I wish to "predict" the number of bugs my company will cause using a linear regression. For any given piece of development, the prediction is not great ($r^2 = .2$). However, my real task is to predict, company-wide, how many bugs we'll produce in a week. My errors are roughly zero-mean, so in this aggregated data set my $r^2$ is closer to .8.

How should I report this? My coworkers (and admittedly myself) are not terribly familiar with statistics, but if I say something like "20% of the variance is 'explained' by the model" that makes it sound as if there is a long way to go. If 80% is explained, that seems quite good.

Am I unfairly increasing my perceived correctness if I report the higher $r^2$ from the aggregate model?

  • 2
    $\begingroup$ You might appreciate reading about the Ecological Fallacy. $\endgroup$
    – whuber
    Dec 19, 2012 at 16:30
  • 1
    $\begingroup$ @whuber: very cool, thanks! As seems to be the usual here, I get an answer to a question I didn't even ask! $\endgroup$
    – Xodarap
    Dec 19, 2012 at 16:47

1 Answer 1


As long as you make it clear what you are predicting (bugs in a given week rather than bugs in a given project), you should use the $R^2$ from the aggregate model.

It would be dishonest to present the model predicting the number of bugs in a given project but give the $R^2$ from the aggregate model, of course.


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.