Dear all,

I am running a FMM using a GLM with poisson and log link. My dependent variable is log transformed: ln(Y+1).

However, when I transform back the predictions, only by reversing the transformation - "exp(yhat)-1", I have positive bias in the prediction for one of the classes (for the other one, the Y is basically zero, so the prediction is alright).

I tried already multiply (exp((`e(rmse)'^2)/2)) while back transforming as estimate of the error variance, but the result was basically the same I had before. I also tried regress the transformed back predicted values on the original Y variable, with no intercept, and then use the coefficient as adjustment term (coefficient*(exp(yhat)-1)), but again without success.

Any thoughts how could I fix this bias in my predicted values when I transform back?

Thanks.