The relationship between therapist and patient during psychotherapy, referred to as alliance and measured by a number of instruments such as the Working Alliance Inventory (WAI) [1], is considered an important common factor shared by diverse treatments [2]. Alliance is conceptualized as requiring three distinct processes: agreement on therapeutic goals, consensus on tasks that make up therapy, and bond between therapist and patient [3]. Alliance is also understood to occur in stages, first with identifying the therapist as a source of encouragement, warmth, and support and, in later stages, the development of a client’s faith and investment in the therapeutic process, a collaborative working together [4]. Some alliance measures, such as the Helping Alliance Questionnaire (HAq) [5], are an authoritative operationalization of their authors’ key theories; however, the boundaries of the popular WAI were never formalized [2] and researchers continue to refine the measure as they seek to develop a holistic understanding of alliance for their particular treatment.

Virtual reality exposure therapy (VRET) is a technology-based method of delivering exposure-based treatments; it has been tested in clinical trials for some 20 years, with positive results [6], but has only recently become available to consumers [7]. VRET works by using head-mounted displays to present computer-animated graphical stimuli or 360-degree video, interactive to head movements, while occluding the outside world [8]. Few studies, however, have assessed alliance during VRET treatments, perhaps due to an expectation that alliance will be poor in treatments that prevent face-to-face contact with a therapist [9]. Contrary to this expectation, studies of patients treated with VRET and in vivo exposure therapy for social anxiety disorder found no significant difference between interventions as measured by the WAI [10-12]. Using augmented reality exposure therapy (ARET), a related technology [13], the same lack of difference in WAI scores was found following treatment for small-animal phobia [14]. In a systematic review of VRET process and outcome research, only two studies were identified that captured alliance information [15]: a study of fear of flying, but not an acrophobia study, found a significant correlation with outcome.

Automated VRET treatments—replacing in vivo stimuli with virtual stimuli in addition to replacing human therapist with virtual therapist [16]—are an example of autonomous digital treatments, and being self-guided may be among the most scalable of psychological interventions [17,18]. However, previous research suggests that treatments conducted without therapist support suffer from lower adherence and efficacy as compared to treatments that include contact [19,20]. There is also evidence that more contact and better contact (ie, frequency of interactions and persuasive design elements, such as dialogue support) may result in more patients completing treatment [21]. Therefore, simulating key elements of face-to-face therapist interaction with a virtual agent manifesting visually and/or through auditory instruction may improve adherence and bridge the efficacy gap between guided and unguided interventions [22]. This is particularly so if they are successful in replicating common factors in therapy and nonspecific relational elements, such as empathy and therapist attention [23], which some have suggested are better conceived of as an active ingredient and should be intentionally used in treatment to facilitate better outcomes [24]. In low-income regions, they may also offer an opportunity to disseminate evidence-based cognitive behavioral therapy (CBT) where literacy skills are low and even therapist-supported online treatments are not appropriate [25].

To date, three automated VRET treatments have been tested in clinical trials: two for use with acrophobia [26,27] and one for use with spider phobia [28]. A treatment for social anxiety disorder has also been tested as a small pilot study [29]. Freeman et al [27] used what they referred to as a virtual coach, physically embodied in virtual reality (VR) by a trained actor, to assist patients to conduct behavioral experiments, challenge safety behaviors, repeat key learning points, and provide empathic encouragement. In the study by Donker et al [26], patients were given background information about phobias, provided case examples, and given motivation using a 2D, animated, mobile phone app-based avatar called Tara. The Hartanto et al [29] system enabled interactive dialogue with an animated virtual health agent via laptop computer to instruct in the use of VRET sessions; provide monitoring data interpretations, based on heart rate, for example; and motivate clients while completing treatments at home. However, the paradigm relied on a personalized treatment plan created by a human therapist versus the three more recent trials, which incorporated game-level-like progression systems similar to an exposure fear hierarchy. We designed our automated treatment for spider phobia [28] with a virtual therapist to deliver guidance and support from within the VRET application using primarily voice-based instructions and, in one version, a graphical representation. Participants were greeted from the beginning of treatment by the virtual therapist, provided information on the purpose of treatment, provided instructions on how to conduct exposure tasks, and were followed throughout treatment with encouragement and positive reinforcement when a task was completed.

