The decov prior

Interfaces rstanarm
1

Operating System: RHEL7
Interface Version: 2.15.3
Compiler/Toolkit: gcc 4.6

Hi!

I am struggling with the decov prior for a random effects intercept + slope model. My situation is that I have multiple (exchangable) studies which do measure a control and trt group. I intend to fit a random effects model for this which I do with:

rate <- 1.5
stan_glmer(event ~ 1 + trt + (1+trt|study), offset=log(n*duration_y),
                   family=poisson, data=qaw_1to5l,
                   prior_intercept=normal(log(1/5), 2, autoscale=FALSE),
                   prior=normal(0, 1, autoscale=FALSE),
                   prior_covariance=decov(1,1,1,1/rate))

What is not clear to me from the documentation of decov is how I can specify different rates for the two random effects. I mean, it is possible to discriminate between the (fixed effect) prior on the intercept and on the treatment effect (which makes sense), but for a correlated intercept+slope random effect I do not see a possibility to put different priors on the between-study standard deviation $\tau$. The only possible way to do this is to split things apart like

stan_glmer(event ~ 1 + trt + (1|study) + (0+trt|study), offset=log(n*duration_y),
                   family=poisson, data=qaw_1to5l,
                   prior_intercept=normal(log(1/5), 2, autoscale=FALSE),
                   prior=normal(0, 1, autoscale=FALSE),
                   prior_covariance=decov(1,1,1,c(1/rate1, 1/rate2)))

However, this is not what I want as I would like to account for the correlations.

The only way I see out of this misery is to scale “trt” such that I would expect it to have the same prior as the intercept. That can be done, but seems not like a straighforward way.

I hope I missed something and this is easier to solve.

Thanks!

Best,
Sebastian

2

Strictly speaking, no, because the parameterization in rstanarm does not make the standard deviations primitive. Instead, the primitives are the proportions of variance. So, you can fiddle with the concentration argument to decov to change the Dirichlet prior on the variance proportions.