Skip to content

P-values from interaction terms using lme4

1 message · Phillip Alday

#
Today is apparently the day for me to follow up on zombie threads:

While Alex is right that mixed-effects models are difficult with small
sample sizes, there seems to be an implicit assumption that a large
number of grouping levels and thus a large number of random intercepts
increases the number of model parameters. This isn't correct. The
variance of the intercepts is a model parameter, not the intercepts
themselves -- the individual intercepts are BLUPs / predictions /
conditional models and not estimates / model parameters. This is the
difference to including subject-id as a covariate in a classical
fixed-effects regression. Indeed, increasing the number of groups
generally helps in mixed-effects regression because variance estimates
are quite sensitive to small sample sizes. (@Ben Bolker and others if
I've messed this up, please correct me!)

Your sample size is still nonetheless quite small if you want to model
e.g. the variance for playback type (which I think could be important in
your model). Using rmANOVA isn't really better than a mixed-effects
model here, because rmANOVA *is* an intercepts-only mixed-effects model
with additional symmetry assumptions (sphericity)!

Compare the output of
Error: Subject
          Df Sum Sq Mean Sq F value Pr(>F)
Residuals 17 250618   14742

Error: Within
           Df Sum Sq Mean Sq F value Pr(>F)
Days        1 162703  162703   169.4 <2e-16 ***
Residuals 161 154634     960

and
Analysis of Variance Table
     Df Sum Sq Mean Sq F value
Days  1 162703  162703   169.4

You can get the residual Mean Sq F in the mixed model by looking at the
variance column of the Residual grouping under Random effects in
summary()-output.

The advantage to mixed-effects models in this case is that you're doing
the regression explicitly.

Best,
Phillip
On 06/13/2016 01:43 AM, Sam Hardman [sah74] wrote: