	0.0003	<0.0001	0.0012
			Effect of treatment group	<0.0001	0.0002	0.0029	<0.0001	0.0124
		Patients	Outcome	>0.9999	0.9992	0.9918	0.9601	>0.9999

Note. YBOCS = Yale-Brown Obsessive Compulsive Scale; QLESQ = Quality of Life Enjoyment and Satisfaction Questionnaire; ICC= Intraclass correlation; MCMC = Markov chain Monte Carlo.

Treatment effects (specific factor)

As reported previously in Simpson et al. (2008) the average decrease on the YBOCS per week in EX/RP treatment group was −1.36 points (t₁₀₅ = −13.13, p < .001) compared to −0.44 points (t₁₀₆ = −4.20, p < .001) in SMT. This difference was significant (b = 0.92, t₁₀₆ = 6.22, p < .001). The average increase of the QLESQ per session in EX/RP was 1.45 points (t₁₇₈ = 5.73, p < .001) compared to 0.77 points (t₁₈₀ = 3.00, p = .003) in SMT. Whereas the QLESQ increased significantly in both groups, the difference between the groups increase did not reach significance (b = −0.67, t₁₇₉ = −1.87, p = .063). In sum, both treatments reduced symptoms and improved life quantity, though symptoms changed considerably more in EX/RP than in SMT.

To calculate an initial estimate of the magnitude of the effect of specific vs common factors, we examined the proportion of change in outcomes shared by both treatments, compared to the proportion of change which differed between treatments, which should be attributed to differences in technique. There was an average reduction of 11.2 points on the YBOCS in EX/RP, and a reduction of 3.6 points in SMT. Thus, of the 11.2 points reduction in EX/RP, one could argue that 3.6 points, or 32%, of the improvement was accounted for by factors beyond specific technique, suggesting that 68% was due to specific techniques employed in EX/RP. Similarly, on the QLESQ, there was an average increase of 12.9 points in EX/RP and an increase of 6.2 points in SMT. Thus, 48% of the improvement can be accounted for by factors other than the specific technique, suggesting that 52% was due to specific technique. However, this is a general estimate, which might be biased if groups differ in common factors (e.g., if groups differed in expectancies and these differences moderated treatment outcome). Therefore, we examined possible differences in common factors, and the degree in which these differences moderated treatment outcome.

Expectancies (common factor I) effects

Although both groups exhibited moderate expectancies, both patients' and therapists' expectancies were significantly higher in EX/RP. The mean patient expectancy in EX/RP was 18.27 (SD = 3.90) compared to 15.57 (SD = 4.64) in SMT (t₉₇ = 3.16, p = .002). The mean therapist expectancy in EX/RP group was 18.00 (SD = 2.64) compared to 13.13 (SD = 3.87) in SMT (t₇₉ = 6.99, p < .001). These differences, according to the common factor model, should explain differences in outcome, therefore we proceeded to explore the degree these differences moderated change in outcome during treatment.

To estimate to what degree early expectancies moderate changes in YBOCS and QLESQ, we examined the time by expectancy interactions. Patient expectancies were not significantly associated with changes in outcome (for YBOCS: b = −0.03, t₁₁₃ = −1.78, p = .078; for QLESQ: b = 0.06, t₁₆₇ = 0.96, p = .338), nor were therapist expectancies (for YBOCS: b = −0.01, t₉₄ = −0.76, p = .451; for QLESQ: b = −0.01, t₁₄₉ = −0.23, p = .815). None of the three way interactions (time by treatment group by expectancy) were significant (all p’s: .356–.787).

