Family-centeredness of childhood obesity interventions: psychometrics & outcomes of the family-centered care assessment tool

At the end of the one-year intervention, we administered the mFCCA to parents. The mFCCA was developed with the input of the authors of the FCCA [9]. Five questions from the original scale were replaced with questions that were more applicable to children with obesity. The original authors provided alternate questions from their previously developed question bank that preserved the psychometrics of the scale. We included questions about addressing parental concerns, helping to find resources, identifying support systems, promoting mental health, and planning for changes in weight management or behavior change routines. These questions replaced items regarding disagreeing with healthcare professionals about recommendations, modifying daily care and treatment routines, ways to pay for things that insurance does not cover, understanding content in the medical record, and discussing hopes for the child’s future. Study investigators reviewed the final questions to ensure content validity for childhood obesity. The modified scale had 24 items that represented principles of family-centered care that resulted in eight different topical areas similar to the original scale. Responses to the scale were ordinal ranging from 1 to 5 with higher scores representing a greater perception of family-centeredness. A “not applicable” response was also provided.

Our first objective was to assess the psychometric properties of the mFCCA using Rasch modeling [12‐15]. The original scale was developed using item response theory, and those analyses were replicated in this study. Only children in the enhanced primary care arm were included in the psychometric analyses as this arm answered questions in relation to services from primary care providers (rather than health coaches) which is more representative of usual care. Because we made modifications to items, we performed an exploratory factor analysis using the principal axis method for factor extraction to confirm the uni-dimensionality of the scale rather than performing a confirmatory factor analysis. Prior to analysis, we determined that any item with a low factor loading < 0.4 would be deleted as was done in the original analyses [9]. We also examined the Scree plot and reviewed the tests of the hypothesis results. We calculated item total correlations to examine the homogeneity of the scale and determined any correlations < 0.3 would be deleted. We then used Rasch analysis using a partial credit model to assess the overall fit of the items [16‐18]. We calculated Chi-square based itemfit statistics that explain how well the data fit the model, including infit and outfit statistics that detected inliers and outliers, respectively. We set our criteria for item inclusion to be within the range of 0.5–1.5 [19, 20]. For each item, we also calculated standard error of the item score and item difficulty. Item difficulty was expressed using a logit scale (centered at 0) that ranges from negative (represents easy items) to positive (represents more difficult items) and measures how challenging it is to implement each item into clinical practice [13]. “Easy” items represent principles of family-centered care that would be easy to incorporate into care (i.e., taking time to address family’s concerns), whereas “difficult” items would be more challenging to incorporate into clinical practice. To test for potential question bias, we performed Differential Item Functioning (DIF) [21, 22], a statistical test that examines whether an item is measuring different abilities for subgroups, and tested for sex, income, race, and ethnicity. For the above analyses, “not applicable” responses were set to missing because they were considered structurally missing and mean imputation was used for other missing responses [23]. We deleted items that did not fall within acceptable predetermined ranges of the described analyses and the Rasch analysis was iterated until all items demonstrated a good fit. Once a final set of items was determined, we calculated person separation reliability to assess the internal consistency of the scale. This metric is equivalent to a Cronbach’s alpha [24, 25].

Our second aim was to determine the family-centeredness of the two intervention arms. First, we described overall participant characteristics according to intervention arm. For each participant, we derived a score by calculating the mean of the final items (items that were determined to show good model fit). Due to missing data, we calculated mean scores rather than a total score as was done in the original paper. We treated “not applicable” and missing responses in a similar manner as we did for the psychometric analyses described above. We used multiple linear regression adjusting for study site to examine the outcomes between the two intervention arms and determine the regression co-efficient and 95% confidence intervals. We set our alpha level at 0.05 to test for statistical significance. All analyses were performed using R version 3.4.4 and the eRM and lordif packages [16, 22, 26].

For the psychometric analysis, 316 responses (enhanced primary care arm only) were included. The results of the exploratory factor analysis revealed the presence of a single factor (eigen value = 11.24) which explained 47% of the variance (see Fig. 1). Individual item factor loadings were all > 0.4 and the item total correlations for the individual items were all > 0.3. Based on the factor loadings and item total correlations, no items were deleted at this stage of the analyses. Rasch item fit statistics revealed that 22 items fell between the range of 0.5–1.5. The two items (items #8 and #13) that were outside of the range were not deleted from the final scale because of their strong factor loadings (0.82 and 0.83, respectively) and strong item total correlations (0.75 and 0.76, respectively). The questions showed a broad range of item difficulty ranging from − 1.22 logits (easiest) to 1.10 logits (most difficult). The DIF analyses did not reveal question bias for sex, income, race, and ethnicity as the pseudo R² measures were < 0.02. In a no DIF condition, we would expect that pseudo R² measures to be < 0.02 [22]. Results of the psychometric analyses are shown in Table 2. Based on the psychometric analyses, we did not remove any items and the final 24 items represented eight core principles of the family-centered care. After finalizing the items of the mFCCA, we calculated the person separation reliability which was 0.99 revealing high internal consistency of the scale.

The mean score (SD) for the final 24 items of the mFCCA was 3.84 (0.95) for the enhanced primary care arm and 4.45 (0.61) for the enhanced primary care plus health coaching arm (Table 3). Figure 2 shows the mean responses for the final items on the mFCCA for both intervention arms. Compared to parents of children in the enhanced primary care arm, those whose children were in the enhanced primary care plus health coaching arm had higher mFCCA scores indicating greater perception of family-centeredness (β = 0.61 units [95% CI: 0.49, 0.73]).