{smcl} {* *! version 1.1.7 feb2020}{...} {vieweralsosee "[R] simulate" "help permute"}{...} {vieweralsosee "[R] bootstrap" "help bootstrap"}{...} {vieweralsosee "[R] jackknife" "help jackknife"}{...} {vieweralsosee "[R] simulate" "help simulate"}{...} {viewerjumpto "Syntax" "ritest##syntax"}{...} {viewerjumpto "Description" "ritest##description"}{...} {viewerjumpto "Options" "ritest##options"}{...} {viewerjumpto "Examples" "ritest##examples"}{...} {viewerjumpto "Author" "ritest##author"}{...} {title:Title} {p2colset 5 20 22 2}{...} {p2col :{bf:ritest} {hline 2}}Randomization inference and permutation tests, allowing for arbitrary randomization procedures{p_end} {p2colreset}{...} {marker syntax}{...} {title:Syntax} {pstd} Compute p values for permutation tests, allowing for arbitrary randomization procedures {p 8 16 2} {cmd:ritest} {it:resampvar} {it:{help exp_list}} [{cmd:,} {it:{help ritest##options_table:options}}] {cmd::} {it:command} {synoptset 27 tabbed}{...} {marker options_table}{...} {synopthdr} {synoptline} {syntab :Main} {synopt :{opt r:eps(#)}}perform {it:#} random permutations; default is {cmd:reps(100)}{p_end} {synopt :{opt le:ft}|{opt ri:ght}}compute one-sided p-values; default is two-sided{p_end} {syntab :Automatic resampling} {synopt :{opth str:ata(varlist)}}permute {it:resampvar} within strata{p_end} {synopt :{opth clu:ster(varlist)}}keep {it:resampvar} constant within clusters{p_end} {synopt :{opth fix:levels(list)}}keep {it:resampvar} fixed for all listed values{p_end} {syntab :File-based resampling} {synopt :{opth samplings:ourcefile(filename)}}take permutations of {it:resampvar} from Stata data file {it:filename}, containing variables named {it:resampvar}1, resampvar2, resampvar3, ...{p_end} {synopt :{opth samplingm:atchvar(varlist)}}merge permutations in {it:samplingsourcefile} with the data using these variables (1:1 or m:1){p_end} {syntab :Program-based resampling} {synopt :{opth samplingp:rogram(it:programname)}}generate permutations of {it:resampvar} by calling user-written program {it:programname}{p_end} {synopt :{opth samplingprogramo:ptions(string)}}optionally pass {it:string} as options to {it:programname} {p_end} {syntab :Reporting} {synopt :{opt l:evel(#)}}set confidence level; default is {cmd:level(95)}{p_end} {synopt :{opt v:erbose}}display full table legend{p_end} {synopt :{opt nodots}}suppress replication dots{p_end} {synopt :{opt noi:sily}}display any output from {it:command}{p_end} {syntab :Advanced} {synopt :{opt null(outcome value)}}specify a null hypothesis of treatment effect {it:value} or {it:varname} for outcome {it:outcome} {p_end} {synopt :{opt kdens:ityplot}}plot the densities of each statistic in {it:exp_list}{p_end} {synopt :{opt kdenstyo:options(string)}}additional options to be passed on to {it:kdensity} {p_end} {synopt :{opt sav:ing(filename)}} save a data set of estimates from the permutations{p_end} {synopt :{opt saver:esampling(filename)}}save all permutations of {it:resampvar} in a file called {it:filename} for later inspection.{p_end} {synopt :{opt seed(#)}}set random-number seed to #{p_end} {synopt :{opt eps(#)}}numerical tolerance; seldom used, default 1e-7{p_end} {synopt :{opt strict}}computes the strict version of the test. p=c/N, with c = #{c -(}|T| > |T(obs)|{c )-}{p_end} {synopt :{opt force}}force {cmd: ritest} to accept weights in {it:command} {p_end} {synopt :{opth reject(exp)}}identify invalid results{p_end} {synoptline} {p2colreset} {marker description}{...