To our knowledge, no clinical trial conducted using an automated VRET treatment with a virtual therapist has reported on alliance, despite the therapist-like human qualities their creators designed them with. There are indications, however, that alliance may be possible toward a virtual therapist. One recent trial of an automated online CBT program for insomnia [30] incorporated an audio-based avatar to guide patients through the application and to provide cognitive restructuring for sleep-related concerns. The authors identified comparable or higher goal and task subscale scores using an adapted WAI scale as compared to a therapist-led, Internet-based CBT (ICBT) program for tinnitus and similar task subscale scores as compared to an outpatient treatment group, but lower goal. Affective bond was stable and relatively high over six sessions according to five items from the Bern Post-Session Report. The authors suggested that a questionnaire designed specifically for automated programs and avatars is needed to provide a clearer picture of alliance relationship with symptom change. Correlations with outcome were found but only for the affective bond subscale.

Just one effort has been directed toward developing a WAI scale intended specifically for VRET and ARET treatment: the WAI applied to virtual and augmented reality (WAI-VAR) [31]. The WAI-VAR collected data from across three studies (n=75) for fear of flying, adjustment disorder (VRET), and cockroach phobia (ARET). Clinically significant recovered and improved participants had large effect size differences with not changed patients according to the WAI-VAR (eta-square=0.32, P<.001) and large Spearman correlations were noted between alliance and clinically significant change (r=.55, P<.001). High scores were also found on the non-VR-specific WAI-short form (WAI-S) [32].

The WAI-VAR [31], however, was developed to address comfort and trust with the virtual environment itself, replacing “my therapist” with “virtual environment” for all items of the Spanish version of the WAI-S. Recent research suggests that using conventional measures and simply replacing “therapist” with another term (eg, “app”) may be insufficient for accurately capturing the subtleties of therapeutic alliance [33]. Given the embodied nature of the virtual therapist in this automated treatment, we developed a novel instrument entitled the Virtual Therapist Alliance Scale (VTAS) to gather information on the patient relationship with the virtual therapist themselves. In addition to phrasing WAI items in terms of “the virtual therapist,” modified items from an empathy questionnaire [34] and copresence scale (ie, “the experience of being with another”) [35] were included to capture additional relational characteristics of working with the virtual therapist. A preliminary psychometric evaluation was conducted using data collected from two independent studies (Miloff et al [28] and Lindner et al, under review), convergent validity assessed according to common process measures (ie, presence and user friendliness), and the scale compared to treatment outcome (ie, self-reported fear). Alliance is considered a generic common factor, with meaning and usefulness expected to be preserved across theoretical orientations [2] and evidence suggests that a range of technology-based interventions can facilitate alliance [36]. Therefore, we hypothesized that the therapeutic relationship with a virtual therapist would correlate significantly with treatment outcome and offer insights on what components of the virtual therapist are most important for treatment efficacy.

Automated VRET treatments using embodied virtual therapists may be a means of increasing access to exposure therapy by supporting patients to conduct treatments on their own; however, little has been known about patient alliance toward the virtual therapist and its relationship to treatment outcome. This study evaluated a novel scale—the VTAS—using data from two samples of spider-fearful patients treated over a single session.

Psychometric evaluation of the VTAS identified a two-factor solution for the scale, with items based on task, goal, and copresence loading primarily on factor 1, and items based on bond and empathy on factor 2. As summarized in Hatcher and Gillaspy [37], a two-factor solution for the WAI scale is not uncommon. Falkenström, Hatcher, and Holmqvist [52] conducted a confirmatory factor analysis of the scale among three large samples in Sweden and the United States and argued that, given the high intercorrelation between task and goal factors, a two-factor solution was more defensible psychometrically. A recently evaluated WAI scale for guided Internet-based treatments had a similar task-goal and bond two-factor solution [53]. Goal and task factors have been referred to as agreement-confidence in the therapist and affective bond as relationship [54]. While this terminology may not be appropriate for use with a virtual therapist, if such a therapist can be conceived of in terms of a guide motivating the patient toward behavior change, then a suitable terminology could be task-oriented versus relationship-oriented guidance [55].

