{smcl}
{* *! version 3.0.0 28March2020}{...}
{cmd: help randcmd}
{hline}
{title:Title}
{phang} {hi:randcmd} {hline 2} Randomization inference based p-values. {p_end}
{title:Syntax}
{phang} {cmd:randcmd} ({opth (varlist)} 1st Stata estimation command) ... ({opth (varlist)} nth Stata estimation command), {opth treatvars(varlist)} [{cmd:,} {it:options}] {p_end}
{title:Options}
{opth strata(varlist)} variables identifying randomization strata.
{opth groupvar(varlist)} variables identifying treatment groups.
{opth seed(#)} set random-number seed to #, default is 1.
{opth saving(filename, replace)} save results to filename (replace to overwrite).
{opth reps(#)} number of attempted randomization iterations, default is 999.
{opt sample} restrict rerandomization to the sample of the estimating equations.
{opth calc1(string)} execute Stata command after each treatment rerandomization.
...
{opth calc20(string)} execute Stata command after each treatment rerandomization.
{p} {opth (varlist)} must appear as right-hand side variables in the Stata estimation command.
{opt treatvars} indicates the base treatment variables that are rerandomized on each iteration, and is not optional.
{opt treatvars} cannot vary within groups identified by {opt groupvar}. Observations for which any of {opt treatvars} are missing are dropped.
{p} If {opt sample} is specified, {opt treatvars} are rerandomized across observations in the estimating equations; otherwise, {opt treatvars} are rerandomized across all non-missing observations.
{p} To update treatment measures that involve interactions with participant characteristics, up to twenty post-treatment calculations are allowed and are executed in numerical order (see examples below).
{p} If the user does not provide a random number seed, the seed is set to 1 (to ensure replicability, provided the data have not been reordered, on subsequent runs).
The seed is restored to its pre-program value on termination. Please note that Stata's random number generator changed with Stata 14, so if using earlier versions of Stata be sure to alert potential replicators.
{title:Description}
{phang} {cmd:randcmd} allows the testing of the statistical significance of treatment effects within a single estimating equation and across multiple equations, depending upon the needs of the user. {p_end}
{phang} {cmd:randcmd} begins by reporting the conventional estimation results and p-values for each Stata command specified by the user.
The Stata command must return estimated coefficients and standard errors for each variable listed in {opth (varlist)} for that estimating equation.
Otherwise, all estimation options associated with any Stata command are allowed.
{cmd:randcmd} then rerandomizes {opt treatvars}, following any strata and grouping specified by the user, performs optional calc1 ... calc20 (see examples below), and then reruns the Stata estimation command.
Conventional coefficient estimates and standard errors, as well as joint test Wald/F statistics, are stored for each iteration and saved if requested by the user.
Randomization p-values are only based upon iterations that produce an estimated coefficient and non-zero standard error for the treatment measures involved.
The number of such iterations is reported.
{p_end}
{phang} {cmd:randcmd} computes randomization-c and randomization-t p-values for individual treatment effects,
and Wald joint and Westfall-Young multiple-testing tests of statistical significance for equations with multiple treatment effects (if any)
and (where there are multiple estimating equations) for the experiment as a whole.
The last can be used as an "omnibus" test of overall experimental significance, i.e. is there statistical evidence to reject the null that the experiment is irrelevant.
The p-value, in each case, is calculated under the null hypothesis that all treatment effects are zero for all participants.
Randomization-t p-values are more robust to deviations from the exact null of zero effects for all participants.
Multiple testing procedures have more power to detect alternatives that lie on the axes (e.g. one treatment effect is non-zero, while others are zero),
while joint tests have more power to detect alternatives that lie within quadrants (e.g. multiple treatment effects are non-zero).
The joint omnibus test is only executed using the randomization-c, as the randomization-t version requires a cross equation estimate of covariance, which is not possible for all Stata commands.
{phang} For details on the differences between the randomization-c and -t and joint and multiple testing, see:
{pmore} Young, Alwyn (2019).
{it:Channelling Fisher: Randomization Tests and the Statistical Insignificance of Seemingly Significant Experimental Results.}
Quarterly Journal of Economics 134 (2): 557-598. {browse "http://personal.lse.ac.uk/YoungA/ChannellingFisherQJE.pdf": {it: Preprint.}} {browse "https://academic.oup.com/qje/article/134/2/557/5195544": {it: Published article.}}
{p_end}
{title:Examples}
{p} Example 1: Basic estimating equation, with sample restrictions and command options.
{phang} randcmd ((treatment) areg outcome treatment covariates [if] [in] [weights], absorb(absorb) cluster(cluster)), treatvars(treatment)
{phang} Comment: Estimating equation can be restricted to subsets of participants (e.g. age > 25), but treatment is rerandomized across the entire experimental population (treatment ~= . in the data file) unless {opt sample} is specified.
{p} Example 2: Treatment variables include interactions with participant characteristics which are not part of treatment and have to be recalculated after treatment is rerandomized.
{phang} randcmd ((treatment treatage) xtreg outcome treatment treatage covariates [if] [in] , re cluster(cluster)), treatvars(treatment) calc1(replace treatage = treatment*age)
{phang} Comment: Calculations could also be coded as calc1(drop treatage) calc2(generate treatage = treatment*age), but use of replace, where appropriate, is simpler.
{p} Example 3: Interactions between treatment measures can be coded with calc or without.