} {title:Description} {pstd} {cmd:ritest} estimates p-values for permutation tests on the basis of Monte Carlo simulations. Unlike permute, ritest allows to specify more complex permutation structures, as generated for example by clustered treatment assignments. {dlgtab:Automatic resampling (permutation)} {pin} {cmd:. ritest} {it:resampvar} {it:exp_list}{cmd:,} {opt reps(#) strata(stratavar) cluster(clustervar)}{cmd::} {it:command} {pstd} randomly permutes the values in {it:resampvar} {it:#} times, respecting strata and clusters, each time executing {it:command} and collecting the associated values from the expressions in {it:exp_list}. Not specifying strata assumes no strata were used, which is equivalent to all observations being in one single stratum. Not specifying clusters assumes no clusters were used, which is equivalent to each observations being a separate cluster. {dlgtab:File-based resampling} {pin} {cmd:. ritest} {it:resampvar} {it:exp_list}{cmd:,} {opt reps(#) samplingsourcefile(filename) samplingmatchvar(varlist)}{cmd::} {it:command} {pstd} merges the data based on {it:varlist} {it:#} times using {it:filename} (1:1, or if ids are not unique m:1). ). It executes {it:command} each time, replacing resampvar iteratively with {it:resampvar}1, {it:resampvar}2, {it:resampvar}3, ..., {it:resampvar}1000, which have to be stored in {it:filename} prior to executing {cmd: ritest}. {dlgtab:Program-based resampling} {pin} {cmd:. ritest} {it:resampvar} {it:exp_list}{cmd:,} {opt reps(#) samplingprogram(progname) samplingprogramoptions(string)}{cmd::} {it:command} {pstd} permutes {it:resampvar} by calling a user-written program {it:progname}, and passing the string {it:samplingprogramoptions} as options to this program. Also, the program will to be passed at two standard arguments: {it:resampvar} is the variable name of the relevant variable and {it:run} is an integer containing the iteration. {dlgtab:General} {pstd} {it:left|right} computes one sided p-values: Pr(T* {ul:<} T) or Pr(T* {ul:>} T). The default is two-sided: Pr(|T*| {ul:>} |T|). Here T* denotes the value of the statistic from a randomly permuted data set, and T denotes the statistic as computed on the original data. {pstd} {it:resampvar} identifies the variable for which values will be randomly permuted. {pstd} {it:command} defines the statistical command to be executed. Most Stata commands and user-written programs can be used with {cmd:ritest}, as long as they follow {help language:standard Stata syntax}. The {cmd:by} prefix may not be part of {it:command}. {pstd} {it:exp_list} specifies the statistics to be collected from the execution of {it:command}. {marker options}{...} {title:Options} {dlgtab:Main} {phang} {opt reps(#)} specifies the number of permutations to perform. The default is 100. {phang} {opt left} or {opt right} requests that one-sided p-values be computed. If {opt left} is specified, an estimate of Pr(T* {ul:<} T) is produced, where T* is the test statistic and T is its observed value. If {opt right} is specified, an estimate of Pr(T* {ul:>} T) is produced. By default, two-sided p-values are computed; that is, Pr(|T*| {ul:>} |T|) is estimated. {dlgtab:Options} {phang} {opth strata(varlist)} specifies that the permutations have to be performed within each stratum defined by the values of {it:varlist}. {phang} {opth cluster(varlist)} specifies that the permutations have to be performed treating each cluster as defined by {it:varlist} as one unit of assignment. {phang} {opth fixvalues(list)} specifies that the permutations have to be performed holding all observations fixed that have the specified values the resampling variable. This can be used for pairwise testing in experiments with more than two treatment arms. {phang} {opth samplingsourcefile(filename)} specifies that permutations of {it:resampvar} have to be taken from Stata data-file {it:filename}. This data set