1
$\begingroup$

I fit a GEE model looking at the relationship between cognitive score and head impacts. Head impacts is a categorical variable separated into 4 quartiles by amount of exposure.

Here are the results:

enter image description here

So there does not seem to be an interaction effect overall between neck strength and head impact quartile (last row). However, the third quartile interaction is significant and fourth quartile interaction gets close (3rd and 2nd row from bottom respectively).

Given these results, is it appropriate to conclude that there could be an interaction in the higher head impact quartiles but not in the lower quartiles despite seeing no significant effect overall?

Improve this question
$\endgroup$
4
  • $\begingroup$ Did you make the quartiles or were they given to you? (That is, do you have the original continuous measure?) $\endgroup$
    – Peter Flom
    Commented Nov 11, 2018 at 11:06
  • $\begingroup$ @PeterFlom Yes I have the original continuous measure. However, given that the number of head impacts is self-reported, I think categorizing into quartiles makes sense and better accounts for the uncertainty. $\endgroup$
    – Hank Lin
    Commented Nov 11, 2018 at 13:06
  • $\begingroup$ I disagree. Quartiles increase uncertainty; if there is already uncertainty in the self report, there will more uncertainty once it is categorized, because it could easily go to the wrong category. $\endgroup$
    – Peter Flom
    Commented Nov 11, 2018 at 19:14
  • $\begingroup$ Hmm ok I see. Another reason for categorizing into quartiles is that the data is very right skewed. Either way though, if the goal is to detect a significant interaction effect between neck strength and # of head impacts, does it matter if the variable is continuous or categorical? $\endgroup$
    – Hank Lin
    Commented Nov 12, 2018 at 1:12

1 Answer 1

Reset to default
2
$\begingroup$

First, as Peter mentioned in the comments, it is not a good idea to categorize a continuous predictor. You lose information by doing so.

Second, regarding your specific question, given that the omnibus test for the interaction is not significant, you should not look at the p-values for the individual categories because these are not automatically corrected for multiple testing.

Improve this answer
$\endgroup$
3
  • $\begingroup$ So what if I did correct for multiple testing, and some of the individual categories were significant but the overall was still not? $\endgroup$
    – Hank Lin
    Commented Nov 13, 2018 at 23:49
  • 1
    $\begingroup$ Typically the omnibus test has a higher power because it more appropriately accounts for the relations between the different categories rather than the multiple-testing-corrected individual tests. $\endgroup$
    – Dimitris Rizopoulos
    Commented Nov 14, 2018 at 8:30
  • $\begingroup$ Ok thank you that makes sense. In regards to categorizing a continuous predictor though, does my argument for the data being strongly right skewed be a good enough reason? Or would transformation be a better option? $\endgroup$
    – Hank Lin
    Commented Nov 17, 2018 at 16:37

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.