Is Computerized CBT Really Helpful for Adult Depression?
Identification and Selection of Studies
All RCTs completed and analysed by 11 July, 2011 were eligible for inclusion in this review. Five bibliographic databases were used [MEDLINE (1948 to July 2011), PsycINFO (1806 to July 2011), EMBASE (1980 to 2011), CENTRAL (Cochrane Library, 2011 latest issue), and CiNii (until July 2011)]. We also searched http://www.controlled-trials.com. Multiple search terms were used (Appendix) and modified for each database, as necessary. The search was performed on 11 July, 2011.
We included 1) randomised trials 2) in which the effects of guided and unguided CCBT specific to depression 3) were compared with one or more control conditions 4) in individuals aged 18 years or older 5) with depression, and in which 6) reliable and standardised rating-scales were equally used both at baseline and follow-up. Also, we only included studies 7) with proper allocation, concealment, and single or greater blinding of outcome assessment; and 8) trials using medications or other psychotherapies were included. We excluded studies on 1) inpatients, because we excluded patients with severe symptoms from self-help intervention, and those with 2) comorbidities such as psychotic disorders, manic status, dementia and severe physical conditions. In fact, we had originally intended to distinguish between patients on waitlists from treatment as usual (TAU), because we considered there to be restrictions on administration of medications to patients on waitlists. Nevertheless, the proportion of subjects taking medication at waitlist baseline was very similar to that with TAU, and medication was mostly not controlled. Therefore, we decided to group together both of these, and checked the influence of this factor on outcomes through a subgroup analysis. This grouping seemed to be justified because the above past five meta-analyses had treated data likewise. Studies had to have a primary endpoint including a measure of depression at the outcome assessment immediately after intervention and at long follow-up (if applicable). We defined long follow-up as follow-up where the final assessment was more than six months after treatment, because this is a recovery period associated with low future recurrence of depression. Function at post-treatment and the number of total dropouts were adopted as secondary endpoints.
Intervention effects were expressed using various types of rating scales for common outcomes, thus the effect sizes using standardized mean differences (SMDs) with 95% confidence intervals for post-treatment were computed, and then incorporated into the meta-analysis and presented with 95% confidence intervals. Where trials used a number of different tools to assess depression, we included the main outcome measure following our hierarchy, including the primary endpoint or endpoint first reported in the results.
Statistical heterogeneity was evaluated through a SMD forest plot. Cochrane's Q statistic (chi-squared test) was performed with a significance level of 0.10. Furthermore, the I-squared (I) statistic for heterogeneity was also used for confirmation of Cochrane's Q statistic. A random-effect model was selected due to the large heterogeneity of each clinical design and participants. All meta-analyses were performed using Review Manager (RevMan ver. 5.1). Subgroup analyses were performed for the type of control (Waitlist and TAU). Also, we re-evaluated the clinical effectiveness through a sensitivity analysis by excluding Beck Depression Inventory I, BDI- I, and II, or according to the difference in attrition rates and imputation techniques. The reason for the former is that, particularly in CCBT studies primarily relying on self-rating scales, measurement bias is suspected due to differences between the scales employed, and it is necessary to avoid this giving rise to underestimation. The reason for the latter is that high dropout rates were expected, thus we also performed an analysis excluding research with attrition rates higher than 20% (such rates probably had an influence on the results irrespective of ITT according to Cochran handbook), significantly higher dropout RRs or non-modern imputation processing.
A funnel plot was used as a test of the main outcome to detect publication or reporting bias through visual inspection. Begg's and Egger's tests were also conducted for statistical checking. When a significant small study effect was noted, we assessed its influence through the trim-and-fill method.
It seems that adequate missing value management is useful in carrying out appropriate CCBT evaluation, because we expect overall attrition rates to be high. Although it was physically unfeasible for us to collect all the original data without imputation, in the present meta-analysis we considered the potential impact on the review result through a sensitivity analysis in terms of the influence of imputation. Also, the modern imputation was defined as an imputation needing more complex processing than classic and comparatively simple imputations such as last observation carried forward (LOCF) or mean imputation (MI).