堆棧內存溢出
  簡體   English   中英

不同版本的R，lme4和OS X在glmer中給出不同的固定效應顯着性結果

[英]Different versions of R, lme4 and OS X give different fixed-effects significance results in glmer

 原文  2014-09-06 11:59:40  1   1    r/ macos/ lme4/ mixed-models

問題描述

我在包lme4中使用glmer（）運行logit混合效果模型。 該實驗使用受試者內部項目設計，其中主題和項目為交叉隨機效應。

我的問題：不同版本的R和lme4（在不同的OS X上運行）對固定效果產生不同的標准誤差估計，因此產生不同的顯着性結果。

這是我的數據的子集（來自最后兩個主題的數據）： 

structure(list(SubjN = c(87L, 87L, 87L, 87L, 87L, 87L, 87L, 87L, 
87L, 87L, 87L, 87L, 87L, 87L, 87L, 87L, 87L, 87L, 87L, 87L, 87L, 
87L, 87L, 87L, 88L, 88L, 88L, 88L, 88L, 88L, 88L, 88L, 88L, 88L, 
88L, 88L, 88L, 88L, 88L, 88L, 88L, 88L, 88L, 88L, 88L, 88L, 88L, 
88L), Items = structure(c(3L, 10L, 11L, 5L, 1L, 12L, 2L, 6L, 
9L, 6L, 3L, 4L, 8L, 11L, 12L, 7L, 8L, 2L, 7L, 10L, 9L, 5L, 1L, 
4L, 10L, 3L, 5L, 11L, 12L, 1L, 2L, 6L, 9L, 6L, 3L, 4L, 8L, 11L, 
12L, 7L, 2L, 8L, 10L, 7L, 9L, 5L, 1L, 4L), .Label = c("a", "c", 
"k", "f", "g", "i", "d", "l", "e", "j", "b", "h"), class = "factor"), 
IV1 = structure(c(1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L, 2L, 2L, 
2L, 2L, 2L, 2L, 2L, 2L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 1L, 
1L, 1L, 1L, 1L, 1L, 1L, 1L, 2L, 2L, 2L, 2L, 2L, 2L, 2L, 2L, 
3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L), .Label = c("N", "L", "P"
), class = "factor"), DV = c(0, 0, 1, 0, 1, 0, 1, 0, 0, 0, 
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 
0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0), 
IV1.h = structure(c(1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L, 2L, 2L, 
2L, 2L, 2L, 2L, 2L, 2L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 1L, 
1L, 1L, 1L, 1L, 1L, 1L, 1L, 2L, 2L, 2L, 2L, 2L, 2L, 2L, 2L, 
3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L), contrasts = structure(c(-1, 
0.5, 0.5, 0, -0.5, 0.5), .Dim = c(3L, 2L), .Dimnames = list(
    c("N", "L", "P"), c("N_vs_L&P", "L_vs_P"))), .Label = c("N", 
"L", "P"), class = "factor"), N_vs_LP = c(-1, -1, -1, -1, 
-1, -1, -1, -1, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 
0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, -1, -1, -1, -1, -1, -1, 
-1, -1, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 
0.5, 0.5, 0.5, 0.5, 0.5, 0.5), L_vs_P = c(0, 0, 0, 0, 0, 
0, 0, 0, -0.5, -0.5, -0.5, -0.5, -0.5, -0.5, -0.5, -0.5, 
0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0, 0, 0, 0, 0, 0, 
0, 0, -0.5, -0.5, -0.5, -0.5, -0.5, -0.5, -0.5, -0.5, 0.5, 
0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5)), .Names = c("SubjN", 
"Items", "IV1", "DV", "IV1.h", "N_vs_LP", "L_vs_P"), row.names = c("3099", 
"3100", "3101", "3102", "3103", "3104", "3119", "3120", "3107", 
"3108", "3109", "3110", "3097", "3098", "3105", "3106", "3115", 
"3116", "3117", "3118", "3111", "3112", "3113", "3114", "3147", 
"3148", "3149", "3150", "3151", "3152", "3167", "3168", "3155", 
"3156", "3157", "3158", "3145", "3146", "3153", "3154", "3163", 
"3164", "3165", "3166", "3159", "3160", "3161", "3162"), class = "data.frame")

