Consider a linear regression

$$ y_i = A_i' \theta + B_i' \psi + \epsilon_i $$

There's a "trick" to find the parameters $\theta$ and $\psi$ in a stepwise fashion.

First minimizing squared error wrt. $\theta$: $$\hat{\theta}(\psi) = (A^t A)^{-1} A^t (y - B\psi) $$

Plugging back in, the residuals are $$ \epsilon = ( I - A(A^t A)^{-1} A^t ) (y - B \psi)$$ so the idea is that to find $\hat{\psi}$ by residualizing both $y$ and $B$ with respect to $A$.

Let $y_{\perp A} = (I - A(A^t A)^{-1} A^t) y $ and $B_{\perp A} = (I - A(A^t A)^{-1} A^t) B$. Then you just do $$ \hat{\psi} = (B_{\perp A}^t B_{\perp A})^{-1} B_{\perp A}^t y_{\perp A}$$

My questions are:

  • What is this called? I've tried searching for "stepwise regression", but that's definitely not the right term
  • Is it possible to extend this argument to a case when the dependence on $\psi$ is non-linear? I'm currently fitting the non-linear model to the residuals of the linear one, but I'm not sure if it's a valid thing to do

Many thanks


Your Answer

By clicking "Post Your Answer", you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.