lcayuela at ugr.es
2009-Jan-07 18:08 UTC
[R] how to estimate overdispersion in glmer models?
Dear all, I am using function glmer from package lme4 to fit a generalized linear mixed effect model. My model is as follows: model1 <- glmer(fruitset ~ Dist*wire + (1|Site), data, binomial) summary(model1) Generalized linear mixed model fit by the Laplace approximation Formula: fruitset ~ Dist * wire + (1 | Site) Data: data AIC BIC logLik deviance 68.23 70.65 -29.11 58.23 Random effects: Groups Name Variance Std.Dev. Lugar (Intercept) 3.5155e-14 1.8750e-07 Number of obs: 12, groups: Lugar, 2 Fixed effects: Estimate Std. Error z value Pr(>|z|) (Intercept) -2.332132 0.856518 -2.723 0.006473 ** Dist 0.001137 0.001141 0.997 0.318902 WireControl 4.710750 1.196550 3.937 8.25e-05 *** Dist:WireControl -0.006180 0.001769 -3.494 0.000475 *** --- Signif. codes: 0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1 Correlation of Fixed Effects: (Intr) Dist WirCnt Dist -0.963 WireContrl -0.716 0.689 Dst:WirCntr 0.621 -0.645 -0.957 My question is, how can I check for overdispersion? In glm models you can check this by comparing the residual deviance with the residual degrees of freedom, but in glmer you don't get this information. Does anyone know how to get information about overdispersion in the model? Thanks in advance for your help, (Ubuntu Intrepid Ibex / R 2.7.1) Luis Cayuela Investigador Post-doctoral Grupo de Ecolog?a Terrestre Departamento de Ecolog?a Centro Andaluz de Medio Ambiente, Universidad de Granada - Junta de Andaluc?a Avda. del Mediterr?neo S/N. 18006. Granada. Espa?a email: lcayuela at ugr.es Fax: +34 958137246 Tel: +34 958241000 (ext. 31202)
<lcayuela <at> ugr.es> writes: [snip]> model1 <- glmer(fruitset ~ Dist*wire + (1|Site), data, binomial) > summary(model1) > > Generalized linear mixed model fit by the Laplace approximation > Formula: fruitset ~ Dist * wire + (1 | Site) > Data: data > AIC BIC logLik deviance > 68.23 70.65 -29.11 58.23 > Random effects: > Groups Name Variance Std.Dev. > Lugar (Intercept) 3.5155e-14 1.8750e-07 > Number of obs: 12, groups: Lugar, 2[snip]> My question is, how can I check for overdispersion? In glm models you can > check this by comparing the residual deviance with the residual degrees of > freedom, but in glmer you don't get this information. > > (Ubuntu Intrepid Ibex / R 2.7.1)a few thoughts -- (1) probably better to ask this question on the R-sig-mixed-models list, which specializes in these problems (2) try lme4:::sigma (3) do you really have just 12 observations in 2 groups? In that case I would strongly recommend just treating group as a fixed factor -- you have no power to estimate variance (note your random effect has a standard deviation of 2 x 10^-7), and you will avoid lots of heartache if you just fit glm(fruitset ~ Dist*wire + Site, data, binomial) [not everyone will agree with me about this ...] (4) I'm a little puzzled that your formula has "Site" as a random effect but your summary lists "Lugar" as a random effect. Did you edit the summary?
Apparently Analagous Threads
- Basic Help with Zoo objects and trading days
- A questionb about the Wilcoxon signed rank test
- use vector to merge multiple xts objects?
- lmer overdispersion
- Is there a function in R to evaluate the adjusted AIC or other statistc where overdispersion existed in GLMs?