Lateef Amusa, Temesgen Zewotir, and Delia North

DOI: 10.14456/sjst-psu.2021.53

pp. 406 - 413

Abstract

We present entropy balancing – a relatively new technique for estimating treatment effects, which has been under-utilised
in the applied biomedical literature. Our objective is to share our experiences learned from using entropy balancing in nonexperimental studies, via Monte Carlo simulations and two empirical examples. We used the inverse probability of treatment
weighting method for benchmarking the performance of entropy balancing. Entropy balancing had remarkably superior
performance in terms of covariate balance and efficient estimation of treatment effects. Entropy balancing does not require
extensive tweaking of the propensity score specification to achieve the optimal covariate balance. Instead, it directly incorporates
covariate balance into the weight function that is applied to the sample units. Notably, the excellent performance of entropy
balancing is not only on the first moment (mean) but also on higher-order moments. However, we also highlighted the situations
where entropy balancing may fail or not have optimal performance. Entropy balancing merits more widespread adoption in
biomedical research.