[R-sig-eco] Question regarding a associating a p-value with the residual variance form a lme mod

Mike Dunbar mdu at ceh.ac.uk
Tue Nov 18 12:16:52 CET 2008


Dear Charlies

I'm the first to admit that I don't entirely understand what lme is doing with its AR structures, but I think a key point is do you have evidence that there is any additional temporal correlation in observations above the correlation between samples from the same tree which is accounted for by the tree random effect, and which is modelled by having a time variable as a fixed effect. You can plot residuals through time for each tree from a model without an AR term to check this, and you'd need to be seeing consistent patterns of under or over prediction.

I also worry about the regular correlation structure if there are any missing observations in the data. I know that Pinhero and Bates note that missing points are allowed, but I'm really not clear how they are handled. Hence a CAR structure seems a bit easier to understand, but you must explicitly tell lme the continuous variable that indicates the (non-integer) time position.

regards

Mike




>>> "Charles Nock" <charles.nock at gmail.com> 11/11/2008 14:42 >>>
Thanks for the reply Mike,

Perhaps the common thread is that in the two papers where I saw a p-value
reported SAS proc mixed was used.  See Table 6. in Uzoh and Oliver 2008.
For. Ecol. Man.,

For the correlation structure, I am trying to model autocorrelation among
years, due to repeated measurements of tree rings from the core samples
taken from each tree. So measurements of wood density close in time are
likely correlated. I think the within group order would be proper as it
would be the same as the order of year (or tree age if you like, logyearcorr
in the model). Would the regular AR be more appropriate ?

On Tue, Nov 11, 2008 at 2:55 PM, Mike Dunbar <mdu at ceh.ac.uk> wrote:

> Dear Charles
>
> It's very difficult to tell without more information on what these papers
> are and what exactly they are saying.
>
> Also with your correlation structure, what are you trying to model? corCAR
> is a continuous AR structure, allowing fractional time points, but you don't
> have a time variable specified, so lme will be using the within-group
> position of the observation in the dataset, which can only be a integer.
>
> regards
>
> Mike
>
>
> >>> Charles Nock <charles.nock at gmail.com> 11/11/2008 11:47 >>>
> I have a lme model of radial variation in wood density which includes
> the fixed effects of the year of the tree ring, annual increment
> (width of the ring), a year*increment interaction, and random effects
> for differences in the slope and intercept of the change in wood
> density with year from tree to tree (see below).
>
> I understand that p-values can be calculated for the individual random
> effects by a likelihood ratio test of models differing by the term of
> interest.
>
> In some papers I have seen a p-value calculated for the residual,
> which I do not understand how to obtain, or what it is telling you.
>
> thanks,
> Charles
> _____________________________
>
> Charles Nock, M.Sc.F
> Doctoral candidate
> Institute of Botany
> University of Natural Resources and
> Applied Life Sciences Vienna
>
>
>
> agexincrement.lme.corr <- lme(densmean ~
> logyearcorr*increment_distrange_cm , random=~ logyearcorr | Tree,
> data=memasterfinalage, method="ML", correlation=corCAR1(form=~1| Tree))
>
>  > summary(agexincrement.lme.corr )
>
> Linear mixed-effects model fit by maximum likelihood
>  Data: memasterfinalage
>        AIC      BIC    logLik
>   3122.368 3154.984 -1552.184
>
> Random effects:
>  Formula: ~logyearcorr | Tree
>  Structure: General positive-definite, Log-Cholesky parametrization
>             StdDev    Corr
> (Intercept) 138.56419 (Intr)
> logyearcorr 115.58319 -0.833
> Residual     64.10158
>
> Correlation Structure: Continuous AR(1)
>  Formula: ~1 | Tree
>  Parameter estimate(s):
>       Phi
> 0.3360245
> Fixed effects: densmean ~ logyearcorr * increment_distrange_cm
>                                        Value Std.Error  DF   t-value
> p-value
> (Intercept)                         450.5252  60.91811 263  7.395587
> 0e+00
> logyearcorr                         177.8543  50.64480 263  3.511799
> 5e-04
> increment_distrange_cm             -124.2732  34.50346 263 -3.601761
> 4e-04
> logyearcorr:increment_distrange_cm  116.2502  33.51491 263  3.468610
> 6e-04
>  Correlation:
>                                    (Intr) lgyrcr incr__
> logyearcorr                        -0.901
> increment_distrange_cm             -0.498  0.495
> logyearcorr:increment_distrange_cm  0.397 -0.443 -0.942
>
> Standardized Within-Group Residuals:
>           Min            Q1           Med            Q3           Max
> -2.0842970533 -0.5809674446  0.0003970132  0.6389420030  2.7177642650
>
> Number of Observations: 277
> Number of Groups: 11
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>         [[alternative HTML version deleted]]
>
> _______________________________________________
> R-sig-ecology mailing list
> R-sig-ecology at r-project.org 
> https://stat.ethz.ch/mailman/listinfo/r-sig-ecology 
>
>
> --
> This message (and any attachments) is for the recipien...{{dropped:15}}



More information about the R-sig-ecology mailing list