每個受試者在3個不同條件下進行24次試驗（因子IV1，水平：N，L，P）。 我記錄他們是否產生了目標語言結構（DV == 1）或不產生（DV == 0）。 在分析中，我只包括那些產生目標結構的受試者至少一個。 盡管如此，他們中的大多數只在極少數情況下產生了目標結構。 這是每個條件下每個主題產生的DV == 1的比例： 

library(plyr)
#dput(ddply(mydata, .(SubjN, IV1), summarise, l = length(DV), y = round(mean(DV),2)))

structure(list(SubjN = c(1L, 1L, 1L, 2L, 2L, 2L, 3L, 3L, 3L, 
4L, 4L, 4L, 5L, 5L, 5L, 6L, 6L, 6L, 7L, 7L, 7L, 8L, 8L, 8L, 9L, 
9L, 9L, 10L, 10L, 10L, 11L, 11L, 11L, 12L, 12L, 12L, 13L, 13L, 
13L, 14L, 14L, 14L, 15L, 15L, 15L, 16L, 16L, 16L, 17L, 17L, 17L, 
18L, 18L, 18L, 19L, 19L, 19L, 20L, 20L, 20L, 21L, 21L, 21L, 22L, 
22L, 22L, 23L, 23L, 23L, 24L, 24L, 24L, 25L, 25L, 25L, 26L, 26L, 
26L, 27L, 27L, 27L, 28L, 28L, 28L, 29L, 29L, 29L, 30L, 30L, 30L, 
31L, 31L, 31L, 32L, 32L, 32L, 33L, 33L, 33L, 34L, 34L, 34L, 35L, 
35L, 35L, 36L, 36L, 36L, 37L, 37L, 37L, 38L, 38L, 38L, 39L, 39L, 
39L, 40L, 40L, 40L, 41L, 41L, 41L, 42L, 42L, 42L, 43L, 43L, 43L, 
44L, 44L, 44L, 45L, 45L, 45L, 46L, 46L, 46L, 47L, 47L, 47L, 48L, 
48L, 48L, 49L, 49L, 49L, 50L, 50L, 50L, 51L, 51L, 51L, 52L, 52L, 
52L, 53L, 53L, 53L, 54L, 54L, 54L, 55L, 55L, 55L, 56L, 56L, 56L, 
57L, 57L, 57L, 58L, 58L, 58L, 59L, 59L, 59L, 60L, 60L, 60L, 61L, 
61L, 61L, 62L, 62L, 62L, 63L, 63L, 63L, 64L, 64L, 64L, 65L, 65L, 
65L, 66L, 66L, 66L, 67L, 67L, 67L, 68L, 68L, 68L, 69L, 69L, 69L, 
70L, 70L, 70L, 71L, 71L, 71L, 72L, 72L, 72L, 73L, 73L, 73L, 74L, 
74L, 74L, 75L, 75L, 75L, 76L, 76L, 76L, 77L, 77L, 77L, 78L, 78L, 
78L, 79L, 79L, 79L, 80L, 80L, 80L, 81L, 81L, 81L, 82L, 82L, 82L, 
83L, 83L, 83L, 84L, 84L, 84L, 85L, 85L, 85L, 86L, 86L, 86L, 87L, 
87L, 87L, 88L, 88L, 88L), IV1 = structure(c(1L, 2L, 3L, 1L, 2L, 
3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 
1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 
2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 
3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 
1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 
2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L,      
3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 
1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 
2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 
3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 
1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 
2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 
3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 
1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 
2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 
3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 1L, 2L, 3L, 
1L, 2L, 3L), .Label = c("N", "L", "P"), class = "factor"), l = c(8L, 
8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 
8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 
8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 
8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 7L, 8L, 7L, 8L, 8L, 
8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 
8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 7L, 8L, 8L, 8L, 8L, 8L, 8L, 
7L, 8L, 6L, 7L, 7L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 
8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 
8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 7L, 8L, 8L, 7L, 7L, 8L, 7L, 8L, 
8L, 7L, 8L, 8L, 7L, 8L, 8L, 7L, 8L, 8L, 7L, 8L, 8L, 7L, 8L, 8L, 
8L, 8L, 8L, 8L, 8L, 8L, 8L, 6L, 8L, 4L, 8L, 8L, 8L, 8L, 8L, 8L, 
8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 7L, 8L, 8L, 8L, 8L, 
8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 
8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 8L, 7L, 
8L, 8L, 7L, 8L, 8L, 7L, 8L, 8L, 7L, 8L, 8L, 7L, 8L, 8L, 7L, 8L, 
8L, 7L, 8L, 8L, 7L, 8L, 8L, 7L, 8L, 8L, 7L, 8L, 7L, 8L, 8L, 8L, 
8L, 8L, 8L, 8L, 8L, 8L, 8L), y = c(1, 0.88, 1, 0.5, 0.25, 0.62, 
0, 0, 0.25, 0, 0.25, 0, 0.12, 0, 0, 0, 0.12, 0, 0, 0.12, 0.12, 
0, 0, 0.12, 0.38, 0, 0.25, 0, 0.12, 0, 0.12, 0, 0.25, 0, 0, 0.12, 
0.5, 0.25, 0.5, 0, 0, 0.12, 0, 0.25, 0.12, 0, 0, 0.12, 0, 0.12, 
0, 0, 0.12, 0.12, 0.12, 0.62, 0, 0, 0.5, 0.25, 1, 0.88, 1, 0, 
0, 0.12, 0, 0.12, 0.12, 0.12, 0.12, 0, 0.62, 0.62, 0.38, 0.5, 
0.88, 0.12, 0.12, 0, 0, 0.12, 0.12, 0, 0, 0.12, 0, 0, 0.12, 0, 
0, 0.12, 0, 0, 0.25, 0, 0, 0.14, 0, 0.5, 0.57, 0.29, 0, 0.12, 
0, 0, 0.12, 0, 0.25, 0.5, 0.25, 0, 0.12, 0.12, 0.25, 0, 0.38, 
0, 0, 0.12, 0, 0, 1, 0.25, 0.12, 0.25, 0, 0.12, 0.12, 0, 0, 0.12, 
0, 0, 0.12, 0.12, 0, 0, 0.12, 0, 0.14, 0.14, 0.12, 0, 0.12, 0, 
0, 0.12, 0.12, 0, 1, 0.88, 1, 0, 0.12, 0, 0.12, 0, 0, 0.12, 0, 
0.12, 0, 0, 0.12, 0.12, 0.12, 0.12, 1, 1, 1, 0.12, 0, 0, 0.12, 
0.38, 0, 0, 0.12, 0, 0, 0, 0.5, 0.5, 0, 0.25, 0, 0.12, 0.29, 
0, 0, 0.38, 0, 0, 0.62, 0.5, 0, 0.12, 0, 0.12, 0.12, 0.25, 0.12, 
0.25, 0.12, 0, 0.12, 0, 0, 0.12, 0, 0, 0.12, 0, 0.12, 0.12, 0, 
0.12, 0.12, 0, 0, 0.12, 0.12, 0.12, 0, 0.38, 0.12, 0.57, 0, 0.12, 
0, 0, 0.12, 0, 0, 0.12, 0, 0, 0.12, 0.14, 0.88, 0.88, 0.86, 0, 
0, 0.14, 0, 0.12, 0.14, 0, 0.12, 0, 0, 0, 0.12, 0, 0, 0.12, 0.38, 
0, 0, 0.5, 0.12, 0)), .Names = c("SubjN", "IV1", "l", "y"), row.names = c(NA, 
-264L), class = "data.frame")

