[R-sig-ME] discrepancy between contrast and pooled levels

espesser robert.espesser at lpl-aix.fr
Fri Mar 6 18:17:57 CET 2009


Dear list members,

I am confused with the following results,
so I greatly appreciate any suggestions or remarks.

The structure of the data:
 > str(eglob_new)
'data.frame': 1200 obs. of 8 variables:
$ subject : Factor w/ 40 levels "letl01","letl03",..: 1 1 1 1 1 1 1 1 1 
1 ...

$ dlmscentre : num -2 -2 -2 -2 -1 -1 -1 -1 0 0 ...
$ nfcompo :    num 2 1 3 2 2 1 3 2 2 1 ...

$ conditionbis : Factor w/ 3 levels "e2b","e2a","e2c": 2 2 2 2 2 2 2 2 2 
2 ...
$ rlet : num 9 7 7 5 8 10 5 6 5 8 ...
$ rnlet : num 8 10 10 12 9 7 12 11 12 9 ...

$ conditionBLOCKVOY : Factor w/ 2 levels "e2ab","e2c": 1 1 1 1 1 1 1 1 1 
1 ...
$ conditionFEEDBACK : Factor w/ 2 levels "e2a","e2bc": 1 1 1 1 1 1 1 1 1 
1 ...


- a first group of 20 subjects experiments the both condition e2b and e2c
- an other group of 20 subjects only experiments the condition e2a

I think the subject labels are correct, ie there is no implicit nesting.

I fit probability of the "let" response with a logit mixed model .
the "best" model I obtained was :

glmer(cbind(rlet, rnlet) ~ dlmscentre*conditionbis + nfcompo + 
(dlmscentre | subject),family=binomial, data=eglob_new)
# model A

 > summary(eglob_new_conditionbis_leger.glmer)
Generalized linear mixed model fit by the Laplace approximation
Formula: cbind(rlet, rnlet) ~ (dlmscentre + conditionbis)^2 + nfcompo 
+      (dlmscentre | subject)
   Data: eglob_new
  AIC  BIC logLik deviance
 2083 2134  -1032     2063
Random effects:
 Groups  Name        Variance Std.Dev. Corr 
 subject (Intercept) 0.22676  0.47619       
         dlmscentre  0.32139  0.56691  0.338
Number of obs: 1200, groups: subject, 40

Fixed effects:
                           Estimate Std. Error z value Pr(>|z|)   
(Intercept)                -0.01225    0.11792  -0.104    0.917   
dlmscentre                 -0.85908    0.12867  -6.677 2.44e-11 ***
conditionbise2a             0.02483    0.15577   0.159    0.873   
conditionbise2c            -0.03907    0.03666  -1.066    0.286   
nfcompo                     0.13241    0.02155   6.144 8.03e-10 ***
dlmscentre:conditionbise2a  0.26804    0.18213   1.472    0.141   
dlmscentre:conditionbise2c  0.17383    0.02828   6.146 7.94e-10 ***


the results are plausible;
there is a decreasing slope (dlmscentre); slope for e2c differs from the 
slope of e2b.

More specifically, I'm interested to test the hypothesis:
Is there a slope difference between e2c vs (e2a pooled with e2b) ?


I first recoded conditionbis into a new factor: conditionBLOCKVOY

#model BLOCKVOY
glmer( cbind(rlet, rnlet) ~ dlmscentre * conditionBLOCKVOY +
nfcompo+(dlmscentre | subject),family=binomial, data=eglob_new)

summary(eglob_new_conditionBLOCKVOY_leger.glmer)
Generalized linear mixed model fit by the Laplace approximation
Formula: cbind(rlet, rnlet) ~ dlmscentre * conditionBLOCKVOY + nfcompo 
+      (dlmscentre | subject)
   Data: eglob_new
  AIC  BIC logLik deviance
 2082 2122  -1033     2066
Random effects:
 Groups  Name        Variance Std.Dev. Corr 
 subject (Intercept) 0.22843  0.47795       
         dlmscentre  0.34013  0.58321  0.336
Number of obs: 1200, groups: subject, 40

