Low intensity technology-delivered cognitive behavioral therapy for obsessive-compulsive disorder: a meta-analysis

In the study samples, participants needed to be adults (mean age > 18 years) and to have OCD symptoms according to a standardized classification system [3], prior diagnosed by a mental health specialist (e.g., [27, 28]), self-reported OCD symptoms (i.e., [29, 30]) or a score measuring OCD symptoms at least on a mild level (i.e., a Y-BOCS score over 7). There were no restrictions in terms of medication use.

Studies were included if they conducted trials with technology-delivered cognitive behavioral therapy such as cCBT, iCBT or bCBT. The latter could be a PDF file sent or provided via a technology (e.g., computer or mobile phone). Interventions were defined as CBT if they included typical CBT elements such as cognitive components (e.g., cognitive restructuring), ERP or components of the third wave of CBT. Moreover, studies needed to provide interventions with low personal contact intensity. According to Newman et al. (2011) our definition of a low contact intensity included category [3] self-help with assistance to ensure engagement with the tool (ASH) and [4] pure self-help (PSH) [17]. According to proposed guidelines by Newman et al. (2011) and Andersson, Calbring, Berger, Almlöv, and Cuijpers (2009) [17, 31], we defined a maximal timeframe of 100 min for category [3] in terms of phone calls. In other words, time spent for therapists’ responses without time latencies (synchronous contact) is not more than 100 min per participant and per intervention independent of the intervention length. Additionally, contacts involved no face-to-face sessions or video calls. No time restrictions were made with regard to responses with time latencies (asynchronous contact) such as written answers or recorded answers (e.g., IVR systems). Asynchronous contacts did not provide crisis interventions or prompt responses and resulted in less personal and situational influences by a therapist as the answers were recorded. Group conditions which provide more than 100 min in total or/and included face-to-face sessions or video calls were defined as an active control group condition.

As explained above, our definition for a low contact intensity included self-help with assistance to ensure engagement with the tool (ASH) and pure self-help (PSH) with a maximal timeframe of 100 min per participant and per intervention independent of the intervention length.

Comparison was drawn between cCBT, iCBTor bCBTand an active or passive control condition. Control conditions with other treatments were face-to-face CBT, video-based CBT, phone-based CBT with phone calls over 100 min during the treatment, only CBT psychoeducation, relaxation or attention treatments or other interventions than CBT. Passive control conditions were considered as wait-list groups or no treatment conditions.

Included studies needed to report outcomes on measures of OCD symptoms (e.g., Yale-Brown Obsessive-Compulsive Scale (Y-BOCS) [32]; Obsessive-Compulsive-Inventory-Revised (OCI-R) [33]; Dimensional Obsessive-Compulsive Scale (DOCS) [34]).The outcomes of interest were standardized mean differences of symptom severity. To calculate effect sizes, studies had to provide pre- and post-treatment means and standard deviations. When these values and information were not provided, authors were contacted.

Studies needed to be published in either German or English. Only studies with randomized controlled trials were included. Thus, open trials, preliminary studies, case reports or trials based on the same data were excluded. Furthermore, studies were required to investigate the efficacy of an intervention. Hence, effectiveness studies were excluded. By definition, studies conducting interventions under real world day-to-day circumstancesare called effectiveness studies. Conducting interventions under ideal, high controlled conditionsis referred to as efficacy studies [35]. To adequately compare trials, we focused solely on efficacy trials. However, until now the research status does not provide many effectiveness studies. Additionally, the question arises if we can appropriately distinguish between effectiveness and efficacy studies since internet interventions can be conducted in clinical settings and in an extensive public health setting [36]. However, in order to attempt the distinction between effectiveness and efficacy studies, we focused during the full-text screening on references to effectiveness studies or otherwise called pragmatic trials.

All analyses were conducted in the programming language R (version 3.6.1; R Core Team, 2019) and specifically with the help of the R package metafor [42]. For all statistical tests, we set a 5% significance level. When directed hypotheses were obtainable, tests were one-tailed.

