Exploring the Role of Persuasive Design in Unguided Internet-Delivered Cognitive Behavioral Therapy for Depression and Anxiety Among Adults: Systematic Review, Meta-analysis, and Meta-regression

Background: Internet-delivered cognitive behavioral therapy (ICBT) is an effective treatment that can overcome barriers to mental health care. Various research groups have suggested that unguided ICBT (ie, ICBT without therapist support) and other eHealth interventions can be designed to enhance user engagement and thus outcomes. The persuasive systems design framework captures most design recommendations for eHealth interventions, but there is little empirical evidence that persuasive design is related to clinical outcomes in unguided ICBT. Objective: This study aims to provide an updated meta-analysis of randomized controlled trials of unguided ICBT for depression and anxiety, describe the frequency with which various persuasive design principles are used in such interventions, and use meta-regression to explore whether a greater number of persuasive design elements predicts efficacy in unguided ICBT for depression and anxiety. Methods: We conducted a systematic review of 5 databases to identify randomized controlled trials of unguided ICBT for depression and anxiety. We conducted separate random effects meta-analyses and separate meta-regressions for depression and anxiety interventions. Each meta-regression included 2 steps. The first step included, as a predictor, whether each intervention was transdiagnostic. For the meta-regression of ICBT for depression, the first step also included the type of control condition. The number of persuasive design principles identified for each intervention was added


Introduction Background
Depression and anxiety are highly prevalent and represent the leading and the sixth leading causes of disability worldwide, respectively [1]. Despite the demonstrated efficacy of psychotherapeutic and pharmacological interventions for depression and anxiety [2][3][4], many people face structural barriers to accessing mental health care (eg, financial barriers, transportation barriers, inconvenience, and limited availability of services) [5,6]. Internet-delivered cognitive behavioral therapy (ICBT) is the most common type of internet intervention and an effective treatment for several common mental health problems, including depression and anxiety [7]. Unlike traditional cognitive behavioral therapy (CBT), ICBT enables users to access treatment materials privately at a time and location that is convenient for them, allowing it to be administered economically on a large scale and circumvent barriers to traditional forms of mental health care [8][9][10]. ICBT can be therapist guided or unguided. Guidance appears to improve adherence and clinical outcomes [11], but unguided ICBT is economical, highly scalable, and believed by many researchers to have considerable potential for improving public health [12][13][14][15].
Since the early 2000s, various research groups have suggested that eHealth interventions such as unguided ICBT can be designed in ways that improve user engagement and thus outcomes. In 2003, Fogg [16] presented the functional triad principle, suggesting that technology can function as a tool, a medium for relaying content, and a social actor to help facilitate behavior change. In 2009, Oinas-Kukkonen and Harjumaa [17] developed the persuasive systems design (PSD) framework, which elaborated on the functional triad and included 28 recommended design principles to produce more persuasive and engaging technological systems. They divided these principles into 4 categories: (1) primary task support principles, which facilitate the completion of the primary tasks of an intervention or other system; (2) dialogue support principles, through which an intervention or other system supports a user to help them enact their target behavior; (3) system credibility support principles, which facilitate a more credible and persuasive intervention or other system; and (4) social support principles, which leverage principles of social psychology to help users of an intervention or other system motivate one another. The 28 principles are described in Multimedia Appendix 1 [17].
In 2012, Kelders et al [25] used the PSD framework to assess whether the persuasive design principles used in 83 eHealth interventions for chronic conditions, lifestyle changes, and mental health predicted adherence. They conducted a meta-regression, finding that a greater number of dialogue support principles predicted greater adherence to eHealth interventions. However, to our knowledge, there is no empirical research demonstrating a relationship between persuasive design and symptom change in eHealth interventions.

Objectives and Hypothesis
This study aims to (1) present a systematic review and meta-analysis of randomized controlled trials of unguided ICBT for depression and anxiety among adults, (2) systematically examine the frequency with which various persuasive design principles are used in such interventions, and (3) use meta-regression to examine the extent to which persuasive design could explain the variability in effect sizes identified through the meta-analysis. Thus, the overarching objective of this study is to review the efficacy, the use of persuasive design, and the relationship between efficacy and persuasive design in unguided ICBT for depression and anxiety. We hypothesized that using a greater number of persuasive design principles would predict greater efficacy among the included studies.

