gms | German Medical Science

Propensity score (PS) analyses are becoming the quasi-standard for analyzing non-randomized studies of treatment effects, which is due to their statistical and epistemological advantages as compared to standard outcome regression. PS analyses are performed in two steps. In the first step, the propensity score, defined as the probability of the treatment conditional on the subject's covariates, is estimated. In the second step, this PS is used for estimating the treatment effect, which is the actual quantity of interest.

The ultimate aim in the first step model is balancing covariates in the treatment groups [1], [2]. However, the methods regularly used to this task, standard logistic regression or more evolved machine learning methods, aim for optimizing the prediction for the respective outcome, effectively ignoring covariate balance. As such, methods have been proposed which explicitly aim for minimizing covariate balance in first step models [3], [4], and they have shown to be superior to standard models in simulations.

We here propose z-balancing, a new method which uses the idea of minimizing weighted z-differences [5] as an optimality criterion for covariate balance in first step models. This method improves on previous methods by modelling all covariates on their original (continuous, binary, ordinal, or nominal) scale. In addition, not only standard inverse probability weights can be used (which frequently have problems with extreme weights compromising model fits), but also other weights, e.g., the recently proposed matching weights [6]. In the talk we introduce the method and present first simulation results that compare z-balancing with its competing methods.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.