In addition to examining the extent in which specific and common factors moderate outcome, it is important to compare the two models directly. Thus, we compared the fit of specific vs. common factor models to the data by computing model comparison statistics. As seen in Table 2, when predicting symptom change (YBOCS), the data fit considerably better with the specific factor model (treatment effects; model 2 and model 5) compared to the common factor models (patient expectancies: model 1; therapist expectancies: model 4) in accordance with hypothesis 1. Moreover, the combined models (including both common and specific factors simultaneously: models 3 and 6) did not improve fit over the specific factor models, and may very well represent a case of overfitting, evident from the lack of decrease in information criteria (AIC and BIC). In terms of predicting change in quality of life (QLESQ), the data fit slightly better with the specific factor models (models 8 and 11) compared to the common factor model (model 7 and model 10). The specific factor models were not superior to the common factor model in variance explained when modeling QLESQ with patient expectancies, nor with therapist expectancies, contrary to hypothesis 5.

Table 2.

Model comparison table for prediction of Yale-Brown Obsessive Compulsive Scale (YBOCS) and of Quality of Life Enjoyment and Satisfaction Questionnaire (QLESQ) from common factor model (expectancy), specific factor model (treatment), and combined factors model (expectancy and treatment).

Outcome measure	Model	Predictors	Np	R2	MC	χ2	p-value	ΔAIC	ΔBIC	ΔR2
YBOCS	1	Time, PER, Tx, Time * PER	8	.31	2 vs. 1			−23.70	−23.70	.07
	2	Time, PER, Tx, Time * Tx	8	.38	3 vs. 1**	26.83	<0.001	−24.83	−21.16	.09
	3	Time, PER, Tx, Time * PER, Time * Tx	9	.40	3 vs. 2	3.13	.077	−1.13	2.54	.02
	4	Time, TER, Tx, Time * TER	8	.30	5 vs. 4			−12.93	−12.93	.04
	5	Time, TER, Tx, Time * Tx	8	.34	6 vs. 4 **	14.11	<0.001	−12.11	−8.54	.05
	6	Time, TER, Tx, Time * TER, Time * Tx	9	.35	6 vs. 5	1.18	.277	0.82	4.39	.01
QLESQ	7	Time, PER, Tx, Time * PER	8	.12	8 vs. 7			−2.44	−2.44	<.01
	8	Time, PER, Tx, Time * Tx	8	.13	9 vs. 7	2.93	.087	−0.93	2.69	.01
	9	Time, PER, Tx, Time * PER, Time * Tx	9	.13	9 vs. 8	0.49	.483	1.51	5.13	<.01
	10	Time, TER, Tx, Time * TER	8	.09	11 vs. 10			−0.51	−0.51	<.01
	11	Time, TER, Tx, Time * Tx	8	.09	12 vs. 10	1.43	.232	0.57	4.09	.01
	12	Time, TER, Tx, Time * TER, Time * Tx	9	.10	12 vs. 11	0.92	.337	1.08	4.60	.01

Note: PER = patients expectancy ratings; TER = therapist expectancy ratings; Tx = treatment group; MC = model comparison; Np = number of parameters in the model; R² = Nakagawa, S, Schielzeth, H. (2013) marginal R²; χ² = chi-square statistic used in the deviance test; p-value= for chi-square distribution with one degree of freedom; ΔAIC = Akaike Information Criterion difference; ∆BIC = Bayesian information criterion difference.

Early Alliance (common factor II) effects

Groups did not significantly differ in early patient alliance scores (CBT: M = 71.35, SD = 8.75; ABM: M = 69.02, SD = 7.59; t₉₁ = 1.37, p = .174). However, therapists’ early alliance scores were higher in EX/RP than SMT (CBT: M = 65.13, SD = 5.90; ABM: M = 62.05, SD = 7.58; t₇₇ = 2.08, p = .040). As with expectancies, these differences, according to the common factor model, should explain differences in outcome. Therefore, we proceeded to explore the degree in which differences in early alliance moderated change in outcome.

Early scores on the WAI-C were not associated with change in outcomes (for YBOCS: b = −0.01, t₉₈ = −0.71, p = .479; for QLESQ: b = −0.02, t₁₆₄ = −1.04, p = .300), nor were early scores on the WAI-T (for YBOCS: b = −0.01, t₉₄ = −0.76, p = .451; for QLESQ: b = −0.01, t₁₄₉ = −0.23, p = .815). None of the three way interactions (time by treatment group by alliance) were significant (all p's: .097–.881).

