使用rjags

时间:2017-08-23 15:26:14

标签: r bayesian bayesian-networks rjags

我正在努力使用rjags定义条件线性高斯贝叶斯网络。 (clg BN由连续子节点(结果)定义,具有连续正常和离散父(预测变量))

对于下面的网,A是离散的,D和E是连续的:

enter image description here

对于rjags模型,我假设我想要的是在值节点E上定义的节点A的参数:伪代码

model {  
  A ~ dcat(c(0.0948, 0.9052 ))
  D ~ dnorm(11.87054, 1/1.503111^2)

  if A==a then E ~ dnorm(6.558366 + 1.180965*D, 1/2.960002^2) 
  if A==b then E ~ dnorm(3.370021 + 1.532289*D, 1/6.554402^2)   
}

我可以通过使用下面的代码获得一些工作,但是它会更快地与更多的预测器和分类级别混淆。

library(rjags)

model <- textConnection("model {  
  A ~ dcat(c(0.0948, 0.9052 ))
  D ~ dnorm(11.87054, 1/1.503111^2)

  int = 6.558366 - (A==2)*(6.558366 - 3.370021) 
  slope = 1.180965 - (A==2)*(1.180965 - 1.532289)
  sig = 2.960002 - (A==2)*(2.960002 - 6.554402)

  E ~ dnorm(int + slope*D, 1/sig^2) 
}")

jg <- jags.model(model, n.adapt = 1000
  

我的问题:我如何简洁地定义这个模型?

数据来自

library(bnlearn)
net = model2network("[A][D][E|A:D]")
ft = bn.fit(net, clgaussian.test[c("A", "D", "E")])

coef(ft) 
structure(list(A = structure(c(0.0948, 0.9052), class = "table", .Dim = 2L, .Dimnames = list(
    c("a", "b"))), D = structure(11.8705363469396, .Names = "(Intercept)"), 
    E = structure(c(6.55836552742708, 1.18096500477159, 3.37002124328838, 
    1.53228891423418), .Dim = c(2L, 2L), .Dimnames = list(c("(Intercept)", 
    "D"), c("0", "1")))), .Names = c("A", "D", "E"))

sigma(ft)
structure(list(A = NA, D = 1.50311121682603, E = structure(c(2.96000206596326, 
6.55440224877698), .Names = c("0", "1"))), .Names = c("A", "D", 
"E"))

1 个答案:

答案 0 :(得分:2)

您只需使用变量A作为索引参数:

library('rjags')

model <- "
model {  
  A ~ dcat(c(0.0948, 0.9052 ))
  D ~ dnorm(11.87054, 1/1.503111^2)

  ints <- c(6.558366, 3.370021)
  int <- ints[A]
  slopes <- c(1.180965, 1.532289)
  slope <- slopes[A]
  sigs <- c(2.960002, 6.554402)
  sig <- sigs[A]

  E ~ dnorm(int + slope*D, 1/sig^2) 
}
"

jg <- jags.model(textConnection(model), n.adapt = 1000)

顺便说一下,由于模型中有很多固定数量,在R中定义这些数据然后将它们作为数据传递给JAGS可能更有意义。这样,您可以调整向量的值和长度(只要catprobs,int,slope和sig的长度匹配),而无需修改JAGS代码。例如(为了方便起见使用runjags,尽管也可以使用jags):

library("runjags")

model <- "
model {  
  A ~ dcat(catprobs)
  D ~ dnorm(Dmu, Dprec)

  int <- ints[A]
  slope <- slopes[A]
  sig <- sigs[A]

  E ~ dnorm(int + slope*D, 1/sig^2) 

  #data# catprobs, Dmu, Dprec, ints, slopes, sigs
  #monitor# A, D, E
}
"

catprobs <- c(0.0948, 0.9052)
Dmu <- 11.87054
Dprec <- 1/1.503111^2
ints <- c(6.558366, 3.370021)
slopes <- c(1.180965, 1.532289)
sigs <- c(2.960002, 6.554402)

results <- run.jags(model)
results

马特