I'm fitting generalized linear mixed models to using several fixed effects
(main effects and a couple of interactions) and a grouping factor (site) to
explain the variation in a dichotomous response variable (family=binomial). I
wanted to compare the output I obtained using PROC GLIMMIX in SAS with that
obtained using lmer in R (version 2.6.1 in Windows). When using lmer I'm
specifying method="PQL" so as to make the estimation method comparable between
lmer and GLIMMIX.
It is difficult to compare the outputs for the interaction terms because SAS
gives the estimates and significance value for each of the categories, whereas
R gives a single estimate for the interaction term. But, from the main effects
it is possible to see very similar estimates obtained with either program.
I am very interested in the interaction term SEX*ELI, and this term comes up as
significant in SAS and nonsignificant in R. Why could this be? It is very
worrisome to think of reporting a significant result that is not validated when
doing a similar analysis using a different program!
Can somebody help me interpret these differences?
Bellow is a summary of the outputs obtained with R and SAS.
Thanks,
Andrea Previtali
Post-doc fellow
Dept. of Biology,
Univ. of Utah.
lmer output:
Generalized linear mixed model fit using PQL
Formula: SURV ~ SEX * ELI + DW * DIST + SEAS + DEN + WT + (1 | SITE)
Family: binomial(logit link)
AIC BIC logLik deviance
1539 1606 -758.7 1517
Random effects:
Groups Name Variance Std.Dev.
SITE (Intercept) 0.27816 0.52741
number of obs: 3104, groups: SITE, 19
Estimated scale (compare to 1 ) 0.9458749
Fixed effects:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -1.144259 0.458672 -2.495 0.012606 *
SEX -0.606026 0.167289 -3.623 0.000292 ***
ELI -0.190757 0.219599 -0.869 0.385034
DW -0.328796 0.175882 -1.869 0.061565 .
DIST -0.117745 0.374148 -0.315 0.752989
SEAS -0.784971 0.158748 -4.945 7.62e-07 ***
DEN -0.013381 0.002585 -5.176 2.27e-07 ***
WT 0.007735 0.019115 0.405 0.685732
SEX:ELI -0.466425 0.461596 -1.010 0.312274
DW:DIST -1.015454 0.404683 -2.509 0.012099 *
-----------------------------------------------------------------------------------
GLIMMIX output:
Model Information
Variance Matrix Blocked By Site
Estimation Technique: Residual PL
Degrees of Freedom Method: Containment
Fit Statistics
-2 Res Log Pseudo-Likelihood: 17868.73
Pseudo-AIC: 17890.73
Pseudo-BIC: 17957.14
Covariance Parameter Estimates
Cov Parm Subject Estimate Std Error
Intercept Site 0.2975 0.1799
Solutions for Fixed Effects
Effect DIST DW ELI SEX SEAS Estimate Std Error DF
t Value Pr> |t|
Intercept -4.6540
0.6878 17 -6.77 F
DIST*DW 3 3077 6.06 0.0004
SEX*ELI 3 3077 6.30 0.0003
WT 1 3077 0.16 0.6918
SEAS 1 3077 24.37
______________________________________________
R-help@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
