ngreifer
commented
9 months ago That's a good question, and one I don't know the answer to. It's more of a statistical question than a question about WeightIt, and I don't know if research has been done on it. It seems like you might conceptualize your situation as having two treatment variables, the numerator and denominator, and you are assuming the functional form of the outcome model involves taking their ratio. That scenario is a bit more complicated than WeightIt is equipped to handle. The problem isn't just about how to specify the weighting model but how to assess balance (i.e., on the ratio, on each component of the ratio individually, all three?). You can always just use a weighting method that doesn't involve a model to balance the covariates. I suppose another challenge is in thinking about whether the denominator is more like a treatment or more like a covariate. Good luck figuring this out; I think it's an interesting question ripe for statistical research.
In one of my models, the exposure is actually a ratio of multiple variables. When fitting an outcome model where the dependent variable is a ratio, I am following the advice found here: I log-transform the outcome, and I include in the RHS of the formula the log of the denominator of the outcome. I was wondering whether it makes sense to do the same when estimating the weights. That is, if the exposure is a ratio, including the log of the denominator of the exposure in the RHS of the formula.