Logarithmic transformations in ANOVA.

A method is presented for choosing an additive constant c when transforming data x to y = log(x + c). The method preserves Type I error probability and power in ANOVA under the assumption that the x + c for some c are log-normally distributed. The method has advantages similar to those of rank transformations--namely, it is easy to use and is resistant to extreme observations. Since the special case c----infinity corresponds in ANOVA to y = x, the method is a useful generalization of least squares.