Fifty-three RCTs (n=9,413 participants, range 20 to 1,656) were included in the analysis. Trials were reported as rigorous with acceptable attrition rates (19.3%), high rates of biologically validated outcomes (75%), and manualised treatment delivery (98%).
Main treatment effect: The pooled effect of cognitive-behavioural treatment (CBT) was found to have a small but significant improvement on the outcome of patients (g=0.144, 95% CI 0.094 to 0.194; fixed-effect model; 53 RCTs). Heterogeneity was reported as high (Q=128.85, p<0.005), so a random-effect model was used (g=0.154, 95% CI 0.066 to 0.242). Subgroup analysis indicated that the time of outcome assessment was an important predictor of between-study variance.
Subgroup moderators: Analysis of subgroup moderators across alcohol or primary drugs indicated that CBT treatment had a moderate pooled effect with marijuana use (g=0.531, 95% CI 0.375 to 0.651; fixed-effect model; six RCTs), whilst other drugs or alcohol showed only a small effect. RCTs across the type of CBT treatment indicated that the best treatment was CBT combined with psychosocial treatment (g=0.305, 95% CI 0.116 to 0.493; random-effects model; 19 RCTs), followed by CBT in combination with pharmacological treatment (g=0.208, 95% CI 0.070 to 0.346; fixed-effect model; 13 RCTs) compared to CBT alone (g=0.172, 95% CI 0.053 to 0.292; random-effects model; 21 RCTs). RCTs across the type of comparison treatment indicated that it was better to use CBT than no treatment at all (g=0.796, 95% CI 0.454 to 1.140; random-effects model; six RCTs). Other comparisons had very small effects and that of using CBT as an adjunct had no effect.
Regression moderators: Meta-regression analyses indicated that outcome assessment type (p<0.005) and time of follow-up assessment (p<0.05) were negatively related to CBT effect size. Female participants were more likely to respond to treatment (p<0.05) and shorter treatment sessions were more likely to produce effective treatments (p<0.005).
Publication bias was not found, rank-order correlation was non-significant and fail safe N was 340.