[R-lang] Re: Removing the correlation parameter

[Dale Barr]

Hi Zhenguang, 

I don't have time to give more than a quick reply, but what you are seeing is some weirdness with how variables defined as data type "factor" behave in regression models. You seem to have four levels of Prime, and if you remove the correlations between the levels you should have three slopes.  So, neither syntax is giving you the right answer!  What you can do in this case is just explicitly define three separate dummy variables, eg.  PrimeC, PrimeE, PrimeF, and then explicitly spell out (0 + PrimeC | Subject) + (0 + PrimeE | Subject) etc.. 

-Dale

Sent from Samsung Mobile

-------- Original message --------
From: Zhenguang Cai <zhenguangcai@gmail.com> 
Date: 30/07/2013  13:23  (GMT+00:00) 
To:  
Cc: "<ling-r-lang-l@mailman.ucsd.edu>" <ling-r-lang-l@mailman.ucsd.edu> 
Subject: [R-lang]  Removing the correlation parameter 
 
Hi all,

Barr et al. (2013) recommended a design-driven (i.e., maximal random 
structure) approach for LME. However, as some of you may have 
experienced, LME with maximal random structure on categorical data would 
often result in non-convergence. Barr et al. recommended, in these 
cases, removing the correlation parameter (i.e., correlation between a 
random intercept and the corresponding random slope). However, I found 
that Barr et al. (e.g., p.262, Table 1) and Baayen et al. (2008, p.395) 
used different R scripts to do this. I wonder which is the correct way 
to do. I use my own example below.

Barr et al: (But still correlations between levels of the fixed 
predictor in the results)
Generalized linear mixed model fit by the Laplace approximation
Formula: ResponseC ~ Prime + (1 | Subject) + (0 + Prime | Subject) + (1 
|      Item) + (0 + Prime | Item)
    Data: E1
    AIC   BIC logLik deviance
  798.3 876.1 -382.1    764.3
Random effects:
  Groups  Name        Variance   Std.Dev.   Corr
  Subject PrimeC      1.2124e+00 1.1011e+00
          PrimeE      1.4114e+00 1.1880e+00 0.583
          PrimeF      1.2177e+00 1.1035e+00 0.737  0.979
  Subject (Intercept) 1.7926e-02 1.3389e-01
  Item    PrimeC      2.1542e-09 4.6413e-05
          PrimeE      4.6549e-02 2.1575e-01 -0.001
          PrimeF      3.3489e-03 5.7870e-02  0.000  1.000
  Item    (Intercept) 2.7090e+00 1.6459e+00
Number of obs: 720, groups: Subject, 35; Item, 24

Fixed effects:
             Estimate Std. Error z value Pr(>|z|)
(Intercept)  -0.2973     0.4257  -0.698   0.4849
PrimeE        0.5507     0.3044   1.809   0.0705 .
PrimeF        1.3644     0.2787   4.896  9.8e-07 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Correlation of Fixed Effects:
        (Intr) PrimeE
PrimeE -0.348
PrimeF -0.350  0.674


Baayen et al: (here the correlations are gone)
Formula: ResponseC ~ Prime + (1 | Subject) + (1 | Prime:Subject) + (1 
|      Item) + (1 | Prime:Item)
    Data: E1
    AIC   BIC logLik deviance
  781.7 813.7 -383.8    767.7
Random effects:
  Groups        Name        Variance Std.Dev.
  Prime:Subject (Intercept) 0.16358  0.40445
  Prime:Item    (Intercept) 0.00000  0.00000
  Subject       (Intercept) 1.02334  1.01160
  Item          (Intercept) 2.62534  1.62029
Number of obs: 720, groups: Prime:Subject, 105; Prime:Item, 72; Subject, 
35; Item, 24

Fixed effects:
             Estimate Std. Error z value Pr(>|z|)
(Intercept)  -0.2984     0.4196  -0.711   0.4769
PrimeE        0.5401     0.2595   2.081   0.0374 *
PrimeF        1.3539     0.2605   5.197 2.03e-07 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Correlation of Fixed Effects:
        (Intr) PrimeE
PrimeE -0.334
PrimeF -0.336  0.54


Thanks,
Zhenguang Cai

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.ucsd.edu/pipermail/ling-r-lang-l/attachments/20130801/809f99b1/attachment-0001.html