PHS Monday Seminar: Frank E Harrell Jr, PhD “Estimating Rather than Checking Predictor Transformations, and the Futility of Variable Selection”

Professor of Biostatistics Vanderbilt University School of Medicine Abstract I will explain why statistical inference is distorted by trying different transformations of predictor variables in regression, and show the value of flexibly estimating such transformations with honesty about the number of parameters involved, using regression splines.  Then I’ll address the impossible task of empirically selecting the “right” predictor variables for a model, showing examples of (1) failure in a very simple situation, (2) inability to estimate relative predictor importance with precision, and (3) failure of smart selection/shrinkage methods such as lasso to select the right variables with a decent probability. Bio Dr. Harrell’s bio can be found here.