Study Design
This study consisted of a systematic review, 2 meta-analyses, and 2 meta-regressions. The methods used in each phase of the study are described in the following sections. We registered the methodological protocol for this study on PROSPERO on October 24, 2019 (ID: 153466), before commencing the literature search, and kept a log of revisions to the original protocol throughout the course of this research (Multimedia Appendix 3 [26][27][28]). We followed the guidelines outlined in the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) statement in the preparation of this paper [29].

Eligibility Criteria
We searched for randomized controlled trials of unguided ICBT interventions for symptoms of depression and/or anxiety among adults that had been published in English in academic journals since 2000. We included trials of ICBT targeting symptoms of any type of depressive or anxiety disorder, as defined in the fifth edition of the Diagnostic and Statistical Manual of Mental Disorders [30], with various kinds of control conditions (eg, waitlist, treatment as usual, and active control). Studies involving samples with a mean age of less than 18 years were excluded.
Although we excluded studies in which ICBT was delivered with guidance from a therapist or coach, we did not exclude studies involving diagnostic interviews or contact of a logistical nature between participants and research teams. Interventions that used a CBT model of treatment and were delivered via the internet were considered ICBT interventions regardless of whether the authors of trials identified them as such. We included interventions using third-wave CBT approaches (eg, mindfulness-based CBT and acceptance and commitment therapy) [31] because prior research has not demonstrated significant differences in outcomes between traditional CBT and third-wave approaches [32,33].

Literature Search
On October 29, 2019, we conducted a literature search on MEDLINE, PsycINFO, PubMed, Web of Science, and PsycArticles. To be identified, articles were required to include the words "CBT," "internet," "trial," and "depression" or "anxiety" or one of several similar phrases for each of these terms in their titles, keywords, or abstracts. The search terms are shown in detail in Multimedia Appendix 4. This search was updated on July 2, 2020.

Study Selection
After removing duplicates of studies identified in 2 or more databases, HCM and CRFS independently screened the studies in 3 stages: by title, by abstract, and by full text. Wherever the 2 screeners reached different decisions about whether to retain or exclude a study, that study was included in the next stage of screening. Differences in decisions on the full-text screening were resolved through discussion.

Data Extraction
HCM extracted several types of data from each study: study characteristics (eg, type of control condition and time between pretreatment and posttreatment measures), risk of bias [34], general intervention characteristics (eg, target symptoms and medium of delivery), persuasive design principles [17] as operationalized by Kelders et al [25], and efficacy data. Consistent with the approach of Kelders et al [25], we did not code principles in the system credibility support category of the PSD framework because they were reported very infrequently and would have been challenging to code objectively (eg, a system should have a "competent look and feel" and "provide endorsements from respected sources" [17]). In most cases, we coded persuasive design principles as present or absent based on the descriptions of interventions in the included studies, although we consulted other available sources of information when possible (eg, intervention websites and study protocols). The complete list of data items is provided in Multimedia Appendix 5. The persuasive design principle tunneling, which refers to the sequential presentation of treatment elements in a structured, linear manner, was not counted toward the total number of persuasive design principles in this study. This is because researchers have recently proposed that eHealth interventions can be made more engaging by providing users with greater flexibility and control concerning the modules or features they wish to use [19,22], which contrasts with the principle of tunneling.

Risk of Bias Assessment
We assessed the risk of bias among included studies using the Cochrane risk of bias tool [34]. We did not assess the risk domain blinding participants and personnel because it is not possible for participants to be blind to their conditions in psychotherapy research [35]. Furthermore, we did not assess the risk domain blinding of outcome assessment because all outcome measures were self-report measures, and participants could thus not be blinded. Self-report measures are generally considered equivalent to blind clinical observers in psychotherapy research, and research suggests that they do not result in inflated effect sizes [35].

Meta-analysis Methods
We conducted meta-analyses using Comprehensive Meta-Analysis software (Biostat Inc) [36]. As prior research suggests that ICBT for generalized and social anxiety is more efficacious than ICBT for depression [7,37], we conducted separate meta-analyses of ICBT for anxiety and ICBT for depression. Given the availability of symptom change data for both anxiety and depression, trials of ICBT designed to treat both conditions were included in both meta-analyses. We measured heterogeneity in the effect sizes of the included studies using the I 2 index and formally tested the degree of heterogeneity using the Q statistic [38]. In each of the 2 meta-analyses, we used a random effects model, used between-groups effect size (Hedges g) as the summary measure, and weighted each study by the inverse of the within-study variance of the primary outcome measure plus the between-study variance. Several studies evaluated 2 unguided ICBT interventions; in such cases, we treated the evaluation of each intervention as a separate study, except we divided the control group sample size by 2, such that each control group participant was included only once in the analyses [39]. We evaluated the risk of publication bias using funnel plots and accounted for publication bias using the trim and fill technique [40]. We explored the influence of study-level bias on outcomes by repeating the meta-analyses without studies deemed to be at high risk on one or more dimensions of the Cochrane tool for assessing risk of bias [34].