It should be noted, however, that not all goal and task items loaded on the first factor and vice versa. For example, item 11—“It felt comforting to have the virtual therapist there with me,” which suggests a bond category—loaded quite highly on factor 1 (0.791), and item 3—“I felt that the virtual therapist gave clear instructions,” suggesting a task category—loaded quite highly on factor 2 (0.614). In the context of using a foreign technical application (ie, VRET), a relationship between clarity of instruction and development of affective bond could be suggested by this outcome (item 3 in factor 2); also, a relationship between receiving comfort and aid and accomplishing the tasks and goals of therapy (item 11 in factor 1). Alternatively, identification of the factor structure could be based on how items were written. Factor 2 items, for the most part, attributed a human quality to the virtual therapist—“I felt/experienced that the virtual therapist was...” friendly (item 1), warm (item 2), supportive (item 4), trustworthy (item 10), and gave clear instructions (item 3)— whereas factor 1 items were all written as an observation on the nature of the relationship or the benefits accruing from the presence of the virtual therapist. Previous authors of text-based, self-help treatments have argued that alliance is toward the application, as summarized in Heim et al [30]; however, embodied avatars do contain more realistic, human-like qualities and, therefore, the concept of alliance here may be closer to the traditional one. Further research will be needed, such as comparing different forms of embodiment (eg, audio-only instruction versus audio plus physical embodiment), using a rating scale sensitive to this. Preliminary results from this study suggest patients rate factor 2 bond items highly (ie, significantly higher than those of factor 1, P<.001); also, factor 2 items explain a sizeable additional total variance in the scale over five items (38.8% vs 14.0%). It is conceivable that bond is strengthened in interventions with an embodied avatar as compared to text-based, self-help applications [35] or, as suggested by Heim et al [30], bond may occur earlier in treatment. In certain treatments, low bond (ie, trust and faith) may be sufficient, whereas in other treatments, such as exposure therapy, high bond may be needed to motivate treatment completion.

Relational bond between patient and mobile self-help apps was explored recently in a mixed qualitative-quantitative study [33]. Qualitative analysis suggested patients do form a personal bond even with their nonavatar, text-based app, referring to it as a “therapist in their pocket” and a friend in the app they could turn to for reassurance and encouragement. Quantitative analysis using a purpose-built questionnaire for measuring alliance in self-guided programs—the mobile Agnew Relationship Measure (mARM)—evaluated one iteration of their scale that simply exchanged “therapist” for “app” on all items of the Agnew Relationship Measure (ARM). Their findings suggest that while certain item terminology could risk inappropriately anthropomorphizing an app—therefore, making it harder to relate to—where human-like qualities are concerned, participants are more likely to endorse a relationship item that includes a qualifying perception (eg, “I feel the app...”) rather than a statement of fact (eg, “The app seems bored or impatient with me”). The former was done for all items attributing human characteristics to the virtual therapist in this study. Internal consistency was lower on factor 2 (Cronbach alpha=.710) and below recent cutoff recommendations by Clark and Watson (alpha=.80) [49]; however, a high coefficient alpha is difficult to obtain in scales with few items, given its relationship to length (ie, 12 versus five items on factor 2) [49]. Factor 1 and overall VTAS consistency was high (Cronbach alpha=.931 and .921, respectively).

Convergent validity findings provide additional support for the independence of factors. The demographic variables age and gender did not have a significant relationship with VTAS scores (P>.43 for both). Questionnaires capturing the subjective experience of presence in a virtual environment were found to correlate significantly more with factor 1 (P<.001), in which all copresence items loaded. Scores of user friendliness correlated moderately with both factor 1 and factor 2 (P<.01 for both). However, of the two measures, only presence significantly predicted factor 1 VTAS results in a multiple linear regression model, other than FSQ follow-up scores, and only user friendliness predicted factor 2 VTAS results. It should be noted that user-friendless scores were between good (71.4) and excellent (85.5), according to generic SUS adjective descriptors (mean 82.8) [56]. This is positive, considering that good system usability may be particularly important in automated treatments that are intended to be self-administered [29].

