A psychometric analysis and revalidation of the Yale-Brown Obsessive Compulsive Scale modified for Binge Eating in adults with binge eating disorder

Item response distributions were evaluated by examining the response frequencies for each item at baseline: the sparseness of response option endorsement and floor and ceiling effects (operationally defined as a threshold of 20% of participants endorsing the lowest or highest response option). Inter-item polychoric correlations were estimated at baseline across all items to evaluate provisional domain patterns and identify item redundancies.

Although the original validation study examined the domain structure of the Y-BOCS-BE [1], the domain structure was re-examined in this manuscript using the two short-term efficacy studies of LDX described above. Domain specification followed a two-stage procedure in which evidence obtained from both EFA and IRT was used to arrive at an empirically justifiable domain solution. Selection of the final domain solution was based upon balancing model fit, clinical input, interpretability of the solution, and scoring statistics [13]. To control for the risk that the two-stage procedure might arrive at a solution due either to over-fitting or capitalization on sample-specific characteristics [14], the solution obtained from study 1 was independently replicated in study 2. The reproducibility of the solution was compared between the two studies to determine whether the solution was stable [15]. Model fit indices for EFA model selection were obtained from limited information estimation. Specifically, the mean and variance adjusted weighted least squares (WLSMV) estimator was used. Reported EFA item parameter estimates, item response theory (IRT) model fit, and IRT item parameters were all obtained from full-information marginal maximum likelihood estimation. Details of the modern psychometric methods are given in the Supplemental Materials. Two competing domain models were examined. One model was a MIRT with domains equal to those obtained from the final EFA solution, and the other was a bifactor model [16], which used the same domain structure as the MIRT augmented by a general domain reflecting overall BED severity.

After agreement on the interpretability and clinical relevance of the best fitting solution was achieved, scoring statistics were computed at baseline to determine how best to use the Y-BOCS-BE to characterize binge eating severity. Scoring statistics computed at baseline compared whether domain scoring (based on the MIRT) or a total score (based on the bifactor) was more appropriate for characterizing binge eating severity. Scoring statistics employed included the ?, ?_H, the ratio of ?_H to ?, the explained common variance (ECV), and the H statistic [13].

Meaningful change was computed on Y-BOCS-BE change scores computed between baseline and week 12 using both distribution- and anchor-based estimates of minimal clinically important improvement (MCII). Distribution-based estimates included 1 and 0.5 baseline standard deviation (SD) and baseline standard error of the measurement (SEM). In addition, following Harvill 1991 [17], the baseline SD and SEM were used to index the amount an individual needed to improve beyond the average change score (??) to reflect meaningful change. These difference estimates are noted as ???Baseline SD and ???Baseline SEM, respectively. Anchor-based MCII was estimated by comparing Y-BOCS-BE change score point and interval estimates for minimal improvement defined by the CGI-S (1-point improvement between baseline and week 12) and CGI-I (response category 3, with descriptor of ?minimally improved?). In addition to point estimates, cumulative distribution functions (CDFs) stratifying on CGI-I and CGI-S were plotted; separation of CDFs at the meaningful change estimate location between treatment arms was tested.

Mixed-effects models for repeated measures (MMRM) were used to assess treatment efficacy using Y-BOCS-BE change from baseline scores based on the full analysis set (FAS), as defined for each study [6, 8]. An unstructured error covariance matrix was used for the repeated measures. Efficacy estimates were based on least squares (LS) mean contrasts at each post-baseline visit. The MMRM was adjusted for baseline Y-BOCS-BE scores and the baseline Y-BOCS-BE score by treatment interaction.

For the maintenance-of-efficacy study (study 3), efficacy analyses differed slightly owing to the randomized-withdrawal study design. Y-BOCS-BE scores obtained during the randomized-withdrawal period were modeled as outcomes in a linear mixed model, with the primary predictors of time, treatment, and treatment by time. Covariate adjustments included the RWB Y-BOCS-BE scores and time by RWB score interaction. An unstructured residual covariance matrix was used in this model.

The missing data rate was negligible (baseline, 0%; week 4, 0.96%; week 12, 0.32%). Item distribution analyses indicated there were no floor or ceiling effects. Inter-item correlations were weak at baseline (range, 0.0?0.70), with only two correlations exceeding 0.6, and were acceptable at week 12 (range, 0.62?0.88).

Model fit indices indicated that the Y-BOCS-BE domain structure was optimally characterized with a three-factor model interpreted after an oblique quartimax rotation (?²P value?<?0.0001; RMSEA [95% CI]?=?0.09 [0.07?0.11]; CFI?=?0.98; TLI?=?0.94; SRMR?=?0.04); this structure was carried forward into the IRT model estimation. The three domains of obsessive/compulsive, restraint, and control were defined, respectively, by time spent on, disruption of life due to, and distress associated with BE thoughts and actions (items 1?3, 6?8); ability to resist BE thoughts and actions (items 4 and 9); and ability to control BE thoughts and actions (items 5 and 10). These domains were modestly correlated (r?=?0.24, 0.22, 0.32, respectively).

