0

Consider the Poisson regression model $$ Y=\exp\left( X_{1}\beta_{1}+X_{2}\beta_{2} +\varepsilon \right). $$ Is there something like the Frisch-Waugh-Lovell theorem for this model that would allow me to estimate $\beta_{2}$ by regressing $\exp (X_{2})$ on $\tilde{Y}$, where $\tilde{Y}$ are the residuals of a regression of $\exp (X_{1})$ on $Y$?

Aristide Herve
  • 103
  • 1
  • 6
  • 1
    This is not a correct expression of Poisson regression: it is a [nonlinear additive-error model.](https://stats.stackexchange.com/questions/148638/how-to-tell-the-difference-between-linear-and-non-linear-regression-models/148713#148713) The [correct expression](https://en.wikipedia.org/wiki/Poisson_regression) for Poisson regression is $$E[Y] = \exp(X_1\beta_1+X_2\beta_2).$$ The importance of this is in revealing that the "residuals" have no clear definition. – whuber Oct 09 '19 at 13:08
  • Thanks. I have corrected the expression for the Poisson regression. Please let me know if the question is still unclear. – Aristide Herve Oct 14 '19 at 14:22
  • You might find it instructive to contemplate what your re-expression of the model implies about the $\varepsilon$ terms. Regardless, since you now posit $Y$ is a function of $\exp(X_1),$ how could regressing $Y$ on $X_1$ possibly be relevant? – whuber Oct 14 '19 at 14:25
  • 2
    You're right, the questions was not well written. I hope it is clearer now. From the comments I understand that because of the non-linearity, there is probably no way we can run the regression in two steps as in the linear (OLS) case. – Aristide Herve Oct 15 '19 at 07:59

0 Answers0