Meta-regression Methods
We conducted 2 meta-regressions using Comprehensive Meta-Analysis [36]-one for depression interventions and one for anxiety interventions-to determine the degree to which persuasive design principles could explain variance in effect sizes among studies. Paralleling the approach taken to the meta-analyses, we included trials of ICBT designed to treat both depression and anxiety in both meta-regressions, given the availability of symptom change data for both conditions. We also weighted each study by the inverse of the within-study variance of the primary outcome measure plus the between-study variance, as in the meta-analyses.
We used 3 predictor variables. Our main predictor of interest was the total number of persuasive design principles identified for each intervention. We were unable to include the number of persuasive design principles in each category of the PSD framework as separate predictors, as Kelders et al [25] did, because of the risk of overfitting, given the limited number of included studies. We also input a binary variable reflecting whether each intervention was transdiagnostic (ie, designed to treat symptoms of both depression and anxiety). We did this to account for the possibility that unguided ICBT focused on treating a narrower range of symptoms (ie, anxiety or depression) may be more efficacious for treating those symptoms than transdiagnostic unguided ICBT designed to treat a broader range of symptoms (ie, both depression and anxiety). Our final predictor was a binary variable reflecting whether each study used a control condition with active elements (eg, psychoeducation and mood monitoring) because a previous meta-analysis of unguided ICBT found a large mean effect size among studies using passive control conditions and a small mean effect size among studies using active control conditions [41]. However, the control condition type was not included as a predictor in the meta-regression of ICBT for anxiety because there were insufficient studies to justify an additional predictor variable (eg, because of the risk of overfitting), following most recommendations concerning acceptable subjects per variable ratios in linear regression analyses [26].
We conducted each meta-regression in 2 steps. The first step included transdiagnostic status and, for the meta-regression of ICBT for depression, the control condition type. In both meta-regressions, the number of persuasive design principles identified was then added in the second step. This 2-step approach was used to reveal the amount of additional variance persuasive design explained in the second step after accounting for the other variables in the first step.
We conducted 5 assumption tests at each step of each meta-regression. First, we examined Pearson r correlations and scatterplots to test the assumption of linearity of the relationship between each continuous predictor variable and Hedges g [42]. Second, we checked Cook distance values to identify any outlier studies that had unduly large influences on the results [43]. Third, we inspected the distribution of studentized residuals using a histogram to ensure that the residuals were normally distributed [42]. Fourth, we inspected scatterplots plotting studentized residuals against predicted values to test the assumption of homoscedasticity [42]. Finally, we examined variance inflation factors to check for multicollinearity [42].

Study Selection
Between the original and updated literature searches, we identified 4471 articles, 39 of which were found eligible for analysis. Having found another 7 eligible articles through a hand search, we included a total of 46 articles. The flow of studies through the study selection process is shown in

Risk of Bias
We evaluated the risk of bias among included studies using 5 of the 7 domains in the Cochrane risk of bias tool [34]. Of the 46 included studies, 14 (30%) were identified to be at high risk of bias in at least one domain, whereas only 4 (9%) were found to be at low risk of bias in all domains assessed. Most studies (28/46, 61%) were found to be at low or unclear risk in each domain. The risk of bias identified in each study is presented in Table 2.

Persuasive Design
On average, interventions included 4.95 (SD 2.85) persuasive design principles (excluding tunneling). The total number of persuasive design elements ranged from 1 to 13. Principles in the primary task support category were the most common (mean 2.86, SD 1.32), followed by principles in the dialogue support category (mean 1.27, SD 1. 19) and social support category (mean 0.81, SD 1.60). The number of interventions in which each persuasive design principle was identified is presented in Table 4.  (8) Shows users who adopted target behavior Recognition a These descriptions were adapted from the operational definitions provided by Kelders et al [25].

Meta-analysis of Unguided ICBT for Depression
We conducted a meta-analysis of 37 comparisons across 34 trials of unguided ICBT for depression. There was statistically significant heterogeneity in Hedges g among the studies (Q=89.85, df=36; P<.001). An I 2 statistic of 59.93 indicated that a moderate proportion of variability was attributable to true heterogeneity rather than sampling error [90,91]. The weighted mean between-subjects effect size was small to moderate (Hedges g=0.31; SE 0.04; 95% CI 0.24-0.38). The forest plot for this meta-analysis is shown in Figure 2. Weighted mean effect sizes after excluding studies deemed to be at high risk of bias and after adjusting for publication bias using the trim and fill technique [40] are presented in Table 5.

