Entropy Balancing in Panel Data Setting

Hi everyone,
I want to examine the effect of a treatment on an observation group with a Diff-in-Diff approach in a panel data setting in Stata. To avoid differences across subsamples I additionally using entropy balancing. Now, I have a methodical question.

I have already read some papers on this setting and noticed that the calculated weights of the pre-period are also used for the post-period. Why they not use the exact calculated weights in this case? I think in some ways this is data manipulation and I don't understand the theoretical explanation.

Here is an example of a paper in which Entropy Balancing is applied with panel data: http://dx.doi.org/10.2139/ssrn.3403486

Can anyone help me? Thank you very much!
Last edited:


Less is more. Stay pure. Stay poor.
DiD is big in the econ field and it seems your example aligns with that and the general terminology used. I am not overly familiar with the speak, but I would imagine the entropy is pseudo-comparable to balancing background imbalances via propensity scores. The weights function to balance covariates, so that the only difference is the intervention/policy being investigated.

In randomized control trials you randomize subjects to treatment groups to balance potential background differences (confounders). However in observational research, groups need to be balanced analytically.
Hi, thank you very much for your answer.
Yes, you can compare the two methods because they are used for the same purpose.

However, my question is now more about their use in the panel data set. For example, if there are several observation rows for one company (several years), the weight of the pre-period is used for the entire company (also for the post-period).

Exemplary in the above mentioned paper: Matching was performed based on the pre-reform (before 2013)

Why is the weight of the pre-period used for the entire company and they don't use the calculated weight for each period?


Less is more. Stay pure. Stay poor.
I will acknowledge that I have not read the linked article thoroughly. But if I were creating weights to balance confounder, I would want more than one set of weights if their were effect heterogeneity between subgroups. Also, I believe in marginal structural models, which have repeated treatments, but one outcome, weights are calculated for each treatment period across time (and I believe multiplied together). So I would agree given YOUR synopsis that intuitively it seem a little weird.