[R-sig-ME] lmer L.ratio

[Robert Kushler]

On 4/14/2011 10:22 PM, Sujal Phadke wrote:
> Cross-posting here because it was suggested so.
>
> I am using the following model
>
> model1=lmer(PairFrequency~MatingPair+(1|DrugPair)+(1|DrugPair:MatingPair), data=MateChoice, REML=F)
>
> 1. After reading https://stat.ethz.ch/pipermail/r-sig-mixed-models/2009q1/001966.html, I have learned that the above
> code is the right way to analyze a mixed model with the MatingPair as the fixed effect, DrugPair as the random effect
> and the interaction between these two as the random effect as well. Please confirm if that seems correct.

      That's correct syntax for two crossed factors, one of which is random.


>
> 2. Assuming the above code is correct, I have model2 in which I remove the interaction term, model3 in which I remove
> the DrugPair term and model4 in which I only keep the fixed effect of MatingPair.
>
> 3. I would like to know how would the same model(s), especially the interaction term, be written in lme rather than
> lmer.  (MatingPair|DrugPair) was suggested but I am under the impression that this represents nesting. Please
> confirm. *Most posts I have found deal with nesting but there is absolutely no nesting in my data*
>
> or do I have to use pdBlocked and pdIdent (Pinheiro-Bates 2000, P.163-167)...I still could not figure it out? or will
> the following do?
>
> lme(PairFrequency~MatingPair, random=~(1|DrugPair)+(1|DrugPair:MatingPair), data=MateChoice, method="ML")...is this a
> right way for lme?

      The syntax is different in lme.  The following should work:     random=~1|DrugPair/MatingPair

>
> Sujal. P. p.s: If it matters how data are arranged, then I have one vector called MatingPair which has 3 levels and
> another vector DrugPair which also has 3 levels. The PairFrequency data is a count data and is normally distributed.

      You mean "approximately" normally distributed?  By definition a count cannot be normally distributed.
      Perhaps you should be considering generalized linear mixed models.


> The data are huge, hence I am not able to post it here. Levels of DrugPair represent three randomly chosen genotypes
> from a population of many.

      Estimates of variance components with only three levels of the factor are very unreliable.


>
> lmer gives me the following output which I guess suggests that model1 gives best fit?
>
> Data: MateChoice Models: model2: PairFrequency ~ MatingPair + (1 | DrugPair) model3: PairFrequency ~ MatingPair + (1
> | DrugPair:MatingPair) model1: PairFrequency ~ MatingPair + (1 | DrugPair) + (1 | DrugPair:MatingPair)
>
>        Df    AIC    BIC  logLik  Chisq Chi Df Pr(>Chisq)
   model2  5 274.90 282.82 -132.45
   model3  5 282.44 290.36 -136.22  0.0000      0    1.00000
   model1  6 276.90 286.40 -132.45  7.5443      1    0.00602 **

  --- Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1


     Model 1 beats model 3 according to the LR test.  Change the order to test model 2 vs 1.
     Models 2 and 3 are not comparable with a test (but according to AIC/BIC 2 is best).

     Since you used "REML=FALSE" when fitting you can also test your model 4.


>
>
>
> [[alternative HTML version deleted]]
>
>
>
>
> _______________________________________________ R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models