1
$\begingroup$

I am currently fitting a linear mixed-effects model to my data where the outcome variable can have both positive and negative values (integers).

The issue now is, that my outcome data is skewed with a heavier tail in the positive direction. As a consequence, my residuals (as plotted with a qqnorm plot) deviate from a qq line in both directions.

Usually, I would go ahead and do something like a log-transformation, but in this case this is not possible as I have both positive and negative data. Using a generalized linear mixed-effects model does not solve my problem as well due to similar reasons (negative data).

A further thing to note is that I have a lot of data (multiple millions of samples).

$\endgroup$
  • $\begingroup$ 1. If your data is ordinal, "negative" makes no sense. You can just add a constant because the numbers are meaningless. 2. Data cannot be skewed in both directions. 3. Log transformation of ordinal data makes no sense. $\endgroup$ – Peter Flom Oct 7 '15 at 21:26
  • $\begingroup$ 1.) You are right, it is not ordinal per-se; it is a simple integer variable with positive and negative values. 2.) Skewed positively $\endgroup$ – apehead Oct 7 '15 at 22:06
3
$\begingroup$
  • The most likely effect of heavy-tailed residuals will be to give you unbiased point estimates with overly optimistic (too-narrow) confidence intervals and (too-small) p-values. Depending on your goals, it might not matter that much if you have millions of data points ...
  • If you actually have ordinal data, you should consider fitting an ordinal model! e.g ordinal::clmm() in R.
  • The robustlmm package in R fits robust LMMs (see the vignette, i.e. vignette("rlmer",package="robustlmm") once you have downloaded/installed/loaded the packages).
| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ I have updated my question regarding that I am not actually have ordinal data. Regarding your first point: I am mostly interested in the coefficients and p-values. $\endgroup$ – apehead Oct 7 '15 at 22:07

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.