I have regression data in which both x and y variables require a log transformation, ie, the appropriate equation is ln(y) = a + b ln(x) + e Predictions of y from this equation will be biased because of the transformation. In fact it will provide predictions of the geometric mean of y for a given x. My inclination is to add half the mean square error to the intercept to correct for the bias. Is this a valid procedure? Is there any other approach that will give unbiased predictions in this situation?