options("mboost_indexmin") and bhistx() gives errors in coef()

[sbrockhaus]

If internally an index is used for model fitting for a model containing bhistx() some methods, e.g., coef(), no longer work:

library(FDboost)

options("mboost_indexmin")

require(refund))
## simulate some data from a historical model
## the interaction effect is in this case not necessary
n <- 100
nygrid <- 35
data1 <- pffrSim(scenario = c("int", "ff"), limits = function(s,t){ s <= t }, 
                 n = n, nygrid = nygrid)
data1$X1 <- scale(data1$X1, scale = FALSE) ## center functional covariate                  
dataList <- as.list(data1)
dataList$tvals <- attr(data1, "yindex")

## create the hmatrix-object
X1h <- with(dataList, hmatrix(time = rep(tvals, each = n), id = rep(1:n, nygrid), 
                              x = X1, argvals = attr(data1, "xindex"), 
                              timeLab = "tvals", idLab = "wideIndex", 
                              xLab = "myX", argvalsLab = "svals"))
dataList$X1h <- I(X1h)   
dataList$svals <- attr(data1, "xindex")


#################################

options("mboost_indexmin" = 10000)

## do the model fit with main effect of bhistx() and interaction of bhistx() and bolsc()
mod <- FDboost(Y ~ bhistx(x = X1h, df = 5, knots = 5), 
               timeformula = ~ bbs(tvals, knots = 10), data = dataList)

coef_mod <- coef(mod)


###################################

options("mboost_indexmin" = 10)

## do the model fit with main effect of bhistx() and interaction of bhistx() and bolsc()
mod2 <- FDboost(Y ~ bhistx(x = X1h, df = 5, knots = 5), 
               timeformula = ~ bbs(tvals, knots = 10), data = dataList)

### breaks within predict
coef_mod2 <- coef(mod2)

[davidruegamer]

I still haven't figured out, where the rub is, but maybe we could just set options("mboost_indexmin" = 10000) in coef ? Because this does not throw an error:

options("mboost_indexmin" = 10000)

coef_mod2 <- coef(mod2)

even if mod2 was fitted using options("mboost_indexmin" = 10) 🙃

[davidruegamer]

PS: The error is due to the fact, that get_index does not work with hmatrix objects. In detail, if we use the mentioned option, we try to apply get_index to an object like this:

'data.frame':	1600 obs. of  4 variables:
 $ X1h    : AsIs [1:1600, 1:2] 0 0 0 0 0 0 0 0 0 0 ...
  ..- attr(*, "dimnames")=List of 2
  ..- attr(*, "x")= num [1:40, 1:40] 39 0 0 0 0 0 0 0 0 0 ...
  ..- attr(*, "argvals")= num  0 0.0256 0.0513 0.0769 0.1026 ...
  ..- attr(*, "timeLab")= chr "t"
  ..- attr(*, "idLab")= chr "wideIndex"
  ..- attr(*, "xLab")= chr "x"
  ..- attr(*, "argvalsLab")= chr "s"
  ..- attr(*, "class")= chr  "AsIs" "hmatrix" "matrix"
 $ tvals  : num  0 0 0 0 0 0 0 0 0 0 ...
 $ ONEx   : num  1 1 1 1 1 1 1 1 1 1 ...
 $ ONEtime: num  1 1 1 1 1 1 1 1 1 1 ...
 - attr(*, "varnms")= chr  "svals" "tvals"
 - attr(*, "xm")= num  0 0.0256 0.0513 0.0769 0.1026 ...
 - attr(*, "ym")= num  0 0.0256 0.0513 0.0769 0.1026 ...
 - attr(*, "myargsHist")=List of 23

But you said something like that it isn't a problem for bhist (without x) baselearners? That would be odd.

[davidruegamer]

Since the error is caused by bl[[w]]$predict(ens[ix], newdata = newdata, aggregate = agg) in bl_lin, I think we do not have a choice, but to either redefine indexmin automatically when predict is called for such objects or throw an error and recommend to set indexmin to a larger value. All other options involve a complete restructuring of our hmatrix concept. A generic get_index function for hmatrixobjects, for example, won't work, as the object, which we pass to get_index is an ordinary data.frame.

[sbrockhaus]

For models with response in long format, "FDboostLong", even more methods break. bhist() is no problem as long as the model has a response matrix and not a response in long format.
As the option is not helpful for functional covariates (it uses ties in the data to speed up computations and ties seem very unlikely in case of functional covariates), we decided to set the option to +Inf within FDboost.

[sbrockhaus]

ea31b33