Skip to content

using residuals of binomial GLM

1 message · Yuval Sapir

Yuval Sapir
# 
Hi all,
This is more a question in statistics, but I hope to get also the R 
practice for my question:
I have an ancova model where the response variable is flowering (plant 
has a flower = 1, no flower = 0). The explanatory variables are leaf 
length, leaf thick (both continuous variables), and soil type (factorial 
with three levels):
 > model<-glm(flower~(thick+length)*soil,family="binomial")
 >summary(aov(model))
In the aov summary I find a significant effect of all variables, and a 
significant interaction between thick and soil, so I want to explore 
this interaction after "cleaning" the effect of length. I thought of two 
possible ways to extract the residuals:
 > res.thick<-resid(update(model,~.-thick-soil-thick:soil))
or:
 > res.thick<-resid(glm(flower~length+length:soil,family="binomial"))
I validated that the two methods give the same results. Anyhow, now I 
want to compare the effect of thick on flowering probability,separately 
for each soil. But the residuals extracted are not 0 or 1 anymore.  
Linear glm, such as
 > model1<-glm(res.thick1~thick*soil)
doesn't seem to be right, and, moreover, I am interested in the 
estimated coefficients and their interpretation (say - plotting a 
meaningful graph). How can I get a logistic regression from residuals? 
Do I NEED logistic regression? How should I understand the coefficients 
I get in summary of the residuals model? How can I use the results of 
the residuals model for plotting the separate lines for the probability 
(logistic) curve?
Thanks in advance
Yuval