The Importance of Accuracy

Abstract

With this issue of the Journal, we are pleased to introduce Statistically Speaking—a new department that provides a forum for the editor for statistics and evaluation and other invited authors to offer highlights and guidance from the world of statistics. We often take for granted the minor miracles that today’s computer power and statistical software perform for us in mere seconds—tasks that were not even possible 10 years ago. These advances, however, come with several little-understood consequences. I focus here on 3 considerations to help guide the ongoing content of this column. First, although we have software programs that will happily produce results once the button is pushed, we often don’t completely understand the applications, assumptions, and interpretations of the more advanced methods (e.g., hierarchical linear models, generalized linear mixed models, structural equation modeling, graphic information systems). Through a series of articles, special issues, and friendly ongoing advice in this column, we hope to make these methods both less mysterious and more aptly applied and interpreted. Second, because these “higher level” analytic methods are now so readily accessible, many of the appropriate simple analyses are often set aside, making the digestion of the content and meaning of many of our articles more difficult. Unfortunately, this tends to then limit the dissemination of potentially important public health findings. Although we hope to help make these higher-level methods interpretable by a greater number of readers, we also urge our authors, as space and appropriateness allow, to proceed analytically from the simple to the complex—but only to the degree of complexity necessary to answer the question at hand. Unnecessary complexity can be an obstacle to understanding. The best analysis is the simplest one that directly addresses the question of interest (think more Occam’s Razor, less Rubik’s Cube). Third, our reliance on point-and-click computer analyses often means that we take whatever is printed on the output as gospel and transfer it verbatim to the tables in our articles. One aspect of this practice has resulted in an administrative change in our “Instructions for Authors”: The Journal will now adopt a uniform practice for reporting results from regression analyses. Although several computer programs produce regression coefficients with the label “Beta,” it is incorrect to refer to these as such. On our output are estimates of the population parameter (β), not β itself, regardless of what the software labels say. In our text and tables, these should be termed “parameter estimates,” denoted by the roman letter b, or “standardized parameter estimates,” denoted by the roman letter B, as appropriate. Although these changes may seem minor, they are important distinctions and are essential to our understanding of statistical inference. Our goal is to maintain the highest standards in the quality of contributions to the Journal, and excellence starts with careful attention to important details. We begin with these. We hope you continue to support our ongoing efforts and tune back in to this column for more analytic updates and suggestions.