Skip to content
Prev 2386 / 20628 Next

zero variance query

Dear Emmanuel and Ben

Many thanks for your advice. Unfortunately, I don't think that I can offset 
with log(area), given that each area is the same. My rationale for 
converting to m2 was to standardise abundances to 1 m2 as I have other 
parameters which were measured to different areas. I had previously 
attempted to normalise my data by logging but felt that it did not improve 
the distribution. I just hadn't tried it in my modelling. Logging my count 
data dramatically improved the fit of the model (AIC 116.7 v 312.5), 
however the variance still remains low. Does this appear acceptable?
Furthermore, can I assess model fit of different transformations of the 
same dataset using AIC values, i.e. compare log(Count) and inverse 
transformed Count?

lncount<-log(Count+1)
m1<-m1<-lmer(lncount~Treatment+(1|Month)+(1|Block),family=quasipoisson)
summary(m1)
Generalized linear mixed model fit by the Laplace approximation
Formula: lncount ~ Treatment + (1 | Month) + (1 | Block)
   AIC   BIC logLik deviance
 116.7 135.1 -52.33    104.7
Random effects:
 Groups   Name        Variance   Std.Dev.
 Month    (Intercept) 1.8937e-14 1.3761e-07
 Block    (Intercept) 3.5018e-02 1.8713e-01
 Residual             3.9318e-01 6.2704e-01
Number of obs: 160, groups: Month, 10; Block, 6

Fixed effects:
                   Estimate Std. Error t value
(Intercept)         -0.4004     0.1239  -3.232
Treatment2.Radiata   0.4596     0.1305   3.522
Treatment3.Aldabra   0.4295     0.1334   3.220

Correlation of Fixed Effects:
            (Intr) Trt2.R
Trtmnt2.Rdt -0.581
Trtmnt3.Ald -0.577  0.530

I used quasipoisson as my data is overdispersed. It was further improved by 
an inverse transformation (AIC 43.54). Again I have small variances.

invcount<-1/(Count+1)
m3<-lmer(invcount~Treatment+(1|Month)+(1|Block),family=quasipoisson)
summary(m3)
Generalized linear mixed model fit by the Laplace approximation
Formula: invcount ~ Treatment + (1 | Month) + (1 | Block)
   AIC BIC logLik deviance
 43.54  62 -15.77    31.54
Random effects:
 Groups   Name        Variance  Std.Dev.
 Month    (Intercept) 0.0000000 0.000000
 Block    (Intercept) 0.0021038 0.045867
 Residual             0.0926225 0.304339
Number of obs: 160, groups: Month, 10; Block, 6

Fixed effects:
                   Estimate Std. Error t value
(Intercept)        -0.51644    0.05411  -9.545
Treatment2.Radiata -0.36246    0.08401  -4.314
Treatment3.Aldabra -0.29319    0.08197  -3.577

Correlation of Fixed Effects:
            (Intr) Trt2.R
Trtmnt2.Rdt -0.566
Trtmnt3.Ald -0.580  0.372

Log(Abundance) did not solve the problem of zero variance. If quasipoisson 
errors are not acceptable to use with abundance, i.e. non-integers, is 
there a family of errors that would be recommended? Or should I simply 
multiply abundance to obtain whole numbers?

Many thanks in advance,
Christine
--On 01 June 2009 23:17 -0400 Ben Bolker <bolker at ufl.edu> wrote:

            
----------------------
Christine Griffiths
School of Biological Sciences
University of Bristol
Woodland Road
Bristol BS8 1UG
Tel: 0117 9287593
Fax 0117 925 7374
Christine.Griffiths at bristol.ac.uk
http://www.bio.bris.ac.uk/research/mammal/tortoises.html