To compute effects sizes, we used the following procedure. The standard deviations of the post – pre differences per group were computed using basic mathematical rules of (co) variances and assuming a correlation of r = 0.5 between pre and post scores (see online supplement for the corresponding code). Correlations around r = 0.5 are commonly found in intervention studies and have been used in other meta-analyses before (e.g., [43]). In a sensitivity analysis, we set the correlation to r = 0, which leads to the largest standard deviations and thus to the most conservative estimates as the uncertainty in the effect sizes is very likely overestimated (results were numerically slightly different but qualitatively equivalent; see online supplement). Based on the obtained post-pre mean differences and corresponding standard deviations per group, Hedges’ g was computed between experimental and control group. Hedges’ g corrects for the small positive bias in Cohen’s d. For both effect sizes, the following classification are typically used: small effect (d = 0.2), medium (d = 0.5) and large (d = 0.8) effects [44]. Intention to treat (ITT) data was used when available, however, completer data was used when ITT was not available.

Regarding OCD instruments, there is a correlation of 0.92 between YBOCS and DOCS scores, providing strong evidence of convergent validity for assessing OCD instruments (r = 0.92 [45]). Thus, if studies reported both DOCS and YBOCS measures, a mean effect estimate of both OCD instruments was calculated. The correlation between OCI-R and Y-BOCS indicates a medium degree (rs = .43 [46]). Abramowitz and colleagues (2010) supposed that the lower degree between OCI-R and Y-BOCS might be due to the measurement of distress associated with specific OC symptoms, whereas the Y-BOCS measures the global symptom severity [34]. Therefore, OCI-R was analysed separately. Comparison between the self-report rating scale of YBOCS versus YBOCS interview report a strong agreement [47] thus both types were analyzed together.

Throughout, we were applying random effects meta-analytic models as studies could a-priori be assumed to differ systematically in their true effect sizes because of differences in, for example, study design or investigated samples [48]. This assumption was later confirmed in the obtained results as we found significant heterogeneity between effect sizes. Heterogeneity between studies occurs as studies vary regarding experimental design, outcome assessment, and treatment characteristics [35]. Statistical heterogeneity results when the true effects of the varying studies lead to a higher discrepancy than expected coincidentally. We applied the statistical heterogeneity by using I² method by Higgins and Thompson (2002) and also investigated the between study and between effect size variance estimates directly [35]. Forest plots were used visualize results and detect potential outliers.

Moderator analyses using meta-regression [30] for comparison between remote treatments versus passive controls were performed with study characteristics of underlying theoretical interest. The following moderators were coded: contact with a therapist/ clinician, OCD diagnosis assessed by a measurement based on DSM, meanage, gender, type of technology, duration of intervention and study quality. For the variables support condition and OCD diagnosed according to a DSM version with a clinical structured interview, we coded dummy variables.

Research articles with significant results are likelier published than with insignificant or otherwise less convenient results [49]. Thus, publication bias can lead to an overestimation of effects. We tried to minimize potential for publication bias as much as possible by actively searching for unpublished results, in particular dissertations. However, dissertation papers (with unpublished data) were not included as we could not find eligible dissertations. To investigate the likelihood and the extent of a publication bias (or other sources of skewness in the effect size distribution), we generated Funnel plots and applied Egger’s regression tests [50].

Table 2 outlines the group-by-time interaction effect sizes for the studies included in the quantitative synthesis. Group-by-time interaction indicates whether low intensity technology-delivered CBT results in significant improvement in OCD symptoms from pre- to post-treatment compared to the corresponding pre-post changes in the control group. Outcomes were analyzed separately for passive controls and active controls. Regarding remote treatment versus passive controls for Y-BOCS only, a total of 10 comparisons were drawn. The pooled effect size indicated a significant moderate to large effect of − 0.66 (95% CI = [− 1.08, − 0.25]; p = 0.006, Q = 23.76, df = 9, p = 0.005, I² = 68.62%; Fig. 3). Referring to the results of remote treatment versus passive controls for Y-BOCS and DOCS combined, there was a significant pooled effect size of − 0.57 (95% CI = [− 0.85, − 0.30], p = 0.001, Q = 12.28, df = 9, p = 0.20, I² = 30.74%). For OCI-R only there was a small significant effect size (g = − 0.30, 95% CI = [− 0.56, − 0.04], p = 0.03, Q = 3.41, df = 6, p = 0.76, I² = 0%).

The overall effect size of the comparisons between remote treatment versus controls with other treatments was not statistically significant and there was an evidence for the presence of heterogeneity for Y-BOCS only (g = − 0.13, 95% CI = [− 1.12, 0.87], p = 0.77, Q = 57.24, df = 6, p < 0.0001, I² = 96.23%) and Y-BOCS and DOCS combined (g = − 0.23, 95% CI = [− 1.09, 0.63], p = 0.55, Q = 61.01, df = 7, p < 0.0001, I² = 95.63%). There were no studies measuring OCI-R.

