glmer 中的交互(统计建议)

Interaction in glmer (stats advice)

我需要帮助来理解和跟进使用 lme4 的 glmer() 获得的交互。

数据来自一项语言处理实验,该实验研究三个分类变量 (control/copula/gender) 对二项式响应(首选或不首选)的影响。每个实验因素都有两个水平: 控制(subject/object) 联结 (ser/estar) 性别(masculine/feminine)。

我运行以下型号:

model1= glmer(preferences~control*copula*gender+(1|participant), family=binomial, data=data2)

这些是我获得的结果:

Generalized linear mixed model fit by maximum likelihood (Laplace Approximation) ['glmerMod']
 Family: binomial  ( logit )
Formula: preferences_narrow ~ control * copula * gender + (1 | participant)
   Data: data2

     AIC      BIC   logLik deviance df.resid 
  1208.6   1261.1   -595.3   1190.6     2517 

Scaled residuals: 
    Min      1Q  Median      3Q     Max 
-8.6567  0.1970  0.2337  0.2883  0.5371 

Random effects:
 Groups      Name        Variance Std.Dev.
 participant (Intercept) 0.254    0.504   
Number of obs: 2526, groups:  participant, 105

Fixed effects:
                                    Estimate Std. Error z value Pr(>|z|)    
(Intercept)                           2.5034     0.2147  11.660  < 2e-16 ***
controlsubject                        0.4882     0.3172   1.539  0.12380    
copulaser                             0.4001     0.3237   1.236  0.21646    
gendermasc                           -0.4524     0.2659  -1.701  0.08888 .  
controlsubject:copulaser             -1.0355     0.4526  -2.288  0.02215 *  
controlsubject:gendermasc             0.5790     0.4430   1.307  0.19121    
copulaser:gendermasc                  1.7343     0.5819   2.980  0.00288 ** 
controlsubject:copulaser:gendermasc  -1.3121     0.7540  -1.740  0.08181 .  
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Correlation of Fixed Effects:
             (Intr) cntrls coplsr gndrms cntrlsbjct:c cntrlsbjct:g cplsr:
contrlsbjct  -0.602                                                      
copulaser    -0.588  0.401                                               
gendermasc   -0.724  0.488  0.479                                        
cntrlsbjct:c  0.415 -0.701 -0.716 -0.342                                 
cntrlsbjct:g  0.432 -0.716 -0.287 -0.599  0.502                          
cplsr:gndrm   0.332 -0.223 -0.556 -0.457  0.397        0.274             
cntrlsbjc::  -0.252  0.421  0.430  0.352 -0.600       -0.588       -0.772

有两个显着的交互作用 controlsubject:copulasercopulaser:gendermasc

我使用 emmeans 跟进了第一次互动:

emmeans(model1, list(pairwise ~ control + copula), adjust = "tukey")

结果似乎表明多重对比正在推动互动(当我为第二次互动做同样的事情时会发生类似的事情):

NOTE: Results may be misleading due to involvement in interactions
$`emmeans of control, copula`
 control copula   emmean        SE  df asymp.LCL asymp.UCL
 object  estar  2.277256 0.1497913 Inf  1.983670  2.570841
 subject estar  3.054906 0.1912774 Inf  2.680009  3.429802
 object  ser    3.544448 0.2697754 Inf  3.015698  4.073198
 subject ser    2.630568 0.1752365 Inf  2.287110  2.974025

Results are averaged over the levels of: gender 
Results are given on the logit (not the response) scale. 
Confidence level used: 0.95 

$`pairwise differences of control, copula`
 contrast                       estimate        SE  df z.ratio p.value
 object,estar - subject,estar -0.7776499 0.2215235 Inf  -3.510  0.0025
 object,estar - object,ser    -1.2671927 0.2910689 Inf  -4.354  0.0001
 object,estar - subject,ser   -0.3533119 0.2088155 Inf  -1.692  0.3279
 subject,estar - object,ser   -0.4895427 0.3138092 Inf  -1.560  0.4017
 subject,estar - subject,ser   0.4243380 0.2396903 Inf   1.770  0.2877
 object,ser - subject,ser      0.9138807 0.3048589 Inf   2.998  0.0145

Results are averaged over the levels of: gender 
Results are given on the log odds ratio (not the response) scale. 
P value adjustment: tukey method for comparing a family of 4 estimates 

然而,注意是什么意思?

NOTE: Results may be misleading due to involvement in interactions

这是跟踪这些互动的好程序吗?

提前致谢! :)

如注释中所示,显示的估计值是控制、copula 和性别组合的预测平均值,按性别取平均值。同时,该模型包括性别与其他两个因素之间的相互作用,这表明这些平均值可能没有意义。您可以通过构建 3 向预测图来形象化这一点:

emmip(model1, gender ~ control * copula)

如果一个案例与下一个案例的预测比较完全不同,那么它们的平均值将是无稽之谈。但如果它们比较起来几乎相同,那么将它们取平均就可以了。这就是警告的内容。

我猜你确实开会担心与性别的互动——在这种情况下你应该单独进行比较:

emmeans(model1, pairwise ~ control * copula | gender)