[R-sig-ME] Unrealistic coefficient values from an MCMCglmm mixed model

[Ben Bolker]

  Hard to say without more information, but it also looks like you have
extremely wide confidence  on your GeneticTypeA estimate (-4.5345,
15.9627).  A few questions/things that look fishy:

* MCMCglmm is reporting results for both "GeneticTypeA" and
"GeneticTypeB", which suggests that it is using a *third* level (maybe
some sort of blank level/typo?) as the baseline.  What is
levels(data$GeneticType) (or table(data$GeneticType) ?)

* is one of your variables (e.g. Temp) a continuous predictor whose mean
is far from zero, in which case the main effects will be reported at an
unrealistic level?


On 2019-04-01 7:56 a.m., Ronan James Osullivan wrote:
> Dear forum,
> 
> I am struggling with the interpretation of the coefficients from a glmm
> implemented using MCMCglmm. My data set has lifetime reproductive success
> (LRS) for individual fish and associated climatic variables (NAO and
> Temperature) are indexed to each fish. I also know the genetic type of each
> fish (A or B). In total, I have 2938 observations with 1321 A fish and 1671
> B fish.
> 
> I ran the following model:
> 
> model<- MCMCglmm(LRS~GeneticType*NAO+
>                              GeneticType *Temp,
>                           random = ~Year_of_Spawning,
>                           family = "poisson",
>                           data = data,
>                           verbose = TRUE,
>                           nitt = 1010000, burnin = 1000, thin = 1000)
> 
> Which gave the following summary:
> 
>  G-structure:  ~Year_of_Spawning
> 
>                                  post.mean     l-95% CI        u-95% CI
>   eff.samp
> Year_of_Spawning       0.792            0.1521            1.934
> 1111
> 
>  R-structure:  ~units
> 
>                               post.mean         l-95% CI        u-95% CI
>  eff.samp
>        units                  1.257                 1.029
> 1.488         1447
> 
>  Location effects: LRS ~  GeneticType *NAO + GeneticType * Temp-1
> 
>                                      post.mean      l-95% CI      u-95% CI
>    eff.samp    pMCMC
> GeneticTypeA                 5.5510        -4.5345        15.9627
>  1009.0      0.2577
> GeneticTypeB                -2.8334      -10.6020         8.0720
> 1009.0     0.4916
> NAO                                0.6995        -0.6520         1.9512
>         918.3     0.2220
> Temp                              -8.2807      -18.7522         1.8865
>     1009.0     0.0991 .
> GeneticTypeB:NAO       -0.7729         -1.2302       -0.3119
>  1009.0   <0.001 ***
> GeneticTypeB:Temp       9.8697          4.7849        15.4317
>  1009.0   <0.001 ***
> 
> My issue is that the predicted LRS values for Genetic Type A are far too
> high. The intercept for A fish is 5.551, and exp(5.551) = expected mean for
> H fish when Temp=0 and NAO = 0 is 244.7. When I solve for LRS for Type A
> fish at a given NAO or temperature (holding the other one constant), I get
> incredibly high values.
> 
> The predicted LRS for Genetic Type B is exp(-2.8334)=0.05881255 which is
> far more realistic for my study system
> 
> Am I somehow mis-specifying the model or mis-calculating LRS with respect
> to each genetic type?
> 
> Cheers,
> Ronan
> 
> 
>