Bayesian Multinomial Regression using rjags package

Question

I am trying to fit a multinomial logistic regression model using rjags. The outcome is a categorical (nominal) variable (Outcome) with 3 levels, and the explanatory variables are Age (continuous) and Group (categorical with 3 levels). In doing so, I would like to obtain the Posterior means and 95% quantile-based regions for Age and Group.

I am not really great at for loop which I think is the reason why my written code for the model isn't working properly.

My beta priors follow a Normal distribution, βj ∼ Normal(0,100) for j ∈ {0, 1, 2}.

Reproducible R code

library(rjags)

set.seed(1)
data <- data.frame(Age = round(runif(119, min = 1, max = 18)),
                   Group = c(rep("pink", 20), rep("blue", 18), rep("yellow", 81)), 
                   Outcome = c(rep("A", 45), rep("B", 19), rep("C", 55)))

X <- as.matrix(data[,c("Age", "Group")]) 
J <- ncol(X)
N <- nrow(X)

## Step 1: Specify model
cat("
model {
for (i in 1:N){

    ##Sampling model
    yvec[i] ~ dmulti(p[i,1:J], 1)
    #yvec[i] ~ dcat(p[i, 1:J])  # alternative
    for (j in 1:J){
      log(q[i,j]) <- beta0 + beta1*X[i,1] + beta2*X[i,2] 
      p[i,j] <- q[i,j]/sum(q[i,1:J])  
    } 
    
    ##Priors
    beta0 ~ dnorm(0, 0.001)
    beta1 ~ dnorm(0, 0.001)
    beta2 ~ dnorm(0, 0.001)
}
}",
file="model.txt")

##Step 2: Specify data list 
dat.list <- list(yvec = data$Outcome, X=X, J=J, N=N) 

## Step 3: Compile and adapt model in JAGS 
jagsModel<-jags.model(file = "model.txt",
                      data = dat.list,
                      n.chains = 3,
                      n.adapt = 3000
)

Error message:

Sources I have been looking at for help:

http://people.bu.edu/dietze/Bayes2018/Lesson21_GLM.pdf

Dirichlet Multinomial model in JAGS with categorical X

Reference from http://www.stats.ox.ac.uk/~nicholls/MScMCMC15/jags_user_manual.pdf, page 31

I have just started to learn how to use the rjags package so any hint/explanation and link to relevant sources would be greatly appreciated!

Saprophyte

I think you just need to move the priors to the outside of the for loops. So just move one of the } from after the priors to before them. – Blanchard

Thank you that's a good spot! I tried your solution but I am now running into another error message. "Unable to resolve the following parameters: X[1,2] (line 7) X[2,2] (line 7) X[3,2] (line 7)...." – Saprophyte

Answer 1

I will include an approach to your issue. I have taken the same priors you defined for coefficients. I only need to mention that as you have a factor in Group I will use one of its levels as reference (in this case pink) so its effect will be taken into account by the constant in the model. Next the code:

library(rjags)
#Data
set.seed(1)
data <- data.frame(Age = round(runif(119, min = 1, max = 18)),
                   Group = c(rep("pink", 20), rep("blue", 18), rep("yellow", 81)), 
                   Outcome = c(rep("A", 45), rep("B", 19), rep("C", 55)))

#Input Values we will avoid pink because it is used as reference level
#so constant absorbs the effect of that level
r1 <- as.numeric(data$Group=='pink')
r2 <- as.numeric(data$Group=='blue')
r3 <- as.numeric(data$Group=='yellow')
age <- data$Age
#Output 2 and 3
o1 <- as.numeric(data$Outcome=='A')
o2 <- as.numeric(data$Outcome=='B')
o3 <- as.numeric(data$Outcome=='C')
#Dim, all have the same length
N <- length(r2)

## Step 1: Specify model

model.string <- "
model{
for (i in 1:N){ 

## outcome levels B, C
o1[i] ~ dbern(pi1[i])
o2[i] ~ dbern(pi2[i]) 
o3[i] ~ dbern(pi3[i]) 

## predictors
logit(pi1[i]) <- b1+b2*age[i]+b3*r2[i]+b4*r3[i]
logit(pi2[i]) <- b1+b2*age[i]+b3*r2[i]+b4*r3[i]
logit(pi3[i]) <- b1+b2*age[i]+b3*r2[i]+b4*r3[i]

} 
## priors
b1 ~ dnorm(0, 0.001)
b2 ~ dnorm(0, 0.001)
b3 ~ dnorm(0, 0.001)
b4 ~ dnorm(0, 0.001)
}
"
#Model
model.spec<-textConnection(model.string)

## fit model w JAGS
jags <- jags.model(model.spec,
                   data = list('r2'=r2,'r3'=r3,
                               'o1'=o1,'o2'=o2,'o3'=o3,
                               'age'=age,'N'=N),
                   n.chains=3,
                   n.adapt=3000)

#Update the model
#Update
update(jags, n.iter=1000,progress.bar = 'none')
#Sampling
results <- coda.samples(jags,variable.names=c("b1","b2","b3","b4"),n.iter=1000,
                        progress.bar = 'none')
#Results
Res <- do.call(rbind.data.frame, results)

With the results of chains for parameters saved in Res, you can compute posterior media and credible intervals using next code:

#Posterior means
apply(Res,2,mean)

         b1          b2          b3          b4 
-0.79447801  0.00168827  0.07240954  0.08650250

#Lower CI limit
apply(Res,2,quantile,prob=0.05)

         b1          b2          b3          b4 
-1.45918662 -0.03960765 -0.61027923 -0.42674155

#Upper CI limit
apply(Res,2,quantile,prob=0.95)

         b1          b2          b3          b4 
-0.13005617  0.04013478  0.72852243  0.61216838 

The b parameters belong to the each of the variables considered (age and the levels of Group). Final values could change because of the mixed chains!

Shannashannah

Thank you so much for the clear explanation of each line of code and clarification for categorical explanatory variable. This is very helpful and I understand all the bits so thank you again! – Saprophyte