Skip to content
Prev 16723 / 20628 Next

[FORGED] Re: Using variance components of lmer for ICC computation in reliability study

Dear Pierre,

Again, thanks for referring me to the paper (and thanks to Ben for the review article). It's becoming clearer now.

When I refer to residuals, as I understand from lmer modelling, it is the level 1 variance. Back to my design, I have many subjects, two fixed raters, sessions nested within a subject, and trials nested within sessions. Hence, level 1 variance would typically reflect inter-trial variance.

If I were to use lmer modelling, I do not have to specify the level 1 random effects (i.e. ~1| SUBJ:SESSION:TRIAL), as it is in the residuals. 

However, using mcmcglmm, there appears to be an even lower level of random effects which is fixed (as you mentioned). So my question is, should I specify also the level 1 random effects?

Sample formula using your suggestion of probit (I just like to know if the priors and model are appropriate):
mcglmm_mod = MCMCglmm(vas ~ RATER, 
                      random =  ~ SUBJ +
                        SUBJ:SESSION +
                        SUBJ:SESSION:TRIAL,
                      data = as.data.frame (df %>% filter (SIDE == "R")),
                      family = "ordinal",
                      prior=list(R=list(V=1, fix=1), G=list(G1=list(V=1, nu=0),
                                                            G2=list(V=1, nu=0),
                                                            G3=list(V=1, nu=0))))
Many thanks again for your (and everyone's) help thus far.

Kind regards,
Bernard
-----Original Message-----
From: pierre.de.villemereuil at mailoo.org <pierre.de.villemereuil at mailoo.org> 
Sent: Saturday, June 16, 2018 7:28 PM
To: Bernard Liew <B.Liew at bham.ac.uk>
Cc: r-sig-mixed-models at r-project.org
Subject: Re: [R-sig-ME] [FORGED] Re: Using variance components of lmer for ICC computation in reliability study

Dear Bernard,
Glad this was helpful.
You can find an explanation for this in the Supplementary File on our article on quantitative genetics interpretation of GLMMs:
http://www.genetics.org/content/204/3/1281.supplemental

It is fairly technical. Trying to put it simply: there is an equivalence between a binomial/probit model and the threshold model. This is because a probit link is actually the CDF of a normal distribution (same goes for a logit and a logistic distribution): using a probit link and a binomial is equivalent to using a threshold model after adding a random noise normally distributed (this is Fig. S1 in our Suppl. File above). 

And the variance of this random noise is thus the variance of a standard normal distribution which is 1 (or the variance of a standard logistic distribution, which is (pi^2)/3, for a logit link).

Computing the ICC on the liability scale "as if" a threshold model was used thus only requires to add that extra-variance on the denominator.
It all depends what you call "residual". In MCMCglmm, there is for example a "residual" variance called "units" which value should be fixed to e.g. 1 when running families such as ordinal. This variance is required to be added in the denominator of the ICC. If you have only a random effect, you would have:
Vrand / (Vrand + Vunits + 1)

Note that there is also the "threshold" family in MCMCglmm, which is special in the sense that (to say it quickly) the "units" variance is the variance of the probit link and should be fixed to 1. In that case, you would have:
Vrand / (Vrand + Vunits)

Sometimes this "extra-variance" (1 or (pi^2)/3) is indeed called "residual variance". This is because, as per my explanation above, probit and logit links can be seen as "adding a random noise then taking a threshold", this random noise is sometimes seen as a "residual error" in the model.

I hope this will clarify things for you. I'm afraid it's difficult to explain clearly the whys without going too much into the technicality of the models.

Cheers,
Pierre.