Meta-analysis of Unguided ICBT for Anxiety
We included 19 studies that reported 21 comparisons in a meta-analysis of unguided ICBT for anxiety. The results indicated statistically significant heterogeneity of Hedges g among studies (Q=68.47, df=20; P<.001). The corresponding I 2 statistic of 70.79 suggested that a substantial proportion of the variability represented true heterogeneity [90,91]. The weighted mean between-subjects effect size was moderate (Hedges g=0.45; SE 0.06; 95% CI 0.33-0.56). A forest plot is shown in Figure 3. Additional weighted mean effect sizes accounting for publication-and study-level bias are presented in Table 5.

Meta-regression of Unguided ICBT for Depression
The meta-regression of ICBT for depression, like the meta-analysis of ICBT for depression, included 34 studies reporting 37 comparisons. We used 3 predictors in this meta-regression: the total number of persuasive design principles (mean 3.90, SD 2.33), whether each intervention was designed to treat symptoms of both depression and anxiety (19/37, 51%) or only depression (18/37, 49%), and whether each study used an active control condition (13/37, 35%) or a passive control condition (24/37, 65%). The results for both steps of the meta-regression are presented in Table 6. With the possible exception of very minor heteroscedasicity of residuals at one or both steps, all assumptions were met, as detailed in Multimedia Appendix 8.

Meta-regression of Unguided ICBT for Anxiety
Similar to the meta-analysis of ICBT for anxiety, the meta-regression of ICBT for anxiety included 19 studies reporting 21 comparisons. We used 2 predictors: the total number of persuasive design principles (mean 5.05, SD 3.17) and whether each intervention was designed to treat symptoms of both depression and anxiety (8/21, 38%) or only anxiety (13/21, 62%). The results for both steps of the meta-regression are presented in Table 7. The assumption of normality of residuals may not have been met fully at both steps, although the residuals roughly approximated normal distributions. The assumption of homoscedasticity of the residuals was violated in step 1. The assumption tests for this meta-regression are detailed in Multimedia Appendix 9.

