calc_power_finite: Calculate power or properties of quantile confidence...

In li-xinran/RIQITE: Randomization Inference for Quantiles of Individual Treatment Effects

Calculate power or properties of quantile confidence intervals for a finite-sample dataset.

Description

Given a fixed schedule of potential outcomes and a proportion to assign to treatment, calculate (via simulation) either the power to reject a null hypothesis or the distribution of all the confidence intervals.

Usage

calc_power_finite(
  Y0,
  tau,
  p_tx,
  R = 100,
  percentile = 1,
  alpha = 0.05,
  c = 0,
  alternative = "greater",
  method.list = list(name = "Stephenson", s = 10),
  score = NULL,
  stat.null = NULL,
  use_pval_bound = FALSE,
  nperm = 1000,
  k.vec = NULL,
  quantile_n_CI = NA,
  targeted_power = TRUE,
  summarise = TRUE
)

Arguments

R	Number of permutation replicates for power calculation.
percentile	Which quantile is of interest (default is 1, the max).
alpha	Alpha for testing (default of 0.05)
c	Threshold for null (see pval_quantile)
alternative	A character takes value "greater", "less" and "two.sided", indicating the alternative hypothesis.
method.list	A list specifies the choice of the rank sum test statistic. For example, list(name="Wilcoxon") means the Wilcoxon rank sum statistic, and list(name = "Stephenson", s = 10) means the Stephenson rank sum statistic with parameter s=10.
score	An n dimensional transformed ranks, i.e., (phi(1), phi(2), ..., phi(n)), where phi() denotes the rank transformation function.
stat.null	An vector whose empirical distribution approximates the randomization distribution of the rank sum statistic.
nperm	A positive integer representing the number of permutations for approximating the randomization distribution of the rank sum statistic.
quantile_n_CI	What quantile of the distribution of number of significant units should be reported?
If	TRUE, calculates the overall bound on impacts using pval_bound() (which allows for the Diff in Means test statistic and other stats that do not need to be distribution free). Otherwise uses pval_quantile().

Details

If targeted_power = FALSE, this method will estimate (via simulation) the distributions all simultaneous confidence intervals that can then be post processed to determine their variability, and also the CI for the number of positive units, etc.

Otherwise, it will, by simulation, calculate the chance of rejecting a null by repeatedly rerandomizing the data and conducting a hypothesis test on the results, via using pval_quantile() (or pval_bound() if use_pval_bound is TRUE)

Value

If targeted_power=TRUE, return a small tibble with statistics of the power calculation (true effect being tested, power to detect, sample size, etc.). sd_Y0, sd_Y1, sd_tau are the standard deviations of the potential outcomes and individual treatment impacts, and r is the correlation of Y0 and the treatment effects. Otherwise, if summarise = FALSE, return a matrix with each column being a simulation run and each row being a quantile, with the listed confidence interval limit for each entry. Otherwise a tibble with a row for each quantile specified in k.vec. The "n" column is a repeated value of the median number of quantiles found significant across simulations.#'

li-xinran/RIQITE documentation built on July 1, 2023, 6:58 p.m.