[R-sig-ME] unbalanced data in nested lmer model

[Jana Bürger]

Dear Christian,

thank you for your answer, it was helpful and does reflect my own ways 
of thought.
Now you point out that the random effects are somewhat irrelevant 
becuase the fixed effects are the thing of interest.
In fact, the region and the farm are not just features of my study 
design but have practical relevance for the response variable. Just I 
don't want to know the exact parameters for the different regions and farms.

Am I wrong thinking that the random effects part of the model makes it 
possible to compare magnitude of influence of the region and the farm, 
and also of the sum of my fixed predictor variables (as given in Residuals)?

In this context, what are the consequences of the sructure of my data, 
with a couple of regions with only one farm?

Any more helpful thoughts are appreciated.

Christian Ritz schrieb:
> Dear Jana,
> 
> let me try to answer your questions:
> 
> 1) No there are no restrictions on how much imbalance there can be. In fact, it's one of
> the advantages of the mixed model approach!
> 
> However, cells with 0 cases will not be used in the model fit. The mixed model will
> "automatically" take the varying number of cases into account in estimates and standard
> errors.
> 
> 2) The imbalance in several random factors will indirectly influence the standard errors
> of the fixed effects estimates (as the fixed effects estimates are obtained through a
> weighted least squares approach with weights reflecting the random effects structure in
> the model).
> 
> As the structure of your data is hierarchical with fields, farms, and regions it makes
> sense to me to have random effects
> 
> (1|region) and (1|region:farm)
> 
> 
> in the model. Actually I don't see the point in testing a feature of the data that was
> imposed by the design of the study or experiment. Just leave both terms in the models and
> proceed to evaluate the fixed effects.
> 
> Random effects reflect structure in the data that is imposed by how the data were
> collected or the underlying experiment designed and therefore it's rarely relevant to test
> this part of the model (it's simply the framework within which we examine some interesting
> explanatory variable). I guess, however, that other statisticians might a different
> opinion about this issue.
> 
> I hope this explanation is useful?!
> 
> Christian
> 

-- 
Jana Bürger

Universität Rostock
Agrar-  und Umweltwissenschaftliche Fakultät
FG Phytomedizin
Satower Straße 48
18059 Rostock

Tel. 0381-498 31 71
Fax.0381-498 31 62