Improved approximate confidence intervals for the mean of a log-normal random variable

Data analysts often compute approximate 100(1 - alpha) per cent confidence intervals for the mean of a log-normal random variable due to the computational effort required for exact intervals. We evaluate two simple approximations and demonstrate that the probabilities with which the intervals fail to capture the population mean (that is, the coverage error) can range from well above the desired level, alpha, to very near zero in small to moderate sample sizes (n less than or equal to 100). The performance of a more sophisticated approximation, implemented via numerical integration or bootstrap sampling, is noticeably improved, but also suffers from coverage errors that are too large when n less than or equal to 25. A new procedure is developed which outperforms existing approximations. Computing these improved intervals requires the integration of standard distribution functions. The calculations are straightforward, however, and lead to satisfactory coverage errors for n as small as 5. A related method that avoids the integration step generally outperforms existing simple approximations for n less than or equal to 100, while maintaining the coverage error at or below alpha. Programs to implement the new procedures are provided in an Appendix. Copyright (C) 2002 John Wiley Sons, Ltd.