LAUSR

SELECTING PREDICTORS OF INTEREST BASED ON STATISTICAL SIGNIFICANCE Observational studies published in Cancer may seek to identify risk factors for an outcome of interest by using multivariable models or multiple univariable models if the sample size is limited. Among multiple candidate predictor variables, emphasis is commonly placed on those explanatory variables achieving statistical significance, which is generally defined as a P value below a threshold of .05 (P < .05); within this group, those with the smallest P values are often highlighted. This article illustrates ways in which this practice can be misleading and endorses confidence intervals (CIs), graphical approaches, and prespecified effects to facilitate a more complete understanding of findings. First, let us consider some recent history. The biomedical research community has voiced serious concerns about inappropriate use of P values and statistical significance to identify important research findings. The American Statistical Association codified some of these concerns in a policy statement aimed at improving the interpretation of statistical findings. The policy statement included 6 principles. Here, drawing on principles 3 and 5, I address problems that arise when predictors in a model are highlighted on the basis of statistical significance. Principles 3 and 5 are as follows: