In Stan125/ghp: Generalisation of hierarchical partitioning

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "figures/"
)
library(ghp)

ghp

GHP stands for General Hierarchical Partitioning. ghp is an implementation of the technique of hierarchical partitioning first mentioned by Chevan and Sutherland (1991). This method fits all possible models for a set of covariates and then extracts a goodness of fit (e.g. $R^2$ for linear models) to obtain independent and joint contributions of the independent variables on the selected figure.

This package is an extension of the hier.part R package, developed by C. Walsh and R. Mac Nally in 2003. While hier.part is fast and simple at what it does, it is limited in the range of possible models as well as goodness of fit figures. Specifically, the motivation of this package is the ability to do deviance partitioning.

Installation

You can install ghp from github with:

# install.packages("devtools")
devtools::install_github("Stan125/ghp")

Example: Partitioning of rsquared in linear regression

Just call the ghp function with the name of the dependent variable (arg: dep) and a data.frame with all relevant variables to obtain the independent and joint effects of the explanatory covariates.

india <- ghp::india
results_lm <- ghp(depname = "stunting", india, method = "lm", gof = "r.squared")
results_lm

The first dataframe captures the actual mean influence of the variable on the goodness-of-fit. Also, joint effects are calculated. The second dataframe shows the percentage influence. We can see that cage has the highest influence with (~43%).

Example: Partitioning of deviance in gamlss models

It is now possible to do deviance partitiong of gamlss models. Gamlss models can model multiple parameters of a distribution. ghp can handle up to two modeled parameters, so you can find out what influence covariates have on the second modeled parameter (e.g. the variance).

results_gamlss <- ghp("stunting", india, method = "gamlss", 
                      gof = "deviance", npar = 2)
results_gamlss

Example: Variable grouping

Since 0.3.0 you can specify variable groups. The partitioning is now not happening with specific variables, but by testing all group combinations. In the given ghp::india dataset, which captures the nutrition of children in india we can now divide all covariates into to groups: those that give information about the child, and those that give information about the mother. Let's try that out:

# Specifying the groups should happen in a data.frame 
groupings <- data.frame(varnames = colnames(india), 
                        groups = c("0", "child", "child", "mother", 
                                   "child", "mother", "mother", "mother"))
results_groups <- ghp(depname = "stunting", india, method = "lm", gof = "r.squared",
                      group_df = groupings)
results_groups

We can now see that both groups have almost the same amount of influence on the $R^2$.

Bar Plots

To get a bar plot of the percentage independent effects, use plot_ghp():

plot_ghp(results_lm)
plot_ghp(results_gamlss) +
  ggplot2::scale_fill_grey()

Comparison with hier.part

Since 0.4.0, ghp is almost as fast as its counterpart hier.part, because the core partitioning was written with C++. A quick comparison:

hp <- function()
  hier.part::hier.part(india$stunting, dplyr::select(india,-stunting),
                       gof = "Rsqu", barplot = FALSE)
ghp <- function()
  ghp::ghp("stunting", india, method = "lm", gof = "r.squared")
microbenchmark::microbenchmark(hp, ghp)