How do I apply a multiple comparison method to linear mixed effect model results?

I have a data file (available for download at http://www.bodowinter.com/tutorial/politeness_data.csv and based upon the research of Winder Grawunder, 2012) that gives conversation voice frequency (Hz) as a function of gender (M/F), attitude (formal/informal register), conversation scenario, e.g. asking a question vs. making a demand, etc., (broken into 7 scenario categories). The data was collected from six different individuals. For each individual, I can store the results in a table, T, having column headers attitude, gender, scenario, and frequency.

If I treat frequency as the numeric response, attitude and gender as fixed categorical effects, and scenario as a categorical random effect, a mixed linear mixed effect model corresponding to this data for each individual is given by

LME = fitlme(T,'frequency~attitude+gender+(1|scenario)');

The command anova(LME), gives me a p-value for each fixed effect, indicating its significance in determining the frequency. However, this p-value varies among the six individuals.

How do I combine all of the data across the six individuals to determine whether a fixed effect, such as gender, is significant in determining the pitch overall?

I realize that this is a multiple comparison problem. Can this be implemented using the multcompare command? If so, what steps are involved, starting from a single file that also accounts for the individuals? (It's stored in a table with column headers, gender, scenario, attitude, frequency, and individual.)

[yuval]

Shouldn't it be analyzed as a hierarchical linear model? As in scenarios nested in individuals? I wouldn't fit a different model for each individual, rather than add individuals as another level to the model. If you fit a model per person, doesn't gender just represents the intercept?

Answer 1

[dimitris]

Run an ANOVA on the LME object (the output of fitlme) eq, anova(LME) It will give you a summary with the F statistics and the corresponding p-values for each fixed factor you included in your GLM.

[Paul]

Running anova() on output from fitlme() does not produce acceptable output for multcompare().

[Pawel81]

Have you found a solution how to perform the post-hoc tests (alternative for multcompare() ) to the lme results?

Answer 2

[Sébastien Puma]

Hello,

I have the exact same question using fitglme function. glme = fitglme (data + pred 1 + pred 2) will provide an output that you can use as input for an Anova : stats = anova (glme).

Yet, multcompare is the only multiple comparisons function I found and does not accept anova' stats as an input.

Since the other anova functions are not valid methods for lme or glme tables, how is it possible to compare groups while using linear mixed effects models ?