Question

我正在尝试使用Jaggs到R的贝叶斯因子在线性回归模型之间做出决定。为了简单起见，我正在研究mtcars数据集，并且这些模型是：

（1）仅在截获时mpg （2）重量MPG （3）显示的mpg （4）mpg在wt和disp上

我使用的是先验先验，并尝试通过以下代码实现模型比较，但遇到一些问题和错误：

reg.bf.model <- "model{
    M ~ dcat(model.p[])
    model.p[1] <- 0.25
    model.p[2] <- 0.25
    model.p[3] <- 0.25
    model.p[4] <- 0.25

  # Likelihood
  for (i in 1:n){
      y[i] ~ dnorm(mu[i,M],tau[M])
        mu[i,1] <- alpha[1]
        mu[i,2] <- alpha[2] + beta1[1] * wt[i]
        mu[i,3] <- alpha[3] + beta1[2] * disp[i]
        mu[i,4] <- alpha[4] + beta1[3] * disp[i]+ beta2* wt[i]
  }

  for (m in 1:4){
      alpha[m] ~ dnorm(0,0.0001)
      sigma2[m] <- (1/tau[m])
  }

  for (j in 1:3) {
      beta1[j] ~ dnorm(0,0.0001)
  }

    beta2 ~ dnorm(0,0.0001)

    tau[1] ~ dgamma(prior.shape1[M],prior.rate1[M])
    tau[2] ~ dgamma(prior.shape2[M],prior.rate2[M])
    tau[3] ~ dgamma(prior.shape3[M],prior.rate3[M])
    tau[4] ~ dgamma(prior.shape4[M],prior.rate4[M])
}"

jags.data <- list(y=mtcars$mpg,n=nrow(mtcars), 
                  wt = mtcars$wt, disp = mtcars$disp,
                  prior.shape1=c(0.0001,1),prior.rate1=c(0.0001,1),
                  prior.shape2=c(1,0.0001),prior.rate2=c(1,0.0001),
                  prior.shape3=c(1,0.0001),prior.rate3=c(1,0.0001),
                  prior.shape4=c(1,0.0001),prior.rate4=c(1,0.0001))

jags.bf <- jags.model(file=textConnection(reg.bf.model),
                      #                      inits=jags.inits,
                      data=jags.data, n.chains=3)

update(jags.bf, 100)
jags.bf.out <- coda.samples(jags.bf,
                            variable.names=c("alpha","beta1","beta2"
                                             ,"sigma2","M"),
                            n.iter=50000, thin=400)

但是我收到以下错误消息：

“ update.jags（jags.bf，100）错误：逻辑错误： SimpleRange :: leftOffset。索引超出允许范围请将错误报告发送至martyn_plummer@users.sourceenter code hereforge.net“

和

“ jags.samples（模型，variable.names，n.iter，thin，类型=“ trace”中的错误，：未设置有效的监视器另外：警告消息： 1：在FUN（X [[i]]，...）中：无法设置监视器。没有模型！

2：在FUN（X [[i]]，...）中：无法设置监视器。没有模型！

3：在FUN（X [[i]]，...）中：无法设置监视器。没有模型！

4：在FUN（X [[i]]，...）中：无法设置监视器。没有模型！

5：在FUN（X [[i]]，...）中：无法设置监视器。没有模特！”

任何朝着正确方向提出的想法或想法都会受到赞赏！

Answer 1

作为更新，显然，需要使用贝叶斯因子作为两个模型之间的常客主义假设检验的替代方法。因此，下面的代码调整使事情起作用。

reg.bf.model <- "model{
  M ~ dcat(model.p[])
  model.p[1] <- prior1
  model.p[2] <- 1-prior1


  # Likelihood
  for (i in 1:n){
    y[i] ~ dnorm(mu[i,M],tau[M])
    mu[i,1] <- alpha[1]
    mu[i,2] <- alpha[2] + beta1[1] * wt[i]
  }

  for (m in 1:2){
      alpha[m] ~ dnorm(0,0.0001)
      sigma2[m] <- (1/tau[m])
  }

  for (j in 1:1) {
    beta1[j] ~ dnorm(0,0.0001)
  }

  beta2 ~ dnorm(0,0.0001)

  tau[1] ~ dgamma(prior.shape1[M],prior.rate1[M])
  tau[2] ~ dgamma(prior.shape2[M],prior.rate2[M])

}"
prior1 = 0.5
jags.data <- list(y=mtcars$mpg,n=nrow(mtcars), prior1 = prior1,
                  wt = mtcars$wt, disp = mtcars$disp,
                  prior.shape1=c(0.0001,1),prior.rate1=c(0.0001,1),
                  prior.shape2=c(1,0.0001),prior.rate2=c(1,0.0001))

jags.bf <- jags.model(file=textConnection(reg.bf.model),
                      #                      inits=jags.inits,
                      data=jags.data, n.chains=3)

update(jags.bf, 100)
jags.bf.out <- coda.samples(jags.bf,
                            variable.names=c("alpha","beta1","beta2"
                                             ,"sigma2","M"),
                            n.iter=50000, thin=400)

# calculating bayes factor
posterior.M <- unlist(jags.bf.out[,"M"])
posterior.odds <- mean(posterior.M==1)/mean(posterior.M==2)
prior.odds <- prior1/(1-prior1)
bayes.factor <- posterior.odds/prior.odds
bayes.factor

Jaggs在R中使用贝叶斯因子

1 个答案: