Primer on Entropy Balancing

Question

I've scoured the internet, but I can't for the life of me, find a comprehensive primer on entropy balancing.

I am currently in the process of cleaning data in order to create weights for the part of the population that is in a cohort study, and that I will analyze. I am currently toying with the WeightIt package in R. The vignette contains some great information, but I'm looking for more.

I have a few general questions about the process, for instance:

• The main "problem" I have is that I have access to endless population-wide register data (education, SES, living conditions, psychiatric diagnoses, doctor visits, etc). So I'm not sure about what is best to include and how to operationalize the variables (and potentially what is redundant or even harmful). In essence: are there any general tips here that can guide my process?

• What is the tradeoff between including many variables, which will eventually lead to more rows with NAs, and selecting fewer variables (many NAs seem to lead to larger variance in the weights).

• How to cross-validate the weights? I'm thinking of looking at whether the cohort sample has similar correlations as the population for a range of variables.

• Are there other ways to assess how well the weights are performing?

I know that these questions must eventually be resolved by me in the context of my study, but that's why I am looking for some primer that can help me guide the process.

Thanks in advance.

Answer 1

I think part of the difficulty is that entropy balancing isn't a specific statistical method; it is a "flavor" of a well-understood method known as propensity score weighting (PSW) or inverse probability weighting (IPW). IPW can be used for many purposes, and you haven't been clear about for what purpose you want to use it. It seems a bit like you have a "hammer" (entropy balancing) and don't have a clear "nail" (the substantive problem you want to solve). You need to be clear about what exactly you are trying to do on a substantive level, and then we can help advise on a specific statistical method to solve that problem. It may be that IPW is the right method to solve your problem, in which case we can ask the secondary question of which flavor of IPW would be best suited for your task and dataset. It may be that entropy balancing is well suited to your task, but you haven't clarified that yet.

Looking up primers and tutorials for IPW will answer all of the questions you asked. These are matters separate from the flavor of IPW you end up using. Note that because IPW can be used for many purposes (e.g., estimating treatment effects in observational studies, adjusting for missing data, adjusting for censoring, generalizing or transporting estimates to a different population, etc.), you will need to be clear about the specific use case you have. That will determine the selection of variables and the optimal use of the weights.

I have written about what entropy balancing is here. Note that WeightIt (of which I am the author) is best suited for estimating weights for adjusting for confounding in observational studies, but not really for other uses of IPW. However, if that is your task, there are many weighting methods available beyond entropy balancing that are worth looking into. It is typically straightforward to "hack" WeightIt to be able to estimate weights for whatever purpose you have in mind.