我運行以下模型，包括IV1作為具有helmert對比度編碼的固定效果; 第一對比：N vs. L＆P，第二對比：L vs. P. 

m1 <- glmer(DV ~ IV1.h + (1 + IV1.h|SubjN) +  (1|Items) + (0 + N_vs_LP|Items) + (0 + L_vs_P|Items), family ='binomial', mydata)

該模型不允許逐項隨機變量之間的相關性（我通過為兩個對比創建單獨的斜率來實現這一點），因為當允許相關時它們是完全相關的（我將其解釋為過度參數化的標志） 。

1）結果使用os x 10.8.5山獅R版3.0.2（2013-09-25）lme4_1.0-5（我運行的原始分析） 

Generalized linear mixed model fit by maximum likelihood ['glmerMod']
 Family: binomial ( logit )
Formula: DV ~ IV1.h + (1 + N_vs_LP + L_vs_P | SubjN) + (1 | Items) + (0 + N_vs_LP | Items)     + (0 + L_vs_P | Items) 
   Data: mydata 

      AIC       BIC    logLik  deviance 
1492.5408 1560.2050 -734.2704 1468.5408 

Random effects:
 Groups  Name          Variance  Std.Dev. Corr       
 SubjN    (Intercept)   2.3885505 1.54549             
          N_vs_LP       0.4394195 0.66289  -0.69      
          L_vs_P        1.9287559 1.38880   0.04  0.08
 Items    (Intercept)   0.0531518 0.23055
 Items.1  N_vs_LP       0.0001950 0.01396
 Items.2  L_vs_P        0.0003619 0.01902             

