FDboost does not deal correctly with `is.bdfamily` families.

[fabian-s]

These now expect matrix valued responses cbind(#success, #failures) (gamboostLSS>1.5), but FDboost treats matrix valued responses as functions and shit's all fucked up.

Code examples in supplement to Statmod discussion paper, too long & secret to post here.

[fabian-s]

Still not working with boost-R/mboost@0326a2a --

set.seed(12) 
n <- 100
grid <- seq(0, 1, l = 20)
xfun <- t(replicate(n, dbeta(grid, runif(1, 3, 10), runif(1, 2, 7))))
xscalar <- rnorm(100)
y_success <- rbinom(n, size = 30, p = plogis(xscalar))
data <- list(y = cbind(y_success, y_fail = 30 - y_success), xfun = xfun, xscalar = xscalar)
m <- FDboost(y  ~ bbs(xscalar) + bsignal(xfun, grid),
  timeformula = NULL, family=Binomial(type="glm"),
  data = data, control = boost_control(mstop = 50, nu = 0.05))

yields:

Error in FDboost(y ~ bbs(xscalar) + bsignal(xfun), timeformula = NULL,  : 
  The length of the response and number of observations of xfun do not match.

because the response matrix for the scalar responses is converted to a long vector in
response <- matrix(response, ncol = 1) https://github.com/boost-R/FDboost/blob/master/R/FDboost.R#L566

[sbrockhaus]

Thanks for the bug report and the MWE. Works now :-)