FragiliTidy

Tidyverse-compatible, high-performance fragility metrics for two-arm clinical trials — for both dichotomous outcomes (Fragility Index / Reverse Fragility Index) and continuous outcomes (Continuous Fragility Index and Reverse Continuous Fragility Index).

FragiliTidy is designed to be fast (~25x faster than stats::fisher.test() / stats::chisq.test()) incorporating rejection sampling and an iterative Welch t-test substitution algorithm for the continuous indices. Everything plugs directly into tidyverse syntax.

Installation

# install.packages("remotes")
remotes::install_github("tomdrake/fragilitidy")

Functions

Function	Purpose
fragility_index()	Add a fragility-index column to a data frame (dichotomous outcomes).
reverse_fragility_index()	Add a reverse-fragility-index column to a data frame.
fragility_index_vec() / reverse_fragility_index_vec()	Vectorised forms for dplyr::mutate().
continuous_fragility_index()	Add a Continuous Fragility Index column to a data frame.
reverse_continuous_fragility_index()	Add a Reverse Continuous Fragility Index column to a data frame.
continuous_fragility_index_summary()	CFI from summary statistics (mean, SD, n per arm). Vectorised, so it works directly on scalars or on data frame columns inside dplyr::mutate() / dplyr::case_when().
reverse_continuous_fragility_index_summary()	Reverse CFI from summary statistics. Also vectorised for direct use in mutate() / case_when().
continuous_fragility_index_raw()	CFI from raw per-patient outcome vectors.
continuous_fragility_index_vec() / reverse_continuous_fragility_index_vec()	Aliases of the _summary() forms, kept for backward compatibility.

Quick start

Dichotomous outcomes

library(dplyr)
library(FragiliTidy)

trials <- tibble::tribble(
  ~study,    ~ie, ~ce, ~in_, ~cn,
  "Trial A",  10,  20,  100,  100,
  "Trial B",   5,  15,   80,   80,
  "Trial C",  30,  30,  200,  200
)

trials |>
  fragility_index(ie, ce, in_, cn) |>
  reverse_fragility_index(ie, ce, in_, cn)

Continuous outcomes

trials_continuous <- tibble::tribble(
  ~study,    ~intervention_mean, ~control_mean, ~intervention_sd, ~control_sd, ~intervention_n, ~control_n,
  "Trial X",  70,                 50,            10,               10,          50,              50,
  "Trial Y",  60,                 55,            15,               15,          40,              40
)

trials_continuous |>
  continuous_fragility_index(intervention_mean, control_mean, intervention_sd, control_sd, intervention_n, control_n) |>
  reverse_continuous_fragility_index(intervention_mean, control_mean, intervention_sd, control_sd, intervention_n, control_n)

Or, for a single trial from summary statistics — note the argument structure matches fragility_index_vec() (intervention/control pairs per statistic), just with mean/SD/n instead of event counts, and it's vectorised the same way:

continuous_fragility_index_summary(
  intervention_mean = 70, control_mean = 50,
  intervention_sd   = 10, control_sd   = 10,
  intervention_n    = 100, control_n   = 100,
  seed = 1
)

reverse_continuous_fragility_index_summary(
  intervention_mean = 55, control_mean = 50,
  intervention_sd   = 10, control_sd   = 10,
  intervention_n    = 30, control_n    = 30,
  seed = 1
)

Background

• The Fragility Index (Walsh et al., 2014) is the minimum number of event reassignments in the smaller-event arm required to flip a statistically significant dichotomous result to non-significance.
• The Reverse Fragility Index is the analogous quantity for non-significant dichotomous results.
• The Continuous Fragility Index (Caldwell et al., 2021) extends the concept to continuous outcomes compared via Welch's t-test, via an iterative substitution algorithm.
• The Reverse Continuous Fragility Index here estimates how many additional participants per arm would have been required to drive a non-significant continuous outcome to significance, given the observed mean and SD per arm.

See vignette("FragiliTidy") for a walkthrough.

References

• Walsh M, Srinathan SK, McAuley DF, et al. The statistical significance of randomized controlled trial results is frequently fragile. J Clin Epidemiol 2014;67:622-628.
• Caldwell JE, Youssefzadeh K, Limpisvasti O. A method for calculating the fragility index of continuous outcomes. J Clin Epidemiol 2021;136:20-25.

License

GPL-3. See LICENSE.