Covariate-adjusted reference intervals for diagnostic data

The analysis of extreme diagnostic measurements in clinical trials relies on reference intervals that help drug developers quickly determine whether a particular value is typical or atypical. The distribution of diagnostic variables is often greatly influenced by various covariates and it is important to properly account for this influence in the analysis of extreme measurements. This paper discusses three approaches to constructing covariate-adjusted reference intervals for quantitative diagnostic data: global quantile smoothing, local quantile smoothing, and stepwise quantile approximations based on recursive partitioning. A detailed review of methods for optimizing the quantile estimation procedures is provided. The paper presents algorithms for selecting the degree of a polynomial approximation in global smoothing, bandwidth parameter in local smoothing, and number of strata in recursive partitioning. The described methods for computing covariate-adjusted reference intervals are applied to the analysis of electrocardiographic data.