Additionally, we compared the fit of specific vs. common factor models to the data by computing model comparison statistics. As seen in Table 3, when predicting symptom change (YBOCS), the data fit considerably better with the specific factor model (treatment effects; model 14 and model 17) compared to the common factor models (patient alliance: model 13; therapist alliance: model 16), in accordance with hypothesis 2. Moreover, the combined models (including both common and specific factors simultaneously: models 15 and 18) did not improve fit over the specific factor models, and may very well represent a case of overfitting, evident from the lack of decrease in information criteria (AIC and BIC). In terms of predicting change in quality of life (QLESQ), the data fit slightly better with the specific factor models (models 20 and 23) compared to the common factor model (models 19 and 22). The specific factor model was not superior to the common factor model in variance explained when modeling QLESQ with patient alliance, nor with therapist alliance, contrary to hypothesis 5.

Table 3.

Model comparison table for prediction of Yale-Brown Obsessive Compulsive Scale (YBOCS) and of Quality of Life Enjoyment and Satisfaction Questionnaire (QLESQ) from common factor model (alliance), specific factor model (treatment), and combined factors model (alliance and treatment).

Outcome measure	Model	Predictors	Np	R2	MC	χ2	p-value	ΔAIC	ΔBIC	ΔR2
YBOCS	13	Time, WAI-C, Tx, Time * WAI-C	8	.24	14 vs. 13			−31.52	−31.52	.13
	14	Time, WAI-C, Tx, Time * Tx	8	.37	15 vs. 13**	32.02	<.001	−30.02	−26.44	.13
	15	Time, WAI-C, Tx, Time * WAI-C, Time * Tx	9	.37	15 vs. 14	0.50	.478	1.50	5.08	<.01
	16	Time, WAI-T, Tx, Time * WAI-T	8	.26	17 vs. 16			−29.41	−29.41	.13
	17	Time, WAI-T, Tx, Time * Tx	8	.39	18 vs. 16 **	29.98	<.001	−27.98	−24.49	.13
	18	Time, WAI-T, Tx, Time * WAI-T, Time * Tx	9	.39	18 vs. 17	0.57	.450	1.43	4.92	<.01
QLESQ	19	Time, WAI-C, Tx, Time * WAI-C	8	.14	20 vs. 19			−1.75	−1.75	<.01
	20	Time, WAI-C, Tx, Time * Tx	8	.14	21 vs. 19	2.83	.093	−0.83	2.70	<.01
	21	Time, WAI-C, Tx, Time * WAI-C, Time * Tx	9	.14	21 vs. 20	1.08	.299	0.92	4.45	<.01
	22	Time, WAI-T, Tx, Time * WAI-T	8	.09	23 vs. 22			−2.93	−2.93	<.01
	23	Time, WAI-T, Tx, Time * Tx	8	.09	24 vs. 22	2.98	.084	−0.98	2.45	<.01
	24	Time, WAI-T, Tx, Time * WAI-T, Time * Tx	9	.09	24 vs. 23	0.06	.814	1.94	5.38	<.01

Note: WAI-C = patients alliance ratings; WAI-T = therapist alliance ratings; Tx = treatment group; MC = model comparison; Np = number of parameters in the model; R² = Nakagawa, S, Schielzeth, H. (2013) marginal R²; χ² = chi-square statistic used in the deviance test; p-value= for chi-square distribution with one degree of freedom; ∆AIC = Akaike Information Criterion difference; ∆BIC = Bayesian information criterion difference.

Changes in alliance as predictors of change in outcome

Mean regression coefficients for change per session for the WAI-C and WAI-T. Patient alliance did not increase significantly in either group (for EX/RP: b = 0.33, t₈₆ = 1.85, p = .068; for SMT: b = 0.18, t₈₈ = 1.02, p = .309), nor was the difference between groups significant (b = −0.14, t₈₇ = −0.58, p = .566). Therapist alliance did not increase significantly in either group (EX/RP: b = 0.10, t₇₅ = 0.54, p = .591; SMT: b = 0.29, t₈₃ = 1.44, p = .154), nor was the difference between groups significant (b = 0.19, t₇₉ = 0.71, p = .482). Although linear changes over time were not significant, alliance may still be associated with changes in outcome. Therefore, we explored the alliance-outcome association using simultaneous and cross-lag models.

