The development instrument to measure the risk level for EDs in the population of children and adolescents considering the IRT started with the use of 10 items (variables). These items are considered observable facets of this trait. The answer options for the items were dichotomous, where a “0” meant “no” and “1” meant “yes.” However, in the sample of 323 children and adolescents who answered the 10 items, for one of them (“Do you vomit or make yourself vomit to get rid of food after eating?”), there was no variability in the answers, so it had to be eliminated when developing the measure further. The nine items that compose the instrument trait are presented in Table 1.

The two-parameter logistic model was adjusted to create the measure with IRT.12,13 In the context of this study, the two-parameter logistic model predicts the probability of child j, with risk symptoms of eating disorder θj, answering category 1 (yes) for item i (items presented in Chart 1), that is, P(Xij=1|θj), as follows14:

Xij is the dichotomous item, which represents the presence or absence of observable eating disorder symptom i, and takes the values of 1 or 014;

θj represents the risk symptoms of an eating disorder of the jth child, estimated on a scale with a zero mean and standard deviation of 1, called “risk symptoms of an eating disorder”;

bi is the positional parameter of item i, estimated on the same scale as θ, and which represents the severity of the symptom measured in item i, that is, the risk symptoms of an eating disorder needed for the child to answer category 1 (yes) for item i with a probability higher than 0.5 (the higher the estimate for parameter bi, the more serious that symptom is)15; ai is the discrimination (or slope) parameter of item i – low values of this parameter indicate that children and adolescents with different levels of an eating disorder have roughly the same probability of answering category 1 for item i. That is, they have roughly the same probability of presenting the symptom measured by item i. Very high values of this parameter divide the children and adolescents into two primary groups: the group that has risk symptoms of an eating disorder below the value of the bi parameter and the group that has risk symptoms of an eating disorder above the value of the bi parameter; D scale factor (constant, equal to 1, or 1.7).15

The two-parameter logistic model has to meet two assumptions in order for its use: local independence and unidimensionality (a single latent trait determines the item responses). These assumptions are related in that if the measuring instrument is unidimensional, then local independence is fulfilled.12,13 The unidimensionality assumption can be relaxed, which is known as sufficient unidimensionality.16,17 For model can be used considering the sufficient unidimensionality, it needs to have a predominant dimension (proportion of explanation of the first dimension at least equal to 20%).12,13 The verification of the sufficient unidimensionality of the measurement instrument was done through exploratory factor analysis using the tetrachoric correlation matrix, as the items have a dichotomous response. The verification of sufficient unidimensionality was performed with the psych package, version 2.0.7, of the R software.

The IRT models’ products evaluated in this study were item characteristic curves (ICCs), test information curve (TIC), and differential item functioning (DIF).18

The ICCs of the two-parameter logistic this study describes how much changes in the latent trait are related to changes in the probability of saying “yes”. Items with high discrimination values (the items should have estimates for the (ai) parameter that are higher than or equal to “1”) present steeper ICCs. Items with higher thresholds that are more to the right in the ICCs graph present greater severity.18

The TIC is a sum of the item information curves. It shows what levels of an eating disorder the measuring instrument created works best; that is, at what levels of an eating disorder in children and adolescents the measuring instrument created estimates the latent trait most accurately.19

DIF is a statistical characteristic of an item that shows the extent to which the item might be measuring different abilities for members of separate subgroups.12,13 The presence of DIF was assessed for the “version” (physical or online instrument), “sex” (male or female), and “age” (up to seven or over seven years old) variables, using Lord's chi-squared method with Bonferroni adjustment. P values above 0.05 indicate the absence of DIF.12,13 5.1 of the difR package in the R software was used. These analyses were performed using version 2.0.7 of the psych package in the R software.

The instrument was assessed in relation to its content by independent expert evaluators with clinical experience in the latent trait being studied. Fifteen professionals participated in this stage, five from each of the areas of psychology, medication, and nutrition, as well as one pedagogue who worked in the classroom with children aged 6–7. In the procedure used to determine the content validity, the professionals assessed the instrument in terms of clarity of language, practical relevance, and theoretical relevance, using a qualitative evaluation (I totally disagree, I partially disagree, I’m indifferent, I partially agree, I totally agree). The instrument was also assessed in relation to its content using a pilot study conducted on 10 families who answered the instrument and made suggestions.20

Reliability was assessed by two coefficients: the Cronbach's Alpha measure and the Spearman–Brown coefficient. The Cronbach's Alpha measure was assessed with a 95% (acceptable values ≥0.7) confidence interval (CI) obtained by bootstrapping using the psych R package.21–23 The Spearman–Brown coefficient had acceptable values ≥0.3.24 The Spearman–Brown coefficient was analyzed using the split-halves method.24 For this method, we adopted the following strategies. First, the items were randomly divided into two halves. A mean of the scale was calculated for each half and then the two sets of means of the instrument were correlated to estimate the split-half correlation.25 The split-half correlation was adjusted using the Spearman–Brown formula to create split-half reliability. This procedure was repeated 1000 times and the mean of the split-half correlations was returned as the best single item reliability estimate.25 Meanwhile, the mean of the split-half reliabilities was returned as the best composite reliability estimate for all the items.25 The analyzes were performed on the item, version 1.1-1, and Multicon, version 1.6 packages of the R software.

Considering the total number of questions contained in the instrument (10 questions) and given that the area of psychometric suggests 10 participants per question, a sample of 100 participants would be sufficient for psychometric analysis.26

Regarding discrimination, items 3 (“When you’re feeling bad, sad, or when you have nothing to do, do you eat to feel better?”) and 8 (“Have you ever wanted to take medication to get thin?”) had the highest discrimination values. Items 4 (“When you do something nice, do you eat something as a prize?”) and 5 (“Have you ever hidden food somewhere to eat afterward?”) had the lowest discrimination values (Table 3). Items 3 and 8 had the steepest curves, and items 4 and 5 had the least steep curves (Fig. 2). Higher discrimination values suggest better psychometric performance.

Figure 2.

Item information curve.

1. Do you eat even if you are not hungry?

2. After you start eating, is it difficult to stop?

3. When you are feeling bad, sad or when you have nothing to do, do you eat to feel better?

4. When you do something nice, do you eat something as a prize?

5. Have you ever hidden food somewhere to eat afterwards?

6. Do you use medicine to get rid of food after you have eaten?

7. Do you exercise to get food after you have eaten?

8. Have you ever wanted to take a medicine to get thinner?

9. Are you afraid of gaining weight and/or getting fat?.

(0.18MB).

Concerning item severity, items 6 (“Do you use medication to get rid of food after you have eaten?”) and 7 (“Do you exercise to get food after you have eaten?”) were the ones that presented the highest values. Items 9 (“Are you afraid of gaining weight and/or getting fat?”) and 1 (“Do you eat even if you are not hungry?”) had the lowest severity (Table 3). In Fig. 2, we can see this same representation, items 6 and 7 were located more to the right, confirming greater item severity, and items 1 and 9 more to the left were less item severe in the continuous latent trait.