should contain variables named [resampvarname]1, [resampvarname]2, [resampvarname]3, ... and a unique id. {phang} {opth samplingmatchvar(varlist)} merge permutations in from {it:samplingsourcefile} with the data using these variables (1:1 or m:1 merge). {phang} {opth samplingprogram(programname)} specifies that permutations of {it:resampvar} be generated by calling user-written {it:programname}. {phang} {opth samplingprogramoptions(string)} specifies that when {it:programname} is called {it:string} is also passed string as options to {it:programname}. {dlgtab:Reporting} {phang} {opt level(#)} specifies the confidence level, as a percentage, for confidence intervals. The default is {cmd:level(95)} or as set by {helpb level:set level}. {phang} {opt nodots} suppresses display of the replication dots. By default, one dot character is displayed for each successful replication. A red 'x' is displayed if {it:command} returns an error or if one of the values in {it:exp_list} is missing. {phang} {opt noisily} requests that any output from {it:command} be displayed. This option implies the {opt nodots} option. {dlgtab:Advanced} {phang} {opt null(outcome value)} allows to test null hypotheses which are different from zero. {it:value} can either be numeric or the name of a variable and will be deducted from the variable {it:outcome} before conducting the randomization test. {phang} {opt eps(#)} defines the numeric tolerance of the test. To account for numeric imprecisions, the p-value is computed as c/N, with c = #{c -(}|T|>=|T(obs)|-eps{c )-}. Default is eps(1e-7). {phang} {opt strict} computes the strict version of the test. By default the p-value is computed based on c = #{c -(}|T| >= |T(obs)|{c )-}. When multiple re-samplings result in the same realization of T, this proceedure yields conservative p-value. The strict version is based on c = #{c -(}|T| > |T(obs)|{c )-} and is weakly overconfident. Requires eps(0). {phang} {opt kdensityplot} produces a density plot of the realizations of expressions in {it:exp_list}. The realization for the expression in the original data is drawn as a vertical line. {phang} {opth reject(exp)} identifies an expression that indicates when results should be rejected. When {it:exp} is true, the resulting values are reset to missing values. {marker examples}{...} {title:Examples} {pstd} Attached to this is help file is a (simulated) data set of schools, classes, and student test scores. The data consists of observations from 2 schools, each school has 4 classes of 10 students each. Classes were randomly assigned to treatment, stratifying by school. The stratification assured that each school had two treatment and two control classes. The outcome are test scores and the researcher wants to conduct fisher randomization test test for the sharp null of no treatment effect based on a regression controlling for age. {p_end} {dlgtab:Automatic resampling} {phang2}{cmd:. ritest treatment _b[treatment], cluster(classid) strata(schoolid): reg testscore treatment age}{p_end} {dlgtab:File-based resampling} {phang2}{cmd: // start with an empty data sets of classes containing }{p_end} {phang2}{cmd: // only schoolid and classid}{p_end} {phang2}{cmd: use example_classes.dta}{p_end} {phang2}{cmd: forvalues i = 1/100 {c -(} //create and save 100 permutations of treatment}{p_end} {phang2}{cmd: tempvar random cutoff}{p_end} {phang2}{cmd: gen `random'=rnormal()}{p_end} {phang2}{cmd: bys schoolid: egen `cutoff' = median(`random')}{p_end} {phang2}{cmd: gen treatment`i' = `random'> `cutoff'}{p_end} {phang2}{cmd: {c )-}}{p_end} {phang2}{cmd: tempfile permutations}{p_end} {phang2}{cmd: save `permutations'}{p_end} {phang2}{cmd: use