{phang} randcmd ((treat1 treat2 treat12) probit outcome treat1 treat2 treat12 covariates [if] [in], robust asis), treatvars(treat1 treat2) calc1(replace treat12 = treat1*treat2)
{phang} randcmd ((treat1 treat2 treat12) probit outcome treat1 treat2 treat12 covariates [if] [in], robust asis), treatvars(treat1 treat2 treat12)
{phang} Comment: Since treat12 is the product of randomized treatment variables, and does not involve interaction with non-randomized participant characteristics,
reallocating pairs of (treat1,treat2) to participants and then calculating treat12 is equivalent to reallocating triplets of (treat1,treat2,treat12) across participants.
{p} Example 4: Treatment is stratified and applied in groups.
{phang} randcmd ((treat1 treat2) areg outcome treat1 treat2 covariates [if] [in] [weights], absorb(village) cluster(village)), treatvars(treat1 treat2) strata(strata) groupvar(village)
{phang} Comment: treat1 and treat2 cannot vary within groupvar. The pair (treat1, treat2) applied to each village is rerandomized across villages (within strata).
{p} Example 5: Variables appearing in {opt treatvars} do not have to appear in {opth (varlist)} for the estimating equation. Consider the case where results are analysed using dif-in-dif, with post denoting the post-treatment period.
{phang} randcmd ((treatpost) reg outcome treatpost treat post, cluster(village))), treatvars(treat) groupvar(village) calc1(replace treatpost = treat*post)
{phang} Comment: The coefficients on variable treat are used to determine average differences in the pre-treatment period, but the treatment effects are actually measured by treatpost.
I am interested in testing the hypothesis that treatment has no effect on post-treatment outcomes, so only treatpost is listed in {opth (varlist)}.
{p} Example 6: Running randcmd with multiple equations so as to get a multiple testing result on overall experimental significance.
{phang} randcmd ((treat1 treat2) areg outcome treat1 treat2 covariates [if] [in] [weights], absorb(village) cluster(village))
((treat1 treat1age) xtreg outcome treat1 treat1age covariates [if] [in], re cluster(village))
((treat1 treat1age) xtmixed outcome treat1 treat1age covariates || participant: || village:, iterate(20)),
treatvars(treat1 treat2) strata(strata) groupvar(village) reps(1999) seed(20) saving(results, replace) calc1(replace treat1age = treat1*age) {p_end}
{phang} Comment: I have three estimating equations and wish to test whether, after accounting for multiple testing, I can reject the null that the experiment is irrelevant, i.e. has zero effects everywhere.
{cmd: randcmd} will rerandomize (treat1 treat2) across the entire experimental dataset (stratified by strata, grouped by village), estimate all three equations on each iteration, and report individual
treatment effect p-values, joint tests and multiple testing results for each equation, and a final Westfall-Young multiple testing test of all treatment effects.
The Westfall-Young procedure uses the joint distribution of p-values across all equations so as to minimize the loss of power brought about by the multiple testing adjustment (see paper referenced above). {p_end}
{p} Example 7: Estimating treatment p-values for Stata commands which produce multiple coefficients for each right-hand side variable requires a careful listing of desired treatment variables in {opth (varlist)}.
{phang} randcmd ((treat select:treat) heckman outcome treat covariates, select(treat covariates othercovariates) twostep), treatvars(treat)
{phang} randcmd ((3:treat 7:treat) mlogit outcome037 treat covariates, baseoutcome(0)), treatvars(treat) {p_end}
{phang} Comment: In the heckman estimating equation I identify that I am interested in the significance of the treat variable in both the outcome and selection equations.
In the mlogit estimating equation I have a dependent variable that takes on three values (0,3,7).
I specify that 0 is the base outcome (treat coefficient automatically set to 0) and calculate p-values for the treat coefficients for outcomes 3 and 7.
A simple rule of thumb is to observe how Stata names the coefficient estimates when reporting results and use those names in {opth (varlist)}.
{title:Reported results}
{phang} In comparing potential outcomes, randomization inference always produces "ties", as the original experimental outcome (at a minimum) must be considered a tie with itself (see paper referenced above).
P-values are uniformly distributed between minimum and maximum values determined by these ties.
{cmd:randcmd} reports these minimum and maximum values and a "final" randomized p-value based upon a random draw from the uniform distribution. The width of the max-min interval can be reduced by increasing the
number of iterations.
However, if the experiment only allows for a small number of potential outcomes, ties will occur within the randomized distribution and the width of the min-max interval will not systematically fall with the number of iterations. {p_end}
{phang} The relevance of the experiment's tie with itself in determining rejection at different levels can be eliminated by selecting a number of replications such that the desired level*(number of replications + 1) equals an integer.
The default number of 999 replications, and numbers such as n-thousand plus 999, meet this condition for the .01, .05, and .10 levels.
{phang} {cmd:randcmd} stores the following matrices in {cmd:e()}: {p_end}
{cmd:e(RCoef)} Randomization-c and -t p-values for individual coefficients.
{cmd:e(REqn)} Randomization-c and -t p-values for the joint-test of the significance of treatment measures in each equation and in the experiment as a whole.
{cmd:e(RMult)} Randomization-c and -t p-values for the Westfall-Young multiple testing test of the significance of {it:any} treatment measure in each equation and in the experiment as a whole.
{title:Author}
{phang} Alwyn Young, London School of Economics, a.young@lse.ac.uk. {p_end}