Refreg: An R Package for Estimating Conditional Reference Regions

Abstract

Multivariate reference regions (MVR) represent the extension of the reference interval concept to the multivariate setting. A reference interval is defined by two threshold points between which a high percentage of healthy subjects' results, usually 95%, are contained. Analogously, an MVR characterizes the values of several diagnostic tests most frequently found among non-diseased subjects by defining a convex hull containing 95% of the results. MVRs have great applicability when working with diseases that are diagnosed via more than one continuous test, e.g., diabetes or hypothyroidism. The present work introduces refreg, an R package for estimating conditional MVRs. The reference region is non-parametrically estimated using a multivariate kernel density estimator, and its shape allowed to change under the influence of covariates. The effects of covariates on the multivariate variable means, and on their variance-covariance matrix, are estimated by flexible additive predictors. Continuous covariate non-linear effects can be estimated by penalized spline smoothers. The package allows the user to propose, for instance, an age-specific diagnostic rule based on the joint distribution of two non-Gaussian, continuous test results. The usefulness of the refreg package in clinical practice is illustrated with a real case in diabetes research, with an age-specific reference region proposed for the joint interpretation of two glycemia markers (fasting plasma glucose and glycated hemoglobin). To show that the refreg package can also be used in other, and indeed very different fields, an example is provided for the joint prediction of two atmospheric pollutants (SO$_2$, and NO$_x$). Additionally, the text discusses how, conceptually, this method could be extended to more than two dimensions.