distreg.vis (version 1.7.5)

moments: Compute distributional moments from the parameters

Description

This function takes (predicted) parameters of a response distribution and calculates the corresponding distributional moments from it. Furthermore, you can specify own functions that calculate measures depending on distributional parameters.

Usage

moments(par, fam_name, what = "mean", ex_fun = NULL)

Arguments

par

Parameters of the modeled distribution in a data.frame form. Can be Output of preds, for example.

fam_name

Name of the used family in character form. Can be one of distreg.vis::dists$dist_name. All gamlss.dist and exported bamlss families are supported. To obtain the family from a model in character form, use fam_obtainer.

what

One of "mean", "upperlimit", "lowerlimit". If it is mean (which is also the default), then the mean of the parameter samples is calculated. 2.5 for lowerlimit and upperlimit, respectively.

ex_fun

An external function function(par) {...} which calculates a measure, whose dependency from a certain variable is of special interest.

Details

With the exception of betareg, the distributional families behind the estimation of the distributional regression models are represented by own objects, e.g. GA or lognormal_bamlss. We worked together with both the authors of gamlss and bamlss such that the functions to compute the moments from the parameters of the underlying distribution is already implemented in the family functon itself. As an example, try out gamlss.dist::BE()$mean, which shows one example. The function moments() utilizes this fact and ensures that the outcome is always in the right format: Two columns named `Expected_Value` and `Variance` detailing the first two moments. One exception appears when an external function is specified, at which point there are three columns.

Each row details one `scenario` meaning one covariate combination for which to predict the moments. moments() is heavily used in plot_moments, where moments are calculated over the entire range of one variable.

If target distribution stems from a bamlss model, moments() can also utilize the samples from the preds function to transform them. This is important for correct estimates, as just taking the mean of the samples and then using those means to estimate the moments can lead to inaccurate results. moments() knows when samples of predicted parameters were specified in the par argument, and then transforms the samples to the moments, before taking averages. Only through this procedure we even get credible intervals for the expected moments (see "upperlimit" and "lowerlimit" as possible outcomes of argument what).

Examples

# Get some artificial data
gamma_data <- model_fam_data(fam_name = "gamma", nrow = 100)

# Estimate model
library("bamlss")
model <- bamlss(list(gamma ~ norm2 + binomial1,
                     sigma ~ norm2 + binomial1),
                     data = gamma_data,
                     family = gamma_bamlss())

# Get some predicted parameters in sample and without sample form
pred_params <- preds(model, vary_by = "binomial1")
pred_params_samples <- preds(model, vary_by = "binomial1", what = "samples")

# Now calculate moments - with samples more correct estimates come out
moments(pred_params, fam_name = "gamma", what = "mean")
moments(pred_params_samples, fam_name = "gamma", what = "mean")

# Now with specifying an external function
my_serious_fun <- function(par) {
  return(par[["mu"]] + 3*par[["sigma"]])
}
moments(pred_params_samples,
        what = "mean",
        fam_name = "gamma",
        ex_fun = "my_serious_fun")