Fit of Inf or -Inf does not raise an error

[ja-thomas]

When an glmboost model fits values of +/-Inf it does not raise an error, but sets an baselearner to NA. So there can be mboost models with mstop > 1 but every baselearner is NA and no coefficients are fitted (coef(model) is NULL) but no error or warning is shown.

If you fit it with gamboost and bols baselearner an error is raised.

[hofnerb]

Can you please supply a minimal example?

Am 15.10.2015 um 15:25 schrieb ja-thomas:

When an glmboost model fits values of +/-Inf it does not raise an error,
but sets an baselearner to NA. So there can be mboost models with mstop

1 but every baselearner is NA and no coefficients are fitted
(coef(model) is NULL) but no error or warning is shown.

If you fit it with gamboost and bols baselearner an error is raised.

—
Reply to this email directly or view it on GitHub
#18.

[ja-thomas]

After some additional testing, it seems to have something to do with the offset value of the family. The problem first appears in gamboostLSS where offset values are written in the family objects.

library(gamboostLSS)
###generate data###


set.seed(15102015)
n = 1000
v1 <- rnorm(n, mean = 1)
v2 <- rnorm(n, mean = 1)
v3 <- rnorm(n, mean = 1)
v4 <- rnorm(n, mean = 1)
v5 <- rnorm(n, mean = 1)
v6 <- rnorm(n, mean = 1)


inf <- cbind(1, v1,v2,v3,v4,v5,v6)

mu_c <- c(1,2,1,5,1,2,8)
sigma_c <- c(2,1,1,5,1,12,6)

mu <- exp(inf %*% mu_c)
sigma <- exp(inf %*% sigma_c)


y <- rnbinom(n, size = sigma, mu = mu)

data <- data.frame(y, inf[,-1])


###models###


model <- glmboostLSS(y ~. , data = data, families = NBinomialLSS())

model[[2]] # broken model


model_sigma <- glmboost(y ~. , data = data, family = NBinomialSigma(stabilization = "none"))
model_sigma # works just fine




#calculate the fit as mu value for the sigma model
model_mu <- glmboost(y ~ ., data = data, family = NBinomialMu(stabilization = "none"),
                      control = boost_control(mstop = 1))


#sigma model with offset
model_sigma_offset <- glmboost(y ~. , data = data, 
                        family = NBinomialSigma(stabilization = "none", 
                                                mu = fitted(model_mu)))

model_sigma_offset # also broken, at least with warnings

In the second case a warning is given while calculating the model.

[hofnerb]

With gamboost we get an error. Yet, this error is not very informative (and actually seems to be a bug as well):

>  gamboost(y ~. , data = data, family = NBinomialSigma(stabilization = "none", mu = fitted(model_mu)))

Fehler in if (all(tsums < 0)) stop("could not fit base learner in boosting iteration ",  : 
  Fehlender Wert, wo TRUE/FALSE nötig ist