[R-sig-ME] Massive difference in random effect's posterior mean between chains

[Ronan James Osullivan]

Hi all,

Apologies if this question is trivial or if I'm over-looking something
obvious.

I am trying to investigate the effect of Genetic Type (A or B) and climate
on the lifetime reproductive success (LRS) of a wild-spawning fish species.
I am interacting Genetic Type with seven different climatic variables. I
have included a nested random effects structure to account for
pseudoreplication among years (~Year_of_Spawning), and to account for some
of the fish spawning across multiple years (~Year_ID2). I also interact
Genetic_Type with the relative survival of B compared to A to correct for a
survival bias between the two genetic types (they are sampled at different
points in the spawning season).

The model is specified in MCMCglmm as follows:
Model_a<- MCMCglmm(LRS ~
                   (Genetic_Type-1)*climate1+
                   (Genetic_Type -1)* climate2+
                   (Genetic_Type -1)* climate3+
                   (Genetic_Type -1)* climate4+
                   (Genetic_Type -1)* climate5+
                   (Genetic_Type -1)* climate6+
                   (Genetic_Type -1)* climate7+
                    Genetic_Type *relative_survival,
                 random = ~Year_of_Spawning+ Year_ID2,
                 family = "poisson",
                 data = data1,
                 prior = prior.exp_1,
                 nitt = 110000,
                 burnin = 10000,
                 thin = 100)

The above model was run to illustrate the problem, hence the low number of
iterations.

The parameter expanded prior is as follows:
prior.exp_1<- list(R = list(V = 1, nu = 2),
G = list(G1 = list(V = 1, nu = 0.002, alpha.mu = 0, alpha.V = 1000),
             G2 = list(V = 1, nu = 0.002, alpha.mu = 0, alpha.V = 1000)))

If I run 2 chains of the model, then the results for the fixed effects are
similar, allowing for Monte Carlo error. However, the posterior mean
estimate for the 'Year_of_Spawning' random effect changes massively, with
the posterior mean larger than the upper 95% credible interval:

MODEL 1:
 G-structure:  ~Year_of_Spawning

                                post.mean          l-95% CI     u-95% CI
 eff.samp
Year_of_Spawning      5.49              7.832e-05      15.85         1000

               ~Year_ID2

                               post.mean          l-95% CI       u-95% CI
 eff.samp
Year_ID2                  0.5262              0.000721         0.9471
87.03

 R-structure:  ~units

                             post.mean            l-95% CI       u-95% CI
 eff.samp
units                        0.5792               0.1383            1.082
      87.61

MODEL 2:
G-structure:  ~Year_of_Spawning

                               post.mean         l-95% CI         u-95% CI
 eff.samp
Year_of_Spawning     39.38             9.65e-05            15.52     1000

               ~Year_ID2

                              post.mean         l-95% CI          u-95% CI
  eff.samp
Year_ID2                 0.4724             4.8e-06              0.9168
    81.07

 R-structure:  ~units

                             post.mean         l-95% CI          u-95% CI
  eff.samp
units                        0.6403             0.1838             1.126
      92.06


I imagine that the problem exists in my random effects structure but I am
unsure of where the error lies. Any and all help is massively appreciated.

Cheers,
Ronan


-- 
Ronan O'Sullivan | Ph.D student | School of Biological, Earth and
Environmental Sciences, University College Cork, Ireland |
http://fisheye.ucc.ie/toms-team/

EMPSEB 26 - Chair of Organizing Committee

Irish Ecological Association - Ordinary Committee Member

	[[alternative HTML version deleted]]