如何在nlme与lme4中指定不同的随机效果?

时间:2016-04-15 09:46:07

标签: r lme4 mixed-models nlme

我想使用nlme::lme(底部的数据)在模型中指定不同的随机效果。随机效应是:1)interceptposition变化超过subject; 2)intercept变化超过comparison。使用lme4::lmer

可以直截了当
lmer(rating ~ 1 + position + 
     (1 + position | subject) + 
     (1 | comparison), data=d)

> ...
Random effects:
 Groups     Name        Std.Dev. Corr 
 comparison (Intercept) 0.31877       
 subject    (Intercept) 0.63289       
            position    0.06254  -1.00
 Residual               0.91458      
 ...

但是,我想坚持lme,因为我也想模拟自相关结构(position是一个时间变量)。 如何使用lme执行与上述相同的操作?我的下面尝试将效果嵌套,这不是我想要的效果。

lme(rating ~ 1 + position,
random = list( ~ 1 + position | subject,
               ~ 1 | comparison), data=d)

> ...
Random effects:
 Formula: ~1 + position | subject
 Structure: General positive-definite, Log-Cholesky parametrization
            StdDev     Corr  
(Intercept) 0.53817955 (Intr)
position    0.04847635 -1    

 Formula: ~1 | comparison %in% subject    # NESTED :(
        (Intercept)     Residual
StdDev:   0.9707665 0.0002465237
...

注意:有关SO和简历hereherehere的类似问题,但我要么不理解答案或建议是使用lmer,这里不算数;)

示例中使用的数据

d <- structure(list(rating = c(2, 3, 4, 3, 2, 4, 4, 3, 2, 1, 3, 2, 
2, 2, 4, 2, 4, 3, 2, 2, 3, 5, 3, 4, 4, 4, 3, 2, 3, 5, 4, 5, 2, 
3, 4, 2, 4, 4, 1, 2, 4, 5, 4, 2, 3, 4, 3, 2, 2, 2, 4, 5, 4, 4, 
5, 2, 3, 4, 3, 2), subject = structure(c(1L, 1L, 1L, 1L, 1L, 
1L, 1L, 1L, 1L, 1L, 2L, 2L, 2L, 2L, 2L, 2L, 2L, 2L, 2L, 2L, 3L, 
3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 4L, 4L, 4L, 4L, 4L, 4L, 4L, 
4L, 4L, 4L, 5L, 5L, 5L, 5L, 5L, 5L, 5L, 5L, 5L, 5L, 6L, 6L, 6L, 
6L, 6L, 6L, 6L, 6L, 6L, 6L), .Label = c("1", "2", "3", "4", "5", 
"6", "7", "8", "9", "10", "11", "12", "13", "14", "15", "16", 
"17", "18", "19", "20", "21", "22", "23", "24", "25", "26", "27", 
"28", "29", "30", "31", "32", "33", "34", "35", "36", "37", "38", 
"39", "40", "41", "42", "43", "44", "45", "46", "47", "48", "49", 
"50", "51", "52", "53", "54", "55", "56", "57", "58", "59", "60", 
"61", "62", "63"), class = "factor"), position = c(1, 2, 3, 4, 
5, 6, 7, 8, 9, 10, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 1, 2, 3, 4, 
5, 6, 7, 8, 9, 10, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 1, 2, 3, 4, 
5, 6, 7, 8, 9, 10, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10), comparison = structure(c(1L, 
7L, 9L, 8L, 3L, 4L, 10L, 2L, 5L, 6L, 2L, 6L, 4L, 5L, 8L, 10L, 
7L, 3L, 1L, 9L, 3L, 9L, 10L, 1L, 5L, 7L, 6L, 8L, 2L, 4L, 4L, 
2L, 8L, 6L, 7L, 5L, 1L, 10L, 9L, 3L, 5L, 10L, 6L, 3L, 2L, 9L, 
4L, 1L, 8L, 7L, 6L, 5L, 2L, 10L, 4L, 3L, 8L, 9L, 7L, 1L), contrasts = structure(c(1, 
0, 0, 0, 0, 0, 0, 0, 0, -1, 0, 1, 0, 0, 0, 0, 0, 0, 0, -1, 0, 
0, 1, 0, 0, 0, 0, 0, 0, -1, 0, 0, 0, 1, 0, 0, 0, 0, 0, -1, 0, 
0, 0, 0, 1, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, 1, 0, 0, 0, -1, 0, 
0, 0, 0, 0, 0, 1, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 1, 0, -1, 0, 
0, 0, 0, 0, 0, 0, 0, 1, -1), .Dim = c(10L, 9L), .Dimnames = list(
    c("1", "2", "3", "4", "5", "6", "7", "8", "9", "10"), NULL)), .Label = c("1", 
"2", "3", "4", "5", "6", "7", "8", "9", "10"), class = "factor")), .Names = c("rating", 
"subject", "position", "comparison"), row.names = c(1L, 2L, 3L, 
4L, 5L, 6L, 7L, 8L, 9L, 10L, 111L, 112L, 113L, 114L, 115L, 116L, 
117L, 118L, 119L, 120L, 221L, 222L, 223L, 224L, 225L, 226L, 227L, 
228L, 229L, 230L, 331L, 332L, 333L, 334L, 335L, 336L, 337L, 338L, 
339L, 340L, 441L, 442L, 443L, 444L, 445L, 446L, 447L, 448L, 449L, 
450L, 551L, 552L, 553L, 554L, 555L, 556L, 557L, 558L, 559L, 560L
), class = "data.frame")

