Statistics in review; Part 1: graphics, data summary and linear models

Abstract

Statistics and biomedical literature have historically had an uneasy alliance. A critical approach to the application of statistics is developed. Initially, we survey graphical data display and trace the historical development of the "testing" statistical paradigm, and the contributions of A R Fisher and J Neyman and E Pearson. The nuances of data summary and testing are illustrated by way of population versus sample estimation. The importance of the normality assumption is stressed, and the recurring contrast of parametric (t test) versus non-parametric (Mann-Whitney) approaches to summary statistics is discussed. The t test is found to be adequate. Effect measures are outlined, and we demonstrate the utility of the unpaired t test for binary data analysis. The theory of linear models is introduced, and the underlying assumptions of the standard ordinary least squares regression are presented. The implications of transformations, in particular log transformation, are detailed, and we conclude with an overview of the principles of model selection.