[R-sig-ME] pvals.fnc + lmer, discrepancy between pvalues
Nadine Klauke
nadine.klauke at biologie.uni-freiburg.de
Tue Aug 7 21:32:10 CEST 2012
Dear R list members,
I´m trying to fit a mixed model with log transformed count
data in R (version 13.0). Here is the model specification:
m1<-lmer(Surv~H*E+HER+G+(1|ID),data=rep1)
The variables are:
Surv ##log-transformed count data
E ## 2-level categorical
H ##count data
HER ##continous between 0 and 1
G ##count data
random effect: individual ID (individuals repeatedly
measured in differnet years)
The response variable is log-transformed because of
underdispersion with poisson-distribution.
When I inspected the p-values of the model given by
pvals.fnc() I realized that the pMCMC values are completely
different from the pvalues based on a t-distribution (R
output see below). I am aware that
pMCMC values are more reliable for mixed models.
Nevertheless, as far as I get it form the help lists etc,
the values usually do not differ so much. Or am I wrong?
When I calculate pvalues through logliklihood estimations
with anova() pvalues look more similar to those of the
t-distribution. Furthermore, the density plots of the fixed
effects given by pvals.fnc
look normally distributed. Might the different pvalues be
due to small sample size? Should I rather rely on the
logliklihood estimation than on pMCMC? Any advices would be
appreciated.
Thanks a lot.
Nadine
Results given by R:
m1<-lmer(Surv~H*E+HER+G+(1|ID),data=rep1)
summary(m1)
pvals.fnc(m1)
Linear mixed model fit by maximum likelihood
Formula: Surv ~ H * E + HER + G + (1 | ID)
Data: rep1
AIC BIC logLik deviance REMLdev
9.428 19.49 3.286 -6.572 17.59
Random effects:
Groups Name Variance Std.Dev.
ID (Intercept) 0.090871 0.301447
Residual 0.003334 0.057741
Number of obs: 26, groups: ID, 19
Fixed effects:
Estimate Std. Error t value
(Intercept) -0.87556 0.47112 -1.858
H 0.04135 0.01389 2.976
Eun -0.52552 0.11116 -4.728
HER 3.45757 0.90229 3.832
G -0.04023 0.03136 -1.283
H:Eun 0.22855 0.03812 5.996
Correlation of Fixed Effects:
(Intr) H Enrfhr HER Gelege
H 0.467
Eun 0.364 0.337
HER -0.950 -0.485 -0.573
G 0.562 0.222 0.777 -0.767
H:Eun -0.550 -0.361 -0.864 0.712 -0.815
pvals.fnc(m1)
$fixed
Estimate MCMCmean HPD95lower HPD95upper
pMCMC Pr(>|t|)
(Intercept) -0.8756 -0.6317 -2.1442 0.9813
0.4034 0.0779
H 0.0414 0.0366 -0.0518 0.1262
0.3992 0.0075
Eunerfahren -0.5255 -0.1302 -0.5614 0.3020
0.5394 0.0001
HER 3.4576 1.6306 -0.7388 4.0931
0.1766 0.0010
Gelege -0.0402 0.1261 0.0005 0.2523
0.0516 0.2143
H:Eunerfahren 0.2286 0.0506 -0.1330 0.2172
0.5524 0.0000
$random
Groups Name Std.Dev. MCMCmedian MCMCmean HPD95lower
HPD95upper
1 ID (Intercept) 0.3014 0.0320 0.0464
0.0000 0.1422
2 Residual 0.0577 0.3015 0.3076
0.2124 0.4149
m1<-lmer(Surv~H*E+HER+G+(1|ID),REML=F,data=rep1)
m2<-lmer(Surv~H+E+HER+G+(1|ID),REML=F,data=rep1)
anova(m2,m1)
#Models:
#m2: Surv ~ H + E + HER + G + (1 | ID)
#m1: Surv ~ H * E + HER + G + (1 | ID)
#Df AIC BIC logLik Chisq Chi Df Pr(>Chisq)
#m2 7 17.8965 26.703 -1.9482
#m1 8 9.4276 19.492 3.2862 10.469 1 0.001214 **
# ---
# Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1
1
m1<-lmer(Surv~H*E+HER+G+(1|ID),REML=F,data=rep1)
m3<-lmer(Surv~H*E+G+(1|ID),REML=F,data=rep1)
anova(m3,m1)
#Models:
#m3: Surv ~ H * E + Gelege + (1 | ID)
#m1: Surv ~ H * E + HER + Gelege + (1 | ID)
#Df AIC BIC logLik Chisq Chi Df Pr(>Chisq)
#m3 7 16.2181 25.025 -1.1090
#m1 8 9.4276 19.492 3.2862 8.7905 1 0.003028 **
---
# Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1
1
m1<-lmer(Surv~H*E+HER+G+(1|ID),REML=F,data=rep1)
m4<-lmer(Surv~H*E+HER+(1|ID),REML=F,data=rep1)
anova(m4,m1)
#m4: Surv ~ H * E + HER + (1 | ID)
#m1: Surv ~ H * E + HER + G + (1 | ID)
#Df AIC BIC logLik Chisq Chi Df Pr(>Chisq)
#m4 7 8.3834 17.190 2.8083
#m1 8 9.4276 19.492 3.2862 0.9559 1 0.3282
More information about the R-sig-mixed-models
mailing list