RLiesel

The goal of RLiesel is to provide an R interface for the probabilistic programming framework Liesel.

RLiesel assists the user in expressing semi-parametric regression models as probabilistic graphical models (PGMs). It supports generalized additive models for location, scale and shape (GAMLSS) with different response distributions, spline-based smooth terms and shrinkage priors. In a second step, the models can be transferred to Python for the further analysis.

Installation

You can install the development version of RLiesel by running the following commands:

# install.packages("devtools")
devtools::install_github("liesel-devs/rliesel")

Usage

library(rliesel)

The liesel() function is the workhorse of RLiesel and can be used to set up semi-parametric regression models. We illustrate the flexibility of the function by means of two examples.

A linear model

In the first example, we assume a linear relationship between the mean and the log-standard deviation of a response variable y and an explanatory variable x.

# generate some data
set.seed(1337)

n <- 250
x <- runif(n)
y <- rnorm(n, mean = x, sd = exp(x))

The regression predictors of the distributional parameters are specified with the predictor() function. The predictors are then passed to the predictors argument of the liesel() function as a named list. Finally, the conditional distribution of the response variable needs to be defined via the distribution argument:

linear_mod <- liesel(
  response = y,
  distribution = "Normal",
  predictors = list(
    loc = predictor(~x, inverse_link = "Identity"),
    scale = predictor(~x, inverse_link = "Exp")
  )
)

The hierarchical structure of a Liesel model can be visualized with the generic plot() function:

plot(linear_mod)

The attributes of the Python model object such as the nodes or the log-probability can be accessed in R with the $ operator, e.g. linear_mod$nodes$loc_p1_beta$value or linear_mod$log_prob.

A generalized additive model (GAM)

In the second example, both the location and the scale parameter are connected to the explanatory variable x through a smoothing spline:

y <- rnorm(n, mean = sin(2 * pi * x), sd = exp(sin(2 * pi * x)))

spline_mod <- liesel(
  response = y,
  distribution = "Normal",
  predictors = list(
    loc = predictor(~s(x), inverse_link = "Identity"),
    scale = predictor(~s(x), inverse_link = "Exp")
  )
)

## plot(spline_mod)

Users who are familiar with the mgcv package may recognize the syntax from the gam() function, which can be used to set up an equivalent model as follows:

library(mgcv)
gam(list(y ~ s(x), ~s(x)), family = gaulss())

Finally, the Liesel model can be saved to a pickle file with the save_model() command:

save_model(spline_mod, "spline-mod.pickle")