1 个答案:

答案 0 :(得分:6)

我一直想试着弄清楚这一点。如果没有更多的工作,我认为我不能得到与lme4完全相同的模型,但我可以接近。

## source("SO36643713.dat")
library(nlme)
library(lme4)

这是您想要的模型,subject的完整随机斜率项(相关斜率和截距)和comparison的随机截距:

m1 <- lmer(rating ~ 1 + position + 
               (1 + position | subject) + 
               (1 | comparison), data=d)

我可以弄清楚如何在lme中复制:独立拦截和斜坡。 (我不特别喜欢这些模型,但它们作为人们简化过于复杂的随机效应模型的一种方式而被广泛使用。)

m2 <- lmer(rating ~ 1 + position + 
               (1 + position || subject) + 
               (1 | comparison), data=d)

结果:

VarCorr(m2)
##  Groups     Name        Std.Dev.
##  comparison (Intercept) 0.28115 
##  subject    position    0.00000 
##  subject.1  (Intercept) 0.28015 
##  Residual               0.93905 

对于这个特定的数据集,无论如何估计随机斜率的方差为零。

现在让我们为lme设置它。关键(???)洞察力是pdBlocked()矩阵内的所有术语必须嵌套在同一分组变量中。例如,Pinheiro和Bates的pp.163ff的交叉随机效应示例具有块,块内的行和块内的列作为随机效应。由于没有comparisonsubject嵌套的分组因子,我只想构成一个dummy“因子”,其中包含单个块中的整个数据集:

d$dummy <- factor(1)

现在我们可以适应模型了。

m3 <- lme(rating~1+position,
          random=list(dummy =
                pdBlocked(list(pdIdent(~subject-1),
                               pdIdent(~position:subject),
                               pdIdent(~comparison-1)))),
          data=d)

我们在随机效应方差 - 协方差矩阵中有三个块:一个用于subject,一个用于position - by - subject交互,一个用于{{1} }。如果没有定义一个全新的comparison类,我无法找到一种简单的方法来允许每个斜率(pdMat)与其相应的截距(position:subjectXX)相关联。 (您可能认为可以使用subjectXX结构进行设置,但我认为没有任何方法可以将pdBlocked对象中的多个块的方差估计值限制为相同。)

结果几乎完全相同,尽管它们的报道不同。

pdBlocked