我有一个混合效果模型，我在RMC的包MCMCglmm中运行。我有兴趣探索几种可能的预测器。 如何进行模型简化以从模型中排除无信息预测变量？
在其他情况下，我使用逐步向后回归，即以＆＃34; full＆＃34;包含我感兴趣的所有预测变量的模型，并使用drop1()查找无信息变量，然后update()将其从模型中删除。
这些功能似乎都不适用于MCMCglmm对象。什么是我最好的选择？我想drop1()不起作用的一个原因是MCMCglmm对象没有AIC值。 DIC值是否可以以类似的方式使用？我不太了解DIC值，但在下面的示例中，它们似乎不像AIC那样：从模型中删除变量后，DIC减少了100以上。
对于update()，显然可以从头开始再次指定模型，但它涉及复制/粘贴相当长的函数调用，并且对于包含交互的较长模型公式会变得更加复杂。我想知道是否有更简单的替代方案。
我还想知道是否有一个原因是没有为这些函数编写的方法来处理MCMCglmm对象 - 我试图以某种方式做什么无效？
使用部分发明数据的可重复示例：
set.seed(1234)
library(MCMCglmm)
data(bird.families)
n <- Ntip(bird.families)
# Create some dummy variables
d <- data.frame(taxon = bird.families$tip.label,
X1 = rnorm(n),
X2 = rnorm(n),
X3 = sample(c("A", "B", "C"), n, replace = T),
X4 = sample(c("A", "B", "C"), n, replace = T))
# Simulate a phenotype composed of phylogenetic, fixed and residual effects
d$phenotype <- rbv(bird.families, 1, nodes="TIPS") +
d$X1*0.7 +
ifelse(d$X3 == "B", 0.5, 0) +
ifelse(d$X3 == "C", 0.8, 0) +
rnorm(n, 0, 1)
# Inverse matrix of shared phyloegnetic history
Ainv <- inverseA(bird.families)$Ainv
# Set priors
prior <- list(R = list(V = 1, nu = 0.002),
G = list(G1 = list(V = 1, nu = 0.002)))
# Run model
model <- MCMCglmm(phenotype ~ X1 + X2 + X3 + X4,
random = ~taxon,
ginverse = list(taxon=Ainv),
data = d,
prior = prior,
verbose = FALSE)
summary(model)
# drop1(model, test = "Chisq") # doesn't work
# update(model, ~.- X2) # doesn't work
model2 <- MCMCglmm(phenotype ~ X1 + X3 + X4,
random = ~taxon,
ginverse = list(taxon=Ainv),
data = d,
prior = prior,
verbose = FALSE)
summary(model2)
完整型号输出：
Iterations = 3001:12991
Thinning interval = 10
Sample size = 1000
DIC: 145.2228
G-structure: ~taxon
post.mean l-95% CI u-95% CI eff.samp
taxon 1.808 0.2167 3.347 17.88
R-structure: ~units
post.mean l-95% CI u-95% CI eff.samp
units 0.6712 0.0003382 1.585 20.98
Location effects: phenotype ~ X1 + X2 + X3 + X4
post.mean l-95% CI u-95% CI eff.samp pMCMC
(Intercept) -0.6005279 -1.5753606 0.2689091 1000.0 0.198
X1 0.7506249 0.5220491 1.0033785 506.8 <0.001 ***
X2 -0.0339347 -0.2452923 0.2085228 900.7 0.712
X3B 0.6276852 0.1238884 1.1727409 1000.0 0.022 *
X3C 1.0533919 0.4244092 1.5095804 1000.0 <0.001 ***
X4B -0.0002833 -0.5099753 0.5103896 478.3 0.978
X4C 0.0723440 -0.4435520 0.6193011 1000.0 0.776
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
减少模型输出：
Iterations = 3001:12991
Thinning interval = 10
Sample size = 1000
DIC: 296.4755
G-structure: ~taxon
post.mean l-95% CI u-95% CI eff.samp
taxon 1.542 0.07175 3.1 26.33
R-structure: ~units
post.mean l-95% CI u-95% CI eff.samp
units 0.8155 0.0004526 1.661 21.89
Location effects: phenotype ~ X1 + X3 + X4
post.mean l-95% CI u-95% CI eff.samp pMCMC
(Intercept) -0.606920 -1.380267 0.335225 1000.0 0.172
X1 0.750929 0.540667 1.011041 1000.0 <0.001 ***
X3B 0.633123 0.149282 1.243790 1000.0 0.032 *
X3C 1.047085 0.518664 1.570641 1000.0 <0.001 ***
X4B 0.007019 -0.528080 0.512161 509.2 0.998
X4C 0.063093 -0.490103 0.590965 1000.0 0.818
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1