Dans une régression polytomique dans laquelle la variable explicative est à plusieurs classes, comment peut-on tester l’effet global ? Je pensais qu’on pouvait utiliser un drop1, mais il me donne des résultats complètements différents d’un chi2…
Ici, un petit exemple, le drop1 donne en global 0.2569 alors que le chi2 est de 0.5769… Quelqu’un sait ou ça coince ?
Code : Tout sélectionner
### Au passage, tres cool cette intruction pour poster... Merci Pierre
options( prompt=" ")
### Préparation des données
f<-c("O","O","O","N","O","N","O","N","O","N","O","N","O","N","O","N","O","N")
fo<-c("OO","OO","OO","N","OO","N","OO","N","OO","N","OO","N","OO","N","O","N","OO","N")
m<-c("C","C","S","MS","S","MS","MS","S","S","C","C","C","MS","S","S","MS","S","C")
m<-ordered(c(m,m),levels=c("S","C","MS"))
Dn <- data.frame(f=c(f,fo),m=m)
table(Dn)
m
f S C MS
N 4 6 6
O 6 3 2
OO 4 3 2
### Regression polytomique
summary(polr(m~f,data=Dn)->polResult)
Re-fitting to get Hessian
Call:
polr(formula = m ~ f, data = Dn)
Coefficients:
Value Std. Error t value
fO -1.1714844 0.7556789 -1.550241
fOO -0.8013447 0.7787466 -1.029018
Intercepts:
Value Std. Error t value
S|C -1.0316 0.5118 -2.0155
C|MS 0.4656 0.4834 0.9632
Residual Deviance: 75.71249
AIC: 83.71249
### drop1 pour tester l'effet global
drop1(polResult,.~.,test="Chisq")
Single term deletions
Model:
m ~ f
Df AIC LRT Pr(Chi)
<none> 83.712
f 2 82.430 2.718 0.2569
### chi2 global
chisq.test(table(Dn))
Pearson's Chi-squared test
data: table(Dn)
X-squared = 2.7721, df = 4, p-value = 0.5967
Warning message:
l'approximation du Chi-2 est peut-être incorrecte in: chisq.test(table(Dn))