Glht is the right one as far as I know. Have you seen the following web page?

http://thebiobucket.blogspot.com/2011/06/glmm-with-custom-multiple-comparisons.html

Thank you Mehmet and Niklas for your answers. I have one question regarding the same. If my response variable is non-normal, then is it alright to use Tukey for the comparisons?

As you're going to carry out multiple comparisons, there is an increased probability of Type I error. In these cases is a good practice to apply the Bonferroni correction or Holm–Bonferroni one.

Is your response variable continuous?

Niklas I couldn't find any functions in the nparcomp package which does multiple pairwise comparisons for regression models. Can you help regarding the same?

Mehmet my response variable is continuous. It is the duration in minutes required to perform a particular task. But I am asking for methods in general, and not only for the current dataset I am working with.

Yang and Niklas I am using Gamma and Lognormal distributions for fitting to the response variable (different datasets). It doesn't fit the normal distribution and is very right skewed.

So is the first option, "Changing the reference level using 'relevel' and then doing the GLMM again to see the test-statistic and p-values of the levels compared to the new reference level" not a good option?

Niklas Schandry, thanks for responding to my comment. Would it be better to change the levels multiple times or is that looked down upon?

The reason for changing the levels would be to compare the different levels of a predictor that is a categorical variable with multiple levels.

I thought the problem of resetting the levels is that that way we don't do a correction of the p values, as done in posthoc tests?

Okay, so if you run a model again after resetting the levels, you do a p-correction? And if you do a post-hoc comparsion you also need to do a p-correction?

Míša Hejná Niklas Schandry

The posthoc comparison should already correct for multiple comparisons (that's why it's so handy), whereas the level resetting in the models would then require some sort of correction. As far as I know. The emmeans package cited above is user friendly.

Míša Hejná and Niklas Schandry your guidance is really helpful! Am I also thinking along the correct lines when I think this -- I have four levels I am interested in and one run of the model compares 3 levels to the one. I'm then interested in comparing two of the remaining levels to each other. So I wouldn't need to use emmeans() to compare all the levels of the factor together (in fact I wouldn't want to) and instead just compare the two levels I'm interested in (also hopefully using emmeans).

I agree with Niklas. This is more efficient, and the code for these posthoc tests isn't difficult (I managed to do it! And I'm not that great with R).

I'm going to shamelessly "hack" the discussion here to ask a question which I've asked elsewhere (with little useful feedback) but which is very relevant to the discussion going on here: https://www.researchgate.net/post/What_are_the_differences_between_these_ways_to_carry_out_posthoc_tests_on_mixed_effects_models Emmeans isn't the only way to do posthoc tests with mixed effects models. So, there's difflsmeans, emmeans, and glht, but how should one choose which of these three to pick? I've noticed most people use emmeans, but I don't want to make my choice based on popularity. Would anyone know what to bear in mind when deciding between these three methods?

Thank you for this! You're the first person who's been able to tell me something useful about this, so it's much appreciated.

I was recommended lsmeans (the precursor for emmeans) as a general solution, too. Thanks for this useful discussion!

Hi Anna,

Here is one option (m6.glmer is my glmm object and Age is the variable for which I want the post-hoc tests):

library (multcomp)

summary(glht(m6.glmer, mcp(Age="Tukey")))

Best,

Martin

I agree with the solutions provided here. You could use emmeans (pretty much the latest version of lsmeans): https://cran.r-project.org/web/packages/emmeans/index.html

Or,

multcomp (https://cran.r-project.org/web/packages/multcomp/index.html )

etc.

Cheers!

The emmeans () function embedded in the emmeans R package will be helpful.

Thanks, Bernard!

emmeans, should work out fine. It is an updated package of lsmeans. Better still, you can use a package called interactions. It's pretty new and cool for such.

lsmeans