Study sample and data sources
This is a cross-sectional study with data collected from a previous study  comprised of parents and their adolescents enrolled in Florida’s Children’s Medical Services Network program (KidCare) in 2005. KidCare is a public insurance program that provides coverage for children who are uninsured under the age of 19 years and whose family has incomes up to 200 % of the federal poverty level. All adolescents in this sample were also enrolled in Medicaid. In Florida, Medicaid is a medical assistance program that is managed by the Agency for Health Care Administration to provide health care services to low-income individuals and families . University of Florida’s Institutional Review Board (IRB) approved the study protocol. Per University’s IRB, we obtained a waiver of collecting written informed consent by collecting verbal agreement from all study participants over the phone when we enrolled them.
We identified a statewide random sample comprised of 700 adolescents from the enrollment files maintained by the Florida Children’s Medical Services Network. The use of this sampling frame was on the basis of sample size needed (at least 230 dyads of adolescents and parents) for psychometric analyses in our previous study , which is also appropriate for the current study. A telephone survey using a structured questionnaire was conducted for families with an adolescent 15 through 18 years old living with them between 12/2005 and 03/2006. Multiple callbacks (at a maximum of 10 times) were performed if phone numbers were busy or not answered. Eleven percent of the families had disconnected phone numbers or did not answer the calls; 25 % of parents reported that their adolescents were physically or mentally unable to complete the survey; 6 % refused to allow their adolescents to be interviewed; and 4 % of the parents subsequently refused to participate after initially granting permission. As a result, the study sample consisted of 376 dyads of adolescents and their parents who completed the survey (survey response rate: 54 %). Thirty-seven dyads were excluded from the final statistical analyses since they had more than 50 % of items missing in HRQoL survey .
The adolescent’s weight and height were self-reported by the parent. BMI was calculated as the weight in kilograms divided by the height in meters squared. Age-and-sex growth charts developed by the U.S. Centers for Disease Control and Prevention  were used to categorize each adolescent into one of the following categories: obese (BMI ≥95th percentile), overweight (BMI ≥85th and <95th percentile), normal weight (BMI ≥5th and <85th percentile), and underweight (BMI <5th percentile). We excluded 16 adolescents with underweight since the mechanisms by which excess body weight (overweight and obese) and underweight that influence adolescents’ health and HRQoL might be different, leaving 323 dyads for the final analyses . People who are underweight may experience poor HRQoL. The possible mechanisms through which underweight affects HRQoL are malnutrition or poor health conditions (e.g., depression and cancer) .
PedsQL Core 4.0 for HRQoL measure
PedsQL Core 4.0 is a widely used validated generic instrument for pediatric HRQoL assessment [23–25]. We used both adolescent self-reports and parent proxy-reports to measure pediatric HRQoL. The PedsQL is comprised of 23 items covering four domains: physical (eight items) (e.g., “In the past month … It is hard for me to walk more than one block”), emotional (five items) (e.g., “In the past month … I feel afraid or scared”), social (five items) (e.g., In the past month … I have trouble getting along with other kids”), and school functioning (five items) (e.g., “In the past month … It is hard to pay attention in class”). A five-point response category for each item is utilized (from 0 = “never a problem” to 4 = “almost always a problem”). The specific domain score is calculated as the sum of the item responses divided by the number of items answered and scores are transformed which range from 0 to 100. The total HRQoL score is computed as the sum of all item responses divided by the number of items answered on all the domains. Higher item and domain scores indicate better HRQoL [23–25]. In our study sample, reliability coefficients (Cronbach’s alpha) of the adolescent self-reports were 0.82, 0.77, 0.76, 0.65, and 0.88 for the domains of physical, emotional, social, school, and total HRQoL, respectively. Reliability coefficients of the parent proxy-reports were 0.87, 0.79, 0.79, 0.73, and 0.91 for the domains of physical, emotional, social, school, and total HRQoL, respectively.
