我使用R中的jags，stan，glm基本方法对库“ Stanarm”中的“ Roaches”数据集执行经典贝叶斯泊松回归，并获得相同的最终结果。为了获得更好的输出，我尝试使用R2jags并发生奇怪的事情：当我设置正常的先验适应阶段时（没有进度条带有交叉），并且每次链收敛到相同的错误值（如果我设置了例如柯西分布先验）或均匀分布先验（在R2jags中），我可以获得正确的值。
我尝试更改迭代次数，稀疏度，正常的先验精度，但什么也没做。
在完整的运行代码下面：
library(R2jags)
library(rstanarm)
data<-roaches
colnames(data)[1]<-"Y"
data.list <- as.list(data)
data.list$N<-nrow(data)
#normal priors
modelString_1="
model {
for (i in 1:N) {
Y[i] ~ dpois(lambda[i])
log(lambda[i]) <-offset[i]+beta_0+ beta_1*roach1[i]+beta_2*treatment[i]+beta_3*senior[i]
offset[i]<-log(exposure2[i])
}
beta_0~dnorm(0.0,tau)
beta_1~dnorm(0.0,tau)
beta_2~dnorm(0.0,tau)
beta_3~dnorm(0.0,tau)
tau<-1/100000
}
"
writeLines(modelString_1,con='classical_poisson_1.txt')
#cauchy priors
modelString_2="
model {
for (i in 1:N) {
Y[i] ~ dpois(lambda[i])
log(lambda[i]) <-offset[i]+beta_0+ beta_1*roach1[i]+beta_2*treatment[i]+beta_3*senior[i]
offset[i]<-log(exposure2[i])
}
beta_0~dt(0,0.00001,1)
beta_1~dt(0,0.00001,1)
beta_2~dt(0,0.00001,1)
beta_3~dt(0,0.00001,1)
}
"
writeLines(modelString_2,con='classical_poisson_2.txt')
params<-c("beta_0","beta_1","beta_2","beta_3")
Model_1<-jags(model.file='classical_poisson_1.txt', parameters.to.save=params,n.iter=20000,data=data.list,n.chains=3,inits = NULL)
#correct values from MCMC
Model_2<-jags(model.file='classical_poisson_2.txt', parameters.to.save=params,n.iter=20000,data=data.list,n.chains=3,inits = NULL)
R输出：
Compiling model graph
Resolving undeclared variables
Allocating nodes
Graph information:
Observed stochastic nodes: 262
Unobserved stochastic nodes: 4
Total graph size: 1877
Initializing model
|**************************************************| 100%
Inference for Bugs model at "classical_poisson_1.txt", fit using jags,
3 chains, each with 20000 iterations (first 10000 discarded), n.thin = 10
n.sims = 3000 iterations saved
mu.vect sd.vect 2.5% 25% 50% 75% 97.5% Rhat n.eff
beta_0 4.038 0.044 3.990 4.023 4.040 4.055 4.088 1.003 3000
beta_1 0.005 0.000 0.005 0.005 0.005 0.005 0.005 1.002 1100
beta_2 -0.740 0.028 -0.793 -0.758 -0.740 -0.722 -0.688 1.001 3000
beta_3 -0.047 0.038 -0.117 -0.072 -0.046 -0.022 0.025 1.001 3000
deviance 17417.012 275.140 16959.270 17261.288 17414.916 17578.130 17919.022 1.001 3000
For each parameter, n.eff is a crude measure of effective sample size,
and Rhat is the potential scale reduction factor (at convergence, Rhat=1).
DIC info (using the rule, pD = var(deviance)/2)
pD = 37874.0 and DIC = 55291.0
DIC is an estimate of expected predictive error (lower deviance is better).
Compiling model graph
Resolving undeclared variables
Allocating nodes
Graph information:
Observed stochastic nodes: 262
Unobserved stochastic nodes: 4
Total graph size: 1876
Initializing model
|++++++++++++++++++++++++++++++++++++++++++++++++++| 100%
|**************************************************| 100%
Inference for Bugs model at "classical_poisson_2.txt", fit using jags,
3 chains, each with 20000 iterations (first 10000 discarded), n.thin = 10
n.sims = 3000 iterations saved
mu.vect sd.vect 2.5% 25% 50% 75% 97.5% Rhat n.eff
beta_0 3.089 0.021 3.046 3.074 3.089 3.104 3.130 1.001 3000
beta_1 0.007 0.000 0.007 0.007 0.007 0.007 0.007 1.001 3000
beta_2 -0.516 0.025 -0.566 -0.533 -0.516 -0.500 -0.468 1.001 2200
beta_3 -0.379 0.034 -0.446 -0.402 -0.379 -0.356 -0.314 1.001 3000
deviance 12187.995 2.839 12184.462 12185.888 12187.338 12189.416 12195.288 1.001 3000
For each parameter, n.eff is a crude measure of effective sample size,
and Rhat is the potential scale reduction factor (at convergence, Rhat=1).
DIC info (using the rule, pD = var(deviance)/2)
pD = 4.0 and DIC = 12192.0
DIC is an estimate of expected predictive error (lower deviance is better).