stat_break: Find smallest subset to exclude from sample for...

In hannesrosenbusch/StatBreak: Examine robustness of sample statistics to case deletions

Description

The function iteratively learns which observations should at least be excluded from the data to reach a conservative 'goal value' for the statistic of interest. It does so by relying on a genetic algorithm, which efficiently explores the (usually vast) space of possible subsets. The result can uncover impactful subsamples and fuel discussions of robustness. Necessary arguments include the dataframe, a function to compute the statistic of interest ('statistic_computation' see examples), and the goal value of interest.

Usage

stat_break(
  data = NULL,
  goal_value = NULL,
  statistic_computation = NULL,
  max_exclusions = NULL,
  pop = 1000,
  max_generations = 2000,
  exclusion_cost = 0.01,
  prop_included_cases = 0.95,
  chance_of_mutation = 0.02,
  stop_search = 200,
  random_seed = 42,
  large_sample_drops = FALSE
)

Arguments

data	A data.frame containing the observations as rows.
goal_value	This conservative value (e.g., small effect size) is targeted.
statistic_computation	A formula which has 'data' as input and returns the statistic of interest.
max_exclusions	maximum number of cases to be excluded
pop	Number of 'individuals' in each generation of the genetic algorithm.
max_generations	Maximum number of generations that the algorithm generates.
exclusion_cost	Used to calibrate fitness function.
prop_included_cases	Initial proportion of included cases (e.g. .90).
chance_of_mutation	Chance that a gene mutates, higher is slower but more accurate (e.g. .02).
stop_search	After how many generations without change is the 'converged' result returned.
random_seed	Seed for replicability.
large_sample_drops	drops one row from best (helps converge with large samples)

Value

named list. includes how many and which rows were excluded plus the original and new statistic

Examples

coefficient_computation <- function(data){
statistic <- cor(data$Sepal.Length, data$Petal.Width)
return(statistic)}

filter <- stat_break(data = iris, statistic_computation = coefficient_computation, goal_value = 0.78, max_generations = 500)
print(filter)

hannesrosenbusch/StatBreak documentation built on Feb. 12, 2020, 10:35 a.m.