distreg.vis

What is distreg.vis?

distreg.vis is built upon the R package bamlss (CRAN Link) and gamlss (CRAN Link), which fit distributional regression models. distreg.vis requires a fitted bamlss or gamlss object to achieve two goals:

1. See and compare the expected distribution for chosen sets of covariates
2. View the direct relationship between moments of the response distribution and a chosen explanatory variable, given a set of covariates.

The implementation is an interactive application using the Shiny framework.

Install

distreg.vis is now on CRAN. Run the following code to install the package and all its extensions:

install.packages("distreg.vis")

Or, if you would like to have the newest version, install the GitHub variant:

devtools::install_github("Stan125/distreg.vis")

Supported distributions

At the moment, the following distributions are fully supported:

dists_char <- dists %>%
  filter(moment_funs) %>%
  dplyr::select(dist_name, class)

# GAMLSS Families
dists_char[dists_char$class == "gamlss", "dist_name"]
#>  [1] "BE"       "BEINF"    "BNB"      "DEL"      "EGB2"     "exGAUS"  
#>  [7] "EXP"      "GA"       "GB2"      "GEOM"     "GEOMo"    "GG"      
#> [13] "GIG"      "GPO"      "GT"       "GU"       "IG"       "IGAMMA"  
#> [19] "JSU"      "JSUo"     "LG"       "LO"       "LOGNO"    "NBF"     
#> [25] "NBI"      "NBII"     "NO"       "NO2"      "NOF"      "PARETO2" 
#> [31] "PARETO2o" "PE"       "PE2"      "PIG"      "PO"       "RG"      
#> [37] "SHASHo"   "SICHEL"   "SN1"      "SN2"      "SST"      "ST2"     
#> [43] "ST3"      "ST3C"     "ST4"      "ST5"      "TF"       "TF2"     
#> [49] "WEI"      "WEI2"     "WEI3"     "ZAGA"     "ZALG"     "ZANBI"   
#> [55] "ZAP"      "ZAPIG"    "ZASICHEL" "ZAZIPF"   "ZIBNB"    "ZINBI"   
#> [61] "ZIP"      "ZIP2"     "ZIPF"     "ZIPIG"    "ZISICHEL"

# BAMLSS Families
dists_char[dists_char$class == "bamlss", "dist_name"]
#>  [1] "beta"        "binomial"    "cnorm"       "gamma"       "gaussian"   
#>  [6] "gaussian2"   "glogis"      "gpareto"     "lognormal"   "multinomial"
#> [11] "poisson"

The rest of the distributions are mostly supported, but only in displaying their predicted distributions, and not their moments.

Show-Case

Let’s show an example using the Wage dataset, taken from the ISLR R package. It depicts the wage of 3000 male workers in the Mid-Atlantic region of the US, related to a couple of socio-economic variables.

First, let’s fit a bamlss using the log normal distribution (Wages cannot be negative):

library(gamlss)
Wage <- ISLR::Wage
wage_model <- gamlss(wage ~ ps(age) + race + year + education,
                     ~ ps(age) + race + year + education,
                     data = Wage, family = LOGNO())

Both parameters of the modeled wage distribution are now modeled as being dependent on the socio-economic variables.

Starting the application

The app can be started as an RStudio Add-In or running distreg.vis::vis(). Afterwards, select the fitted model like so:

After the model was selected, you can head to the “Scenarios” tab.

Expected Distribution

At the Scenarios tab, you can choose specific observations of explanatory variables for which to predict the distribution. In this case, we select socioeconomic variables and get the expected income distribution for this combination:

The real power of distreg.vis is visible when we add multiple “scenarios” for comparing different covariate combinations:

Influence of Covariates

Furthermore, we can easily view the “influence” of a covariate on the expected income distribution. For this, head to the “Properties” tab:

Obtaining Code

All plots which were created can be reproduced with the code that appears when you press “Obtain Code”: