Jeromy Anglim's Notes

I was fitting a random coefficient quadratic model in jags with all three regression coefficient parameters as varying. See the following extract of the JAGS/BUGS code:

for (i in 1:length(y)) {
    mu[i] <- beta1[subject[i]] + beta2[subject[i]] * trial[i] +
        beta3[subject[i]] * pow(trial[i], 2);
    y[i]  ~ dnorm(mu[i], tau.c)
}

# Random coefficients
for (i in 1:N) {    
    beta1[i] ~ dnorm(beta1.mu, beta1.tau);
    beta2[i] ~ dnorm(beta2.mu, beta2.tau);
    beta3[i] ~ dnorm(beta3.mu, beta3.tau);
}

I then ran the model

mod1 <- jags.model("model.txt", data=jagsdata, n.chains=4, n.adapt=1000)
update(mod1, 200) # burn in

# monitor
mod1.samples <- coda.samples(model=mod1,
                             variable.names=c('beta1.mu', 'beta1.sigma', 
                                              'beta2.mu', 'beta2.sigma',
                                              'beta3.mu', 'beta3.sigma',
                                              'sigma.c'),
                             n.iter=1000)
summary(mod1.samples) # print descriptive statistics of posterior densities for parameters

and received the warning:

Warning message:
glm.fit: algorithm did not converge 

Solution

The warning is discussed here. Martyn Plummer suggests that the warning can indcate poor mixing, and that one strategy is “to do quite a long run with a large thinning interval”.

When I increased the iterations and incorporated thinning, the warning went away:

mod1.samples <- coda.samples(model=mod1,
                             variable.names=c('beta1.mu', 'beta1.sigma', 
                                              'beta2.mu', 'beta2.sigma',
                                              'beta3.mu', 'beta3.sigma',
                                              'sigma.c'),
                             n.iter=10000, thin=10)