A new paper out in Ecology by Xiao and colleagues (in press, here) compares the use of log-transformation to non-linear regression for analyzing power-laws.
They suggest that the error distribution should determine which method performs better. When your errors are additive, homoscedastic, and normally distributed, they propose using non-linear regression. When errors are multiplicative, heteroscedastic, and lognormally distributed, they suggest using linear regression on log-transformed data. The assumptions about these two methods are different, so cannot be correct for a single dataset.
They will provide their R code for their methods once they are up on Ecological Archives (they weren't up there by the time of this post).