Suggested modification of presentation of stroke trial results

Abstract

The modified Rankin Scale (mRS) at 90 days after stroke onset has become the preferred outcome measure in acute stroke trials, including recent trials of interventional therapies. Reporting the range of modified Rankin Scale scores as a paired horizontal stacked bar graph (colloquially known as “Grotta bars”) has become the conventional method of visualizing modified Rankin Scale results. Grotta bars readily illustrate the levels of the ordinal modified Rankin Scale in which benefit may have occurred. However, complementing the available graphical information by including additional features to convey statistical significance may be advantageous. We propose a modification of the horizontal stacked bar graph with illustrative examples. In this suggested modification, the line joining the segments of the bar graph (e.g. modified Rankin Scale 1–2 in treatment arm to modified Rankin Scale 1–2 in control arm) is given a color and thickness based on the p-value of the result at that level (in this example, the p-value of modified Rankin Scale 0–1 vs. 2–6)—a thick green line for p-values <0.01, thin green for p-values of 0.01 to <0.05, gray for 0.05 to <0.10, thin red for 0.10 to <0.90, and thick red for p-values ≥0.90 or outcome favoring the control group. Illustrative examples from four recent trials (ESCAPE, SWIFT-PRIME, IST-3, ASTER) are shown to demonstrate the range of significant and non-significant effects that can be captured using this proposed method. By formalizing a display of outcomes which includes statistical tests of all possible dichotomizations of the Rankin scale, this approach also encourages pre-specification of such hypotheses. Prespecifying tests of all six dichotomizations of the Rankin scale provides all possible statistical information in an a priori fashion. Since the result of our proposed approach is six distinct dichotomized tests in addition to a primary test, e.g. of the ordinal Rankin shift, it may be prudent to account for multiplicity in testing by using dichotomized p-values only after adjustment, such as by the Bonferroni or Hochberg-Holm methods. Whether p-values are nominal or adjusted may be left to the discretion of the presenter as long as the presence or absence is clearly stated in the statistical methods. Our proposed modification results in a visually intuitive summary of both the size of the effect—represented by the matched bars and their connecting segments—as well as its statistical relevance.