Robustness in real life: a study of clinical laboratory data

Previous studies have evaluated the performance of robust estimators of location using computer-generated distributions. This paper describes how distributions commonly seen in biomedical research differ from the simulated distributions, and compares the performance of six families of estimators using samples from four biomedical distributions. The biomedical distributions can be asymmetric, and have shorter tails and other anomalies, such as digit preferences, which are not modeled in simulation studies. The four biomedical distributions studied differ markedly from the models previously studied, and these differences strongly affect the robust estimators. In such situations the trimmed mean provides an improvement over the sample mean, and its performance is more stable than that of estimators whose psi functions have steeply-descending tails. The central value and dispersion of each of the estimators are evaluated.