Abstract

Real world data has recently attracted attention in the medical field for its part in establishment of new treatment methods and development of new medicines. Treatment effect, the difference in outcome between the target and standard treatments, is estimated as a treatment evaluation measure. Average treatment effect is the estimation of the entire population, but the results do not always fit to all subjects. The idea of heterogeneous treatment effect (HTE) examines the point that the treatment effect on each subject will likely vary depending on the covariates. However, HTE cannot be calculated in some clinical trials because, for example, an individual’s results are obtained from only one study arm. To solve this problem, various treatment effect models have been proposed. When focusing on tree-based approaches, causal forest (Warger and Athey, 2018) and Bayesian Additive Regression Trees (BART) (Chipman et al., 2010) are included, which directly estimate treatment effect. However, it is unclear which covariates or rules affect the results, because these methods are constructed “black boxed” model. RuleFit method was proposed by Freidman and Popescu (2008) as an interpretable method, and then Lee et al. (2020) applied it to clinical study, but they only adapted the rule ensemble method to the HTE estimated by different methods.
We propose a novel framework of rule ensemble directly applied to the data to estimate interpretable treatment effects. The proposed method, which is based on the RuleFit method, has two features. First, the proposed model has a main effect term that estimates the impact on outcomes independent of the treatment effect. If main effects in a true model are present, estimating the treatment effect without their consideration may lead to bias. This feature is expected to enhance the estimation accuracy of the treatment effect by considering the main effect. A second feature of the proposed method is that it uses group lasso to select the same rules. The original RuleFit method uses lasso but may select different rules between the treatment arms, so comparability cannot be guaranteed. To overcome this, our proposed method uses group lasso, which allows estimation under the same rules across both treatment groups. The proposed method appears to be useful in terms of prediction accuracy compared to existing treatment effect models (causal forest, BART) through numerical simulations. Furthermore, we verify its usefulness of the proposed method by the application to real data.