Note: This question is analagous to the question I asked here except instead of a removing column, I am adding it.

I am interested in a linear regression on the model;

$Y= X\beta + \epsilon$

And I have computed the OLS estimator $\beta$, $\hat{\beta}$;


I realize now though that my design is missing the first column $x_1$ I should actually be using matrix $\tilde{X}=[x_1, X]$.

Unfortunately though I no longer have access to my $Y$ data! Is there a way for me to update $\hat{\beta}$ to be based on $\tilde{X}$ rather than $X$ when I don't have access to $Y$ anymore? i.e. can I update the OLS solution when I add a column to the design matrix?

  • 1
    $\begingroup$ I assume you also don't have the residuals $\epsilon$ from your original model (which would allow you to recover the $Y$)? $\endgroup$ – Stephan Kolassa May 7 '19 at 13:22

I am afraid you are out of luck.

Consider the simplest possible case. Assume $X$ contains no predictors at all, only the intercept column. Then the initial model would contain just a single coefficient estimate $\hat{\beta}_0$, which would be estimated by the mean of $Y$. So your question in this case becomes:

Given the mean $\hat{\beta}_0$ of my data, but not the data $Y$ themselves, can I estimate a regression slope $\hat{\beta}_1$ for $Y$ on a predictor $x$?

And of course, it is obvious that this cannot be done. This slope $\beta_1$ crucially depends on the original data $Y$, and by changing $Y$ (without affecting the mean), we can change $\beta_1$ to be any value we choose.

The argument is analogous in higher dimensions.

| cite | improve this answer | |

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.