Entropy Balancing in Panel Data Setting

[Yolanda Schmidt]

Hi everyone,
I want like to quantify a treatment effect in a model by measuring the differential effect of a treatment on a 'treatment group' versus a 'control group'. Because it is a panel data set, the diff-in-diff approach is used . Nevertheless, I would like to introduce weights to mitigate potential effects of differences across subsamples on my empirical estimation. I am using entropy balancing for the control sample to equalize the distribution of determinants across treatment and control samples.

However, I do not understand how entropy balancing is applied to the panel data set. In some papers I have read that the weights are only calculated on the basis of the pre-period, so for each unit the weights of the pre-period are used.
The question I ask myself is, for what reason uniform weights are used for each unit and why they dont use different weights for each observation of a unit (the exact calculated weight for each observation)?
For example, if there are several observation rows for one company (several years), the weight of the pre-period is also used for observations of the company as a whole (also for the post-period).
Here is an example of a paper in which entropy balancing is applied in a panel data setting: http://dx.doi.org/10.2139/ssrn.3403486
In the paper they mention: Matching was performed based on the pre-reform (before 2013)

I would also like to know how to implement such a regression in stata ?

Thank you very much for your help!

[Nick Cox]

This is a repost of https://www.statalist.org/forums/for...ropy-balancing with more detail.

There is advice on your situation -- you asked a question and no one replied -- at https://www.statalist.org/forums/help#adviceextras

In general, the best tactic is to expand on an unanswered post in the same thread. I have never heard of "entropy balancing", which is why I don't have an answer.

[Yolanda Schmidt]

This is a repost of https://www.statalist.org/forums/for...ropy-balancing with more detail.

There is advice on your situation -- you asked a question and no one replied -- at https://www.statalist.org/forums/help#adviceextras

In general, the best tactic is to expand on an unanswered post in the same thread. I have never heard of "entropy balancing", which is why I don't have an answer.

Thanks for the feedback.
Sorry, I thought I could delete the old post and give a better description of my problem. I will consider it for the next time.

In my model I want to use entropy balancing, but the question is also applicable to other weighting methods. Suppose there is a panel data set (for each company there are observations over several years) and a treatment takes place in any year. How can weights be applied, for example in the form of a propensity score? Are the weights for each company determined uniformly from one period? And if so, why?
In the paper mentioned above, only the year 2013 is used as an example to calculate weights, so that the control group and treatment group are adjusted. The question I ask myself is why the panel data set only takes one year into account for calculation the weight of a company (and especially the year before the treatment).

Thanks for your help!

[Nick Cox]

As explained in the FAQ Advice a poster can't delete old posts even if they now seem pointless. There's never been that scope since the forum started for any post that starts a thread. . There could be a much longer discussion of the principles, but I stop there.

Otherwise you have interesting questions and I hope that someone else has good answers, as they go way beyond anything I've ever done or wanted to do with panel data. That's in no sense a criticism.

[Ben Jann]

The answer depends on what exactly you are interested in. If your goal is to adjust for pre-treatment confounding, then obtaining balancing weights based on the first wave of your panel dataset seems appropriate. That is, you make the groups comparable at the start of your timeline. More complicated are cases where there might be post-treatment confounding, or more generally, if you study a dynamic/endogenous process. In this case marginal structural models may be an answer, see for example https://scholar.harvard.edu/xzhou/pu...ights-marginal. Don't know whether there are any Stata implementations of such models...
ben

[Fabio Motoki]

I'm coming a bit late for the discussion, but you can try taking a look at these two papers:

McMullin, J. L., & Schonberger, B. (2020). Entropy-balanced accruals. Review of Accounting Studies, 25(1), 84–119. https://doi.org/10.1007/s11142-019-09525-9
McMullin, J., & Schonberger, B. (2022). When Good Balance Goes Bad: A Discussion of Common Pitfalls When Using Entropy Balancing. Journal of Financial Reporting, 7(1), 167–196. https://doi.org/10.2308/JFR-2021-007