Analysis of large and small samples of biochemical and clinical data

Statistical software often offers a list of various descriptive statistics of location and scale, but rarely selects an efficient estimate that is sta tistically adequate for an actual univariate sample. The sample interval es timate for a specified degree of uncertainty seems to be more meaningful if it covers an unknown value of the population parameter. The concept of an interval estimate in medicine is then used for medical decisionmaking. The proposed methodology, which uses the S-Plus algorithm for biochemical, biol ogical and clinical data analysis contains the following steps: (i) Explora tory data analysis identifies basic statistical features and patterns of th e data, the distributions of which are mostly non-normal, non-homogeneous a nd often corrupted by outliers. (ii) Sample assumptions about data, indepen dence of sample elements, normality and homogeneity are examined. (iii) Pow er transformation and the Box-Cox transformation to improve sample symmetry and stabilize the spread. (iv) Classical and robust statistics for both la rge (n>30) and medium-sized samples (15<n<30), point and interval estimates for the parameters of location, scale and shape. For an analysis of small samples (4<n<20) the Horn procedure of pivot measures is recommended. The p roposed methodology is demonstrated in two case studies, a large sample ana lysis of mean pregnenolone concentrations in the umbilical blood of newborn s, and a small sample analysis of mean haptoglobin concentrations in human serum.