Say I have a continuous variable that for some reason needs to be binned into a dichotomous variable. In such a case, I pick some cutoff point and encode everything above that cutoff as 1 and everything below as 0.

Now I run a Pearson's point biserial correlation with both versions of that variable. Generally I'll expect a fairly high correlation, but still below 1.

Why is this and how should I interpret it?


1 Answer 1


In effect you are asking why the chart below involves a correlation of about 0.83 rather than 1

The answer is that the points do not lie on a straight line, though high values of $x$ are still associated with high values of $y$

enter image description here

  • $\begingroup$ Would then it be such that the lower the standard deviation within each bin (and thus the more centralized the data), the more you could draw a straight line, and thus the higher the expected correlation? $\endgroup$
    – Josh
    Commented May 6, 2021 at 12:42
  • $\begingroup$ @Josh - if the standard deviations of the original data were much smaller within each group were much smaller than overall then the correlation would be higher: this would involve clustering near the top and bottom of the original data as opposed to being spread out $\endgroup$
    – Henry
    Commented May 6, 2021 at 14:40

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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