MCMCglmm gives very different results from lme4 - how to diagnose issue?

Question

I am taking the plunge into Bayesian analysis for some new projects. I have some yes/no data, and three fixed effects and for the time being, I'm simply taking random intercepts (I will worry about random slopes later). I'm also sticking with the default prior settings, as I'll worry about that once I've settled on the library that I'm going to use, etc.

If I run a glmm using lme4, I get:

summary(glmer(gesture~vis*comm*task+(1|subject), dat, family="binomial"))
Generalized linear mixed model fit by maximum likelihood (Laplace
Approximation) [glmerMod]
Family: binomial  ( logit )
Formula: gesture ~ vis * comm * task + (1 | subject)
Data: dat

AIC      BIC   logLik deviance df.resid 
513.0    569.6   -243.5    487.0      562 

Scaled residuals: 
Min      1Q  Median      3Q     Max 
-3.9970 -0.4008 -0.0662  0.3445  3.9459 

Random effects:
Groups  Name        Variance Std.Dev.
subject (Intercept) 7.527    2.744   
Number of obs: 575, groups:  subject, 48

Fixed effects:
                        Estimate Std. Error z value Pr(>|z|)    
(Intercept)                 -3.41133    0.83651  -4.078 4.54e-05 ***
visvisible                  -0.03572    0.73199  -0.049   0.9611    
commtape                     0.87192    1.13579   0.768   0.4427    
taskact                      5.75031    0.91292   6.299 3.00e-10 ***
taskani                      3.27341    0.76841   4.260 2.04e-05 ***
visvisible:commtape          0.02226    1.03044   0.022   0.9828    
visvisible:taskact          -1.27673    1.00445  -1.271   0.2037    
visvisible:taskani           0.42952    0.95662   0.449   0.6534    
commtape:taskact            -1.91736    1.13234  -1.693   0.0904 .  
commtape:taskani            -0.86535    1.02607  -0.843   0.3990    
visvisible:commtape:taskact  1.13058    1.36033   0.831   0.4059    
visvisible:commtape:taskani -0.08530    1.32874  -0.064   0.9488    

So now i try using MCMCglmm:

m = MCMCglmm(
gesture ~ vis * comm * task, 
random=~subject, 
data=dat, 
family="categorical",
verbose=F)

And get the following output:

Iterations = 3001:12991
 Thinning interval  = 10
 Sample size  = 1000 

 DIC: 24.22676 

 G-structure:  ~subject

        post.mean l-95% CI u-95% CI eff.samp
subject     25172    487.6    53908    2.488

 R-structure:  ~units

      post.mean l-95% CI u-95% CI eff.samp
units      9486    295.7    17854    2.103

 Location effects: gesture ~ vis * comm * task 

                            post.mean  l-95% CI  u-95% CI eff.samp  pMCMC    
(Intercept)                 -173.2727 -334.1856  -30.9605    4.813 <0.001 ***
visvisible                     4.3220  -67.1333   82.1748   58.509  0.910    
commtape                      47.8088  -64.3913  194.5895   46.842  0.380    
taskact                      282.8265   49.6004  465.3049    2.952 <0.001 ***
taskani                      169.3790   23.3087  303.2590    6.197 <0.001 ***
visvisible:commtape           -7.5823 -130.7870   89.6066   54.957  0.878    
visvisible:taskact           -52.1963 -158.5566   67.1533   44.944  0.252    
visvisible:taskani            18.4001  -86.3001  136.4961   48.813  0.708    
commtape:taskact             -94.0063 -234.2362    8.6195   15.732  0.062 .  
commtape:taskani             -52.2950 -179.9766   37.5422   32.639  0.320    
visvisible:commtape:taskact   48.8714  -97.5667  216.0025   30.265  0.434    
visvisible:commtape:taskani    0.2211 -139.2106  157.8038   71.036  0.990

So the "good news" is that the pattern of statistical significance is the same under both methods. But I was under the impression that the advantage of moving to these new methods was to provide more robust estimations of effect sizes. And if we look at the estimates in both cases, there's a huge difference.

So, something is going wrong.

Questions:

• If I just looking at the MCMCglmm output, what should I be looking out for as a warning sign?

• More generally, is there a good source for information about this package and how to interpret the output?

• What should be my next step in my analysis? I tried increasing nitt, but that, if anything, makes the problem worse. I'm guessing this means that I'm getting nowhere close to convergence, and I need to increase the burn in parameter, and possibly increase thin?

Thank you for you help

Answer 1

For family="categorical" in MCMCglmm, the response variable must be a factor. You don't mention that you converted your response to a factor when switching from lme4 binomial model to MCMCglmm categorical model.