95) also satisfies the condition 1  b a > 1  b, there is no obvious reason for selecting them. Moreover, some of the sample sizes presented in Table 2 by Flahault et al. [1] are incorrect. For example, when P 0 Z 0.99 and P L Z 0.85, 0.86, and 0.87, n 1 Z 30, 32, and 35, and n 2 Z 40, 43, and 47, respectively, which are not equal to 50 as presented in the table. Moreover, when P 0 Z 0.98 and P L Z 0.85, n 1 Z 40 and n 2 Z 50. Thus the value presented in the table was n 2 , not n 1 for this example. Finally, it is also important to note that n 1 Z n 2 in some cases. For example, when P 0 Z 0.98, P L Z 0.87, a Z 0.05, and 1  b Z 0.95, n 1 Z n 2 Z 58 as presented in Table 2 by Flahault et al. Because the actual power never goes below the required power for any n larger than n 2 , and n 2 is always equal to or larger than n 1 , n 2 is more conservative than n 1 . In some instances, the difference between n 2 and n 1 is substantial. For instance, n 2 can be as much as 34% higher than n 1 (e.g., 47 is 34.2% higher than 35). Therefore, we caution readers when using the tabulated sample sizes given in Flahault et al. [1] for designing a study on diagnostic test assessment. Instead, we recommend using the conservative n 2 for the exact sample size calculations when designing studies on diagnostic test assessments, especially when the expected sensitivity and specificity are high. Haitao Chu Stephen R. Cole Department of Epidemiology The Johns Hopkins Bloomberg School of Public Health 615 N. Wolfe Street Baltimore, MD 21205, USA E-mail address: hchu@jhsph.edu (H. Chu) References [1] Flahault A, Cadilhac M, Thomas G. Sample size calculation should be performed for design accuracy in diagnostic test studies. J Clin Epide- miol 2005;58:859e62. [2] Chernick MR, Liu CY. The saw-toothed behavior of power versus sample size and software solutions: single binomial proportion using exact methods. Am Stat 2002;56:149e55. Appendix SAS code to carry out power calculation using exact methods proc power; ods output plotcontent=PlotData; onesamplefreq test=exact sides = 1 alpha = 0.05 nullproportion = 0.85 proportion = 0.95 ntotal = 93 102 power = .; plot x=n min=80 max=120 step=1 yopts=(ref=.95) xopts=(ref=93 102); run; doi: 10.1016/j.jclinepi.2006.09.015 Author reply: sample size calculation using exact methods in diagnostic test studies We thank Haitao Chu and Stephen Cole for pointing out that in our article we have indeed overlooked the nonmono- tonic character of power as a function of sample size due to discreteness of the binomial distribution. The tables thus give the smallest sample size meeting the specified require- ments, but larger sample sizes may indeed violate these requirements, as shown by H. Chu and S. Cole. Antoine Flahault* Michel Cadilhac Guy Thomas Inserm; UPMC; UMR-S 707 27 rue Chaligny F-75571 Paris cedex 12 France E-mail address: flahault.a@wanadoo.fr (A. Flahault) * Corresponding author: Tel.: þ336 0766 6959, fax: þ331 4473 8454. doi: 10.1016/j.jclinepi.2007.07.001 1202 Letter / Journal of Clinical Epidemiology 60 (2007) 1201e1202