Participants were recruited from patients admitted to the Toronto General Hospital's inpatient or day hospital treatment program between September 2010 and August 2012. In a pre-admission orientation session, patients were asked by a clinical team member if they would like to be contacted by our research team to find out about a study being conducted on psychosocial functioning and eating disorders. Of the 172 patients admitted into treatment over the recruitment period, 130 agreed to be contacted. Of these, 26 patients did not return our phone calls and 104 patients agreed to meet with a research assistant within a few days of their admission. Of these, 97 patients agreed to participate upon learning more about the study and reviewing the consent form. All participants gave written informed consent.
Our final sample of 97 was predominantly female (97%) and Caucasian (79.2%), with 4.5% of participants identifying themselves as East Asian, 1.4% as South Asian, 2.8% African-Canadian, 10.8% Latino, and 1.5% as mixed race. The mean age in our sample was 28 years (SD = 9.6), and participant ages ranged from 17 to 57 years. All participants were assessed using the Eating Disorder Examination, and met DSM-IV-TR criteria for an eating disorder. The diagnostic breakdown of our sample was as follows: 27.2% AN restricting type (AN-R), 18.5% AN binge-purge type (AN-BP), 29.6% BN, and 24.7% EDNOS. The mean BMI in our sample was 21 (SD = 5.5) at admission, and ranged from 12.6 to 44. Of those patients who participated in our study, 27.8% were admitted to our inpatient unit, and 72.2% were admitted to our day hospital.
Both of these specialized treatment programs are group-therapy based. Groups in each program are ongoing, with patients entering and leaving at different times. A multi-disciplinary team consisting of psychiatrists, psychologists, nurses, dieticians, social workers, and occupational therapists runs the treatment programs. Treatment goals include medical stabilization, weight restoration in the case of underweight patients, nutritional rehabilitation, and normalized eating through staff-supported meals and snacks, and eradication of binge eating, purging, and excessive exercise. Although the underlying orientation of both programs is cognitive-behavioral, patients attend a variety of manual-based groups on psycho-education, relationships and sexuality, expressive arts, anxiety management, dialectical behavior therapy, and cognitive-behavioral therapy. Self-compassion is implicitly encouraged in some of the groups, but there is no group in either program devoted primarily to building self-compassion or reducing shame.
Patterns of Missing Data
The present analyses were performed on 97 patients admitted to one of our two specialized eating disorder treatment programs. The EDE-Q was administered as part of a routine clinical intake assessment. Responses at Time 0 (i.e., baseline) were available for 83 of the 97 participants, with missing data accounted for by eight patients refusing to complete or simply not returning the questionnaire, and six patients leaving treatment before returning their responses. Both other questionnaires (i.e., the ESS and SCS) were administered at admission as part of our specific study survey, and were thus available for all participants 97 at Time 0.
Over the course of their treatment, participants were asked to complete online questionnaires on five occasions—namely, at the time of their admission, and after 3, 6, 9, and 12 weeks of treatment. The mean number of online assessments completed by each participant was 3.5 questionnaires over 12 weeks. Sixty-three of the 97 participants completed three or more questionnaires, and 34 completed less than three. Of those participants who completed one or two assessments only, 17 were active patients in the treatment but simply failed to complete several study questionnaires; 16 dropped out of treatment prematurely either because the clinical team asked them to leave for reasons such as poor program compliance, or because the participant chose to drop out early; and one participant successfully completed what the team considered to be a sufficient “dose” of treatment before the 12-week mark.
All analyses were conducted using SAS 9.2 (SAS Institute, 2012). We tested our primary hypotheses with multilevel modeling using maximum likelihood estimation. Multilevel modeling is the recommended statistical approach when observations are nested within participants, as in longitudinal treatment datasets. It models change trajectories over time (i.e., slopes, or rates of change) without requiring fixed data collection schedules. It has as an advantage that it retains data from participants for whom observations are missing, provided that these are missing at random (MAR).MAR means that the “missingness” of the data is unrelated to the unobserved value(s) but may be related to the missing observation(s) through other variables in the model for which observations are not missing.
We tested the MAR assumption in the present dataset with pattern-mixture variables based on Hecker and Gibbon's (1997) recommendations. We created two categorical variables to represent the two primary reasons for missing data (i.e., early discharge or poor study compliance) and assigned participants a value (yes–no) based on their pattern of missing data. We found no effect for either of these variables, or their interaction with time, in predicting changes in our dependent variables. As such, patterns of missing data did not seem to influence outcomes in our dataset, in support of the MAR assumption. Because only one participant left treatment (and thus our study) early due to adequate progress, we did not control for this scenario when testing the MAR assumption. Furthermore, the “good-enough level” of change assumption—that patients who leave treatment early due to progress have faster rates of improvement –appeared irrelevant in our dataset (Baldwin, Berkeljon, Atkins, Olsen, & Nielsen, 2009).
We tested our primary hypotheses in two multilevel models in which the dependent variables were patients' global scores on the EDE-Q and ESS between the time of admission to the program and week 12 of treatment. It is important to include all data points when modeling slopes as dependent variables, since doing so provides more precise estimates of participants' rates of change. Although participants were asked to complete their assessments within 48 h of receiving a notification to do so (sent precisely at 0, 3, 6, 9, and 12 weeks), they sometimes completed their questionnaires late. Nevertheless, all completed surveys were date stamped, which allowed us to identify the precise time, relative to patients' admission date, at which the survey was completed. To maximize the precision of our data analyses, we entered the precise time point in each participant's treatment course (i.e., 3 weeks, 4.5 weeks etc.) at which they completed each of their surveys used these times in our multilevel models described below. This approach, rendered possible through multilevel modeling, is recommended when measurement occasions are variably spaced, even if they were intended to occur at fixed times, as it more precisely models participants' within-person trajectories over time.[45, 46]
In each of our models, we included a fixed- and random-effects portion for effects considered to be constant and variable across participants respectively. Initial models included a random intercept for patients and an unstructured error covariance structure. We then included a random effect for time. This was significant in both models and improved the model fit according to the AIC criterion; it was thus retained in our analyses. Fixed effects in both multilevel models included time, baseline levels of the relevant dependent variable and its interaction with time, and diagnosis and its interaction with time. Diagnosis was a dummy-coded categorical variable representing patients' eating disorder diagnosis at admission (i.e., AN-R, AN-BP, BN, or EDNOS).
Patient demographic characteristics, including age, marital status, and living circumstances, were also included in initial models but were not significant and were therefore dropped. Because of the fluctuating composition of the treatment groups, we were unable to represent patients as nested within groups. We did, however, control for patient treatment program (i.e., inpatient vs. day hospital) in our initial models. There was no significant effect for Program or Program × Time in either model so these terms were removed from the final models. To check whether our dependent variables changed in a non-linear fashion, we tested for quadratic and cubic effect of time on eating disorder symptoms and shame; however, results supported a linear model of change for both dependent variables.
Early Change Variables
To test our hypothesis that larger early changes in shame and self-compassion would predict faster symptom decreases, we calculated difference scores to reflect the magnitude of change participants experienced in shame and self-compassion between their first and second assessment points. First assessments were completed within a few days of admission, and second assessments were generally completed between weeks 3 and 5 (Mean = 4.3, SD = 1.1). For ease of interpretation, differences scores were calculated such that positive values indicated improvements in the variable in question in the desired direction (i.e., an increase for self-compassion, a decrease for shame). In addition, these variables and all other between-person predictors were standardized before we tested our multilevel models.