83 – Applications of Mordern Regression Methodologies to Health Policy Studies
Regression Analysis of Anthropometry Data: A Simulation Study of a Two-Stage Estimator
Kevin Konty
New York City Department of Public Health and Mental Hygiene
Stuart Sweeney
University of California at Santa Barbara
Regression analysis of anthropometry data has a long history in public health research. Early work relied on conditional mean regression models, but given that most policy interest is in either the lower or upper tail of a distribution, recent studies have utilized either binary outcome regression (logistic or ordinal logistic) or quantile regression. If the errors of the index function underlying binary models have non-constant variance, it is well-known that parameter estimates are inconsistent. We present simulation results of a proposed two-stage estimator to adjust for heteroskedasticity of unknown form. The two-stage estimator appears to substantially reduce bias in both parameter estimates and predictive changes in prevalence.
