I am analysing the accuracy of two diagnostics tests that detect if an individual has disease ("1") or no disease ("0") against a ground truth (gold standard). I want to calculate sensitivity, specificity and likelihood ratios with 95% Confidence Intervals. However, I need to do the analysis with bootstrapping and I haven't been able to manage to do it so far. Would deeply appreciate an example with the following data set or suggestions for which package to use and which functions to code.
this is a fragment of my database
I have already managed to calculate sensitivity, specificity and likelihood ratios with CI using the acc.1test. However, I need to apply bootstrapping to this.
Many thanks!
This is an example of bootstraping using the R package rsample. There are many other possibilities like boot, bootstrap or manual coding but I like this approach since it's very flexible. First of all I load the packages
library(dplyr)
library(purrr)
library(rsample)
library(yardstick)
#> For binary classification, the first factor level is assumed to be the event.
#> Set the global option `yardstick.event_first` to `FALSE` to change this.
That warning means that we need to set options(yardstick.event_first = FALSE) since the first factor level for gold_std, test1 and test2 is 0.
options(yardstick.event_first = FALSE)
Then i load the data
data <- tibble(
ID = 1:25,
gold_std = factor(c(0, 1, 1, 0, 0, 0, 1, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 1, 0, 1, 1, 1, 0, 0, 0)),
test_1 = factor(c(0, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 0, 1, 1, 0, 0)),
test_2 = factor(c(0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 0, 1, 1, 1, 0, 0, 0, 1, 0, 1, 0, 1, 0, 0, 0))
)
The confusion matrix is:
table(select(data, gold_std, test_1))
#> test_1
#> gold_std 0 1
#> 0 8 7
#> 1 1 9
which means that the sensitivity is equal to 9 / (9 + 1) and specificity is equal to 8 / (7 + 8). Indeed
sens(data, truth = gold_std, estimate = test_1)
#> # A tibble: 1 x 3
#> .metric .estimator .estimate
#> <chr> <chr> <dbl>
#> 1 sens binary 0.9
and
spec(data, truth = gold_std, estimate = test_1)
#> # A tibble: 1 x 3
#> .metric .estimator .estimate
#> <chr> <chr> <dbl>
#> 1 spec binary 0.533
The functions sens e spec are defined in the R package yardstick. I can use the R package rsample to create 30 bootstrap replicates of the data:
set.seed(1)
bootstrap_data <- bootstraps(data, times = 30)
and this is the result
bootstrap_data
#> # Bootstrap sampling
#> # A tibble: 30 x 2
#> splits id
#> <list> <chr>
#> 1 <split [25/9]> Bootstrap01
#> 2 <split [25/13]> Bootstrap02
#> 3 <split [25/11]> Bootstrap03
#> 4 <split [25/7]> Bootstrap04
#> 5 <split [25/10]> Bootstrap05
#> 6 <split [25/10]> Bootstrap06
#> 7 <split [25/8]> Bootstrap07
#> 8 <split [25/7]> Bootstrap08
#> 9 <split [25/9]> Bootstrap09
#> 10 <split [25/7]> Bootstrap10
#> # ... with 20 more rows
bootstrap_data is a new dataframe with 30 rows (as many as the number of boostrap replications) and 2 columns: splits e id. The column splits records the number of observations excluded from each bootstrap replicate while id is simply and id column. For example in the first replicate of the bootstrap samples 9 observations were excluded. Moreover we can use the function analysis to retrive the observations sampled in each bootrstrap replicate. This is a more efficient way to store the indexes of a bootstrap sampling.
Now I can define two function which take as input a split object and return the estimate of sensitivity and specificity for that boostrap sample.
bootstrap_sens <- function(splits) {
x <- analysis(splits)
sens(x, truth = gold_std, estimate = test_1)$.estimate
}
bootstrap_spec <- function(splits) {
x <- analysis(splits)
spec(x, truth = gold_std, estimate = test_1)$.estimate
}
I use the notation $.estimate since I need just the estimate of sensitivity and specificity (and you can see from the output at the beginning that they return a dataframe). I can apply these functions to the bootstrap replicates using map_dbl (since the output of both functions is an object of type double) and this is the result:
bootstrap_data <- bootstrap_data %>%
mutate(
sens = map_dbl(splits, bootstrap_sens),
spec = map_dbl(splits, bootstrap_spec)
)
bootstrap_data
#> # Bootstrap sampling
#> # A tibble: 30 x 4
#> splits id sens spec
#> * <list> <chr> <dbl> <dbl>
#> 1 <split [25/9]> Bootstrap01 0.727 0.571
#> 2 <split [25/13]> Bootstrap02 1 0.471
#> 3 <split [25/11]> Bootstrap03 1 0.312
#> 4 <split [25/7]> Bootstrap04 0.9 0.533
#> 5 <split [25/10]> Bootstrap05 0.8 0.533
#> 6 <split [25/10]> Bootstrap06 1 0.364
#> 7 <split [25/8]> Bootstrap07 0.9 0.6
#> 8 <split [25/7]> Bootstrap08 1 0.533
#> 9 <split [25/9]> Bootstrap09 0.667 0.625
#> 10 <split [25/7]> Bootstrap10 0.667 0.75
#> # ... with 20 more rows
Created on 2019-07-22 by the reprex package (v0.3.0)
You can use the same approach for test_2. I'm not sure what you mean by likelihood ratios but you can define your own function to evalute it and apply it in the same way as bootstrap_sens or bootstrap_spec. You can find much more details about rsample at the package website.