Number of obs: 2077, groups: SubjN, 88; Items, 12

Fixed effects:
                              Estimate Std. Error z value Pr(>|z|)    
(Intercept)                    -2.2998     0.1964 -11.710  < 2e-16 ***
IV1.hN_vs_L&P                   0.3704     0.1378   2.689  0.00717 ** 
IV1.hL_vs_P                     0.2060     0.2320   0.888  0.37459    
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Correlation of Fixed Effects:
              (Intr) IV1.N_
IV1.hN_vs_L&P -0.388       
IV1.hL_vs_P    0.014  0.019

2）結果使用：OS X 10.9.4 Mavericks R 3.1.1版（2014-07-10）lme4_1.1-7優化器'bobyqa' 

Generalized linear mixed model fit by maximum likelihood (Laplace Approximation)     ['glmerMod']
 Family: binomial  ( logit )
Formula: DV ~ IV1.h + (1 + N_vs_LP + L_vs_P | SubjN) + (1 | Items) + (0 +  
    N_vs_LP | Items) + (0 + L_vs_P | Items)
   Data: mydata
Control: glmerControl(optimizer = "bobyqa")

     AIC      BIC   logLik deviance df.resid 
  1492.5   1560.2   -734.3   1468.5     2065 

Scaled residuals: 
    Min      1Q  Median      3Q     Max 
-2.4174 -0.3364 -0.2595 -0.1706  4.6028 

Random effects:
 Groups  Name        Variance Std.Dev. Corr       
 SubjN   (Intercept) 2.38791  1.5453              
         N_vs_LP     0.43935  0.6628   -0.69      
         L_vs_P      1.92629  1.3879    0.04  0.07
 Items   (Intercept) 0.05319  0.2306              
 Items.1 N_vs_LP     0.00000  0.0000              
 Items.2 L_vs_P      0.00000  0.0000              
Number of obs: 2077, groups:  SubjN, 88; Items, 12

Fixed effects:
              Estimate Std. Error z value Pr(>|z|)    
(Intercept)    -2.2998     0.2095 -10.975   <2e-16 ***
IV1.hN_vs_L&P   0.3703     0.1892   1.958   0.0503 .  
IV1.hL_vs_P     0.2063     0.2679   0.770   0.4413    
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Correlation of Fixed Effects:
            (Intr) IV1.N_
IV1.hN__L&P -0.379       
IV1.hL_vs_P -0.001  0.003

我真的不知道應該相信哪個結果。 任何幫助將非常感謝。

PS。 對不起，如果事情不明確 - 這是我的第一篇帖子:)

非常感謝！

1 個解決方案

解決方案1
 6 已采納  2014-09-06 12:41:26

來自lme4的NEWS文件 ，版本為1.1-4

現在默認情況下從近似Hessian計算固定效應的標准誤差（參見vcov.merMod中的use.hessian參數）; 當隨機和固定效應參數的估計相關時，這給出了更好（正確）的答案（Github＃47）

問題的描述在這里

您應該能夠通過sqrt(diag(vcov(fitted_model,use.hessian=FALSE)))從較新的（1.1-7）模型中檢索舊的標准錯誤，但新版本更可能是正確的。

對於更精確的置信區間/ p值​​，您可以進行似然比檢驗（使用anova比較嵌套模型）和/或使用confint(fitted_model,which="beta_")計算輪廓置信區間confint(fitted_model,which="beta_") 。

暫無
暫無

聲明:本站的技術帖子網頁，遵循CC BY-SA 4.0協議，如果您需要轉載，請注明本站網址或者原文地址。任何問題請咨詢:yoyou2525@163.com.

相關問題 不同 R/lme4 版本的奇異擬合的不匹配結果 混合效應邏輯回歸：MASS和lme4的不同結果 nlme和lme4中的不同隨機效應 excel 的等效項 如果 r 中的 lme4 glmer 出現錯誤 在 R 中指定邏輯混合 model 中重復測量的隨機效應：lme4::glmer 如何在nlme與lme4中指定不同的隨機效果？ 如何遍歷列以評估連續 lme4 混合模型中的不同固定效應並提取系數和 P 值？ R 中 function 的時間固定效應 不同的lme4版本給出不同的結果 如何從lme4中提取固定效果的一部分摘要？
相關標簽
 
粵ICP備18138465號  © 2020-2024 STACKOOM.COM