Это выглядит слишком хорошо, чтобы быть правдой.
Не понимаю, почему метод так не популярен в биомедицине.

suppressPackageStartupMessages({
    library(tidyverse)
    library(logistf)
})

set.seed(15321)

n <- 150
p1 <- 0.015
p2 <- 0.03

x <- rep(0:1, each = n)
probs <- rep(c(p1, p2), each = n)

sim <- function() {
    df <- data.frame(
        x = x,
        y = rbinom(
            n = n * 2, size = 1, 
            prob = probs
        )
    )
    
    fit <- logistf(
        y ~ x, data = df, 
        pl = TRUE, firth = TRUE
    )
    
    c(
        n1 = sum(df$y[x == 0]), 
        n2 = sum(df$y[x == 1]),
        beta = fit$coefficients[[2]],
        lcl = fit$ci.lower[[2]], 
        ucl = fit$ci.upper[[2]],
        lrt = 1 - pchisq(-2 * diff(fit$loglik), df = 1)
    )
}

res <- replicate(10000, sim())

eff <- log((p2/(1 - p2))/(p1/(1 - p1)))

t(res) |>
    as_tibble() |>
    mutate(eff = eff) |>
    summarise(
        coverage = mean(between(eff, lcl, ucl)),
        power = mean(lrt.null < 0.05),
        `S-type` = mean(beta < 0 & lrt.null < 0.05),
        mean = mean(beta),
        true = mean(eff),
        bias = mean - true
    )