I had a mixed effects model where the residuals were not normally distributed, so I log-transformed the DV to meet the assumption.
There is one fixed effect with 3 categorical levels (L1, L2, L3). Before the transform, L2 had a greater beta coefficient than L3. After the transform, I exponentiated the coefficients and subtracted 1 to interpret them as proportional changes. However, now L3 had a greater beta coefficient than L2.
How could this switch? Does it have to do with the fact the residuals were not normal in the first model? Or could it be collinearity in the levels of the predictor?