The comparison of parametric and nonparametric bootstrap methods for reference interval computation in small sample size groups

Abstract

According to the IFCC, to determine the population-based reference interval (RI) of a test, 120 reference individuals are required. However, for some age groups such as newborns and preterm babies, it is difficult to obtain enough reference individuals. In this study, we consider both parametric and nonparametric bootstrap methods for estimating RIs and the associated confidence intervals (CIs) in small sample size groups. We used data from four different tests [glucose, creatinine, blood urea nitrogen (BUN), and triglycerides], each in 120 individuals, to calculate the RIs and the associated CIs using nonparametric and parametric approaches. Also for each test, we selected small groups (m = 20, 30,…, 120) from among the 120 individuals and applied parametric and nonparametric bootstrap methods. The glucose and creatinine data were normally distributed, and the parametric bootstrap method provided more precise RIs (i.e., the associated CIs were narrower). In contrast, the BUN and triglyceride data were not normally distributed, and the nonparametric bootstrap method provided better results. With the bootstrap methods, the RIs and CIs of small groups were similar to those of the 120 subjects required for the nonparametric method, with a slight loss of precision. For original data with normal or close to normal distribution, the parametric bootstrap approach should be used, instead of nonparametric methods. For original data that deviate significantly from a normal distribution, the nonparametric bootstrap should be applied. Using the bootstrap methods, fewer samples are required for computing RIs, with only a slightly increased uncertainty around the end points.