This meta-analytic review was conducted and reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines (37), and was submitted for pre-registration in February 2020 at the International Prospective Register of Systematic Reviews (38). All analyses were planned before the systematic searches, literature review and data extraction were performed.
Search strategy
Electronic databases that were searched were PsycInfo, Embase, Medline, Proquest Dissertations and Theses, and Cinahl. Three search strings were constructed for “eating disorders”, “cognitive behavior therapy”, and “psychodynamic-interpersonal therapy”, respectively. The search strings representing the treatment approaches were first combined with the operator OR, then combined with the construct “eating disorder” with the operator AND. The complete search strategy is attached in appendix B.
Study selection and data management
Reports were pooled across databases and reviewed. Data from included studies were extracted by the two authors independently. As EDs first was included as an independent chapter in DSM-3 in 1980, records older than 1980 were excluded. A pilot-review was conducted to ensure inter-rater agreement on the extracted data. Discrepancy was solved by discussion. When calculating effect sizes, data based on intention to treat analyses were prioritized over data based on completers analysis.
Eligibility criteria
During screening, all references to original clinical trials on the treatment of EDs were considered for full text review. To be considered eligible for final inclusion, the reports had to provide information to calculate a between-group or within-groups effect size for the proportion of patients in remission, or changes in specific or general psychopathology across at least two time-points; for at least one psychotherapeutic intervention that had a cognitive-behavioral focus or a psychodynamic-interpersonal focus; directed to outpatients with a clinical ED. Waitlist/nutritional counseling (WL/NC) conditions were included if present in reports that met other inclusion criteria.
Exclusion criteria in the full-text review were multimodal therapies combining, e.g., milieu therapy, medication, exercise; treatments combining aspects of CBT and PIT; interventions not targeting the cognitive or psychodynamic-interpersonal aspects of EDs, e.g., exposure and response prevention, dietary advice or specialist supportive clinical management; and treatments broader in scope than CBT or PIT, e.g., dialectical behavior therapy, and acceptance and commitment therapy.
Data extraction and coding
Outcome variables. All outcome variables were coded across two time-points: Pre-treatment (t0) and 12 months follow-up (t1). Because relapse rates are high for EDs, 12 months follow up was used to assess treatment effects that can be said to be stable over time. If outcome assessments were available for several time-points after the end of treatment, the time-point closest to 12 months was prioritized. If follow-up assessment was unavailable, end of treatment assessment was used and coded as 0.
The number of patients intended to be treated at t0, and the number of patients in remission at t1 were extracted. For the continuous outcomes, standardized between group and within group changes in the form of Cohen's d, were computed based on means and standard deviations at t0 and t1, or from correlations or p-values for pre-post changes. Computing Cohen's d, the correlation between the pre and post measures were set to .70, which is considered sufficiently close to the test-retest reliability of many psychometric scales (39).
Primary outcome variable. Remission was defined as the proportion of patients in the treated sample that has undergone weight normalization (AN-samples), cessation of compensatory behaviors (AN- and BN-samples), and cessation of bingeing at t1 (BN- and BED-samples).
Secondary outcome variables.Furthermore, two secondary outcome variables were coded. First, specific psychopathology was coded for t0 and t1. Scales such as the Eating Disorder Examination (EDE) were preferred if primary studies reported several measures. These instruments consist of the subscales; restraint, eating concerns, shape concerns, and weight concerns, assumed to encompass the specific psychopathology. In studies where other instruments were used for measuring specific psychopathology, each subscale was evaluated in terms of relevance to its core features. Second, change in general psychopathology from t0 to t1 was quantified using assessment scales for depressive (e.g., BDI, HAM-D) or anxious (e.g., STAI-S, STAI-T, HAM-A) symptomatology. In cases where several subscales were reported, composite change scores were made from subscale scores measuring specific or general psychopathology. Each sample contributed only with one effect size per outcome measure.
Predictor variables. Treatment approaches were coded categorically as either cognitive-behavioral therapy or psychodynamic-interpersonal therapy. The CBT approach was included and coded based on the focus on dysfunctional thoughts, beliefs and attitudes regarding eating, body shape and weight, and how these relate to behavior and emotions. The PIT approach was included and coded according to the definition by Blagys & Hilsenroth (40).
Standardized change scores for specific and general psychopathology were also used as predictors of remission in the analyses.
Three patient variables/moderator variables were coded. First, ED diagnosis was coded as either AN, BN, BED, or mixed samples. Second, personality disorder was coded as the number of patients in the treated sample with a personality disorder diagnosis. Third, mean patient age was coded as a continuous variable.
Additional study-level predictors. To examine the potential moderating role of follow-up time on treatment effect, the number of months from end of treatment to follow up was coded as a continuous variable. This variable will allow an assessment of the stability of treatment effects over time.
Data synthesis and meta-analysis
Meta analyses were performed by using the Comprehensive Meta-Analysis Software version 3. All meta-analytic models were constructed with effect sizes weighted by their inverse variance, assuming random effects, as is recommended when the true treatment effects reported by studies are expected to vary (41).
To answer question 1, computation of main effects of treatment approach were planned in separate between- and within-groups analyses. Synthesis of between-group effect sizes were planned for the relative effects of CBT and PIT versus WL/NC in direct comparisons using odds ratios for remission and Cohen's d for standardized differences in changes in specific and general psychopathology. Within-group summary effect sizes were calculated for individual treatment arms where CBT, PIT or WL/NC were delivered, using event rates for remission and Cohen's d for pre-post changes in psychopathology. Because effect sizes were derived from studies with different designs and patient samples, significant statistical heterogeneity was expected and subjected to examination.
To answer question 2, the impact of change in specific and general psychopathology on remission was assessed. The values of change scores were centered as is recommended for continuous variables used in multiple regression with categorical variables (42). Regression models were made for each treatment approach, where remission rates were independently predicted by change scores. Additional regression models were planned to test whether treatment effects for remission were mediated by change in specific and general psychopathology, according to the model proposed by Baron & Kenny (43). Furthermore, the relative importance of each hypothesized mediator was examined by comparing their respective regression coefficients, the variance explained by, and significance of the addition of this variable to the model.
To answer question 3, regression analyses of remission on patient-characteristics (ED diagnosis, age and comorbidity), for each treatment approach were performed. The strength and direction of relationships, significance-levels, as well as variance explained by the models were examined. Regression analyses of remission rates on treatment approach and patient variables as simultaneous predictors were planned.
Furthermore, analyses were performed to assess the possible significance of design characteristics, i.e., follow-up time and allocation to treatment conditions.
Publication bias assessment
One vulnerability of meta-analyses is the potential presence of publication bias, i.e., if studies with weak or non-significant effects are not published and therefore not included in the analysis. Publication bias has been identified as a problem in both psychological and medical research (44) but is unreliable to test with one method only. The use of several methods is therefore recommended (45). To test for publication bias, we used funnel plots to visually assess the presence of publication bias and Egger’s regression for examining correlations between sample size and estimated effect sizes. Using the Duvall and Tweedie`s Trim and fill method for imputing missing studies, the adjusted effect sizes for the CBT, PIT and WL/NC conditions will be examined for all outcomes.