2026-08-25 19:21 UTC
@wviechtb@fediscience.org Excellent, thanks!
Another question. In reality the proportions I'm trying to estimate are multinomial (e.g., proportions in different race categories). However, in most of the literature it looks like people are estimating these separately as a series of binomials (e.g. Black vs not-Black, then Asian vs not-Asian, etc). When I try this, I wind up with proportions that don't sum to 1. Do I have a realistic better option? Stipulating that I am not going to derive a better statistic myself.
Replies (1)
-
@wviechtb@fediscience.org 2026-08-25 20:02
@emjonaitis@mathstodon.xyz I would then consider doing a multivariate meta-analysis, since you have multiple proportions arising from the same study that are not independent. Also, if there are c classes, then I would only use c-1 one of them as outcomes, since the last one is redundant. Then the estimate for the last class is simply 1 minus the sum of the estimates for the c-1 classes.