[R-sig-ME] Effect sizes for mixed-effects models

[Phillip Alday]

On 10/10/2019 23:41, Schäfer, L. (Lena) wrote:
> We are uncertain how to best estimate the dfs in our mixed-models. We considered using Kenward-Roger approximated dfs but this does not seem feasible since we only have access to parts of the raw data-sets used to derive dw and V{dw}.


You have 30-60 observations per participant, but how many participants?
If it's the typical 20+ in psychology, I would use the easiest
approximation of all for denominator degrees of freedom: treat them as
infinite, i.e. treat the t values as z values. The Kenward-Roger
approximation really doesn't really change your results for non trivial
datasets and the implementation in R (in pbkrtest, which lmerTest and
car::Anova() use internally) computes a matrix inverse for an n x n
matrix, where n is the total number of observations. This is
computationally painful for non trivial n with minimal benefit.

Best (from next door at the MPI),

Phillip