print.ritest: A print method for ritest objects

In grantmcdermott/ritest: Randomisation Inference Testing

Description

Printed display of ritest objects. Tries to mimic the display of the equivalent Stata routine.

Usage

## S3 method for class 'ritest'
print(x, verbose = FALSE, ...)

Arguments

x	An ritest object.
verbose	Logical. Should we display the original model summary too? Default is FALSE.
...	Currently ignored.

Examples

#
## Example 1: Basic functionality
#

# First estimate a simple linaer regression on the base 'npk' dataset. For
# this first example, we won't worry about strata or clusters, or other
# experimental design complications.
est = lm(yield ~ N + P + K, data = npk)

# Conduct RI on the 'N' (i.e. nitrogen) coefficient. We'll do 1,000
# simulations and, just for illustration, limit the number of parallel cores
# to 2 (default is half of the available cores). The 'verbose = TRUE'
# argument simply prints the results upon completion, including the original
# regression model summary.
est_ri = ritest(est, 'N', reps = 1e3, seed = 1234L, verbose = TRUE)

# Result: The RI rejection rate (0.021) is very similar to the parametric
# p-value (0.019).

# We can plot the results and various options are available to customise the appearance.
plot(est_ri)
plot(est_ri, type = 'hist')
# etc

# Aside: By default, ritest() conducts a standard two-sided test against a
# sharp null hypothesis of zero. You can can specify other null hypotheses as
# part of the 'resampvar' string argument. For example, a (left) one-sided
# test...
plot(ritest(est, 'N<=0', reps = 1e3, seed = 1234L, pcores = 2L))
# ... or, null values different from zero.
plot(ritest(est, 'N=2', reps = 1e3, seed = 1234L, pcores = 2L))


#
## Example 2: Real-life example
#

# Now that we've seen the basic functionality, here's a more realistic RI
# example using data from a randomized control trial conducted in Colombia.
# More details on the dataset -- kindly provided by the study authors -- can
# be found in the accompanying helpfile ("?colombia"). The most important
# thing to note is that we need to control for the stratified (aka "blocked")
# and clustered experimental design.

data("colombia")

# We'll use the fixest package to estimate our parametric regression model,
# specifying the strata (here: treatment-control pairs) as fixed-effects and
# clustering the standard errors by location (here: city blocks).
library(fixest)
co_est = feols(dayscorab ~ b_treat + b_dayscorab + miss_b_dayscorab |
                 b_pair + round2 + round3,
               vcov = ~b_block, data = colombia)
co_est

# Run RI on the 'b_treat' variable, specifying the strata and clusters.
co_ri = ritest(co_est, 'b_treat', strata='b_pair', cluster='b_block',
               reps=1e3, seed=123L)
co_ri
plot(co_ri, type = 'hist', highlight = 'fill')

# This time, the RI rejection rate (0.11) is noticeably higher than the
# parametric p-value (0.024) from the regression model.

grantmcdermott/ritest documentation built on Jan. 8, 2022, 12:07 a.m.