[R-sig-ME] Very large HPDintervals, mcmcsamp, lmer

[Ben Bolker]

Colin Wahl <biowahl at ...> writes:

> I am using lmer to describe patterns in stream variables between
> streams draining watersheds with different land uses. I need to make
> some kind of determination of significance. I understand that mixed
> models cannot provide accurate p values because the denominator df is
> unknown. As per D. Bates recommendation in his post "lmer, p-values
> and all that" I used mcmcsamp and HPDinterval to get 95% "confidence
> intervals." These intervals were very large and suggest that there are
> no significant differences in variables where differences seem
> obvious. I'll use one of my variables as an example: total stream
> nitrogen. There are 12 streams and 24 sites (two riparian types per
> stream).
> 
> model<-lmer(TN ~ wsh*rip + (1|stream), data=all24)
>    Data: all24
>    AIC   BIC logLik deviance REMLdev
>  78.47 90.25 -29.24    65.15   58.47
> Random effects:
>  Groups   Name        Variance Std.Dev.
>  stream   (Intercept) 4.10443  2.0259
>  Residual             0.20885  0.4570
> Number of obs: 24, groups: stream, 12

  Sorry for all the snippage, but: since you have two riparian
sites per stream, can you try something much simpler and reduce
this to a series of paired t-tests and 1-way ANOVAs?

  You say you have 24 sites, divided into 12 streams and
two riparian sites per stream.  It sounds like streams are
nested within watersheds, so you have (an average of) 3 streams
per watershed/land use type (cultivated, D, F, G).

  If you average the two sites per stream you can do a
one-way ANOVA on the effect of land use type.

  You can do a paired t-test on riparian types within streams.

  If you take the differences within streams you can do a one-way
ANOVA on the differences across land use types to test the
interaction between land use type and riparian type.

  Modern mixed models are great but sometimes there is a simpler
and less error-prone way to do the job ...