Evidence for the relationship with outcome indicates that perceived alliance toward a virtual therapist had a small, nonsignificant (r=.213) correlation with outcome at posttreatment, but a moderate and significant (r=.310) correlation at follow-up. Unlike the study by Heim et al [30], which showed a correlation between bond-factor and outcome and is the only other known study to measure alliance to a therapist avatar in an automated treatment format, only factor 1 (ie, task, goal, and copresence) correlated with outcome in this study. The lack of significant relationship with outcome immediately after treatment is not surprising, given the small-to-moderate sample size and limited power to detect a small association: a post hoc power analysis indicated that 70 participants would have 80% power to detect a correlation of above .33. However, the significant relationship at follow-up may suggest that those who had a good alliance with the virtual therapist better understood and interpreted the treatment instructions correctly or received adequate reinforcement to apply them in their daily lives. Further research is needed here, as well as a better understanding of why particular factors correlate with outcome (eg, as compared to Heim et al [30]). A recent narrative review of therapeutic alliance in Internet interventions found no other study that identified a relationship between the bond-factor and outcome, but many with a meaningful and statistically significant relationship between task-goal and outcome [9]. The author suggests that this may be due to a ceiling effect from high scores on the bond dimension in Internet interventions, similar to this study, or possibly the relative lack of importance of bond in treatments with little-to-no face-to-face interaction. Nevertheless, overall results in this study compare favorably to current meta-analytic best evidence for the relationship between alliance and outcome [57]. Across 39 different alliance measures and 306 studies, the alliance correlation with outcome was .278 (P<.0001); over 18 Internet-based studies, the alliance correlation with outcome was .275 (P<.0001).

Limitations of this research include the relatively small sample size in the two studies included and small differences between the automated treatments, such as language used and whether the treatments were carried out with a technician present. Together, this meant that measurement invariance had to be assumed but could not be formally tested. The use of a single-session treatment format prevented the exploration of cross-lagged and mediation models to predict symptom reduction following a session [58]. This exploration is important, given the limited capacity of alliance research to show an experimentally causal association with treatment outcome [57]. In the Heim et al [30] study, multiple treatment sessions also provided the authors evidence that a human therapist was increasingly missed over time, although this may be due to the therapeutic techniques used by the virtual avatar in later sessions. Finally, this study explored only client-reported ratings of alliance, versus therapist, observer, and friend or family ratings, whereas other studies, such as a recent blended Internet-based CBT treatment for depression, found only therapist-rated alliance results to be significant with outcome [59].

In the future, research on automated treatments may benefit from evaluating different forms of embodiment for virtual therapists and exploring their effect on alliance, as mentioned earlier, and relationship to treatment outcome and treatment adherence. The virtual therapist in this study primarily used voice-based instruction and was noninteractive; however, it is conceivable that a quality implementation of visual presence, behavioral realism, and interactivity may improve outcomes [35]. Heim et al [30] suggest this may be possible through improved credibility and expectancy, which mediation models indicate function in a bidirectional relationship with alliance to improve outcomes [60]. Previous research on embodied graphically generated agents have included SimCoach [61], which used rich visual representations of humans, gestures, and emotions, as well as natural conversational speech to comfort, guide, and motivate armed service members toward psychological health and to solicit anonymous information. As summarized by Lucas et al [62], rapport may be improved by relying on verbal and nonverbal behavior, such as specific verbal patterns, intonation, and choice of language as well as welcoming gestures, open posture, and attentive eye gaze. Disclosure of personal backstories and intimate personal details by an automated therapist has been shown to increase self-disclosure by highly anxious patients [63]. Relational behaviors, such as calling the patient by name, use of humor when appropriate, including information discussed during past interactions, and engaging in a social chat at the beginning of a new interaction, have all been shown to increase how much a patient feels cared for and how useful the information they received felt to them [25]. Adherence to psychoeducation, operationalized as successful task execution, may be improved by attitude toward a virtual agent, such as their level of trust of the agent, its realism, and its amiableness [64].

In conclusion, this is the first study evaluating an alliance scale for use with virtual therapists in an automated VRET treatment format. Preliminary evidence suggests that this novel VTAS instrument has sound psychometric properties, with one exception being the somewhat low internal consistency of factor 2. Nevertheless, despite this and other limitations, alliance toward a virtual therapist was found to be positively correlated with treatment outcome at follow-up. Further exploration of the predictive abilities of this scale on treatment outcome and adherence is warranted, particularly in conjunction with alternative forms of embodiment, in larger samples of patients, and across different treatment programs and treatment formats.