[R-sig-ME] no-normal dependent variable

[Ben Bolker]

Iasonas Lamprianou <lamprianou at ...> writes:

> Hi all, a very quick question.


> In the context of mixed effects linear models, the theory tells us
> that we should not worry if the distribution of the dependent
> variable is not normal (e.g. if it is uniform, or skewed) as long as
> the residuals are normally and randomly distributed (and in a
> homoscedastic way): the estimates should be consistent, but the
> standard errors might be inflated. So, if a reviewer worries too
> much about the standard errors, is there a way to use lmer or lme to
> compute robust standard errors for my fixed and/or random effects?
> It seems that neither lme4 nor nlme currently support this, but
> there might be some other software out there to do the trick.

  I may be missing something here: if so, sorry.

  As I understand it the theory of mixed models doesn't say *anything*
about the *marginal* distribution of the response variable (I prefer
"predictor"/"response" to "independent"/"dependent"); it *only* refers
to the conditional distribution (= distribution of the residuals)
(and to the distribution of the random effects, but that's a tangent).

  But perhaps you are referring to the conditional distribution
being non-normal?

  You might try seeing whether the vcovHC function from the sandwich
package can be adapted, although I was unsuccessful at my first
attempt ...