Win statistics were two-sided. Models were stratified according to diabetes status (and according to trial in the pooled dataset). Adjustments were not made for multiple comparisons. The exact P values were 0.00001 in the pooled dataset and 0.0002 in the DELIVER dataset.