Simultaneous models

Mean regression coefficients for alliance measures predicting outcome simultaneously and cross-lagged models are presented in Table 4. In EX/RP, patient alliance significantly covaried with YBOCS (b = −0.25, t₁₆ = −2.21, p = .041) but not with QLESQ (b = 0.26, t₃₃ = 0.85, p = .404). In SMT, patient alliance did not covary significantly with the YBOCS (b = −0.08, t₄₄ = −0.61, p = .546) or with the QLESQ (b = 0.4, t₅₃ = 1.25, p = .217). The difference between groups in the association of WAI-C to YBOCS was not significant (b = −0.17, t₂₇ = −1.03, p = .312). Therapist alliance did not covary significantly with YBOCS or QLESQ in either group, and the differences between the groups were not significant (all p’s: .255 – .940).

Table 4.

Mean regression coefficients and 95% confidence interval for simultaneous and cross lag models of the association between alliance and outcome.

Outcome measure	Alliance measure	Alliance predicting outcome simultaneously	Lag alliance predicting outcome	Lag outcome predicting alliance
		EXRP
YBOCS	WAI-C	−.25 [−.49, −.01]*	.09 [−.23, .40]	−.34 [−.68, .00]*
	WAI-T	−.16 [−.45, .13]	.08 [−.23, .39]	−.21 [−.50, .07]
QLESQ	WAI-C	.26 [−.37, .90]	.09 [−.98, 1.16]	.05 [−.07, .18]
	WAI-T	.09 [−.42, .60]	−.58 [−1.32, .15]	−.14 [−.31, .04]
		SMT
YBOCS	WAI-C	−.08 [−.33, .18]	−.27 [−.56, .01]	−.24 [−.73, .25]
	WAI-T	−.09 [−.46, .27]	−.28 [−.63, .07]	.04 [−.46, .53]
QLESQ	WAI-C	.40 [−.24, 1.03]	−.23 [−1.10, .64]	−.13 [−.35, .08]
	WAI-T	.03 [−.67, .72]	.01 [−.76, .78]	−.05 [−.27, .17]

Note. YBOCS = Yale-Brown Obsessive Compulsive Scale; QLESQ = Quality of Life Enjoyment and Satisfaction Questionnaire; EXRP = Exposure and ritual prevention; SMT = Stress management training; WAI-C = Working Alliance Inventory - Client; WAI-T = Working Alliance Inventory - Therapist.

Cells within rows with different superscripts are different from one another at the p < .05 level.

^*/** significantly different from zero at the p < 0.05/0.01 level.

Cross-lag models

As seen in Table 4, In EX/RP, changes in previous YBOCS scores predicted subsequent changes in WAI-C scores (b = −0.34, t₆₆ = −2.00, p = .050), but previous changes in WAI-C did not significantly predict subsequent changes on YBOC scores (b = 0.09, t₁₆ = 0.58, p = .568). This finding suggests that changes in YBOCS precede the changes in WAI-C in EX/RP partially supporting hypothesis 3. In SMT, changes in previous YBOCS scores did not predict subsequent changes in WAI-C scores (b = −0.24, t₆₃ = −0.99, p = .325), nor did previous changes in WAI-C predict subsequent changes on YBOC scores (b = −0.27, t₃₁ = −1.93, p = .062). No significant relationships were found when predicting QLESQ in either of the groups or regarding the WAI-T on either measure (all p’s: .101 – .986). Moreover, differences between groups were not significant (all p’s: .094 – .740).

