1
$\begingroup$

I ran a model and got the following residuals: original residual plot

I proceeded to log the fitted values to get an idea of what's happening at the lower end of my predicted values: logged fitted values residual plot

It was then I saw on the left side the banded values. Their formation in stripes made me wonder what phenomenon could be producing this? Should I be adjusting my model somehow or just chalk this up to a quirk when logging fitted values?

These are from a negative binomial mixed effect model in R, if that helps.

Improve this question
$\endgroup$
4
  • 5
    $\begingroup$ The fact that you have count data in your response is what causes that. The leftmost "stripe" in the bottom plot will almost certainly be the 0's, the next one the 1's, and so on. This is addressed in several questions on site. I'll try and find you some. $\endgroup$
    – Glen_b
    Commented Dec 29, 2015 at 3:49
  • 1
    $\begingroup$ Here's one. A similar issue can be seen in multiple regression if you have a discrete response. $\endgroup$
    – Glen_b
    Commented Dec 29, 2015 at 5:10
  • 1
    $\begingroup$ The easy way into this is to note that as residual $=$ observed $-$ fitted, particular observed values (0, 1, ....) define lines with negative unit slope and differing intercept. Your log scale for fitted warps them into curves. $\endgroup$
    – Nick Cox
    Commented Dec 29, 2015 at 11:29
  • 1
    $\begingroup$ (but with modifications for using Pearson residuals rather than raw residuals). $\endgroup$
    – Nick Cox
    Commented Dec 29, 2015 at 12:31

1 Answer 1

Reset to default
0
$\begingroup$

Answered in comments:

The fact that you have count data in your response is what causes that. The leftmost "stripe" in the bottom plot will almost certainly be the 0's, the next one the 1's, and so on. This is addressed in several questions on site. I'll try and find you some.

– Glen_b

Here's Interpreting plot of residuals vs. fitted values from Poisson regression. A similar issue: Parallel straight lines on residual vs fitted plot can be seen in multiple regression if you have a discrete response.

The easy way into this is to note that as residual = observed − fitted, particular observed values (0, 1, ....) define lines with negative unit slope and differing intercept. Your log scale for fitted warps them into curves.

– Nick Cox

Improve this answer
$\endgroup$

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