Fixed effects:
                                  Estimate Std. Error z value Pr(>|z|)   
(Intercept)                      0.0001614  0.0891725   0.002    0.999   
dlmscentre                      -0.7257465  0.0936049  -7.753 8.95e-15 ***
conditionBLOCKVOYe2c            -0.0371947  0.0363233  -1.024    0.306   
nfcompo                          0.1324147  0.0215497   6.145 8.02e-10 ***
dlmscentre:conditionBLOCKVOYe2c  0.1702476  0.0281450   6.049 1.46e-09 ***


The interaction is significant: slope for e2c is different from the 
slope for pooled (e2a ,e2b)

b)
I test the hypothesis by setting a specific contrast for conditionbis .

eglob_new$conditionbis -> eglob_new$conditionBLOCKVOYcontr
ginv( cbind(1,1,-2)) -> contrasts(eglob_new$conditionBLOCKVOYcontr)
 > contrasts(eglob_new$conditionBLOCKVOYcontr)
[,1] [,2]
e2b 0.1666667 -7.071068e-01
e2a 0.1666667 7.071068e-01
e2c -0.3333333 -9.877082e-17

# model BLOCKVOYcontr

 > summary(eglob_new_conditionBLOCKVOYcontr_leger.glmer)
Generalized linear mixed model fit by the Laplace approximation
Formula: cbind(rlet, rnlet) ~ dlmscentre * conditionBLOCKVOYcontr + 
nfcompo + (dlmscentre | subject)
   Data: eglob_new
  AIC  BIC logLik deviance
 2083 2134  -1032     2063
Random effects:
 Groups  Name        Variance Std.Dev. Corr 
 subject (Intercept) 0.22676  0.47619       
         dlmscentre  0.32138  0.56691  0.338
Number of obs: 1200, groups: subject, 40

Fixed effects:
                                   Estimate Std. Error z value Pr(>|z|)   
(Intercept)                        -0.01699    0.09165  -0.185    0.853   
dlmscentre                         -0.71179    0.09540  -7.461 8.60e-14 ***
conditionBLOCKVOYcontr1             0.10297    0.16385   0.628    0.530   
conditionBLOCKVOYcontr2             0.01756    0.11014   0.159    0.873   
nfcompo                             0.13241    0.02155   6.144 8.03e-10 ***
dlmscentre:conditionBLOCKVOYcontr1 -0.07966    0.18615  -0.428    0.669   
dlmscentre:conditionBLOCKVOYcontr2  0.18949    0.12878   1.471    0.141   
---

Correlation of Fixed Effects:
            (Intr) dlmscn cBLOCKVOY1 cBLOCKVOY2 nfcomp d:BLOCKVOY1
dlmscentre   0.291                                               
cnBLOCKVOY1 -0.247 -0.097                                        
cnBLOCKVOY2 -0.262 -0.101  0.896                                 
nfcompo     -0.469 -0.004  0.000      0.000                      
d:BLOCKVOY1 -0.089 -0.299  0.294      0.316      0.000           
d:BLOCKVOY2 -0.091 -0.308  0.307      0.320      0.001  0.953

dlmscentre:conditionBLOCKVOYcontr1, the interest interaction is not 
significant.
the slope for (e2a+e2b) cannot be distinguished from the slope for e2c

I did'nt expect  the same results for the model BLOCKVOY and the model 
BLOCKVOYcontr,
but not such an opposite result .

There is a high correlation (0.953) between the two interaction 
coefficients,
which means this last model is badly specified ?
Do I misunderstand something about contrasts ?

 > sessionInfo()
R version 2.8.1 (2008-12-22)
i386-pc-mingw32
......
other attached packages:
[1] lme4_0.999375-28 Matrix_0.999375-21 lattice_0.17-17

Sorry for this long email, and thank you in advance
Regards

R. Espesser
Laboratoire Parole et Langage,CNRS et Université de Provence
13100 Aix-en-provence, France




More information about the R-sig-mixed-models mailing list