|
Abstract:
|
Studies on childhood overweight and obesity often use the body mass index (BMI) z-score as the outcome variable. The BMI z-scores are normalized on gender and age allowing for direct comparisons between different groups. Studies on childhood overweight and obesity typically restrict participation to those children whose BMI z-scores sit in the upper 15%. This inclusion restriction implies the distribution of the BMI z-scores of concern are right skewed. Often, the length of time of follow-up does not shift the distribution shape appreciably. The skewness of the data and the small samples sizes often employed call into question the use of traditional asymptotically normal models. This study uses simulated and real parallel design data sets to compare model fit between linear, log-normal, generalized skewed t-distribution, quantile regression, Pearson type IV, etc distribution models using mean square error and the coefficient of determination. In addition, empirical bias, power and type I error rate for each technique are computed. This study seeks to determine which of the methods are the most precise, interpretable, implementable, and meaningful from a research perspective.
|
