1

I was trying to fit several mixed linear models, however for 2 (out of 4) I am getting quite strange results.

The point is that I've collected data from two courses (a few students were the same in both courses). The idea is to predict students' position in social graph based on certain linguistic characteristics. Graph metrics (closeness, betweenness, etc.) are considered dependent variables, while measures of linguistic characteristics are considered independent variables (and fixed effects). I defined 1+course|student as a random effect and it seems that this model is better than 1|student (compared by ICC values). However, the model for betweenness centrality yields the following results:

Random effects:
 Groups   Name           Variance Std.Dev. Corr 
 student  (Intercept)    8563     92.54        
          courseCourse2  8521     92.31   -1.00
 Residual                27714    166.47        
Number of obs: 3066, groups:  student, 1353

What could be a cause for this perfect negative correlation? Should I remove course from random effects? Any suggestions what could be wrong with the data/analysis?

amoeba
  • 93,463
  • 28
  • 275
  • 317
Srecko
  • 177
  • 1
  • 11
  • 1
    This is a standard symptom of an overfitted model. what to do about it (suppress correlation, remove random slope term, penalize using `blme` package) isn't as clear. http://glmm.wikidot.com/faq *might* have some useful information – Ben Bolker Sep 22 '14 at 22:28
  • Thanks Ben, I will try with blme, if not, than probably will have to remove the random slope. – Srecko Sep 22 '14 at 23:29

0 Answers0