Between-group effect sizes at post-treatment for the included studies in the quantitative synthesis are presented in Table 3. Passive controls and active controls were analyzed separately. According to passive controls, a total of 10 comparisons were drawn for Y-BOCS only and for Y-BOCS and DOCS combined and a total of 7 comparisons for OCI-R. Heterogeneity was present for Y-BOCS only (Q = 20.98, df = 9, p = 0.01, I² = 67.56%). For Y-BOCS and DOCS combined (Q = 14.97, df = 9, p = 0.09, I² = 47.39%) as well as for OCI-R only (Q = 3.57, df = 6, p = 0.74, I² = 0%), there was statistically no evidence for heterogeneity. The significant pooled effect sizes were moderate to large for Y-BOCS only effect (g = − 0.59, 95% CI = [− 0.99, − 0.18], p = 0.01; Fig. 4) and Y-BOCS and DOCS combined (g = − 0.55, 95% CI = [− 0.87, − 0.24], p = 0.003) and small for OCI-R (g = − 0.36, 95% CI = [− 0.62, − 0.09], p = 0.02).

According to controls with other treatments, a total of 7 comparisons were drawn for Y-BOCS only and a total of 8 comparisons for Y-BOCS and DOCS combined with a presence of heterogeneity (Y-BOCS: Q = 47.69, df = 6, p < 0.0001, I² = 95.22%; Y-BOCS and DOCS combined: Q = 49.55, df = 7, p < 0.0001, I² = 94.32%). The pooled effect sizes were not significant (Y-BOCS: g = − 0.14, 95% CI = [− 1.02, 0.74], p = 0.71; Y-BOCS and DOCS combined: g = − 0.22, 95% CI = [− 0.97, 0.53], p = 0.51). There were no studies measuring OCI-R.

The following results are only presented according to YBOCS outcomes and passive controls (n = 10). Sample age did not significantly moderate the overall effect of remote CBT with low personal contact intensity on OCD symptoms (b = − 0.03, SE = 0.06, 95% CI = [− 0.18, 0.12], p = 0.69), whereas women percentage showed a small significant moderating effect (b = − 0.03,SE = 0.01, 95% CI = [− 0.06, − 0.002], p = 0.04). Results demonstrated no significant moderating effect bythe duration of intervention (b = − 0.06, SE = 0.08, p = 0.46,95% CI = [− 0.25, 0.12]). Irrespective of the support conditions (yes versus no), we could not find evidence of a difference in symptom outcomes (b = − 0.84, SE = 0.41, 95% CI = [− 1.79, 0.11], p = 0.08). Moreover, we could not find evidence of a difference in symptom outcomes if samples were diagnosed according to DSM or not (b = − 0.51, SE = 0.41, 95% CI = [− 1.45, 0.43], p = 0.25). Study quality of included studies was not a significant moderator of the overall effect (b = 0.14, SE = 0.16, 95% CI = [− 0.22, 0.50], p = 0.41). Conducting a sub analysis including only studies with low risk of bias, results showed larger significant improvement in OCD symptoms from pre- to post-treatment compared to the corresponding pre-post changes in the control group (YBOCS: g = − 0.88, 95% CI = [− 1.56, − 0.21], p = 0.02, Q = 16.01, df = 5, p = 0.007, I² = 74.66%; YBOCS and DOCS combined: g = − 0.70, 95% CI = [− 1.12, − 0.28], p = 0.008, Q = 7.59,df = 5, p = 0.18, I² = 37.90%). However, heterogeneity was present for the analysis with YBOCS only. For between-group effect sizes at post-treatment, results showed as well larger significant effect sizes (YBOCS: g = − 0.84, 95% CI = [− 1.36, − 0.31], p = 0.009, Q = 9.83, df = 5, p = 0.08, I² = 58.01%; YBOCS and DOCS combined: g = − 0.77, 95% CI = [− 1.11, − 0.43], p = 0.002, Q = 5.51,df = 5, p = 0.36, I² = 9.44%).

The forest plot of the main effect of time (change) of Y-BOCS for only the passive control groups is shown in the supplementary material S1 Figure.