Definition and elimination of outliers is a key element for medical laboratories establishing or verifying reference intervals (RIs). Especially as inclusion of just a few outlying observations may seriously affect the determination of the reference limits. Many methods have been developed for definition of outliers. Several of these methods are developed for the normal distribution and often data require transformation before outlier elimination.Methods:
We have developed a non-parametric transformation independent outlier definition. The new method relies on drawing reproducible histograms. This is done by using defined bin sizes above and below the median. The method is compared to the method recommended by CLSI/IFCC, which uses Box-Cox transformation (BCT) and Tukey's fences for outlier definition. The comparison is done on eight simulated distributions and an indirect clinical datasets.Results:
The comparison on simulated distributions shows that without outliers added the recommended method in general defines fewer outliers. However, when outliers are added on one side the proposed method often produces better results. With outliers on both sides the methods are equally good. Furthermore, it is found that the presence of outliers affects the BCT, and subsequently affects the determined limits of current recommended methods. This is especially seen in skewed distributions. The proposed outlier definition reproduced current RI limits on clinical data containing outliers.Conclusions:
We find our simple transformation independent outlier detection method as good as or better than the currently recommended methods.