Hardly ever, in my opinion. The purpose is to minimize the effects of violations of underlying assumptions, such as normality and homogeneity of variance, but procedures such as ANOVA and t-tests are pretty robust if only one of those assumptions is violated - and the violation needs to be really severe.
The other problem is that, once you've transformed the data, the practical interpretation of the results becomes difficult - explaining what a "log" or "natural log" is to a non-stats person is not going to go over well.
the only time i would do it is if i were trying to establish that the data are related by a log or power function. you can test this idea by tranforming the data to log and doing a linear regression if log regession is not availible. been a long time since i've done that though....
you might also do it to allow yourself the opportunity to plot a wide range of data on a single plot, John's caveat applies here too.