Discussion

An augmentation trial for the treatment of OCD trial was selected as a test of the common factor model which contends that therapist factors, expectancy, and alliance are the main contributors of outcomes in all treatments, and should predominantly account for any differences in techniques in any disorder (Wampold & Imel, 2015). This is the first study we are aware of to directly test common vs. specific factors by examining the relative contribution of patient or therapist expectancies, alliance from patient and therapist perspectives, and treatment condition on outcomes. We found that alliance did not differ between the two treatment conditions, that neither alliance nor expectancies explained significant variance in the differential outcomes of EX/RP vs. SMT, and that there were minimal therapist differences in outcomes as well as in alliance. We did find differences in therapist and patient expectancies between the two conditions, but these differences did not account for differences in outcomes. In addition, we did not find that any of these factors had a significant, positive, differential impact on primary OCD symptoms or quality of life. Taken together, these results support the specific factor model over the common factor model in the context of treatment for OCD.

As an initial calculation of the estimate of the contribution of common and specific factors to outcomes in psychotherapy for OCD, 32% of the improvement in symptoms was accounted for by factors other than EX/RP, suggesting that the other 68% was due to specific factors. Similarly, we found 48% of the improvement in quality of life was accounted for by factors other than EX/RP vs. 52% due to specific factors. This estimate varies greatly from previous general estimates of specific factors overall (Lambert et al., 1986; Wampold & Imel, 2015) and for depression (c.f., Cuipers et al., 2012). We conclude that these estimates likely vary greatly by disorder and treatment type, and that more studies are needed to determine the effects for a given treatment for a given disorder. In addition, given the extremely small impact that expectancies and alliance had in moderating both types of outcomes, these common factors did not account for the differential effects of EX/RP vs. SMT in the current study.

In terms of expectancies, we found that they were not related to outcomes, findings that are partially consistent with the literature on OCD (Başoğlu, Lax, Kasvikis, & Marks, 1988; Freeston et al., 1997; Steketee et al., 2011). It is possible that expectancies may have differential impacts on outcome, depending on the disorder. Indeed, differential placebo effects in OCD vs. other disorders support this notion (Sugarman et al., 2017). Alternatively, other factors such as severity, chronicity, or treatment augmentation could moderate the impact of expectations on outcomes.

The findings that neither patient nor therapist early alliance were predictive of outcomes in either treatment condition is largely consistent with the literature on the impact of alliance in treatment of OCD (as reviewed above). Whereas early alliance was a weak predictor in our total sample (accounting for less than 2% of the variance), this effect decreased to close to zero when treatment condition by time was entered into the model (with treatment accounting for 13% of the model). Our data reinforce previous findings in the literature which suggest that the relationship between alliance and outcomes in OCD is more complex than the standard literature on alliance indicates (c.f., Horvath et al., 2011). Indeed, the current data examining alliance over time suggest that alliance is more likely a consequence of symptom change than the other way around (c.f., Mano-Silcha et al., 2014), and is more a patient level variable (c.f., del Rey et al., 2012). All these data converge to suggest that whereas alliance may play a supporting role in effecting change in EX/RP for OCD, the major active ingredients are exposure and response prevention. Given these findings, alliance-outcome relationships appear to differ according to disorder and treatment type. Unfortunately, meta-analyses to date have not examined this question.

We predicted that common factors will have less impact in EX/RP whereas they will play a stronger role in SMT. However, none of our three-way interaction models, whether they be with expectancy or alliance or from patient or therapist perspectives added additional explanatory power to our models. This may be due to the fact that there was little influence of either alliance or expectancy on outcomes.

Overall, there were minimal therapist effects in outcomes (median effect of 0.12%), similar to those reported by Huppert et al., (2014; in CBT for panic disorder) and Zilcha-Mano et al., (2015; short term dynamic therapy for depression) in that therapist effects were non-existent or considerably smaller than patient effects. This is in contrast to some other studies which found the opposite (Baldwin, Wampold, & Imel, 2007; in a mixed outpatient sample with various treatment types combined⁶). Whereas therapist effects were minimized in the current trial through careful training and supervision in the context of a research setting examining a specific treatment for a specific disorder, in the real world, therapist effects likely exist in the treatment of OCD and other disorders (Baldwin & Imel, 2013). Indeed, even in a large, controlled research trial, there were site effects in the treatment of pediatric OCD (POTS, 2004), suggesting that therapists at one site were more effective in delivering the same treatment than those at another site. Note that no site effects were found in the study examined here (see Simpson et al., 2008). However, even when such effects exist, it is not clear whether they are due to competency in administering the specific techniques or due to better common factors (DeRubeis, Brotman & Gibbons, 2005).