Intra-class correlation coefficient (ICC) was estimated to demonstrate the magnitude of the agreement between pediatric HRQoL rated by adolescents and parents. Given that the domain scores of HRQoL were not normally distributed, we conducted Wilcoxon signed rank tests to investigate the differences in pediatric HRQoL rated by adolescents and parents by individual BMI weight categories. The Cohen’s effect size was calculated to quantify the magnitude of the difference and a two-tailed p < 0.05 was deemed the statistical significance. Effect sizes of <0.2, 0.2–0.49, 0.5–0.79, and ≥0.8 indicate “negligible’, ‘small’, ‘medium’, and ‘large’ differences, respectively . We also performed a multiple regression analysis with robust standard errors to examine whether BMI weight categories in adolescents and other factors (such as the adolescent’s age, gender, and parent’s race and educational background) were associated with the differences in pediatric HRQoL rated by adolescents and parents. The selection of covariates was based on evidence from literature and results of our bivariate analyses [7–10]. In the regression analysis, normal weight was treated as a reference category compared to the overweight and obese categories.
We conducted Mann–Whitney U-tests to investigate the differences in pediatric HRQoL between different BMI weight categories. Mann–Whitney U-tests were conducted to compare self-reports (and parent proxy-reports) of pediatric HRQoL for obese adolescents versus normal weight adolescents, obese adolescents versus overweight adolescents, overweight adolescents versus normal weight adolescents, and normal weight adolescents versus excess body weight (defined as a combination of overweight and obese). The effect size was calculated to quantify the magnitude of the difference and a two-tailed p < 0.05 was used to determine the statistical significance.
DIF occurs when the individuals from subgroups (e.g., different BMI weight categories) rate an item unequally given the same underlying HRQoL (e.g., emotional functioning) the item intends to measure. Evidence of DIF in HRQoL items suggests the problematic construct validity of HRQoL measures because DIF implies misinterpreting the meaning of a HRQoL item between subgroups. In this study, we used MIMIC method to identify DIF associated with BMI weight categories in adolescents by incorporating additional background variables (e.g., the adolescent’s age, gender and parents’ race and educational background) into the analysis. The MIMIC model is a special case of structural equation model (SEM) and comprises two parts: a measurement model which defines the relations between a latent variable (a specific HRQoL domain) and its indicators (items measuring a specific HRQoL domain) and a structural model which specifies the relationships among latent variables and BMI weight status. Ideally, the relationships of BMI weight status with individual items of a specific HRQoL domain are explained through the relationship with a specific HRQoL domain. However, if the relationships of BMI weight status with individual HRQoL items exist, it will indicate presence of DIF. The technical merit of MIMIC methodology is the use of SEM framework to test a disparity in the magnitude of parameter for a group variable (e.g., over weight vs. normal weight) associated with a response to an item of emotional domain conditioning on the same underlying HRQoL (e.g., emotional functioning). In SEM, the underlying HRQoL (e.g., emotional functioning) is estimated through specific items (i.e., emotional items) that measure this specific HRQoL by incorporating measurement errors embedded in the items (i.e., variance of emotional items not estimated by the underlying emotional functioning). When the DIF issue was adjusted in the analysis, the comparison of HRQoL between different groups of individuals is regarded as unbiased and reflects the true difference.
In this study, serial tests of nested models, beginning with the most constrained model, sequentially relaxing cross-group equality constraints on the item parameters, and ending up with the least constrained model, were performed to detect DIF . The procedures are iterative and inclusive of the following steps:
Step 1: constraining the relationship between body weight (e.g., overweight and obese) and individual items of the PedsQL to be zero, and examining the modification indices to suggest how much the model fit would be improved if specific relationships were freely estimated, and
Step 2: starting with an item with the largest and significant modification index, and adding individual items of the PedsQL one at a time to the model for freely estimating its relationship with body weight (e.g., overweight and obese) until no modification indices were greater than 3.84 (d.f. = 1).
We performed the DIF analyses using Mplus 6.0 , and conducted the rest of the analyses using SAS 9.1 .