Skip to content

MCMCglmm interaction and posterior mode

4 messages · Kamal Atmeh, Walid Mawass

Original
Kamal Atmeh
# 
Dear all,

I have some questions, which may sound trivial, pertaining to 
interaction models with MCMCglmm.

I am running the following model with a gaussian distribution and a 
3-way interaction between two categorical two-level variables (tactic: 
F/H and period PB/B) and one continuous variable (env):

model <- MCMCglmm(lD ~ tactic*period*env
 ???????????????????????????????????? , random = 
~sp_phylo+species2+phylo_pop+phylo_popY+phylo_pop_id
 ???????????????????????????????????? , family = "gaussian"
 ???????????????????????????????????? , ginverse = list(sp_phylo = 
inv.phylo$Ainv) # include a custom matrix for argument phylo
 ???????????????????????????????????? , prior = prior1
 ???????????????????????????????????? , data = Data
 ???????????????????????????????????? , nitt = 22e+04
 ???????????????????????????????????? , burnin = 20000
 ???????????????????????????????????? , thin = 100
 ???????????????????????????????????? , pr=TRUE)

After looking at the results, I found that the 2-way interaction 
tactic*env from the tactic*period*env interaction was not significant, 
however the 3-way interaction itself was, with the following output in 
the summary:

 >>>?? tacticH:periodB:env????? 0.17831? 0.05360 0.30512???? 5000? 
0.0052 ** (the intercept represents tactic F and period PB)

I tried to run the model again in order to simplify it using ":" and 
therefore remove the non-significant 2-way interaction:

model2 <- MCMCglmm(lD ~ tactic*period + period*env + *tactic:period:env*
 ???????????????????????????????????? , random = 
~sp_phylo+species2+phylo_pop+phylo_popY+phylo_pop_id
 ???????????????????????????????????? , family = "gaussian"
 ???????????????????????????????????? , ginverse = list(sp_phylo = 
inv.phylo$Ainv) # include a custom matrix for argument phylo
 ???????????????????????????????????? , prior = prior1
 ???????????????????????????????????? , data = Data
 ???????????????????????????????????? , nitt = 22e+04
 ???????????????????????????????????? , burnin = 20000
 ???????????????????????????????????? , thin = 100
 ???????????????????????????????????? , pr=TRUE)

When using ":", the output of my model returns the posterior mean for 
each level of the categorical variables instead of one level as before:

tacticF:periodPB:env -0.1668620 -0.3554264? 0.0005143??? 195.0 0.0923 .
tacticF:periodB:env? -0.2018706 -0.3783204 -0.0174366??? 195.0 0.0410 *
tacticH:periodPB:env -0.1561097 -0.2066183 -0.1093840??? 118.2 <0.005 **

How should I define the interaction in the model in order to obtain an 
output similar to the one when the "*" interaction was used 
(tacticH:periodB:env) while simplifying and removing the non-significant 
interaction from the 3-way interaction?

Finally, is there a way to automatically compute the posterior mean of 
the continuous variable for each modality of the interaction?

Thank you and stay safe!

Kamal
Walid Mawass
# 
Hello Kamal,

One way to do this with MCMCglmm is to use the at.level() function. You can
determine for which level of your categorical variable you want the
interaction, ex. at.level(tactic,2):a.level(period,2):env.

However, in doing so you are only estimating the coefficient for that
specific interaction and ignoring the rest. That to me would seem a bit odd
given that for the 2-way interactions you are still including the other
levels, but of course it depends on what assumptions you are making for
your model.

and regarding automating the computation of the posterior mean, I think
(not sure though) that the broom package offers some wrap functions for
MCMCglmm and computing posterior estimates.

Good luck
Kamal Atmeh
# 
Hello Walid,

Thank you for your prompt response!

The at.level() function works perfectly and does exactly what I want. I 
have also modified the other 2-way interactions to only include the 
level of interest (H).

As for the broom package, even broom. mixed, it seems that it mostly 
gives a tidy version of the model's output. I haven't found a function 
that does the automatic computation though, but I will look more thoroughly.

Thanks again!

Cheers,

Kamal

Le 03/03/2021 ? 23:17, Walid Crampton-Mawass a ?crit?:

  
  


  


      

    

    

      
      

        


  

        

      Walid Mawass
    

    

      
      #
    

  


  

      
Another option would be the QGglmm package by Pierre de Villemereuil
(2016), it has multiple functions to extract estimates but I wouldn't say
it is automatic as much as it is faster computationally to compute
estimates.

Good luck