In stufield/power: R Package for Basic Power Analyses

options(width = 100)
Sys.setlocale("LC_COLLATE", "en_US.UTF-8") # ensure common sorting envir
knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "man/figures/README-",
  out.width = "100%"
)
ver <- desc::desc_get_version(".")
ver <- paste0("https://img.shields.io/badge/Version-", ver,
              "-success.svg?style=flat&logo=github")

The power package

The power package contains some simple functions to empirically simulate and estimate statistical power under various statistical test conditions. In general, simulations are performed with known effect sizes or differences, and the proportion of detected significant p-values represents the empirical power, i.e. $1-\beta$ or 1 - TypeII error.

The goal is typically get an idea of the required sample size given an experimental design, statistical test, effect size, and desired power.

---

Installation

The power package is not currently on CRAN but you can install the latest version from github via:

# current dev version
remotes::install_github("stufield/power")

# or a specific version
remotes::install_github("stufield/power@v0.0.1")

Loading power

Loading the power package is as simple as:

library(power)

Plot power curves

The simplest way to generally plot power curves is via plot_power_curves() which uses power.t.test() under the hood:

plot_power_curves(
  delta_vec = seq(0.5, 2, 0.1),
  power_vec = seq(0.5, 0.9, 0.1)
)

Power curves and KS-distance

There is a loose "rule-of-thumb" relationship between Sensitivity/Specificity and KS-distance, and of course, KS is related to effect size. So we can visualize this relationship also via a standard power curve:

ks_tables <- ks_power_table()
ks_tables

# `power` as y-axis
plot(ks_tables)

# `n` as y-axis
plot(ks_tables, plot_power = FALSE)

---

Two-Groups

Empirical Power via Simulation

A more robust (?) empirical calculation of power can be generated via simulation:

# constant effect size (delta)
size_tbl <- withr::with_seed(1,
  t_power_curve(seq(10, 50, 2), delta = 0.75, nsim = 25L)
)
size_tbl

plot(size_tbl)

# constant sample size (n)
delta_tbl <- withr::with_seed(2,
  t_power_curve(seq(0.5, 2.5, 0.1), n = 10, nsim = 25L)
)
delta_tbl

plot(delta_tbl)

Solve for Sample Size

To solve for the sample size given a corresponding power value you must have simulate power keeping "delta" (effect size) constant and varying n. For example, the size_tbl object created above:

solve_n(size_tbl, 0.8)

Fisher's Exact for Count Data

For count data, Fisher's Exact tests assume a 2x2 contingency matrix and equal proportions across the margins (rows x cols):

fisher_power(0.85, 0.75, 200, 200, nsim = 200L)

Fisher's Power Curve

f_tbl <- fisher_power_curve(seq(50, 400, 5), p = 0.85, p_diff = -0.1,
                            nsim = 200L)
f_tbl

Plot the Power Curve

gg_pwr <- plot(f_tbl)
gg_pwr

Solve for n

pwr_n <- solve_n(f_tbl, 0.8)
pwr_n

Visually check the curve and add the solution to the ggplot.

gg_pwr +
  ggplot2::annotate("segment",
    x        = c(pwr_n[["n"]], min(f_tbl$n)),
    xend     = c(pwr_n[["n"]],pwr_n[["n"]]),
    y        = c(min(f_tbl$power), pwr_n[["power"]]),
    yend     = c(pwr_n[["power"]], pwr_n[["power"]]),
    linetype = "dashed", colour = "#00A499")

stufield/power documentation built on June 1, 2025, 7:16 p.m.