Transformation bias in the costs and benefits of housing subsidies

The predictions in levels from non-linear models are biased whenever the regressand is transformed, i.e. E(Y λ )≠[E(Y)] λ . An approximation is provided for the log-transformation bias case, and it is demonstrated how this may be incorporated into the dummy variable modelling strategy for preparing predictions. The transformation bias that exists in predictions involving other non-linear regressand models cannot be handled analytically because a general solution for that bias does not exist. The bootstrap procedure provides non-parametric estimates of predictions and prediction standard errors that are combined with the dummy variable technique of preparing predictions.