2016-04-15 62 views
3

我想在模型中使用nlme::lme(底部的數據)指定不同的隨機效果。隨機效應是:1)interceptpositionsubject之間變化; 2)interceptcomparison而異。這是一個簡單的使用lme4::lmer如何在nlme與lme4中指定不同的隨機效果?

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  
... 

不過,我想堅持lme,因爲我也想自相關結構(position是一個時間變量)模型。 如何使用lme與上述相同?我在下面的嘗試嵌套效果,這不是我想要的。

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:有對SO和CV herehere一些類似的問題,here但我既沒有理解答案或建議是使用lmer其在這裏不計;)

的例子

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") 

回答

3

我一直嘗試使用數據弄清楚了一段時間。沒有更多的工作,我不認爲我可以得到與lme4完全一樣的模型,但我可以接近。

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

這是你想要的型號,採用了全隨機斜坡期限爲subject(相關斜率和截距)和隨機截距comparison

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

這是一個我可以弄清楚如何複製lme:獨立截距和斜坡。 (我並不特別喜歡這些模型,但他們在作爲一種爲人們簡化過於複雜的隨機效應模型相當普遍使用。)

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

結果:

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

對於這個特定的數據集,隨機斜率估計無論如何都有零方差。

現在讓我們設置爲lme。關鍵(???)是pdBlocked()矩陣中的所有術語必須是嵌套在同一分組變量內的。例如,Pinheiro和Bates的pp。163ff上的交叉隨機效應例子具有塊,塊內的行和塊內的列作爲隨機效應。由於沒有聚合因子在其中comparisonsubject都是嵌套的,我只是去彌補,其中包括在一個塊中的整個數據集的dummy「因素」:

d$dummy <- factor(1) 

現在我們能適應該模型。

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

我們在隨機效應方差 - 協方差矩陣三個大塊:一爲subject,一個用於position -by- subject互動,以及一個用於comparison。沒有定義全新的pdMat類,我找不到一個簡單的方法來允許每個斜率(position:subjectXX)與其相應的截距(subjectXX)相關聯。 (你可能會認爲你可以用pdBlocked結構設置的,但我沒有看到任何方式限制的方差估計是一個pdBlocked對象內的多個塊相同。)

結果是幾乎相同儘管他們的報道不同。

vv <- VarCorr(m3) 
vv2 <- vv[c("subject1","position:subject1","comparison1","Residual"),] 
storage.mode(vv2) <- "numeric" 
print(vv2,digits=4) 
        Variance StdDev 
subject1   7.849e-02 2.802e-01 
position:subject1 4.681e-11 6.842e-06 
comparison1  7.905e-02 2.812e-01 
Residual   8.818e-01 9.390e-01