How to nest tables in a column of a dataframe? - r

Question

I read that it is possible to store dataframes in a column of a dataframe with nest:
https://tidyr.tidyverse.org/reference/nest.html
Is it also possible to store tables in a column of a dataframe?
The reason is that I would like to calculate the Kappa for every subgroup of a dataframe with Caret. Although caret::confusionMatrix(t) expects a table as input.
In the example-code below this works fine if I calculate the Kappa for the complete dataframe at once:
library(tidyverse)
library(caret)
# generate some sample data:
n <- 100L
x1 <- rnorm(n, 1.0, 2.0)
x2 <- rnorm(n, -1.0, 0.5)
y <- rbinom(n, 1L, plogis(1 * x1 + 1 * x2))
my_factor <- rep( c('A','B','C','D'), 25 )
df <- cbind(x1, x2, y, my_factor)
# fit a model and make predictions:
mod <- glm(y ~ x1 + x2, "binomial")
probs <- predict(mod, type = "response")
# confusion matrix
probs_round <- round(probs)
t <- table(factor(probs_round, c(1,0)), factor(y, c(1,0)))
ccm <- caret::confusionMatrix(t)
# extract Kappa:
ccm$overall[2]
> Kappa
> 0.5232
Although if I try to do group_by to generate the Kappa for every factor as a subgroup (see code below) it does not succeed. I suppose I need to nest t in a certain way in df although I don't know how:
# extract Kappa for every subgroup with same factor (NOT WORKING CODE):
df <- cbind(df, probs_round)
df <- as.data.frame(df)
output <- df %>%
dplyr::group_by(my_factor) %>%
dplyr::mutate(t = table(factor(probs_round, c(1,0)), factor(y, c(1,0)))) %>%
summarise(caret::confusionMatrix(t))
Expected output:
>my_factor Kappa
>1 A 0.51
>2 B 0.52
>3 C 0.53
>4 D 0.54
Is this correct and is this possible?
(the exact values for Kappa will be different due to the randomness in the sample data)
Thanks a lot!

Answer 1

You could skip the intermediate mutate() that's giving you trouble to do:
library(dplyr)
library(caret)
df %>%
group_by(my_factor) %>%
summarize(t = confusionMatrix(table(factor(probs_round, c(1,0)),
factor(y, c(1,0))))$overall[2])
Returns:
# A tibble: 4 x 2
my_factor t
<chr> <dbl>
1 A 0.270
2 B 0.513
3 C 0.839
4 D 0.555

Answer 2

The above approach is the easiest to get the desired results. But just to show whats possible, we can use your approach with rowwise::nest_by which groups the data set rowwise.
In the approach below we calculate a separate glm for each subgroup. I'm not sure if that's what you want to do.
library(tidyverse)
library(caret)
# generate some sample data:
n <- 1000L
df <- tibble(x1 = rnorm(n, 1.0, 2.0),
x2 = rnorm(n, -1.0, 0.5),
y = rbinom(n, 1L, plogis(x1 + 1 * x1 + 1 * x2)),
my_factor = rep( c('A','B','C','D'), 250))
output <- df %>%
nest_by(my_factor) %>%
mutate(y = list(data$y),
mod = list(glm(y ~ x1 + x2,
family = "binomial",
data = data)),
probs = list(predict(mod, type = "response")),
probs_round = list(round(probs)),
t = list(table(factor(probs_round, c(1, 0)),
factor(y, c(1, 0)))),
ccm = caret::confusionMatrix(t)$overall[2])
output %>%
pull(ccm)
#> Kappa Kappa Kappa Kappa
#> 0.7743682 0.7078112 0.7157761 0.7549340
Created on 2021-06-23 by the reprex package (v0.3.0)