Abstract 2374: A New More Sensitive Method to Assess Balance Among Stroke Trial Populations
Background: Stroke outcome is dependent on baseline factors such as NIHSS and age. Relationships between these variables and outcomes are often non-linear and imbalances can influence outcomes, particularly in subgroup analysis with smaller number of subjects. Balance in baseline variables factors are typically compared by Wilcoxon rank sum, t-test or ANOVA. Because of non-linearity, these tests may be insensitive to important differences in the distribution of these factors and if multiple factors are considered simultaneously. We adapted a multi-dimensional extension of Kolmogorov-Smirnov (KS) test proposed by Fasano and Franceschini (FF) to compare population distributions. The FF algorithm provides a method to calculate KS Distance (KSD) between two distributions in multiple dimensions and a probability value can be obtained. We hypothesized that the FF algorithm would be more sensitive than traditional statistical tests to determine whether baseline factors differed among two trial arms. We further show that matching for baseline variables (nearest neighbor Euclidean matching, pPAIRS©; Mandava Kent Stroke 2010) improves the KSD, indicating closer matched populations.
Methods: The NINDS database was used for this study (ntis.gov). The subgroup of rt-PA and placebo treated normoglycemic subjects with large artery stroke was analyzed. Median and mean NIHSS and age were compared and KSD and a p value were calculated using a custom program, pPOPULATION© written in Matlab®. rt-PA and placebo subjects were then matched using pPAIRS© and outliers eliminated. KSD and p value for the post-matched groups were calculated.
Results: The left half of the table shows the pre-match comparisons. Baseline variables were not different using usual tests. A KSD value of 0.283, however yielded p=0.008, suggesting that the population distributions are indeed different when two variables are considered simultaneously. Right half of the table shows the post-match comparisons of baseline variables. The KSD value, 0.217, is lower and is associated with a p value = 0.175, indicating that the post-matched distributions are similar.aWilcoxon Rank-Sum; bStudent t-test;
Conclusion: We demonstrate here a new application of a 2d version of the KS distance to verify the similarity of stroke populations and show that it is more sensitive than traditional difference testing. This finding is important because baseline imbalances are critical for accurate assessment of outcome. This algorithm can be further extended to additional dimensions (e.g.: glucose). Its relative advantages over other methods will be discussed.
- © 2012 by American Heart Association, Inc.