In an earnings regression (log linear, i.e. coefficient is an approximate % change), with gender why doesn't the coefficient value change (except the sign which obviously reverses) when you alternate between male and female as the base reference category? Gender pay gap % diff should depend on base?