|
Abstract:
|
Outliers in longitudinal studies involving growth data such as height, weight and BMI are quite common and can be either cross-sectional or longitudinal. Influential outliers, when not properly accounted for, can potentially bias parameter estimates, inflate standard errors and thus decrease the power to detect a significant effect. We investigate two alternative methods to detect outliers on statewide longitudinal childhood data on heights and weights collected on public school children. First, we use the Mahalanobis distance based on the scaled residuals from linear mixed models to identify individuals with outliers. We then explore an alternative modeling approach for outlier detection by allowing the residual ? to have a general heavy-tailed distribution. One suitable specification is the Student t-distribution that allows for more extreme values than a Gaussian error. The underlying logic is that the non-outlier observations can be modeled as having t-residuals with large degrees of freedom whereas outliers will have t-residual with small degrees of freedom depending on how extreme they are.
|
