as.families("BEINF") not working

[hofnerb]

The following code does not work:

library("gamboostLSS")
library("gamlss.dist")
data <- data.frame(x1 = runif(400), x2 = runif(400))
data$y <- with(data, rBEINF(400, mu = range((x1 + x2 - min(x1 + x2) + 0.001) / (diff(range(x1 + x2)) + 0.002)), 
                          sigma = sqrt(x1), nu = 0.1, tau = 0.1))
mod <- gamboostLSS(y ~ x1 + x2, data = data, families = as.families("BEINF"))
# Error in FAM$dldm(y = y, mu = FAM$mu.linkinv(f), sigma = sigma, nu = nu,  : 
#   unused arguments (nu = nu, tau = tau)

The reason can be found in

gamlss.dist::as.gamlss.family("BEINF")$dldm
#  function (y, mu, sigma) 
#  {
#      a <- mu * (1 - sigma^2)/(sigma^2)
#      b <- a * (1 - mu)/mu
#      dldm <- ifelse(((y == 0) | (y == 1)), 0, ((1 - sigma^2)/(sigma^2)) * 
#          (-digamma(a) + digamma(b) + log(y) - log(1 - y)))
#      dldm
#  }

which is a function of mu and sigma only.

How can we fix this for this family (and potentially others)? Is a change in gamlss.dist needed or can we fix this ourselves? Are there other families leading to the same or a similar issue?

@mayrandy, can you have a look at this?

[mayrandy]

Thanks for finding this! This should be a problem for all inflated families, e.g. BEOI():

BEINF()$nopar
[1] 4
 BEOI()$nopar
[1] 3

But their corresponding functions for the derivatives BEOI()$dldm have arguments mu and sigma only which leads to the error.

I think we should be able to fix this in our code, I'll take a look.

[hofnerb]

... no markdown available ...

[hofnerb]

Thanks a lot!

Perhaps there are even more families with similar properties? Could check this issue by comparing nopar with the number of parameters in dldm either by simply calling the functions or by computing on

deparse(gamlss.dist::as.gamlss.family("BEINF")$dldm)[1]

for all available families in a loop?

[mayrandy]

OK, all those families have FAM()$type == "Mixed", found that in my own code.

Apparently I had already implemented such cases for 3-parametric families and was only too lazy to do the same also for the 4-paremetric ones. Should be fixed now in "devel".

[hofnerb]

Merci :)