I am new to this form. I am going to run an OLS regression with dependent variable as a correlation coefficient (between -1 and 1). I am aware that this dependent variable may be problematic in the OLS regression. But what can I do in order to implement the regression? Thanks!

  • $\begingroup$ A correlation coefficient between which two variables? How do they enter into the regression equation? $\endgroup$
    – Herr K.
    Sep 27 '18 at 15:47
  • $\begingroup$ Thanks for your reply. I use the correlation coefficient between firm size and productivity as a measure of allocative efficiency of a industry s. Then I want to investigate the correlation of this dependent variable and other industry level variables. Is there any suggestion how should I deal with this dependent variables? $\endgroup$
    – gxxjjj
    Sep 27 '18 at 17:02
  • $\begingroup$ Could you clarify your goal here? "Implementing" a regression is usually not the goal, unless in coursework just for practice. You presumably want some result/information from the data and care about whether the result is valid if you use a given method (like OLS). The goal determines the method. $\endgroup$ Sep 28 '18 at 0:59

Welcome. Many things can be done. Some simple ones are:

Option 1: Do not care and just do OLS.

Option 2: Do some transformation such as $\ln (1+y) - \ln (1-y)$, logit of $\frac12 (y+1)$, to change the range from $(-1,1)$ to $(-\infty, \infty)$ and do OLS. I assumed that there are no $\pm 1$.

Option 3: Change your model. Let $y_{1,ij}$ and $y_{2,ij}$ be the firm-level ($j$=firm) variables from which you obtain the correlation coefficient $y_i$. As $y_i$ is also obtained by regressing $y_{1,ij}$ on $y_{2,ij}$ (up to a scaling factor), you might want to regress $y_{1,ij}$ on $y_{2,ij}$ including other industry-level variables $x_i$ as regressors. But in that case you need to think about which one of $y_{1,ij}$ and $y_{2,ij}$ to use as the dependent variable. If you are worried about endogeneity, you can also consider instrumental variable estimation, but that's a different story.

I would choose Option 1 unless many correlation coefficients in the sample are very close to $\pm 1$. Note that interpretation changes in Option 2.


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.