Because this three-domain structure differed from both the two- and seven-domain solutions reported by Deal et al. [1], confirmatory analyses were performed to assess whether the three-domain structure fits better than the parsimonious two-domain structure reported by Deal et al. [1]. These structures were compared to a bifactor modification to the three-domain structure, which was found to achieve the best fit across all indices (C2 P value?=?0.0001; RMSEA [95% CI]?=?0.09 [0.07?0.11]; AIC?=?8394.36; BIC?=?8630.14) and minimize local dependence. The bifactor solution, emphasizing measurement of BE severity in a total domain, was clinically determined to also possess the greatest verisimilitude and interpretability. Consequently, the bifactor structure was retained in the IRT model estimation. A full explication of the evidence supporting this conclusion is contained in the Supplemental Materials.

Figure 1 depicts the test information function (TIF; proportional to scale reliability) for the final IRT model solution. The TIF is multimodal, with score reliability maximized at approximately 1 SD above and below the mean, which is identified by 0 on the x-axis, for the latent BE severity scale.

Baseline scoring statistics supported the use of total score (??=?0.94; ?_H?=?0.89; ?_H/??=?0.95; ECV?=?0.95; and H?=?0.87). Definitions and interpretations of these statistics are given in Rodriguez et al. [13]. Descriptive statistics characterizing the distributions of these scores for baseline, week 4, and week 12 are presented in Table 1, and reflect the substantial reductions in post-baseline scores.

Table 2 summarizes internal consistency reliability at baseline and week 12. Y-BOCS-BE item to total correlations and item ? were lower at baseline for each item (ranges: correlations, 0.20?0.56; Cronbach?s ?, 0.74?0.78) and for total score (Cronbach?s ?, 0.77) than at week 12 for each item (ranges: correlations, 0.74?0.85; Cronbach?s ?, all 0.95) and total score (Cronbach?s ?, 0.96). However, baseline total score internal consistency achieved the pre-specified acceptability criterion.

Test?retest reliability correlations between baseline and week 4 in participants whose CGI-S ratings did not change between assessments or who had CGI-I ratings of ?no change? between assessments were low (r?=?0.34 and r?=?0.59). Sensitivity analyses of test?retest reliability correlations between week 4 and week 12, conditioned on similar CGI-S and CGI-I anchors, were acceptably higher (r?=?0.77 and r?=?0.90).

At baseline, convergent validity estimates were weak (r???0.4) across all validators (Table 2). At week 4, convergent validity estimates increased for the TFEQ disinhibition and hunger domains but remained low for binge days and the TFEQ cognitive domain (Table 3). Convergent validity estimates were high for all validators (Table 3; all r???0.66) at week 12 except for binge days and the TFEQ cognitive domain.

The discussion section contains an extensive clinical explanation of the relationship between poor BED insight and self-monitoring effects, addressing the low test?retest reliability and convergent validity.

Distribution-based MCII estimates for ???Baseline SD and ???Baseline SEM, respectively, were ?17.04 and ?14.48 in study 1 and ?15.06 and ?12.67 in study 2. Anchor-based score estimates for ?minimal improvement? on the CGI-I or a 1-point improvement on the CGI-S from baseline to week 12, respectively, were ?5.00 and ?6.78 in study 1 and ?4.52 and ?5.63 in study 2. The proportions of participants achieving the anchor-based estimates were excessively high (>?80%). Therefore, the distribution-based estimates (???Baseline SD and ???Baseline SEM) reported above were retained.

Observed treatment-stratified CDF curves with estimated odds ratios corresponding to estimated MCII values are shown in Fig. 2. MCII estimates based on ???Baseline SD or ???Baseline SEM met criteria for optimal MCII for scores (Fig. 2A). Fifty percent of participants in the LDX group and 16.6% in the placebo group achieved a degree of improvement that met or exceeded the ???Baseline SD estimate; 63.5% of participants in the LDX group and 28.0% in the placebo group achieved a degree of improvement that met or exceeded the ???Baseline SEM estimate. Corresponding CDF curves for study 2 (Fig. 2B) also showed substantial separation between treatment groups.

LDX treatment effects across all studies are summarized in Table 4. The LS mean treatment difference for change from baseline (LDX vs placebo) significantly favored LDX over placebo at weeks 4, 8, and 12 (all P?<?0.0001) in studies 1 and 2; these findings are consistent with the primary publication [6]. The LS mean treatment difference for change from randomized-withdrawal baseline (LDX vs placebo) significantly favored LDX over placebo at weeks 16, 20, 24, 28, 32, and 38 (all P?<?0.0001) in study 3, which is consistent with the primary publication [8].