R package dontknow

dontknow provides a gamlss2 family for regression models with an explicit "don't know" response option. The response is represented as Y = cbind(y1, y2), where y1 indicates whether the DK option was chosen and y2 contains the ordinal response among non-DK observations.

DK(k) fits a joint latent bivariate normal model for:

• the DK decision
• the ordinal outcome
• their dependence through a correlation parameter

The package also contains sim_DK() for generating example data from the same model structure.

The package depends on gamlss2, which can be installed from R-universe:

install.packages(
  "gamlss2",
  repos = c(
    "https://gamlss-dev.R-universe.dev",
    "https://cloud.R-project.org"
  )
)

The package itself can then be installed from GitHub:

remotes::install_github("freezenik/dontknow")

Minimal example:

library("gamlss2")
library("dontknow")

d <- sim_DK(500)

## y1: DK indicator, y2: ordinal category among non-DK responses
head(d$Y)

## remove intercepts in the latent mean predictors
f <- Y ~ -1 + s(x1) | -1 + s(x2) | 1
b <- gamlss2(f, family = DK(4), data = d)

summary(b)

## fitted distribution parameters
par <- predict(b, type = "parameter")
head(par$mu1)
head(par$mu2)
head(par$rho)

## ordered thresholds on the latent scale
head(family(b)$alpha(par))

Implementation note:

DK(4) uses compiled code by default for the log-likelihood, score, and observed Hessian. When score and Hessian are both enabled, the family also uses a fused compiled $update() function for faster RS updates in gamlss2.