helper function

[lauravana]

I am currently using the ctm package from R Forge, which also implements model boosting when dealing with ordinal outcomes (i.e., ordered factors). The predict.ctm() function delivers an error probably because of:

function check_newdata() in mboost/R/helpers.R/
line: if (!all(sapply(newdata[nm], class) == sapply(mf, class)))

This will give an error when the outcome variable is of class "ordered" "factor".

[sbrockhaus]

I think I have the same problem with check_newdata() when using bl1 %X% bl2 with a variable that has more than one class and unsing predict() with newdata. I think the error occurs whenever you have more than one variable in the base-learner and at least one of those variables has more than one class. Consider the following minimal example (based on help of mboost):

library(mboost)
data("volcano", package = "datasets")
vol <- as.vector(volcano)
x1 <- 1:nrow(volcano)
x2 <- 1:ncol(volcano)
x <- expand.grid(x1, x2)
# generate a variable with more than one class
x$factorz <- I(gl(2, 2, length=nrow(x)))

modx <- mboost(vol ~ bbs(Var2, df = 3, knots = 10) %X%
                 bols(factorz, df = 3), data = x,
               control = boost_control(nu = 0.25))

# try to predict the data, gives error...
test <- predict(modx, newdata=x)

# ... as this comparison in check_newdata() is not possible, 
# even if the data is the same
sapply(x, class) == sapply(x, class)

[sbrockhaus]

This is an ugly hack, but it workes for my example. I simply override the current version of check_newdata() by a function that does not check the class of the variables in newdata at all!
Just run this when loading mboost:

my_check_newdata <- function(newdata, blg, mf, to.data.frame = TRUE) {
  nm <- names(blg)
  if (!all(nm %in% names(newdata)))
    stop(sQuote("newdata"),
         " must contain all predictor variables,",
         " which were used to specify the model.")
  if (!class(newdata) %in% c("list", "data.frame"))
    stop(sQuote("newdata"), " must be either a data.frame or a list")
  if (any(duplicated(nm)))  ## removes duplicates
    nm <- unique(nm)
  #if (!all(sapply(newdata[nm], class) == sapply(mf, class)))
  #  warning("Some variables in ", sQuote("newdata"),
  #          " do not have the same class as in the original data set",
  #          call. = FALSE)
  ## subset data
  mf <- newdata[nm]
  if (is.list(mf) && to.data.frame)
    mf <- as.data.frame(mf)
  return(mf)
}

library(mboost)
## write my version of check_newdata() into the namespace of mboost
assignInNamespace("check_newdata", my_check_newdata, ns="mboost", envir=as.environment("package:mboost"))

# check whether it worked
mboost:::check_newdata

[lauravana]

Thanks! I did a similar hack by replacing

if (!all(sapply(newdata[nm], class) == sapply(mf, class)))
with
if (!all(sapply(1:length(mf), function(i) is(nd[nm][[i]], class(mf[[i]])) & is(mf[[i]], class(nd[nm][[i]])))))

[sbrockhaus]

Thank you very much, I think your hack is nicer than mine. I just had to replace nd with newdata to make it work, i.e. replacing
if (!all(sapply(newdata[nm], class) == sapply(mf, class)))
with
if (!all(sapply(1:length(mf), function(i) is(newdata[nm][[i]], class(mf[[i]])) & is(mf[[i]], class(newdata[nm][[i]])))))

[hofnerb]

Thanks for the bug report.

I've used a different (but similar) fix as I've already modified check_newdata in pkg/mboostPatch/R/helpers.R. Please use

library("devtools")
install_github("hofnerb/mboost/pkg/mboostPatch")
library("mboost") 

and check your code with this package. Please let me know if the error persists.

[lauravana]

Thanks for the fix. Everything works fine for me!

[sbrockhaus]

Thanks a lot, for me it works as well!