Skip to content
Prev 6717 / 7420 Next
Original Raw

GLMM: AIC worsens but BIC improves

Alexandre F. Souza 
Dear all,

First I would like to say thank you for replying to my last posts, your
replies have halped me a lot and complemented the formal sources I have
been consulting.

I would like very much your opinion on a further issue below, if possible.

I am using glmmTMB to find the fixed and random variables that influence
whether a?ai palms produce fruit or not in the Amazon. So far so good, I
ended upt with a much simpler model that basically contains the year (2
years only), habitat type (3 levels, Grupos_Finais below) and density of
adult palms (R_dens below). This is model 8.

When I try to further simplify this model reducing it to only year +
habitat or year alone, I get an worsened AIC but an improved BIC... When I
fit model 8 with REML, the significance of adult palm density disappear (I
have been making model comparison manually with ML untill now). Together,
these results make me pend to exclude habitat type and adult density, but I
am intrigued with worsened AIC.

What do you think?

modelo8.glmm = glmmTMB(reproducao_bin ~ 0 + Grupos_Finais + ano_medida +
R_dens_total + (1|numero) + (1 |bloco/parcela), family = binomial, data =
dados)
Family: binomial  ( logit )
Formula:
reproducao_bin ~ 0 + Grupos_Finais + ano_medida + R_dens_total +
    (1 | numero) + (1 | bloco/parcela)
Data: dados

     AIC      BIC   logLik deviance df.resid
  1665.5   1707.6   -824.8   1649.5     1420

Random effects:

Conditional model:
 Groups        Name        Variance Std.Dev.
 numero        (Intercept) 3.0944   1.7591
 parcela:bloco (Intercept) 0.7850   0.8860
 bloco         (Intercept) 0.7421   0.8615
Number of obs: 1428, groups:  numero, 762; parcela:bloco, 147; bloco, 10

Conditional model:
                                          Estimate Std. Error z value
Grupos_FinaisDisturbed Unflooded Forests  -0.01490    0.57713  -0.026
Grupos_FinaisFlooded Open Forests          1.92320    0.46866   4.104
Grupos_FinaisOld Growth Unflooded Forests  1.11360    0.72798   1.530
ano_medida2017                            -1.73544    0.18708  -9.276
R_dens_total                              -0.03749    0.01542  -2.430
                                          Pr(>|z|)
Grupos_FinaisDisturbed Unflooded Forests    0.9794
Grupos_FinaisFlooded Open Forests         4.07e-05 ***
Grupos_FinaisOld Growth Unflooded Forests   0.1261
ano_medida2017                             < 2e-16 ***
R_dens_total                                0.0151 *
---
Signif. codes:  0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1

modelo9.glmm = glmmTMB(reproducao_bin ~ 0 + Grupos_Finais + ano_medida +
(1|numero) + (1 |bloco/parcela), family = binomial, data = dados)

modelo10.glmm = glmmTMB(reproducao_bin ~ ano_medida + (1|numero) + (1
|bloco/parcela), family = binomial, data = dados)
modelo8.glmm   8 1665.501
modelo9.glmm   7 1669.163
modelo10.glmm  5 1671.159
modelo8.glmm   8 1707.613
modelo9.glmm   7 1706.011
modelo10.glmm  5 1697.479


Perhaps report bot AIC and BIC and decide for the simpler model?... Thank
you very much for any inputs,

Cheers,

Alexandre