Chapter 5 Comparing Two Groups 163 theoretical results suggesting that when computing a two-sided confidence interval, a symmetric two-sided confidence interval should be used instead. A symmetric two-sided confidence interval can be obtained for the situation at hand by replacing T y given by Eq. (5.14) with ¯ ¯ | X t1 - X t2 | T y = d 1 + d 2 and letting T y represent the value of T y based on the bootstrap sample denoted by C i j . As before, repeatedly generate bootstrap samples yielding T y1 , . . . , T 1B . Now, however, set a = (1 - )B, rounding to the nearest integer, in which case the critical value is c = T y(a) , and the 1 - confidence interval is ¯ ¯ ¯ ¯ ( X t1 - X t1 - c d 1 + d 2 , X t1 - X t2 + c d 1 + d 2 ). (5.17) A variation of this approach was derived by Guo and Luh (2000), which is based in part on a transformation stemming from Hall (1992). The basic idea is to transform Yuen's test statistic so that it is approximated reasonably well by a Student's t-distribution. Results reported by Keselman, Othman, Wilcox, and Fradette (2004) indicate, however, that it is preferable to