Therapist effects in the therapeutic alliance–outcome relationship: A restricted-maximum likelihood meta-analysis

Abstract

Objective

Although the relationship between the therapeutic alliance and outcome has been supported consistently across several studies and meta-analyses, there is less known about how the patient and therapist contribute to this relationship. The purpose of this present meta-analysis was to (1) test for therapist effects in the alliance–outcome correlation and (2) extend the findings of previous research by examining several potential confounds/covariates of this relationship.

Method

A random effects analysis examined several moderators of the alliance–outcome correlation. These included (a) patient–therapist ratio (patient N divided by therapist N), (b) alliance and outcome rater (patient, therapist, and observer), (c) alliance measure, (d) research design and (e) DSM IV Axis II diagnosis.

Results

The patient–therapist ratio (PTR) was a significant moderator of the alliance–outcome correlation. Controlling for several potential confounds in a multi-predictor meta-regression, including rater of alliance, research design, percentage of patient Axis II diagnoses, rater of outcome and alliance measure, PTR remained a significant moderator of the alliance–outcome correlation.

Conclusion

Corroborating previous research, therapist variability in the alliance appears to be more important than patient variability for improved patient outcomes. This relationship remains significant even when simultaneously controlling for several potential covariates of this relationship.

Introduction

The therapeutic alliance is considered to be an important aspect of successful treatment and has been found to be a consistent predictor of therapy outcomes in over 30-years of psychotherapy research (Horvath and Bedi, 2002, Horvath et al., 2011, Horvath and Symonds, 1991, Martin et al., 2000). A recent meta-analysis that examined the impact of the therapeutic alliance found that the alliance was a robust, albeit moderate (r = 0.275) predictor of treatment outcome, accounting for about eight percent of variability in outcomes (Horvath et al., 2011). However, significant variability in this overall alliance–outcome relationship was found, likely due to examined (i.e., moderators, such as time of alliance measurement) and unexamined (e.g., multiple measures of alliance with lack of consensus on alliance operationalization) factors. The therapeutic alliance refers to the collaborative relationship between patient and therapist (Bordin, 1979, Hatcher and Barends, 2006), which is impacted by in-session responsiveness (i.e., process where behavior is influenced by emerging information in the therapy session; see Stiles, 2009, Stiles et al., 1998) between the therapeutic dyad. Although the field has yet to settle on a uniform definition of the concept (Fitzpatrick et al., 2005, Horvath, 2005), there seems to be convergence empirically and theoretically that the central aspects of the therapeutic alliance construct involve the bond between the therapist and the patient as well as agreement about the therapeutic goals and tasks (Hatcher and Barends, 2006, Horvath and Bedi, 2002).

The influence of the alliance has also been examined in a variety of disorders. For example, its impact has been demonstrated in patients with depression (Krupnick et al., 1996, Raue et al., 1997), anxiety (Piper, Boroto, Joyce, & McCallum, 1995), PTSD (Cloitre, Chase, Miranda, & Chemtob, 2004), eating disorders (Constantino, Arnow, Blasey, & Argas, 2005), personality disorders (Andreoli et al., 1993, Klein et al., 2003, Strauss et al., 2006), and a number of other disorders (for an overview see e.g., Castonguay & Beutler, 2006). As well, the alliance and outcome have shown a remarkably robust association (although perhaps not linear; see Stiles et al., 2004), with treatment outcomes across different moderating variables such as measures of the alliance, measures of outcomes (primary symptom measure and non-targeted measures), rating perspectives, type of treatment (e.g., evidence-based, manualized, focused on specific ingredients), and context in which treatment was delivered (e.g., RCT or not; Flückiger et al., 2012, Horvath et al., 2011).

Despite the robust relationship with outcome across a number of different contexts, establishing causality is difficult as the alliance cannot be experimentally manipulated. As DeRubeis, Brotman, and Gibbons (2005) have argued (see also Strunk, Brotman, & DeRubeis, 2010), the alliance–outcome correlation may be due to (a) contributions of the patient, (b) contributions of the therapist, (c) the interaction of therapist and patient (i.e., the match), or (d) early change in functioning. For example, patients with good attachment histories and well developed social skills might well form better alliances and have better prognoses; consequently, the alliance–outcome correlation would be due to the patient's characteristics and not something that the therapist offers in treatment. Indeed, there are a number of studies that have shown that patients with better attachment histories or more adaptive attachment styles report better alliances with their therapists (Mallinckrodt, 1991, Mallinckrodt et al., 1995, Strauß and Schwark, 2007).

In contrast, others have argued that the alliance–outcome correlation is due, to a large extent, to the therapist. That is, therapists who consistently form better alliances with their patients generally achieve better outcomes. Several recent studies have investigated the impact of therapist vs. patient variability in the alliance using mixed effects models and found evidence suggesting that the therapist contribution is more critical than the patient contribution to the alliance–outcome correlation (Baldwin et al., 2007, Dinger et al., 2008, Marcus et al., 2011, Zuroff et al., 2010). For example, Dinger et al. (2008) studied an inpatient population and found significant therapist effects in the alliance–outcome correlation. Similarly, Baldwin , Wampold, Imel (2007) used mixed effects models to examine the impact of the therapist on the alliance–outcome correlation and found that the therapist contribution to the alliance was a statistically significant predictor of outcome, whereas the patient contribution was not. The relationship between therapist and patient contributions to the alliance and outcome found by Baldwin et al. is presented in panel (a) of Fig. 1. In this figure, it is clear that those therapists who generally formed better alliances had better outcomes (correlation between the average alliance scores for a therapist and outcome adjusted for pretest scores was 0.33). On the other hand, variability in alliances among patients within therapists was unrelated to outcome (r = 0). It is important to keep in mind that total alliance–outcome correlation in this study was 0.24, which misrepresents both the within therapist and between therapist correlation (less than the between therapist correlation and greater than the within therapist—i.e. the patient contribution—correlation).

