Bioequivalence for Dichotomous outcomes, and also with nested data.

[Robert Ploutz-Snyder]

Seeking guidance for conducting a bioequivalence (or non-inferiortty) test comparing two groups, but with a dichotomous outcome (essentially a dichotomous version of the TOST test). Then will need to re-run on a dataset with multiple observations nested within person... so a mixed-model of sorts. Any guidance ?

[Joseph Coveney]

Any guidance ?

Are you looking for a method to estimate sample size? Recommendation for an estimation command to fit a regression model or perform some kind of analogue to TOST for binomial data? What guidance are you seeking?

[Robert Ploutz-Snyder]

Are you looking for a method to estimate sample size? Recommendation for an estimation command to fit a regression model or perform some kind of analogue to TOST for binomial data? What guidance are you seeking?

Thanks Joseph--I'm seeking code to run an equiv test similar to TOST, but for binomial data. We have the data in-hand.

[Joseph Coveney]

I'm seeking code to run an equiv test similar to TOST, but for binomial data.

You can fit a generalized linear model to the data with the official Stata glm command, and then examine the confidence interval for the risk difference to determine whether the upper and lower 90% confidence bounds lie within your equivalence criterion.

glm out i.grp, family(binomial) link(identity) level(90)
// ten-percentage-point margin of equivalence (±10%):
assert (-0.1 <= r(table)["ll", "out:1.grp"]) & (r(table["ul", "out:1.grp"] <= 0.1)

There is also a user-written command rdci available for installation from SSC that has several methods of estimating the confidence interval for risk differences of two independent proportions that are better behaved than the conventional Wald interval.

search rdci

It depends upon another user-written command, ridder, for one of its methods of estimation, and so that would have to be installed, too.

search ridder

For a multilevel or mixed-effects model, the official Stata command that's directly analogous to glm is meglm, except that the latter command does not allow the identity link function with the binomial distribution family. Fortunately, there is a user-written generalized linear mixed model command that will allow that combination of distribution family and link function: gllamm, which is older (doesn't have factor variables syntax, for example) and has its own syntax for specifying grouping (clustering) variables and the like.

search gllamm

for where to find it.

[Robert Ploutz-Snyder]

Thank you Joseph... We'll explore these options. Much appreciated