Graphics for Clinical Trials

Frank E Harrell Jr
Vanderbilt Department of Biostatistics
Office of Biostatistics FDA CDER

DIA/FDA Statistics Forum, North Bethesda MD, 2017-04-24

In the never-ending quest to replace tables with graphics, new graphics solutions to common data display problems in clinical trials are becoming available. This short course will focus on high-information graphics that faithfully convey characteristics of data and summary statistics using such tools as extended box plots, dot charts, and spike histograms. With the rapid evolution of HTML5 and html notebooks, new possibilities now exist, and graphics can be less cluttered with more information made available by merely hovering with the mouse or clicking the legend to activate the display of additional data layers.

Some of the learning objectives of this course are

  • learn principles of graph construction
  • learn which features of summary statistics should be emphasized in a graph
  • see examples of increasing information using modern graphics, whether static or interactive
  • understand features of html notebooks for statistical reports
  • learn how to use RStudio to make html notebooks
  • learn to use new functions in the R Hmisc package which use the R plotly package to produce somewhat interactive graphics
  • get ideas for constructing your own interactive graphics for statistical reports by seeing examples of placing supplemental information in initially hidden layers of graphics

Big Picture

Graphics are generally excellent for
  • providing a more interpretable snapshot of study results
  • displaying trends and finding patterns
  • speeding up reviews
  • displaying data that are too messy to model (e.g., sites within countries within regions)
  • displaying marginal (unadjusted) estimates
  • replacing tables having many numbers, especially when continuous variables are involved
  • showing whole distribution of continuous variables
  • finding data problems
  • finding problematic data distributions including excessive ties
  • displaying results of complex statistical models
Tables used to be good for
  • displaying finest levels of details
  • providing exact numeric values for elements of graphs
This can now be better done by drilling down on a graph.

In many cases, statistical models are the best descriptive statistics:
  • best way to handle confounding and outcome heterogeneity
  • accounts for more variables than any expert could understand in a graph
  • handles continuous variables optimally
  • handles censoring, incomplete data, large datasets, and heavy ties in data

Course Material


Topic revision: r13 - 12 Mar 2018, FrankHarrell

This site is powered by FoswikiCopyright © 2013-2017 by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
Ideas, requests, problems regarding Vanderbilt Biostatistics Wiki? Send feedback