Prediction using mboost multinomial logistic regression in R

Question

I am trying to use the mboost package in R to apply a multinomial logistic regression model. I found this example online but I added the "newdata = iris" in the predict function to see how the prediction formula worked in mboost for new data. I am getting an error however. First here is the code:

library(mboost)

### 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)

The error I'm getting is as follows:
Error in [.data.frame(newdata, nm) : undefined columns selected

I just re-used the iris data in the prediction just as a test but has anyone experienced this problem before?

Aye, that was the unedited example taken from the web. I'm looking to predict on new data. I used the iris dataset as a "new" dataset just for test purposes so it is my fault for the confusion. — Lorcan Treanor, Jul 23 '15 at 08:38
I submitted an issue with the developer on [github](https://github.com/hofnerb/mboost/issues/13). — Roman Luštrik, Jul 23 '15 at 08:48

score 3 · Accepted Answer · edited Jun 20 '20 at 09:12

3

The reason why you cannot use the predict function with new data is that you use pre-defined design and penalty matrices in buser(), i.e., X0 and K0. These are not part of the new data set and thus are not available when building new design matrices for prediction.

Sarah Brockhaus postet a solution on github that replaces buser with bols. To do this one needs to convert the data set to a list and add the new dummy to this list. If one really wants to predict with new data, one needs to keep this dummy untouched. See also my post on github.

[Edit] As @Lorcan-Treanor mentions in his comments, the number of factors needed for bols is not always equal to two. Here it actually is nlevels(iris$Species) - 1, i.e., one less factor level than we have classes in the outcome. I've also updated my post on github accordingly.

edited Jun 20 '20 at 09:12

Community

1
1

answered Jul 23 '15 at 13:20

Benjamin Hofner

211
1
9

Looks good, cheers for that. Before I give you the thumbs up, if I can ask one stupid question. Is the factor(1:2) for the dummy used because we have 3-1 classes (setosa, versicolor and virginica) or because I use df = 3-1 in the buser or now bols argument? – Lorcan Treanor Jul 24 '15 at 07:20
1

Actually never mind, sussed it out and it is because of the number of classes. Apologies and thank you and Sarah for your efforts. I like this package because of the success I had with the glmboost function and I am new to boosting techniques so thank you. – Lorcan Treanor Jul 24 '15 at 07:25

Prediction using mboost multinomial logistic regression in R

1 Answers1