American Journal of Clinical Nutrition, Vol 51, 908S-916S, Copyright © 1990 by The American Society for Nutrition

Kernel regression for smoothing percentile curves: reference data for calf and subscapular skinfold thicknesses in Mexican Americans

¹ From the Division of Human Biology, Department of Pediatrics, Wright State University, Yellow Springs, OH and Ross Laboratories, Columbus, OH

A nonparametric kernel regression was used to smooth selected percentile curves for Mexican Americans from Hispanic HANES. Kernel regression can provide good approximations to irregular data when the parametric method is inappropriate due to restricted assumptions on the shape of the curves. The principle of kernel regression is weighted averaging of the observed values. The study sample was categorized into annual ages groups from 1 to 18 y and, for each age group, the 10th, 50th, and 90th percentiles of calf and subscapular skinfold thicknesses were generated for each sex. Each percentile curve comprised a specific percentile level for ages from 1 to 18 y. Kernel regression was applied to these curves using linear, cubic, and quintic polynomials as weight functions with smoothing parameters of 0.1, 0.3, 0.5, 0.7, and 1.0 y. After evaluating the goodness of fit from residual mean square errors, a smoothing parameter of 0.5 y, in combination with cubic polynomials, was chosen. The results from kernel estimation were compared with those from the restricted cubic splines model. It is concluded that the restricted cubic splines model is appropriate for fairly regular data, but for irregular data kernel estimation is a better choice.

Key Words: Kernel • smoothing • percentile curves