The majority of participants in all studies were female (). Age and BMI varied considerably across trials, with the oldest participants in the Look AHEAD trial (Study 1) and the heaviest participants among bariatric surgery candidates (Study 2). Race and ethnicity of the participants across all studies included approximately 70.4% Caucasian, 23.5% African American, 3.4% Hispanic/Latino, 0.6% Asian American, 0.6% Native American, 1.5% other/unknown.
Factor analysis of the dichotomous items revealed a first main factor accounting for 56.3% of the variance, with a first-to-second factor eigenvalue ratio of 3.43:1. Root mean square residual for this model was 0.051, suggesting that the five items entered into the factor analysis are represented adequately by a single unidimensional construct.
Empirical item response functions for each of the five items were calculated to show the proportion of the sample endorsing each item as a function of the number of the remaining four items that were endorsed (i.e., the rest-score). All five items showed clear evidence of monotonicity with increasing item response functions ().
Empirical item response functions.
presents the percent of participants endorsing each of the five items by study. As would be expected, rates of endorsement for all items were lower for the obesity and psychiatric clinic studies (Studies 1–3) and control study (Study 5) than for Studies 4 and 6 that focused specifically on night eating problems.
Frequency and percent of participants endorsing symptoms by study.
Model Specification and IRT Analysis
PARSCALE software was used to evaluate three separate 5-item models. The -2 log likelihood value for Model 1 (2PL common slope model) was 8306.87, compared to 8273.04 for Model 2 (2PL fully unconstrained model) and 8273.01 for Model 3 (3PL fully unconstrained model). Thus, Model 2 provided a significantly better fit than Model 1 (χ2 = 33.83, df = 5, p < .001). However, the fit for Models 2 and 3 was not significantly different 1 (χ2 = 0.03, df = 6, p > .05), suggesting that Model 2, the two parameter fully unconstrained model, provided the most parsimonious fit.
shows the parameter estimates and standard errors for this model. shows the corresponding ICC’s and the IIC’s based upon these parameter estimates. Of note, the difficulty parameters (from ) for three of the items (evening hyperphagia/nocturnal eating, initial insomnia, and night awakenings) are relatively comparable, ranging from 0.53 to 0.82. This suggests that they provide diagnostic information across a comparable range of the night eating severity (). Second, the symptom which provides the most amount of diagnostic information () is evening hyperphagia/nocturnal eating, which are the core features of NES. Finally, two of the items, morning anorexia and delayed morning eating, have considerably lower item response curve slopes () than the remaining items, and consequently, provide considerably less psychometric information ().
Item response curves for the 2PL logistic model.
Item information curves for the 2PL logistic model
Local Item Independence
The lower item response curves for morning anorexia and delayed morning eating led to the evaluation of their local item independence. Analysis of the 2PL fully unconstrained 5-item model characterized across the six samples in produced chi-square ratios greater than 3.0 for 3 of 10 item pairs and 5 of 10 item triplets. All of the pair ratios and all but 2 of the triplet ratios above 3.0 included morning anorexia and/or delayed morning eating. The chi-square ratio for the morning anorexia/delayed morning eating pair was 34.4, and all ratios for triplets containing both items were 17.9 or higher. This pattern of results suggests that one or both of these two items violated the assumption of local independence. Consequently, these items were dropped and a new 3-item (evening hyperphagia/nocturnal eating, initial insomnia, and night awakenings) 2PL fully unconstrained IRT model was reevaluated for local item independence. All of the chi-square ratios for single items, item pairs, and triplets were less than 3.0. These results suggest that the 3-item fully unconstrained model is consistent with the IRT assumption of local item independence.
Studies 4 and 6 had the largest number of participants who met all three criteria, at 49% and 39% of the samples, respectively, followed by Studies 3 (22%), 2 (7.5%), 1 (4.5%), and 5 (2%). An additional 47% and 43% of Studies 4 and 6, respectively, met two of the criteria, followed by Studies 2 (26%), 3 (22%), 1 (11%), and 5 (2%).