Search code examples
rmachine-learningstargazerpartymodelsummary

Generate table with side-by-side node models of `partykit:mob()` object

Let's say I fit a model using partykit:mob(). Afterward, I would like to generate a side-by-side table with all the nodes (including the model fitted using the whole sample). Here I attempted to do it using stargazer(), but other ways are more than welcome.

Below an example and attempts to get the table.

library("partykit")
require("mlbench")
## Pima Indians diabetes data
data("PimaIndiansDiabetes", package = "mlbench")
## a simple basic fitting function (of type 1) for a logistic regression
logit <- function(y, x, start = NULL, weights = NULL, offset = NULL, ...) {
  glm(y ~ 0 + x, family = binomial, start = start, ...)
}
## set up a logistic regression tree
pid_tree <- mob(diabetes ~ glucose | pregnant + pressure + triceps + insulin +
                  mass + pedigree + age, data = PimaIndiansDiabetes, fit = logit)

pid_tree 
# Model-based recursive partitioning (logit)
# 
# Model formula:
#   diabetes ~ glucose | pregnant + pressure + triceps + insulin +
#   mass + pedigree + age
# 
# Fitted party:
#   [1] root
# |   [2] mass <= 26.3: n = 167
# |       x(Intercept)     xglucose
# |        -9.95150963   0.05870786
# |   [3] mass > 26.3
# |   |   [4] age <= 30: n = 304
# |   |       x(Intercept)     xglucose
# |   |        -6.70558554   0.04683748
# |   |   [5] age > 30: n = 297
# |   |       x(Intercept)     xglucose
# |   |        -2.77095386   0.02353582
# 
# Number of inner nodes:    2
# Number of terminal nodes: 3
# Number of parameters per node: 2
# Objective function: 355.4578

1.- Extract summary(pid_tree, node = x) + stargazer().

## I want to replicate this table extracting the the nodes from partykit object.   
library(stargazer)  
m.glm<-   glm(diabetes ~ glucose, family = binomial,data = PimaIndiansDiabetes)

typeof(m.glm)
## [1] "list"
class(m.glm)
## [1] "glm" "lm" 
stargazer(m.glm)
## ommited output.



## Extracting summary from each node
summ_full_data <- summary(pid_tree, node = 1)
summ_node_2    <- summary(pid_tree, node = 2)
summ_node_4    <- summary(pid_tree, node = 4)
summ_node_5    <- summary(pid_tree, node = 5)

## trying to create stargazer table with coefficients
stargazer(m.glm,
          summ_node_2, 
          summ_node_4,
          summ_node_5,title="MOB Results")
##Error: $ operator is invalid for atomic vectors

2.- Extract pid_tree[x] + stargazer().

## Second Attempt (extracting modelparty objects instead)
node_2    <- pid_tree[2]
node_4    <- pid_tree[4]
node_5    <- pid_tree[5]

class(node_5)
##[1] "modelparty" "party"     

stargazer(m.glm,
          node_2, 
          node_4,
          node_5,title="MOB Results")
# % Error: Unrecognized object type.
# % Error: Unrecognized object type.
# % Error: Unrecognized object type.

3.- Not really elegant, I know: Force class to emulate the glm object.

## Force class of object to emulate glm one
class(m.glm)
class(summ_node_2) <- c("glm", "lm") 
stargazer(summ_node_2)
##Error in if (p > 0) { : argument is of length zero

A rather pragmatic solution would be just re-fit the model recovering the rules found by partykit:mob() and then use stargaze() on them, but for sure I am missing something here. Thanks in advance.

Solution

  • It's best to extract (or refit) the list of model objects per node and then apply the table package of choice. Personally, I don't like stargazer much and much rather use modelsummary instead or sometimes the good old memisc.

    If the tree contains the model $objects in the $info (as for pid_tree) you can use nodeapply() for all nodeids() to extract these:

    pid_models <- nodeapply(pid_tree, ids = nodeids(pid_tree), FUN = function(x) x$info$object)

    If you just want to extract the fitted models for the terminal nodes (leaves) of the tree, then you can do so by setting ids = nodeids(pid_tree, terminal = TRUE).

    Alternatively, especially when the model objects are not stored, you can easily refit them via:

    pid_models <- refit.modelparty(pid_tree)

    Here, you could also include node = nodeids(pid_tree, terminal = TRUE) to only refit the terminal node models.

    In all cases you can subsequently use

    msummary(pid_models)

    to produce the model summary table. It supports a variety of output formats and of course you can tweak the list further to change the results, e.g., by changing their names etc. The default output looks like this:

    msummary output