flexibly implements the Westfall-Young step-down resampling method
existing packages are powerful but have difficult to use syntax. This implementation has very simple user-facing syntax.
inspired by twoway, this command allows for a list of estimation commands with the extra option adjust.
We call stepdown by typing sets of estimatation commands, enclosed in parenthesis. We then flag the coefficients with p-values we'd like to adjust with the adjust options. There are no restrictions requiring the estimation commands to be the same, or for the arguments to adjust to be the same. The FWER will be controlled over all coefficients flagged in a particular call to stepdown. In this example, we will control the FWER over estimates of coeffients on x2 and x3 in the first model and coefficients on x2 and x4 in the second model.
stepdown (reg y1 x1 x2 x3, adjust(x2 x3)) (areg y2 x2 x4 x5, absorb(z) adjust(x2 x4))
Except for the adjust option, estimation commands will be submitted exactly as specified here. Any suboptions supported by the underlying estimation command are allowed, like clustering or high dimensional fixed effects with reghdfe. Additionally, adjust supports factor variable syntax. So, the coefficients for particular factor estimates can also be added to the list of coefficients to adjust.
stepdown (reg y1 i.z, adjust(1.z 2.z 3.z)) (areg y2 x2 x4 x5, absorb(z) adjust(x2 x4))
Since stepdown uses a bootstrap procedure, stepdown allows bootstrap options to be appended to the end of the function call.
In particular, stepdown recognizes reps to control the number of bootstrap repitions (1000 is recommended, but the default is only 10). strata cluster idcluster and weight will all be passed through to gsample (see help gsample). Additionally, bs_sample takes the bootstrap sample size (defaults to a bootstrap draw of equivalent size). By default, stepdown runs quietly, but the echo option will cause stepdown to announce each bootstrap iteration.
stepdown (reg y1 x1 x2 x3, adjust(x2 x3)) (areg y2 x2 x4 x5, absorb(z) adjust(x2 x4)), reps(1000) cluster(classroom) strata(state)
stepdown returns two matrices to r(). It returns r(p) which a list of the original, unadjusted p-values, and r(adj_p) which contains the Westfall-Young adjusted p-values.