[R-sig-ME] effect sizes in lmer

[Henrik Singmann]

Do you want the effect size of single parameters? Then the friendly hint of David Winsemius (fixef or ranef) is the way to go (i.e., reporting the size of the parameters).

If you are interested in a value for the whole model, you could have a look at the different alternatives of R² discussed on the faq: http://glmm.wikidot.com/faq#mcmcsamp_status (scroll down a little).

I tend to like Omega²_0 which is given by the following for a model m:
1-var(residuals(m))/(var(model.response(model.frame(m)))

See: Xu, R. 2003. Measuring explained variation in linear mixed effects models. Statist. Med. 22:3527-3541. doi:10.1002/sim.1572

Hope that helps,
Henrik


Am 13.10.2013 00:03, schrieb Joshua Hartshorne:
> On Sat, Oct 12, 2013 at 5:27 PM, David Winsemius <dwinsemius at comcast.net>wrote:
>
>>
>>> Your suggestions would be greatly appreciated. I'm under re-review at a
>>> journal that requires effect size measures.
>>
>> The journal's purpose of requiring effect sizes is to prevent authors from
>> focussing solely getting p-values, and then failing to explain the
>> implications of model estimates. Don't you have access to the fixef and
>> ranef functions?
>>
>
> I doubt that's the purpose for this journal. In my field, researchers are
> pretty much only interested in interpreting the implications of the model
> estimates, and frankly I'm not sure what the alternative is. Quite frankly,
> most researchers in my field don't care about effect size one way or
> another, and those that do are mainly using it in conjunction with
> p-values, etc., to determine the likelihood the null hypothesis is true.
>
> I'm working with a 4x3x2 design. Let's call them factors A, B, and C. What
> I need to report is the size of the effect of A, B, C, and their
> interactions, which I won't get from fixef (so far as I know). I would just
> use an ANOVA but I have an unbalanced design.
>
> (Even if the estimates for the individual fixed effects in the regression
> were theoretically interesting -- they aren't -- large regression tables
> simply aren't published in my field, and it would be a long, probably
> unsuccessful slog trying to get one published.)
>
> 	[[alternative HTML version deleted]]
>

-- 
Dipl. Psych. Henrik Singmann
PhD Student
Albert-Ludwigs-Universität Freiburg, Germany
http://www.psychologie.uni-freiburg.de/Members/singmann