unable to use predict on a (new) dataset

[romunov]

Lorcan Treanor reported on Stackoverflow that he is unable to use predict. I installed the dev version with the same error.

library("mboostDevel")

### fitting multinomial logit model via a linear array model
X0 <- K0 <- diag(nlevels(iris$Species) - 1)
colnames(X0) <- levels(iris$Species)[-nlevels(iris$Species)]
mlm <- mboost(Species ~ bols(Sepal.Length, df = 2) %O%
                buser(X0, K0, df = 2), data = iris,
              family = Multinomial())
round(predict(mlm, type = "response", newdata = iris), 2)

Error in `[.data.frame`(newdata, nm) : undefined columns selected 

[sbrockhaus]

The problem is that predict() does not work with newdata when buser() was used in the model.
If the model is fitted using bols() instead of buser() the prediction with newdata works.

### fitting multinomial logit model via a linear array model
X0 <- K0 <- diag(nlevels(iris$Species) - 1)
colnames(X0) <- levels(iris$Species)[-nlevels(iris$Species)]
mlm <- mboost(Species ~ bols(Sepal.Length, df = 2) %O%
                buser(X0, K0, df = 2), data = iris,
              family = Multinomial())

round(predict(mlm, type = "response", newdata = iris), 2)
pred <- round(predict(mlm, type = "response"), 2)

## set up data that contain a dummy-variable 
## thus you can use bols() instead of buser()
myiris <- as.list(iris)
myiris$dummy <- factor(1:2)

## fitting the model without buser()
mlm2 <- mboost(Species ~ bols(Sepal.Length, df = 2) %O%
                bols(dummy, df = 2, contrasts.arg = "contr.dummy"), data = myiris,
              family = Multinomial())

pred2 <- round(predict(mlm2, type = "response", newdata = myiris), 2)

## compare the predictions of the two models
all(pred2==pred)

### compare design and penalty matrices of the two models
extract(mlm, "design")[[1]][2]
extract(mlm, "penalty")

extract(mlm2, "design")[[1]][2]
extract(mlm2, "penalty")

## look at offset 
## mlm2$offset

[hofnerb]

Thanks, @sbrockhaus.

Just one addition: To make predictions for really new data, one needs to always keep the dummy as is. E.g. we can make a prediction only for the first observation as follows:

newdata <- as.list(iris[1,])
## define a dummy vector with one factor level less than the outcome
newdata$dummy <- factor(1:(nlevels(iris$Species) - 1))
pred3 <- round(predict(mlm2, type = "response", newdata = newdata), 2)

## check results
pred2[1,]
pred3