Although these recent findings that support the importance of therapist effects in the alliance–outcome correlation are informative, the evidence cannot be characterized as conclusive at this point in time for several reasons. Most importantly, the studies that used mixed effects models to disentangle patient and therapist contributions have not been extensively replicated and may be idiosyncratic to the study characteristics (e.g., populations, methods) or simply may be instances of Type I errors (i.e., falsely rejecting the null hypothesis). For example, Baldwin et al. (2007) used a sample derived from college counseling centers and the alliance and outcome measures were from the same perspective (viz., the patient), raising the possibility of mono-method bias (Hoyt, 2002). Dinger et al.'s (2008) sample consisted of inpatients receiving psychodynamic therapy, along with other services, and only the patient rated alliance was used, and thus it is not clear how much of the treatment impact can be attributed to the therapist factors. Zuroff Kelly, Leybman, Blatt, & Wampold (2010) was a reanalysis of the National Institute of Mental Health Treatment of Depression Collaborative Research Program; these data has been analyzed in a variety of ways, and many of these re-analyses have resulted in contradictory conclusions (cf., Elkin et al., 2006, Kim et al., 2006).

A preferred method to synthesize primary studies and investigate aggregate effects, including the impact of moderators, is meta-analysis (Hunt, 1997, Mann, 1994). However, appropriately conducting a meta-analysis can be challenging, as the method requires a sufficient number of studies and for those studies to provide the relevant data to synthesize the results. Unfortunately, until only fairly recently have there been appropriate methods (and knowledge to implement these methods) to examine the relative contributions of therapist and patient variability in the alliance on outcome (i.e., mixed effects models). Therefore, we could only identify a limited number of studies that permitted direct evaluation of the therapist contributions vis-à-vis the patient's contribution to the alliance–outcome correlation. Because of this paucity of studies, we developed an alternative meta-analytic strategy that enabled us to use a more broad selection of investigations providing a better representation of the universe of studies that have examined the correlation of the alliance with outcome.

The meta-analytic strategy in the present study involved examining the ratio of the number of patients to the number of therapists (i.e., patient N divided by therapist N, denoted by PTR) as an indicator of research design properties of the included studies, which in turn allowed for examination of therapist and patient contributions to the alliance–outcome correlation. The PTR refers to the design of the studies that investigated the alliance–outcome correlation, yet is an index of the degree to which therapist and patient variability in the alliance was related to outcome. In Fig. 1, the Baldwin et al. (2007) results (panel a) are illustrated as well as hypothetical studies with high and low PTRs, as illustrated in panels (b) and (c) of Fig. 1. If the ratio is large (many patients and few therapists), then most of variability in the alliance would be due to the patients. The extreme example of a study with a high PTR is illustrated in panel (b) where all N patients are seeing the same therapist. In such a study, there is no therapist variability in the alliance because there is only one therapist; all of the variability in the alliance in such a study is between patient variability (i.e., all variability is within a single therapist). If the Baldwin et al. (2007) result that within therapist variability is not related to outcome is replicated, then the correlation of the alliance and outcome (i.e., the total alliance/outcome correlation) in this study would be zero because all the variability in the alliance is due to the patients. On the other hand, if the ratio is small (i.e., few patients per therapist), then a greater proportion of the variability in the alliance will be due to the therapists. A study with a PTR of 1 (viz., one patient per therapist) is illustrated in panel (c) and variability in the alliance is due to between therapist differences (no variability in the alliance within therapists).¹ If the therapists' contribution to the alliance is critical, as Baldwin et al. found, then it would be expected that the total alliance–outcome correlation in studies with relative small ratios would be relatively large because the variability in the alliance in such studies is mostly due to therapist differences. According to Baldwin et al., in a study in which the PTR approaches 1, the total correlation should approach the value of − 0.33. In this way, the PTR reflects a decomposition of the total alliance–outcome correlation into within- (patient-level correlation) and between-therapist (therapist-level correlation) components.

The purpose of the current meta-analysis was to test therapist effects in the alliance meta-analytically by using the ratio of patients to therapists as a moderator of the strength of the alliance–outcome correlation. It is hypothesized that there will be an inverse relation between the PTR and the size of the alliance–outcome correlation.

Because there are several study characteristics that might be related to the ratio, we attempted to rule out these confounding variables. Specifically, the following variables were also examined to see if they moderated the hypothesized relationship between the patient–therapist ratio and the alliance–outcome correlation: (a) alliance rater (patient, therapist, observer), (b) alliance measure (Working Alliance Inventory [WAI], California Psychotherapy Alliance Scale [CALPAS], Helping Alliance Questionnaire [HAQ], Vanderbilt Psychotherapy Process Scale [VPPS]), (c) research design (randomized controlled trial [RCT] vs. naturalistic design), (d) percentage of DSM IV Axis II diagnosis, and (e) outcome rater (patient, therapist, observer). The RCT variable as a potential moderator was included because an investigation using an RCT design will likely have a smaller number of therapists and more patients per therapist in contrast to naturalistic settings, which introduces a potential confound because the significance of the PTR variables might be due to study characteristics.