1

I have two independent variables to predict my independent variable. For both theoretical and practical reasons, it made sense to log all of the three variables. In R formula:

log(y) ~ log(x1) + log(x2)

The model works great and meets all assumptions of linear regression with flying colors.

Now, I'd like to transform the fitted model including the fitted parameters to its 'unlogged' state, i.e., without the logs. I think this should be:

$$y = exp(\hat{\alpha}) \cdot exp(\hat{\beta_1})^{x_1} \cdot exp(\hat{\beta_2})^{x_2}$$

where each of the three parameters are the estimates from the regression, transformed exponential. Is this correct?

  • Seems correct, but keep in mind that the new model might (and probably is) not be optimal since you minimized mean square error in log space. Although I'm not sure how to optimize on multiplicative model. – Cherny Jun 24 '21 at 08:11
  • @Cherny In many circumstances this is not correct, because it ignores the error term. Exponentiating it produces a multiplicative error *whose expectation is greater than $1.$* For some discussion of this, see threads on [Duan's smearing procedure](https://stats.stackexchange.com/search?q=duan+smear*) and [back-transforming log regressions](https://stats.stackexchange.com/search?q=regression+log+back+transform*). – whuber Jun 24 '21 at 13:05

0 Answers0