bug in predict for two-class cases?

[luca-scr]

When computing predictions for a two-class case there seems to be a mistake.

Here is a reproducible example:

library(polyreg)
library(MLmetrics)

data(kyphosis, package = "rpart")
kyphosis$y <- ifelse(kyphosis$Kyphosis == "absent", 1, 0)
kyphosis$Kyphosis <- NULL
mod <- glm(y ~ ., data = kyphosis, family = binomial())
mod
# Coefficients:
# (Intercept)          Age       Number        Start  
#     2.03693     -0.01093     -0.41060      0.20651  
#
# Degrees of Freedom: 80 Total (i.e. Null);  77 Residual
# Null Deviance:	    83.23 
# Residual Deviance: 61.38 	AIC: 69.38
table(ifelse(predict(mod, type = "response") > 0.5, 1, 0), kyphosis$y)
#    0  1
# 0  7  3
# 1 10 61
Accuracy(ifelse(predict(mod, type = "response") > 0.5, 1, 0), kyphosis$y)
# 0.8395062
table(ifelse(predict(mod) > 0.5, 1, 0), kyphosis$y)
#    0  1
# 0 10  8
# 1  7 56
Accuracy(ifelse(predict(mod) > 0.5, 1, 0), kyphosis$y)
# 0.8148148

data(kyphosis, package = "rpart")
kyphosis <- kyphosis[,c(2:4,1)]
kyphosis$Kyphosis <- as.character(kyphosis$Kyphosis)
pf <- polyFit(kyphosis, deg = 1, use = "glm")
pf$fit 
# Coefficients:
# (Intercept)           V1           V2           V3  
#     2.03693     -0.01093     -0.41060      0.20651  
# 
# Degrees of Freedom: 80 Total (i.e. Null);  77 Residual
# Null Deviance:	    83.23 
# Residual Deviance: 61.38 	AIC: 69.38

Ok the same model is fitted, but computing predictions:

table(predict(pf, kyphosis), kyphosis$Kyphosis)
#         absent present
# absent      56       7
# present      8      10
Accuracy(predict(pf, kyphosis), kyphosis$Kyphosis)
# 0.8148148

seems to be wrong. Looking at the code you can see

# glm case
  if (is.null(object$glmMethod)) { # only two classes
    pre <- predict(object$fit, plm.newdata)
    pred <- ifelse(pre > 0.5, object$classes[1], object$classes[2])
  } 

IMHO the prediction returned is in the link scale (see help(predict.glm)) but it should be on the probability scale, i.e. type = "response", or if in the link scale pre > 0. However, I prefer to resonate in terms of probability scale.