Browsing by Subject "observational data"

In social sciences, as in health sciences, there is an increasing interest in exploring differences in treatment effects amongst subpopulations and even individuals. In many cases, researchers must rely on observational data where the assignment mechanism of the treatment is non-randomized. Nevertheless, by including a sufficient set of covariates in the used model, it is possible to draw a causal inference. However, some causal structures have proved to cause bias in the treatment effect estimates when particular pre-treated variables in them are conditioned. In existing literature there is no consensus as to how to treat these structures, especially in the heterogeneous treatment effect estimation case. The aim of this thesis is to explore how causal structures affect covariate selection in the heterogeneous treatment effect estimation context. The theoretical background of this subject is built on the potential outcomes framework and structural causal models. This thesis provides an overview of heterogeneous treatment effect estimation methods, including a more detailed view on the causal forest method. The second stage of the thesis is carried out by executing a simulation study where the causal forest method is applied with different causal structures. In each simulation, different sets of conditioned covariates are tested. The simulation study results prove almost consistent. In every simulation except one, a higher number of variables implicates improvement in performance. Surprisingly, this result is applicable even to the cases where structural causal models literature suggests not to condition all the variables. According to the results of the simulation study, a practical recommendation would be to include as many relevant pre-treated, non-instrumental variables in the model as possible. The results are in line with practical recommendations given in potential outcomes framework literature.