One of the contributions of the current study is the examination of both primary symptoms (i.e., YBOCS) and quality of life. Most of the significant effects for EX/RP we found were on primary symptoms and not on quality of life. Whereas this could be an argument for the common factors approach (c.f., Wampold & Imel, 2015), we contend it is not. There were no therapist, alliance, or expectancy effects on quality of life and 52% of the outcomes could be explained by technique, all of which contradict the common factor model. In addition, the YBOCS itself includes items regarding functioning within its severity scale, thereby suggesting that a specific treatment (EX/RP) impacts functioning. At the same time, it is possible that we did not see differential improvement in quality of life because improvements in this area were modest. Indeed, quality of life may take more time to improve than symptom reduction.

There are a number of limitations to the study that should be considered. First, whereas the results are from a randomized augmentation trial for OCD, caution should be used in generalizing the findings to all treatment for OCD. However, given that outcomes are similar to those from other trials in the literature (c.f., meta-analysis by Ost et al., 2015), and to effectiveness trials (Franklin et al., 2000), it is likely that the current findings can be extrapolated. Another possible concern is that alliance was high and stable throughout the timepoints data were collected. More frequent measurement of alliance would allow for finer grained analyses that could reveal sequential effects or other associations (c.f. Falkenström, Ekeblad & Holmqvist, 2016). In addition, alliance and outcomes were not measured at exactly the same time, raising the possibility that simultaneous analysis could have found an effect (though this does not apply for the lagged analyses). Outcomes were measured at weeks 0, 4, and 8, equivalent to sessions 1, 8 and 17. Common factors were measured in sessions 1, 2, 10, and 17, thus making the measure proximal to each other. Furthermore, in order to examine impact of alliance and therapists more reliably, more patients per therapist and more measurements of alliance are necessary (Crits-Christoph et al., 2011). Further research is needed to explain the improvements that occurred in SMT given that alliance and expectancy did not account for a significant variance.

It could be argued that therapist allegiance should be controlled in psychotherapy trials by assigning therapists to treatments of their expertise and to which they advocate (e.g., Wampold & Imel, 2015) and that the current study was biased in this aspect. Indeed, therapists and patients alike reported somewhat higher expectancy for EX/RP than SMT. It is important to note, however, that whereas expectancy was higher in EX/RP, it was not a significant predictor of outcome in either OCD symptoms or quality of life. Furthermore, according to general criteria for bona-fide treatments (Wampold & Imel, 2015), SMT was indeed a bona fide treatment, and no other psychotherapy is considered a better candidate to be a bona fide treatment for OCD other than various versions of CBT that overlap significantly with EX/RP. Therefore, it would be close to impossible to find therapists who are adherents of a different approach of psychotherapy as a first line treatment for OCD.

In sum, the current study tested the common vs. specific factor theories using data from a clinical trial of OCD. The data were not supportive of common factors including alliance and expectancy (from either the patient or the therapist perspectives) accounting for treatment variance related to outcomes, whereas specific treatment factor overwhelmed the variance accounted for in symptom improvement. However, improvement in quality of life was largely not predicted by specific or common factors. Therefore, the application of the common factor model to a specific disorder (OCD) was not supported on the basis of the current report. Theories should attempt to integrate findings of common and specific factors that vary by contexts, disorders, and treatments into their models in order to best account for the variability and complexity of the therapeutic endeavor.

Highlights.

• This study directly compares common vs. specific factors in predicting outcomes in CBT for OCD.

• Treatment condition explained a relatively large significant amount of variance in outcomes

• Patient and therapist expectancies/alliance explained relatively little variance of outcomes

• These findings stress the importance of specific treatments for specific disorders