Simulations for Negative Binomial outcomes

Introduction

The sim_power_nbinom function is used to estimate the empirical power of a negative binomial regression model. This method is useful when outcomes consist of over dispersed count data, and the objective is to test whether the relative risk (RR) between two groups differs significantly from a null boundary.

Data are simulated under a negative binomial distribution, a model is fit using glm.nb, and it is determined whether the null hypothesis is rejected based on confidence interval limits relative to a specified boundary. The procedure accounts for whether the true RR is below or above 1, applying the appropriate directional test logic.

Example

What is the empirical power using negative binomial regression to detect efficacy of vaccine with a 95% lower vaccine efficacy boundary above 30% assuming the true vaccine efficacy is 60%? if the sample size is 500 per group average follow-up of 1 year, the event rate is 0.25 episode per person-year in the control group, and the dispersion parameter is 2?

library(ssutil)
set.seed(123)
result <- sim_power_nbinom(
  n1 = 500, n2 = 500,
  ir1 = 0.2, tm = 1,
  rr = 0.4, boundary = 0.7,
  dispersion = 2,
  alpha = 0.05,
  nsim = 1000
)

result
#> Empirical Power Result
#> ----------------------- 
#> Power:       0.7990
#> 95% CI:      [0.7728, 0.8234]
#> Simulations: 1000