atm_diff: Calculate the difference of asymmetric trimmed means between...

In michaelpollmann/parTreat: Efficiently Estimate Treatment Effects Based On A Parametric Model For Treatment Effects

Description

Calculate the difference in asymmetric trimmed means between Y and X. The percent trimmed from left (α) and right ((β)) are chosen adaptively using grid search. If gridsize = 'sqrtn', the grid of α and β values are spaced by 1/sqrt(n) where n is the size of the smaller group. If gridsize = 'n', the grid values are spaced by 1/n. Note that the two-dimensional grid search is then of size n^2 which can be prohibitively slow for large n, while sqrt(n) will be near optimal and feasible even for large samples. Alternatively, specify any integer of at least 3, up to the size of the smaller group.

Usage

atm_diff(
  X,
  Y,
  gridsize = "sqrtn",
  max_alpha = 1,
  max_beta = 1,
  min_frac = 0.25,
  min_obs = 50
)

Arguments

X	numeric vector, the outcomes of the control observations.
Y	numeric vector, the outcomes of the treated observations.
gridsize	either the string 'sqrtn' or 'n' or an integer giving the number of alpha and beta values to try
max_alpha	numeric between 0 and 1, the maximum fraction of observations to trim from the left (default 1)
max_beta	numeric between 0 and 1, the maximum fraction of observations to trim from the right (default 1)
min_frac	numeric between 0 and 1, the minimum fraction of observations to KEEP after trimming (default 0.25)
min_obs	numeric, the minimum number of observations of each treatment state to KEEP after trimming (default 50)

Details

The parameters max_alpha and max_beta allow to set bounds for the amount of trimming from the left and right, respectively. Note that the grid spacing is not affected by this; that is, fewer values of α or β need to be searched over when changing the defaults. To only consider trimming of up to the smallest 3 To only trim the right tail of the data, set max_alpha = 0; forcing the function to not trim the left tail may improve stability when appropriate, and should be much faster computationally. The size of the sample after trimming is 1 - α - β where α and β are the selected trimming percentages.

Value

list of four elements:

tau	the point estimate of the treatment effect
se	the estimated standard error (if calc_se = TRUE)
alpha	the trimming fraction from the left
beta	the trimming fraction from the right

Examples

# draw a random sample with additive treatment effect
X <- rnorm(n=1000)
Y <- rnorm(n=1000)
atm_diff(X,Y)
# for smaller samples, a n-grid may be computationally feasible
X <- rnorm(n=100)
Y <- rnorm(n=100)
atm_diff(X,Y, gridsize='n')

michaelpollmann/parTreat documentation built on Dec. 21, 2021, 5:58 p.m.