[R-sig-ME] Dealing with overdispersion with glmmadmb beta distribution
Heather Kharouba
kharouba at zoology.ubc.ca
Sun Jul 8 22:52:25 CEST 2012
Dear R users,
In the analysis I'm doing, I'm interested in testing the importance of a
factor (with 5 levels). The response variable varies continuously from 0
to 1 so I've used a beta distribution with glmmadmb:
m1<-glmmadmb(AUC~variables+log_area+model+taxa+(1|study), family="Beta",
verbose=TRUE, debug=TRUE, data=mat);
Call:
glmmadmb(formula = AUC2 ~ variables + log_area + model2 + taxa +
(1 | study), data = mat, family = "Beta", verbose = TRUE,
debug = TRUE)
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -1.1568 1.3899 -0.83 0.4052
variables 0.0248 0.0112 2.22 0.0264 *
log_area 0.1976 0.0859 2.30 0.0215 *
model2autologistic regression 2.2526 0.7345 3.07 0.0022 **
model2domain -0.3246 0.5753 -0.56 0.5726
model2GAM 0.1757 0.4768 0.37 0.7125
model2GARP -0.2985 0.5626 -0.53 0.5957
model2gdm -0.5436 0.5691 -0.96 0.3395
model2GLM -0.3243 0.5860 -0.55 0.5800
model2localwghtregression 0.6689 0.4780 1.40 0.1617
model2logregression -0.5179 0.6661 -0.78 0.4369
model2maxent 0.0182 0.5353 0.03 0.9728
model2other -0.4466 0.5114 -0.87 0.3825
taxaHER 0.3200 0.1024 3.12 0.0018 **
taxaINV 0.6967 0.1591 4.38 1.2e-05 ***
taxaMAM -0.1792 0.0859 -2.09 0.0368 *
taxaP 0.1230 0.0616 2.00 0.0459 *
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Number of observations: total=4317, study=20
Random effect variance(s):
Group=study
Variance StdDev
(Intercept) 0.328 0.5727
Beta dispersion parameter: 14.503 (std. err.: 0.35671)
Log-likelihood: 9167.05
The dispersion parameter is 14.503 suggesting the model is overdispersed.
A recent suggestion from http://glmm.wikidot.com/faq and
r-sig-mixed-models archives to account for overdispersion is to include
individual-level random effects. However, if I include this additional
random effect, I get the following error:
Error in glmmadmb(AUC2 ~ variables + log_area + model2 + taxa + (1 |
study) + :
The function maximizer failed (couldn't find STD file)
suggesting that the model cannot be estimated. I'm wondering whether there
are other methods to reduce overdispersion I might have overlooked that
could affect the variance-mean relationship and that would be appropriate
for this type of response variable.
A snapshot of the data:
study taxa AUC model variables log_area Araujo
et al. 2005 BIRD 0.9156878 GAM 7 16.21771
Araujo et al. 2005 BIRD 0.9288596 GAM 7 16.21771
Araujo et al. 2005 BIRD 0.9254065 GAM 7 16.21771
Araujo et al. 2005 BIRD 0.8825593 GAM 7 16.21771
Araujo et al. 2005 BIRD 0.9388894 GAM 7 16.21771
Araujo et al. 2005 BIRD 0.9061483 GAM 7 16.21771
Thanks in advance!
Heather Kharouba
More information about the R-sig-mixed-models
mailing list