Boa tarde pessoal,
Tenho uma váriavel resposta binomial chamada Infestação
(1 presença e 0 ausência de insetos) e dois fatores qualitativos que são
vegetação (pastagem e floresta) e altitude (alta e baixa), além de parcelle que
são as repetições no campo. Gostaria de fazer contrastes utilizando glm (family=binomial),
então primeiro eu fiz:
> m.full<-glm(Infestacao~Altitude*Vegetation*Parcelle,
data=H.termes, family="binomial")# Modelo completo
Mensagens de aviso perdidas:
In model.matrix.default(mt, mf, contrasts)
:
variable 'Altitude' converted to a factor
Já fiquei preocupado com essa mensagem, mais de qualquer
maneira Altitude é um fator mesmo, então fiz:
> summary(m.full)
Call:
glm(formula = Infestacao ~ Altitude * Vegetation * Parcelle,
family = "binomial", data =
H.termes)
Deviance Residuals:
Min 1Q Median 3Q
Max
-1.0793 -0.4643 -0.4542 -0.2530
2.6296
Coefficients:
Estimate
Std. Error z value Pr(>|z|)
(Intercept) -3.2454 0.6030
-5.382 7.36e-08 ***
Altitudebaixa 3.0243
0.6322 4.784 1.72e-06 ***
Vegetationpature 0.9997
0.9678 1.033 0.30161
Parcelle 0.5115
0.2564 1.995 0.04602 *
Altitudebaixa:Vegetationpature -3.0214 1.0505
-2.876 0.00403 **
Altitudebaixa:Parcelle -0.5256 0.2669
-1.969 0.04893 *
Vegetationpature:Parcelle -0.9048 0.4561
-1.984 0.04729 *
Altitudebaixa:Vegetationpature:Parcelle 0.9421
0.5024 1.875 0.06073 .
---
Signif. codes: 0 ‘***’ 0.001
‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’
1
(Dispersion parameter for binomial family
taken to be 1)
Null deviance: 1578.6 on 1532 degrees
of freedom
Residual deviance: 1317.2 on 1525 degrees
of freedom
AIC: 1333.2
Number of Fisher Scoring iterations: 5
Não tem superdispersão então fiz uma anova:
> anova(m.full,test="Chi")
Analysis of Deviance Table
Model: binomial, link: logit
Response: Infestacao
Terms added sequentially (first to last)
Df Deviance
Resid. Df Resid. Dev P(>|Chi|)
NULL 1532
1578.6
Altitude 1 94.613 1531
1484.0 < 2e-16 ***
Vegetation 1 154.761 1530
1329.2 < 2e-16 ***
Parcelle 1 0.248 1529
1329.0 0.61866
Altitude:Vegetation 1 6.511 1528
1322.5 0.01072 *
Altitude:Parcelle 1 1.195 1527 1321.3
0.27440
Vegetation:Parcelle 1 0.491 1526
1320.8 0.48328
Altitude:Vegetation:Parcelle 1 3.577 1525
1317.2 0.05857 .
---
Signif. codes: 0 ‘***’ 0.001 ‘**’
0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Mensagens de aviso perdidas:
In model.matrix.default(object, data =
list(Infestacao = c(0, 0, :
variable 'Altitude' converted to a factor
Então fui reduzindo o modelo ate restarem as variáveis
significativas, sendo:
> anova(m3,test="Chi")
Analysis of Deviance Table
Model: binomial, link: logit
Response: Infestacao
Terms added sequentially (first to last)
Df Deviance Resid. Df
Resid. Dev P(>|Chi|)
NULL 1532
1578.6
Altitude 1 94.613 1531 1484.0 <
2e-16 ***
Vegetation 1 154.761 1530 1329.2 <
2e-16 ***
Parcelle 1 0.248 1529 1329.0
0.61866
Altitude:Vegetation 1 6.511 1528 1322.5
0.01072 *
---
Signif. codes: 0 ‘***’ 0.001 ‘**’
0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Mensagens de aviso perdidas:
In model.matrix.default(object, data =
list(Infestacao = c(0, 0, :
variable 'Altitude' converted to a factor
> summary(m3)
Call:
glm(formula = Infestacao ~ Altitude + Vegetation + Parcelle
+
Altitude:Vegetation, family =
"binomial", data = H.termes)
Deviance Residuals:
Min 1Q Median 3Q
Max
-1.0824 -0.4636 -0.4554 -0.3072
2.4807
Coefficients:
Estimate
Std. Error z value Pr(>|z|)
(Intercept) -2.20627 0.23768 -9.283
< 2e-16 ***
Altitudebaixa 1.91926 0.21846 8.785
< 2e-16 ***
Vegetationpature -0.83836
0.36768 -2.280 0.02260 *
Parcelle 0.01484
0.06585 0.225 0.82173
Altitudebaixa:Vegetationpature -1.10337
0.41007 -2.691 0.00713 **
---
Signif. codes: 0 ‘***’ 0.001
‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’
1
(Dispersion parameter for binomial family
taken to be 1)
Null deviance: 1578.6 on 1532 degrees
of freedom
Residual deviance: 1322.5 on 1528 degrees
of freedom
AIC: 1332.5
Number of Fisher Scoring iterations: 5
Então posso deduzir, se é que agora fiz tudo certo, que
a variável resposta é diferente entre altitude alta e baixa e entre vegetação de
pastagem e floresta, agora gostaria de comparar altitude alta com as
diferentes vegetações, só que não entendi porque a análise separou Vegetationbaixa
e Vegetationpaturage e também não sei que fator a análise fixou no intercepto,
Alguém saberia me explicar o que posso estar fazendo de
errado,
Obrigado,
Alexandre dos Santos
Ingenieur forestier, Msc.
INRA- Biostatistique et Processus Spatiaux (BioSP)
Domaine Saint-Paul
Site Agroparc
84914 - Avignon - France
Tél. : +33 (0)6 87 95 16 29