Accuracy of BMI correction using multiple reports in children

BACKGROUND: Errors in reported height and weight raise concerns about body mass index (BMI) and obesity estimates obtained from self or proxy reports. Researchers have corrected BMI using linear statistical models, primarily with adult samples. We compared the accuracy of BMI correction in children for models that included child or parent reports versus both reports, and models that separately predicted height and weight compared to a single model for BMI. METHODS: Height and weight from child reports, parent reports, and objective measurements for 475 children participating in the Military Teenagers’ Environment, Exercise and Nutrition Study were analyzed. Two approaches were evaluated: (1) separate linear correction models for height and weight versus (2) a single linear correction model for BMI. Each approach considered models for height, weight, or BMI with child reports, parent reports, or both reports, respectively, as predictors, stratified by gender. Prediction accuracy was computed using leave-one-out validation. Models were compared using root mean squared error for BMI, and sensitivity and specificity for overweight and obesity indicators. RESULTS: Models that included both reports provided the best fit relative to a model using either set of reports, with adjusted R(2) of height, weight, and BMI models ranging from 67.1 to 87.6 % in males, and 69.2 to 88.3 % in females. Estimates of BMI from separate models for height and weight had the least prediction error, relative to those derived from a single model for BMI or from uncorrected (child or parent) reports. Cross-validated Root Mean Squared Error (RMSEs) preferred a model that included only parent reports among males and females, compared to models with only child reports or both reports. When assessing sensitivity (true positive) for obesity and overweight/obesity, the results varied by gender and outcomes. Specificity (true negative) was similarly high for all models. CONCLUSION: Objective measurements are more accurate than self- or proxy-reports of BMI. In situations where objective measurement is infeasible, an approach that combines collecting a validation sub-sample including multiple reports of children’s height and weight, with estimation of BMI correction models maybe a cost-effective and practical solution. Correction models generate BMI estimates that are closer to objective measurements than reports.