example_students.dta}{p_end} {phang2}{cmd: //using this file}{p_end} {phang2}{cmd: ritest treatment _b[treatment], r(500) ///}{p_end} {phang2}{cmd: samplingsourcefile("`permutations'") samplingmatchvar(schoolid classid): ///}{p_end} {phang2}{cmd: reg testscore treatment age, cluster(classid)}{p_end} {phang2}{cmd: //redoing the analysis testing the null of the treatment effect being 10(constant)}{p_end} {phang2}{cmd: ritest treatment _b[treatment], r(500) null(testscore 10) ///}{p_end} {phang2}{cmd: samplingsourcefile("`permutations'") samplingmatchvar(schoolid classid): ///}{p_end} {phang2}{cmd: reg testscore treatment age, cluster(classid)}{p_end} {phang2}{cmd: //redoing the same analysis using automatic resampling}{p_end} {phang2}{cmd: ritest treatment _b[treatment], r(500) ///}{p_end} {phang2}{cmd: strata(schoolid) cluster(classid) : ///}{p_end} {phang2}{cmd: reg testscore treatment age, cluster(classid)}{p_end} {pstd}The benefit of the file-based approach is that more complicated randomization structures can be implemented. For smaller samples, it is also straightforward to implement complete (non-stochastic) Fisher tests, which exhaust all possible permutations{p_end} {dlgtab:Program-based resampling} {pstd}Define a program to create random permuations:{p_end} {phang2}{cmd: program permme //define the program}{p_end} {phang2}{cmd: syntax , ///}{p_end} {phang2}{cmd: resampvar(varname) ///<- name of the permutation variable}{p_end} {phang2}{cmd: school(varname) ///<- name of the strata variable}{p_end} {phang2}{cmd: class(varname) ///<- name of the cluster variable}{p_end} {phang2}{cmd: * //<- ritest also passes other things to the permutation procedure (e.g. run(#))}{p_end} {phang2}{cmd: tempvar mr r}{p_end} {phang2}{cmd: // draw one random variable per cluster}{p_end} {phang2}{cmd: qui bys `school' `class': gen `r'=rnormal() if _n==1 // draw one random variable per cluster}{p_end} {phang2}{cmd: // compute median of these random variables within cluster}{p_end} {phang2}{cmd: qui bys `school': egen `mr' = median(`r') if !missing(`r') }{p_end} {phang2}{cmd: // replace the permutation var with the new randomization outcome}{p_end} {phang2}{cmd: replace `resampvar' = cond(`r',`r'>`mr',.,.) }{p_end} {phang2}{cmd: sort `school' `class' `r' }{p_end} {phang2}{cmd: // replace the permutation var with the new randomization outcome for all remaining villages}{p_end} {phang2}{cmd: by `school' `class': replace `resampvar' = `resampvar'[_n-1] if missing(`resampvar')}{p_end} {phang2}{cmd: end}{p_end} {pstd}Call {cmd:ritest} using the program:{p_end} {phang2}{cmd: ritest treatment _b[treatment], r(500) samplingprogram(permme) ///}{p_end} {phang2}{cmd: samplingprogramoptions("school(schoolid) class(classid)") : ///}{p_end} {phang2}{cmd: reg testscore treatment, cluster(classid)}{p_end} {pstd}This method can be used implement any kind of re-randomization.{p_end} {title:See also} {pstd} {help permute}, {help simulate}, {help bootstrap}, {help jackknife} {title:Citation} {pstd} Heß, Simon, "Randomization inference with Stata: A guide and software" {it:Stata Journal} 17(3) pp. 630-651. {p_end} {pstd} {browse "https://raw.githubusercontent.com/simonheb/ritest/master/ritest.bib":BibTeX-file}. {p_end} {title:Author and acknowledgments} {pstd} Simon Heß, Goethe University Frankfurt.{p_end} {pstd} The latest version of ritest can always be obtained from {browse "https://github.com/simonheb/ritest"} or {browse "http://HessS.org"}. I am happy to receive comments and suggestions regarding bugs or ideas for improvements/extensions via {browse "https://github.com/simonheb/ritest/issues"}. {p_end} {pstd} The {cmd:ritest}-code is based on the code for {cmd:permute}. {p_end}