The most common reason is, I think, to make the variable normal. This is a mistake. First, even in OLS regression, the variables do not have to be normal, only the errors (estimated by the residuals) have to be normal. Second, it is better to use a method that fits the data then to make the data fit the method. read more

In terms of affecting the regression and in terms of affecting the variables. read more

Source: stats.stackexchange.com

If we use, say, log(unem) in a regression, where unem is the percentage of unemployed individuals, we must be very careful to distinguish between a percentage point change and a percentage change. read more

Source: stats.stackexchange.com

Wikipedia:

Econometrics and the Log-Log Model

www.dummies.com

Regression example: log transformation

people.duke.edu