Page 29 - The efficacy and effectiveness of psychological treatments for eating disorders - Elske van den Berg

Chapter 2 29
Statistical analysis
For each comparison between psychological treatment and control condition, we calculated the pooled effect sizes (Hedges’ g; weighted by inverse variance) for the primary and secondary outcomes. Effect sizes were calculated by subtracting the mean score at posttest of the psychological treatment group from the mean score of the control group and dividing the result by the pooled weighted standard deviations of the two groups. Effect sizes of 0–0.32 are considered to be small, while effect sizes of 0.33 – 0.55 are moderate and effect sizes of 0.56 – 1.2 are large (Lipsey & Wilson, 1993). When means and standard deviations were not reported, we used other statistics (p value) to compute the effect size (Dare, Eisler, Russell, Treasure, & Dodge, 2001). As weight is assessed in various ways (see Appendix C), we reported a combined effect size. To explore whether this influenced the results, we also performed the meta-analyses including only BMI or kg.
With regard to our secondary outcome measures, we calculated pooled mean effect sizes, using the random effects model in the Comprehensive Meta-Analysis (CMA) software package (version 3; Borenstein & Rothstein, 2009). We also calculated the I2 statistic, where a value of 0% indicates that there is no observed heterogeneity, 50% determines a moderate heterogeneity and 75% a high heterogeneity (Higgins, Thompson, Deeks, & Altman, 2003). The 95% confidence intervals around the I2 was calculated using the non-central χ2-based approach within the heterogi module for Stata (release 15; StataCorp., 2017). Potential publication bias was examined accord- ing to Duval and Tweedie’s trim and fill procedure (Duval & Tweedie, 2000) using Comprehensive Meta-Analysis, which calculates an adjusted effect size taking into account missing studies. The symmetry of the funnel plots were tested using the Begg & Mazumdar rank correlation test (Begg & Mazumdar, 1994) and Egger’s test (Egger, Davey Smith, Schneider, & Minder, 1997).
In this meta-analysis, five studies were included in which two psychological treat- ments were compared with the same control group (Byrne et al., 2017; Dare et al., 2001; Gowers et al., 2007; McIntosh et al., 2005; Zipfel et al., 2014), thus resulting in multiple comparisons in the same analysis. Because those comparisons are not inde- pendent from each other, this may have resulted in an artificial reduction of heter- ogeneity and may have influenced the pooled effect size. We therefore performed a sensitivity analysis by including the largest effect size and smallest effect size for each of these studies separately. Sensitivity analyses were conducted on low risk of bias studies and on outpatient studies.
Subgroup analyses were used to explore the influence of various patient- and study characteristics. We performed subgroup analyses when at least three studies were available per subgroup condition. Subgroup analyses were performed using the