[R-sig-ME] Logarithmic Mixed Models?

Ben Bolker bbolker at gmail.com
Wed Apr 9 03:00:42 CEST 2014

AvianResearchDivision <segerfan83 at ...> writes:

> Hi all,
> I am working with ambient noise level as an environmental predictor
> variable and am wondering if using linear mixed models for this sort of
> data would be incorrect.  It seems since sound pressure level (dB) is on a
> log scale, that performing mixed models on a non-linear scale or
> transforming the response variable might be the correct way to proceed.
> Any thoughts on this?  Anyone have any experience?
> Best,
> Jacob

  I would say that there is no _a priori_ way to define the correct
scale on which to do a statistical analysis (this relates to a parallel
discussion on this list about the dependence of the presence, and
interpretation, of interactions on what scale you are thinking about).

You said that noise level is a predictor variable, but then you
talk about transforming the response variable, so I'm a bit confused.

* If your _response_ variable is on a scale that works well (responses
are linear functions of the predictors, 
distribution of residuals is Normal and heteroscedastic,
interactions might be minimized, variable is interpretable in some
sensible way) then you should just go for it.
* If you are worried about the scaling of a _predictor_ variable,
then you have even less to worry about since the distributional
issues are not important (the theory of linear models doesn't make
any assumptions about the distributions of predictor variables).

  Ben Bolker

More information about the R-sig-mixed-models mailing list