Principal Findings
Recent years have witnessed a proliferation of randomized trials of eHealth interventions, including many trials of unguided ICBT for depression and anxiety. Indeed, most of the studies included in this review were published in or after 2017. There was considerable diversity in the design of both studies (eg, study duration and type of control condition) and interventions (eg, mode of delivery and use of persuasive design principles).
The results of the meta-analysis of unguided ICBT for depression were consistent with the results of previous meta-analyses. We reported 4 mean effect sizes (Hedges g) for unguided ICBT for depression, ranging from 0.22 to 0.31, based on the corrections we made for publication bias and study-level bias. Previous meta-analyses of unguided ICBT for depression have found comparable mean effect sizes (Hedges g or Cohen d) ranging from 0.24 to 0.36 [12,[92][93][94][95]. Our meta-analysis of unguided ICBT for anxiety yielded a mean effect size of 0.45. There was no evidence of publication bias, and the mean effect size was greater (Hedges g=0.54) after excluding studies found to be at a high risk of bias. Several previous meta-analyses of ICBT for symptoms of anxiety disorders found effect sizes between 0.70 and 1.12 [41,[96][97][98]; however, all these meta-analyses included trials of guided ICBT interventions, which likely explains the greater mean effect sizes, at least in part. We are aware of only 1 meta-analysis that has included a subgroup analysis of unguided ICBT for anxiety-social anxiety, specifically-finding mean effect sizes (Hedges g) of 0.78 and 0.19 for studies using passive and active control conditions, respectively [41]. It is worth noting that our review included many transdiagnostic interventions designed to treat symptoms of both depression and anxiety. The meta-regressions showed that these interventions were significantly less efficacious for treating anxiety symptoms compared with interventions designed to treat anxiety symptoms only; however, their efficacy in treating depression did not significantly differ from interventions designed to treat symptoms of depression only.
We identified wide variability in the use of persuasive design in unguided ICBT for depression and anxiety, with several interventions using only 1 persuasive design principle and others using as many as 13. The intervention identified as having the greatest number of persuasive design principles (ie, 13), called Challenger, was specifically designed to be engaging, with many features inspired by the literature on gamification [79,99]. The mean number of persuasive design principles identified across interventions (4.95, excluding the principle of tunneling) was comparable with the mean of 5.4 principles identified by Kelders et al [25] among mental health interventions in their review. The mean number of persuasive design principles identified in the primary task support (mean 2.86, SD 1.32; excluding tunneling), dialogue support (mean 1.27, SD 1. 19), and social support (mean 0.81, SD 1.60) categories were also roughly comparable with the corresponding means identified among mental health interventions by Kelders et al [25] (2.6, 1.6, and 1.3, respectively).
Persuasive design was a significant predictor of effect size in the meta-regression of ICBT for depression. The unstandardized β coefficient (B) of 0.04 suggested that for each additional persuasive design principle an intervention uses, one could predict the effect size (Hedges g) for that intervention to increase by 0.04, compared with a control condition in a randomized trial. However, meta-regression is an inherently observational procedure [100], and the results therefore could not show whether persuasive design caused certain ICBT interventions for depression to be more efficacious than others. Persuasive design did not predict efficacy in the meta-regression of ICBT for anxiety. However, it is worth noting that the meta-regression of ICBT for anxiety included far fewer studies than the meta-regression of ICBT for depression and had limited statistical power to identify an effect. Indeed, persuasive design had an unstandardized β coefficient of 0.03 in the meta-regression of ICBT for anxiety, which-although not statistically significant-was comparable in magnitude with that of the meta-regression of ICBT for depression. The results of the meta-regression of unguided ICBT for anxiety should be interpreted cautiously because assumption tests showed that certain assumptions were unmet. Nonetheless, our results suggest that persuasive design is more closely related to outcomes in interventions for depression than anxiety. Given that persuasive design is purported to motivate engagement in treatment [17] and that lack of motivation is a hallmark of depression, it is possible that persuasive design is particularly important in ICBT for depression.
Overall, our findings support the hypothesis that persuasive design predicts efficacy in unguided ICBT, at least in the treatment of depression. Our findings also support the validity of the PSD framework [17] by showing that it is meaningfully related to treatment outcomes. Although the results do not demonstrate the importance of any specific persuasive design principles, they support the growing body of theory and data suggesting, broadly, that persuasive design matters in eHealth [18][19][20][21][22][23][24]. These findings are encouraging and timely. ICBT has become well established over the last two decades, having now been evaluated in hundreds of trials [101] and currently being funded by many governments around the world [102]. It is clear that ICBT is effective, and a natural next step in ICBT research will be to explore possible avenues for making it more effective. Our findings suggest that enhanced persuasive design may be one such avenue. Notably, because ICBT is highly scalable, particularly when it is unguided, even slight increases in effectiveness can have substantial and wide-reaching implications for public health.

Limitations
This study had several limitations. First, a considerable amount of data was unreported; in particular, it is likely that many interventions used persuasive design principles that were not described in the included studies. Second, although we were able to identify the principles in the PSD framework as present or absent, we did not have access to the interventions themselves, and we were unable to evaluate how effectively persuasive design principles were implemented. Third, we were unable to show, through our meta-regressions, whether specific persuasive design principles predicted efficacy. Finally, only 1 researcher was involved in data extraction; a second extractor would have helped reduce the risk of error, inconsistency, or bias.

Future Directions
Further research will be required to clarify the role of persuasive design in unguided ICBT and other eHealth interventions. First, dismantling studies comparing versions of interventions with and without certain persuasive design principles could evaluate the utility of specific principles. Factorial randomized trials of this kind would allow researchers to efficiently evaluate multiple persuasive design principles in a single study. Second, it would be helpful to explore how intervention users experience persuasive design, which could perhaps be achieved through qualitative research or the development of a self-report questionnaire assessing user experiences of persuasive design. Third, the literature would benefit from a more detailed description of persuasive design in unguided ICBT interventions based on a careful review of the interventions themselves (ie, rather than this study's review of descriptions of interventions from randomized trials). Finally, further research will be required to test our finding that persuasive design predicts efficacy in unguided ICBT for depression but not for anxiety.

Conclusions
The literature on ICBT and other eHealth interventions is evolving rapidly. This review has provided an updated meta-analysis of unguided ICBT for depression and anxiety, generally finding smaller effect sizes for depression than for anxiety. It has also documented the wide variability in the use of persuasive design in unguided ICBT and demonstrated through a meta-regression that persuasive design predicts efficacy in unguided ICBT for depression. Persuasive design is a promising avenue for further optimization of eHealth interventions, including ICBT, and an area of research that is worth investigating further.