I am wondering if anyone can point me to a paper/lecture notes on the rationale behind first running an OLS on a set of variables, and then in a second regression using the residuals of that regression as the dependent variable to regress on several new (but related) independent variables. To specify, this is not aiming for an IV/2SLS approach - there's no instrument here that's uncorrelated with the dependent variable in the "first stage." Instead, it aims to kind of set a standard across the sample set with the first result, and then attribute the differences from the market-wide (ie the residuals) on the 2nd set of variables.

First off, wouldn't this approach by definition limit the first regression to one independent variable for a regular OLS? Otherwise the dependent variable in the 2nd would be an nx2 matrix...

Overall, is there any purpose to such a process? I believe I have seen it done before, but my searches have come up mostly fruitless. Closest conversation I've found is this Stata thread: http://www.stata.com/statalist/archive/2008-03/msg00264.html

  • 3
    $\begingroup$ are you talking about something along these lines? "In analysing multivariable datasets it is common that in looking at the effect of some variable ($x_1$) on a dependent variable of interest ($y$), the effects of a third continuous variable ($x_2$) are to be controlled for, for instance because its effects may confound those of $x_1$. In such circumstances it has become common to perform a regression of $y$ on $x_2$ and use the residuals from this regression in testing for the effects of $x_1$" found it here $\endgroup$
    – Robson
    Jun 30 '15 at 22:41
  • 1
    $\begingroup$ Yes, this should be helpful for background info - thank you. $\endgroup$
    – Z_D
    Jul 1 '15 at 15:34
  • 3
    $\begingroup$ Important to add here that the quote is only paraphrasing what other people say / do - the linked article explains why there is NO justification for this approach. $\endgroup$ Oct 27 '17 at 13:35

What you're describing is called 'two-stage residual inclusion' or 'control function' approach. In a linear model, it gives exactly the same estimator as "traditional" two-stage least squares, as for instance derived by Hansen, pp. 335-38. But in both cases, you need a valid IV. So your idea of doing this without an instrument in the first stage won't work.

  • $\begingroup$ I've tried both methods often enough, obtaining exactly the same results from each, to know that they do work. A reason for using the residual method is to allow for graphs that show a given X's relationship to Y after controlling for other predictors. Sometimes the partial plot one obtains via the more standard, condensed method isn't quite what's desired. $\endgroup$
    – rolando2
    Jul 2 '21 at 19:39

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.